The impact of price frames on consumer decision making

May 2010

OFT1226

 Crown copyright 2010

This publication (excluding the OFT logo) may be reproduced free of charge in any format or medium provided that it is reproduced accurately and not used in a misleading context. The material must be acknowledged as crown copyright and the title of the publication specified.

FOREWORD BY AMELIA FLETCHER This report was commissioned by the Office of Fair Trading (OFT) from London Economics in association with Steffen Huck and Brian Wallace (University College London). It examines the impact on consumer decision-making of different ways in which prices can be framed. Consumers deal with relatively complex price frames on a daily basis, both on the high street and the Internet. Typical examples are'3 for 2' 'price was x now y', 'sale ends tomorrow', and 'price is x while stocks last'. When these offers are genuine they can clearly benefit consumers. But what about when they do not necessarily represent a genuine offer? Can consumers see through the different ways of framing prices, or are they in some way tricked? This report uses a controlled economic experiment to isolate and address precisely this issue. By using a comparative metric the authors are able to rank, in terms of their potential magnitude, five different pricing frames. Further, the authors investigate whether and where in the decision making process any behavioural biases occur and what impact the frames have on a retailer's sales. Finally, repetition enables the authors to analyse the impact of learning. The views of this paper are those of authors and do not necessarily reflect the views of the OFT nor the legal position under existing competition or consumer law which the OFT applies in exercise of its enforcement functions. Rather the aim of the report is to shed some evidence on this interesting issue, and promote economic debate in this area. This report is part of the OFT's Economic Discussion Paper series. If you would like to comment on the paper, please write to me, Amelia Fletcher, at the address below. The OFT welcomes suggestions for future research topics on all aspects of UK competition and consumer policy. Dr Amelia Fletcher, Chief Economist Office of Fair Trading, Fleetbank House, 2-6 Salisbury Square London EC4Y 8JX Amelia.fletcher@oft.gsi.gov.uk

CONTENTS
Chapter/Annexe Page

1 Executive summary 2 Introduction 3 Review of emprical and theory literature on price framing 4 The experiment 5 Experimental results 6 External validity and implications for non-Laboratory markets 7 Policy recommendations A Screen shots from the experiment B Web trawl of pricing practices C Theory: Optimal search strategy D Regression details E Personality test A feedback questionnaire B Experiment instructions C References

5 15 21 36 54 88 92 94 109 126 132 164 165 166 170

1
1.1

EXECUTIVE SUMMARY
This report presents a controlled economic experiment that analyses whether or not the way prices are presented or 'framed' to consumers has effects on consumer decision making and consumer welfare.1 The experiment was completed by London Economics with Steffen Huck and Brian Wallace.2 The experiment was designed to create a comparative metric alongside which each of the frames can be compared with a baseline treatment where sellers show clear per-unit prices. This comparative metric enables a ranking of price frames in terms of the potential magnitude of impact on consumer decisions. The experiment isolates the effect of price frames and as such, is not constructed to explicitly examine their potential benefits. There may, however, be benefits if these price frames reflect genuinely better offers or efficiencies.

1.2

1.3

1.4

The price frames

1.5 The price frames tested in this controlled experiment are the following: A baseline treatment in which consumers see straight per-unit prices.

For information on what experiments in economics are see, OFT 2009 and Ofcom 2010. A brief explanation is provided in chapter 2 of this report. Steffen Huck is a Specialist Consultant with London Economics, and Head of the Economics Department at University College London. Brian Wallace is a Senior Research Fellow at University College London.
2

OFT1226 | 5

Drip pricing where the consumers see only part of the full price up front and price increments are dripped through the buying process. Sales in which a sale price is given and a pre-sale price is also given as a reference to the consumer, 'was 2 is now 1' (actual prices are identical to the baseline treatment). Complex pricing where the unit price requires some computations, '3 for the price of 2'. Baiting in which sellers may promote a special price but there is only a limited number of goods actually available at that price. Time limited offers where the special price is only available for a pre-defined short period of time.

The experiment design

1.6 The experiments were conducted at the University College London (UCL) Experimental Laboratory, and used 166 participants (also called subjects) drawn from across the UCL student population. Each subject participated in the unit-pricing baseline and two of the five price frames. Subjects participated in 30 rounds of the experiment, with each subject experiencing each of his three frames 10 times. This repetition allows subjects to learn about the experiment and adapt their behaviour. This allows us to determine whether any potential effects of the price frames can be overcome through experience. We discuss the effect of learning below. The full design is presented in detail in Chapter 4.

1.7

1.8

Results
1.9 The evidence from the controlled experiment shows that, in contrast to the predictions of standard economic theory, price frames do matter for consumer decision making and welfare. Consumers make more mistakes and achieve lower consumer welfare under the price frames we investigate as compared to straight unit pricing (the baseline).

OFT1226 | 6

1.10

We use a number of measures to compare across the price frames. Each measure is explained in detail in Chapter 5. Here we provide an overview.

Ranking of the price frames 1.11 Using the consumer welfare metrics (Chapter 5), we can rank the price frames (1 - 5) in terms of which cause the worst problems, defined as welfare losses, for consumers. Welfare losses are a measure of actual monetary earnings in each of the 5 price frames and the baseline treatment as compared to potential monetary earnings if subjects behaved optimally. Optimal behaviour is what a fully rational subject that knew the experimental environment would do. The ranking of the price frames, starting with the worst that which causes the greatest welfare loss - is as follows: (1) drip pricing (2) time limited offers (3) baiting (4) sales, and (5) complex pricing. 1.13 If we consider why there are behavioural problems giving rise to consumer welfare losses we can explore two types of errors the subjects may be making. First, subjects may search 'too much' or 'too little' between shops before making a decision to buy. Searching 'too much' or 'too little' is a measure of actual search behaviour compared to optimal search behaviour. Optimal search behaviour is what a fully rational subject that knows the experimental environment would do in each of the frames and the baseline. The optimal search strategy can be derived analytically and is shown in the annexe. Second, consumers may buy 'too many' or 'too few' units of a good given their utility (or pay-off) function. Again 'too many' or 'too few' units is a measure of actual

1.12

OFT1226 | 7

purchasing behaviour in each of the price frames and the baseline as compared to optimal purchasing behaviour. 1.14 We observe that drip pricing and time-limited offers which generated the biggest welfare losses are also the price frames in which subjects make the most errors. The most prevalent error under these two price frames is that subjects buy at the first shop at prices that are too high, that is, at prices where they should continue their search. Purchasing errors are less frequent but do occur. Buying too many or too few units occurs equally often and the overall sales volume is virtually constant across the different frames. We also analyse the impact of search costs on the quality of consumer decision making. There are two main effects. First, search errors decrease as search costs increase. The intuition is that buying at the first shop tends to be optimal for higher search costs. Hence, consumers who tend to buy at the first shop 'come-what-may' behave optimally if search costs are indeed high, but not if search costs are low. Second, however, search costs have the opposite impact on purchasing errors. It seems if consumers go through more costly (longer) search this raises the chance that they buy more units of the good than they optimally should. Intuitively this is because consumers fall prey to the sunk cost fallacy, that is,, they appear to justify earlier incurred (sunk) costs through high levels of consumption. Nonetheless, overall it is the first effect on search errors which dominates implying that price frames have a stronger (worse) effect on in markets with low search costs.3

1.15

1.16

Learning 1.17 If we consider subjects' behaviour over time, we can assess if any learning happens and if errors decline as subjects become more experienced.

With very large search costs the effect of the sunk cost fallacy could, of course, start to dominate such that the overall effect of search costs on consumer detriment from price frames could well be U-shaped.

OFT1226 | 8

1.18

We observe that errors generally do decline as experience increases although at a slowing rate such that learning cannot completely eradicate the problem. Moreover, in the time-limited frame there is no learning at all. Likewise, we observe that in all frames consumer welfare losses decline (that is, subject pay-offs increase) over time except in the time-limited frame.

1.19

Behavioural forces 1.20 What are the behavioural forces driving errors and welfare losses? The data from the experiment allows us to explore this in detail. Further, the responses to the voluntary qualitative questionnaire, conducted at the end of the experiment, also shed some light. In summary the behavioural forces at work can be identified as follows: 1) Drip pricing: If a consumer sees a low base price and they make the decision to buy the good, they shift their reference point because they imagine already possessing the good. Later, when they realise that there are additional costs and charges, it is thus more difficult for them to give up the good which they already have 'in their basket'. Therefore, they purchase the good despite the increase in price. This is in line what is known in the literature as loss aversion or the endowment effect. Subjects reported in the questionnaire feeling disappointed in this frame because they felt they were receiving a good deal when they saw the base price. Subjects reported that they still bought the good after they found out the additional charges, but felt cheated and annoyed because their pay-off was reduced. 2) Time-limited offers: Consumers believe erroneously that if they leave the store then the prices will go up (note: the prices may go up or down in this experiment), as such they have a tendency to buy at the first store they go to; or, if they do not buy at the first store, they will buy at the second store if the price at the second store is at all

1.21

OFT1226 | 9

profitable. This erroneous pre-conception is self-enforcing. As subjects do not return to the first shop they cannot learn that their beliefs are false. The main behavioural issue in this frame is, hence, of a cognitive nature. The responses support this finding, subjects reported that they felt 'compelled to buy' or it was 'something not to be missed'. Some reported they were enticed to buy without further searching. Many reported being confused and those who did venture back to the first shop reported finding it strange that the price upon return to a shop may be lower than when they first visited. 3) Baiting: Consumers choose a shop using the advertising they see (this is the only time subjects see any price information before going to a shop). They choose the store based on the deal they are being offered in the advertisement. This choice raises their willingness to pay because they expect to get a deal and again envisage owning the good. Not buying the good at a higher price would be perceived as a loss. Therefore, it is once again loss aversion or the endowment effect working that is driving consumer behaviour. Subjects reported that this frame enticed then to go quickly to the shop with the best deal so as not to miss out. They reported anger and frustration if they missed out on the offer in the experiment, however, they reported they may buy anyway. 4) Complex pricing: Consumers tend to buy at the first store they visit, and they tend to buy the offer even if it may not be the best thing for them to do (for example, it may be more profitable to simply buy one unit instead of the '3 for 2' deal). These errors can be attributed to cognitive failure. They have a significant yet comparatively small consequences for consumer welfare.4

The effect on consumer welfare loss due to this price frame is small but significant (at five per cent) as compared to the welfare loss incurred in the baseline (Table 5.1). The effect on search and purchasing errors are not significant (Tables 5.3 and 5.4).

OFT1226 | 10

Subjects reported that they did realise that in some instances it may be better not to buy the deal but instead buy just one unit of the good. However, 32 per cent of those who chose to answer this question reported that they always chose to buy the offer. 5) Sales frame: While this frame had no significant effect on consumer welfare there was a small and significant increase in deviations from the optimal search rule. The meaningless information 'was X is now Y' seems to be used by consumers although it clearly should not be. Use of information that should be discarded is a cognitive error that drives behaviours in this frame. 41 per cent of subjects reported that the special offer did influence their decision making. These subjects reported that they were 'getting a good deal', 'receiving a discount', 'saving money' and therefore were more tempted to buy the product.

Sellers
1.22 Shops in this experiment were computerised and as such they did not respond endogenously (within the experiment environment) to the behaviour of the consumers.5 We observe that the overall sales volume of both shops taken together is virtually constant across the different frames (Table 1.1), and that in time-limited offers, one of the most detrimental frames for consumers, firms actually experience reduced sales as compared to the baseline. The effect on shops of the different price frames is in the distribution of sales. We find that the first shop visited gains in drip pricing, time-limited offers, baiting and complex pricing (Table 1.1). Therefore, firms may engage in activities to

It is possible to have both consumers and firms active in experiments.

OFT1226 | 11

encourage consumers to visit their shops first when elaborate pricing frames are used.6 1.23 Sales for complex pricing (Table 1.1) are substantially higher but whether this would translate into higher profits for sellers depends on unit costs which are unmodelled in this experiment.7

Summary of main findings

The main findings for each price frame, as discussed above, are summarised in the table overleaf.

In complex pricing subjects tend to undersearch and buy too often at the first shop they visit. It is not that in this frame subjects buy too many units of a good, but that they buy at all when it would be better for them to continue to search. These errors are small (Table 1.1) but nonetheless our observations suggest that there is an attraction created by complex pricing which entices consumers to buy the first offer they see. The sellers do not have production costs. The prices sellers offer to consumers is selected randomly from a pre-determined distribution of prices. Of course it would be possible to introduce endogenous firms into the experiment in order to test firms' response to consumers' decisions. Not including firms was a design choice made at the beginning of this experiment.
7

OFT1226 | 12

Table 1.1: Summary of main findings

Price frame Significantly more purchasing errors Significantly more search errors Significant consumer welfare losses Number of units sold by shops Sales by shops First shop visited benefits? Behavioural biases Learning helps

Baseline Drip pricing Yes, substantially Yes Large

1.58 1.59

125 124 Yes Endowment effect/loss aversion/maybe sunk cost fallacy Cognitive errors/maybe sunk cost fallacy Endowment effect/loss aversion/sunk cost fallacy Cognitive errors/sunk cost fallacy Cognitive errors/sunk cost fallacy Yes

Time-ltd offers

Yes

Yes

Medium

1.59

121

Yes

No

Baiting

Yes

Yes

Medium

1.61

124

Yes, strongly

Yes

Complex pricing

No

No

Small

2.73

221

Yes

Yes

Sales frame

Yes

Yes

No

1.59

125

No

Yes

Note: The measure 'significantly more purchasing errors' shows if subjects in the experiment either bought 'too many' units or 'too few' units than was optimal in the relevant price frame as compared to purchasing errors that subjects made (buying too few or too many units than was optimal) in the baseline treatment with straight per unit pricing. 'Significantly more search errors' shows if subjects in the experiment either searched 'too little' or 'too much' than was optimal in the relevant price frame as compared to search errors that subjects made (searching too little or too much than was optimal) in the baseline treatment. 'Significant consumer welfare losses' shows if actual earnings compared to optimal earnings in each price frame are less than actual earnings compared to optimal earnings in the baseline. 'Number of units sold by shops' reports the number of units sold by the two shops in this experiment. 'Sales by shops' reports the turnover for both shops in the experiment. 'First shop visited benefits' reports if the distribution of sales in this experiment is to the advantage of the first shop visited under each price frame. 'Behavioural biases' shows which behavioural bias is causing actual behaviour to differ from optimal behaviour in each price frame. 'Learning helps' indicates if errors in purchasing and/or search errors decline over time as subjects gain more experience in the price frames.

OFT1226 | 13

Policy
1.24 From a policy point of view, this experiment has several broad implications. These are the following: Price frames do change consumer behaviour which is in contrast to what orthodox economic theory on consumer behaviour would predict. Furthermore, change in behaviour is for the worse. Price frames do cause consumer detriment. The worst culprits are drip pricing and time-limited offers.8 Overall, the negative effects on consumer welfare are greater when search costs in the market are low. However, with higher search costs consumers become prone to the sunk cost fallacy and this effect could become much stronger if search costs are substantially higher. Due to learning through experience the adverse effects of all frames except time-limited offers are presumably greater in markets for less frequently purchased items while time-limited offers can also do harm in markets for goods that are purchased with high frequency.

1.25

Therefore, policymakers may be particularly worried about drip pricing in markets for goods that are not purchased (particularly) frequently and time-limited offers in markets for goods that are purchased with high frequency.

Particularly when the offer is not a genuine offer. As is discussed in detail in Chapter 5, consumers believe the shop's claim in this price frame, but if the claim is not true, and in fact prices can decrease below the offer (after the offer has expired), consumers never learn this because they do not go away and then return to shop again (in order to learn).

OFT1226 | 14

2
2.1

INTRODUCTION
While prices are assumed to take on a crucial role for economic behaviour, their presentation, or framing, has historically received little focus in economic research because how prices are presented has been deemed irrelevant. Yet, sellers use many different ways of presenting or framing prices in the marketplace and they often change these frames.9 As changing the presentation of prices is costly, we find ourselves in the paradoxical situation where standard economics suggests on the one hand, that price framing is irrelevant for consumer behaviour but must, on the other hand, have some impact on demand because otherwise firms would not spend money on manipulating them. This study is designed to throw some light on this paradox. The question we ask is simple and entirely empirical. We simply set out to analyse whether different price frames do or do not have an effect on consumer choice. Of course, one would expect that the literature should have long since answered this question but as our literature review reveals this is not the case (section 3). Existing evidence on price frames is sparse, scattered and often based on questionable data. In fact, we believe that this is the first study that provides comprehensive evidence on the impact of price on consumer behaviour in a systematic, fully incentivized environment. Specifically, we use an experimental approach studying the impact of price frames on consumer choices in a laboratory environment. The box below provides a brief summary of what experiments in economics are.

2.2

2.3

The annex presents a web trawl of price frames actually used by different shops in seven different consumer good markets: airfares, package holidays, accommodation and hotels, theatre tickets, electronic goods, computers and equipment and furniture. Industries chosen are illustrative only; the experiment is concerned with the impact of pricing frames in general and not just for these industries listed. The web trawl was used as background information to inform the design of the treatments for this experiment.

OFT1226 | 15

Box 1: Experiments in economics

Economic experiments allow policy-makers to: Test if there are problems in markets. For example, do consumers have problems making the best decision under different price frames as tested in this experiment for the OFT. Pre-test policy interventions and remedies to either compare alternative interventions or to fine tune a given intervention. See for example the Ofcom experiment, 2010, used to test the performance of alternative price revelation mechanisms on consumer behaviour in telecommunications markets. Ex-post checking of existing policy interventions in order to identify if the interventions are not working as well as they could, and they could be changed to improve the outcomes in our economies. See for example the HewlettPackard experiment, 200110, used to test the impact of minimum advertised prices on HP retailer behaviour. Experiments are an empirical method, and act as a complement to field data collected via survey methods, and simulations that use computing power to model the behaviour of consumers and firms. The hallmarks of experiments in economics are control, treatment and replication. Control means individual decisions made in the experiment are induced by the incentives created in the experiment and by no other factors. Treatment is the ability to change specific incentives or features of a policy and to identify how individual decisions change as a result, thus, establishing true causality, that is, why behaviour is changing. Replication is the ability for the experiment to be conducted multiple times by the same researcher, by different researchers and across different populations, in order to verify the results. Experiments use real people to make real economic decisions in controlled

10

Charness.G. and Chen.K. (2001).

OFT1226 | 16

environments. Experiments induce behaviour by creating real monetary incentives (and disincentives) that are directly linked to the decisions that participants (also called subjects) make in the experiments. This is called induced decision making and incentivisation which alternative survey methods do not employ. Experiments set-up simplified environments that mirror complex real world environments. Experiments are designed such that they focus on the main behavioural drivers of interest and reduce other confounding factors on behaviour. This simplification allows the experiments to establish causality. Causality is the identification of what is really driving behaviour and why. The two main criticisms of experiments are: (1) representativeness of the participants, and (2) extrapolation beyond the laboratory. In regard to the former, subject pool selection is an area of continued debate and research in experimentation both in psychology and economics. Overall the research suggests that the fundamental underlying behaviour, if it is robust, given the incentives tested in the experiment, will hold across all types of subject pool. However, there may be some difference in the magnitude of outcomes if the subjects have had prior experience, or if the cognitive ability of subjects is different. We discuss subject pool choice for this experiment completed for the OFT in chapter 6. Extrapolation is often addressed by concept of parallelism where the experimenter tackles the criticism through conducting new experiments that investigate the precise nature of the criticism. For example, if the critique is that certain experiments do not have external validity because in real markets there are many more participants, the researcher can carry out new experiments doubling the number of market participants to find out whether this is a real issue or not. Again we discuss extrapolation of this experiment for the OFT in chapter 6.

OFT1226 | 17

2.4

This environment retains all the crucial features of real consumer choice problems: Goods are on offer at multiple shops, consumers might want to buy single or multiple units and they can search among the different shops. At the same time, the environment is simple and highly controlled which allows us to make clean-cut inferences on how price frames affect consumer behaviour and, ultimately, consumers' welfare. Notice that our design isolates the pure effects of framing, that is, the presentation of prices, holding the actual price levels constant. In many real-life markets there are sometimes genuine offers that may or may not be framed in one of the ways that we study here. In these cases the effect of such offers on consumers is the following: there is the direct (positive) effect of lower prices and in addition there may be adverse effects from the price framing. In this study we isolate the latter. (Only in the case of '3 for 2' do we examine actually lower prices.) Similarly, we isolate the pure effects of each price frame. In real-life markets, some of these frames might appear in combination with one another). Moreover, consumers might have to compare different frames which would be another complication we abstract from in this study. Our main result is that how prices are framed (or presented) has a profound impact on consumer choice. We compare simple per-unit pricing ('This good costs 1 per unit') with five other price frames: sales ('was 2 is now 1'), complex pricing ('3 for 2'), time-limited offers ('1 only today'), baiting ('1 while stocks last' with the possibility that stocks are very small such that consumers are very likely to face a substantially higher price once they check out the offer) and drip pricing (where the effective full price of a good is only revealed in 'drips', for example, with shipping and handling charges that have to be paid on top of the basic price). We find that all of these practices have some adverse effect on consumer behaviour and that most of them significantly impact on consumer welfare. Our research also identifies two price frames that are particularly detrimental for consumers: time-limited offers and, even more so, drip pricing. As we will discuss in some detail these are extremely strong

2.5

2.6

2.7

2.8

OFT1226 | 18

results as they have been detected in comparatively weak treatments in a comparatively smart subject pool. If anything, our experiment will, therefore, underestimate the true harm done by these practices. 2.9 We also discuss how our findings relate to known biases in economic behaviour and how they are likely to impact on market outcomes and competition between firms. In our experiment human subjects (UCL students registered on a database for economics experiments) participate in a sequence of oneperson decision tasks that mirror the repeated purchasing of different goods. Before we go into the detail of the experimental design it is perhaps useful to offer a rough overview of how the experiment works. Subjects work on their own on a computer screen and are rewarded for their performance. Specifically, they receive a monetary payoff (reflecting consumption utility; the satisfaction one derives from the consumption of a good) for each unit of each good they buy but, of course, they also have to pay the price that the shops charge. In the entire experiment there are always two shops. A priori, these shops are entirely identical. Subjects who play the role of consumers have to choose which shop to visit first and we will call this shop the first shop and the other the second shop. Going to a shop involves some (monetary) search costs reflecting the time costs of search on the internet or the actual travel costs in the case of visiting physical stores. Our main experimental manipulation is how prices are framed when the consumer subject arrives at a shop. In our baseline subjects see simple per-unit prices. The five more elaborate price frames that we analyse are then modelled on price frames used in the field. The precise wording is explained in detail further below. With this experiment we can examine the effect of price frames on consumers in two ways. First, we can simply examine whether price frames have an effect on consumer welfare (reflected in the experiment by subjects total pay). Since some pricing strategies convey a real advantage or disadvantage to consumers (for example, under '3 for 2' effective prices are really lower) we will always compare the achieved

2.10

2.11

2.12

OFT1226 | 19

welfare to the welfare that would be obtained under optimal behaviour. Optimal behaviour can be precisely computed for all six different environments and we will explain it in some detail in this report. 2.13 The second way of looking at our data is by comparing actual choices with the optimal strategy. This allows us to identify errors in consumer decision making and how they are triggered by the different price frames. A more detailed analysis of errors also allows us to indentify where in the process errors occur which is important for understanding the root causes of the consumer detriment. It turns out that the two main root causes for errors in consumer decision making are to be found in a phenomenon called the endowment effect (where consumers willingness to pay shifts once they believe themselves already in possession of a good) and in cognitive errors where available information is simply not processed properly and consumers beliefs about what shops are actually doing are severely biased. In what follows we shall first review the surprisingly small literature on price frames. We then move on to our experiment, spending some time on the precise setup and design before analysing the data. Finally, we discuss the implications of our findings for markets outside the laboratory and offer some policy advice.

2.14

2.15

OFT1226 | 20

REVIEW OF EMPRICAL AND THEORY LITERATURE ON PRICE FRAMING

From this review, the first thing to notice is that the literature on price framing and the impact on consumer behaviour is surprisingly patchy. In particular, the empirical evidence available is thin. Even within the marketing literature the effects of different price frames have not been systematically explored. Moreover, some of the existing studies suffer from methodological problems. Methodological problems include, for example, using purely hypothetical choice problems or relying simply on perception or recall data. It is well-known that in many contexts the link between hypothetical choice and actual choice is rather tenuous and questionnaire and perception data often unreliable (see, for example Hertwig and Ortmann (2001)).11 We also discuss some of the literature on price transparency which is indirectly related. If price framing makes price comparisons more difficult this could amount to reduced price transparency in markets. Again there is not a large amount of evidence on the role of price transparency but there are some useful sources that we briefly discuss. The relevant theory papers, investigate how firms would optimally choose prices and price frames if consumers struggle with comparing framed prices. However, it is not perfectly clear how safe the assumption that consumers struggle in comparing prices is.

3.1

3.2

3.3

11

For a more detailed discussion of the methodological strengths and weaknesses of different qualitative and quantitative methods for testing consumer behaviour see also 'Road Testing Consumer Remedies', OFT, 2009 (www.oft.gov.uk/advice_and_resources/resource_base/economic-research/)

OFT1226 | 21

3.4

We present this literature below.12

Empirical evidence on price framing

3.5 There is some limited evidence on some forms of price framing. This evidence comes from experimental research and from consumer surveys. Does framing matter? 3.6 On a fundamental level the literature suggests that, in supermarkets, consumers often grab products without checking the price at all (Dickson and Sawyer 1990). Consumers knowledge or recall of prices immediately after shopping appears poor. Of course, if this is indeed true it would mitigate any effect of different price frames. If consumers grab products without checking prices, price frames become irrelevant.13 Partitioned pricing (drip pricing) 3.7 One price frame which has received some attention is partitioned pricing (which covers drip pricing).14 Morwitz, Greenleaf and Johnson (1998) show how partitioned pricing lowers consumers recalled price for the good and increases demand. For example, in a fully incentivised auction experiment in which participants bid for a jar for pennies, one group of participants/buyers was given a bid form which told them that they must

12

Notice that we do not discuss the wider literature on behavioural economics and biases in decision making. While this literature is mostly based on evidence from laboratory experiments there are also some studies that have shown biases in real-life consumer decision making (DellaVigna and Malmendier (2006) demonstrate, for example, the role of biases in time preferences for taking out gym memberships.) For a general overview of field evidence on behavioural biases, see DellaVigna (2009). Notice, however, that the interpretation of such recall data crucially rests on what you are willing to assume about consumers memory. It is, of course, entirely possible that consumers pay close attention to the prices while picking goods in the aisles and then simply forget the prices when they check out. This also covers to opt-in/opt-out pricing.

13

14

OFT1226 | 22

pay 15 per cent in addition to their bid if they win the auction. A second group of participants was given a bid form that told them that the price they will pay if they win will be their bid price. Buyers premiums are often charged by auctioneers, and such a set-up is a form of partitioned or drip pricing because the buyer does not see the total price of the good but instead must calculate it from its parts. Morwitz et., al, found that the partitioned pricing increased demand for the good as compared to the situation where the buyers premium was not separated. In a second experiment, subjects were shown the price of telephones and then later asked to recall the total price of the different phones. One group of participants was shown the total price of the good including any surcharges such as shipping and handling, while a second group was shown partitioned prices where the surcharges were reported separately to the base price for the phones. The surcharges were either reported as the base price plus the surcharge in dollars (that is, $69.85 plus $12.95 for shipping and handling) or base price plus the surcharge in percentage terms (that is, $69.95 plus 18.5 per cent for shipping and handling). Participants in the partitioned price group consistently recalled lower total prices as compared to those in the total price group. Further, those given base price plus surcharge in dollars recalled higher total prices as compared to those given base price with the surcharge in percentage terms. 3.8 One reason for these observations put forward by Morwitz, et al, is the use of heuristics by consumers. Namely, consumers may regard the base price and the surcharge as separate pieces of information. Therefore, instead of calculating the mathematical summation of all the price parts (which may require some cognitive effort) consumers are anchoring to the first piece of information seen which is generally the base price and then attribute less importance to later pieces of information (the surcharges). An interesting field experiment was conducted by Hossain and Morgan (2006). This was a natural field experiment implemented using 'eBay'. Buyers in an eBay auction made bids for CDs and Xbox games. In the auctions there was a reserve price (that is, the minimum price that will be accepted by the seller such that bid prices below the reserve will not

3.9

OFT1226 | 23

be accepted) and separate (fixed) shipping and handling costs. In this field experiment it was observed that when the reserve price (which is the opening bid essentially) is low as compared to the retail price of the good, and the shipping and handling costs are high, then the auction always results in a higher sale price than a situation in which the reserve price is high (relative to retail) and the shipping and handling is low. The authors present a number of theories that may explain this behaviour. Bidders may be loss averse, which is similar to the anchoring argument posed by Morwitz above, namely that consumers treat the base price separately from the shipping and handling fees and focus on the base price in order to win the good they want and have invested effort in bidding for. Therefore, the consumer imagines they have the good, they have 'won' it, and when they subsequently see the additional costs it is hard for them to give the good up (lose it). Further, consumers may simply ignore or miss the additional costs. 3.10 In a follow-up paper Brown, Hossain and Morgan (mimeo 2007) sold different iPod models on Yahoo Taiwan and eBay Ireland. They find that shrouding low shipping charges is a money-losing strategy for the seller but that raising shrouded shipping charges does increase revenue. In this instance shipping charges were either shown on the search page for each auction or they were shrouded because buyers had to read each individual auction page to learn the shipping costs. Brown, Hossain and Morgan attribute their findings to a model where some consumers are fully rational while others are either naive and overlook the shipping charge or they are suspicious in the sense of always assuming that shipping charges are very high. These studies imply that drip pricing may generate bias in behaviour such that consumers end up paying more, and buying more, when priceparts are dripped as opposed to a total price. In addition the relative size of the drips may be important. If the additional surcharges are a large proportion of the total cost then buyers may take greater account of the surcharges when considering the base price, and as such the drip pricing may have less effect on behaviour.

3.11

OFT1226 | 24

Reference pricing 3.12 The marketing literature suggests that consumers use reference prices when they evaluate goods that are on offer. These reference prices may be internal (and be based on advertising, associations the product generates or on remembered prices) or they may be external (prices formerly charged for the product that are still on display). Unfortunately, it is not always clear what such reference prices would mean from an economic point of view and whether they can be reliably inferred from the data. Evidence of reference pricing suggests that consumers are sceptical when they see an offer/sale price and a reference price, but not sceptical enough for the reference price not to be potentially misleading. One such study by Blair and Landon (1981) ran a field experiment in a shopping centre. The experiment was not incentivised, but instead used presentation cards to elicit respondents valuation of savings on electrical goods when exposed to an offer price with a reference price (shown as, for example, list price $69.95, sale price $44.95) as compared to an offer price with no reference (shown as sale price $44.95). Participants that were shown only the offer price perceived that the saving due to the offer was lower as compared to participants that were shown the offer price and the reference price. However, those that saw the reference price did not fully believe the reference. For example, the offer price for a television was shown as $459. The reference price was shown as $544. In this study the authors found that respondents who saw only the offer price reported that they believed the average saving from the full retail price was $58, while those that were also shown the reference price reported a believed saving of $65, but not the full referenced difference of $85. This observation held across well known brands and lesser known brands. Therefore, in this study the authors observe that reference price claims are discounted by about 25 per cent, but do have an impact on consumers' perceptions of the savings they are receiving. In a survey of research on reference pricing Biswas, Wilson, and Licata (1993) also report mixed evidence. While the majority of studies do find

3.13

3.14

OFT1226 | 25

effects of reference pricing on consumer perceptions there are several that do not. 3.15 A fairly recent study by Chandrashekaran and Grewal (2006) shows how subjects internal reference prices (their perceptions of 'fair prices', market prices, and 'acceptable prices') respond to the observation of sales prices. It is shown that subjects internal reference prices are revised in the direction of sales prices. However, these results are not derived from choice data, nor is it entirely clear how these adjustments would affect actual choice. In other words, it is entirely unclear whether this finding has any implication for consumer behaviour and welfare. A study completed by the University of Nottingham Business School for the Office of Fair Trading in 2005, used non-incentivised web-based experiments to test consumers beliefs about different (external) reference prices for four product types televisions, holidays, chocolates and books. The study implemented three different experiments. Experiment 1 compared between time limited (external) reference prices and non time limited. Experiment 2 explored how the reference price was presented - as an absolute discount, a percentage discount or a discount in relation to the recommended retail price. Experiment 3 compared sales in stores to sales via the internet. Five outcome variables were analysed search intentions, purchase intentions, transaction value (the consumers perceived value of the deal), acquisition value (the satisfaction or utility derived from the good as compared to the price paid for the good) and quality (the consumers perceived quality of the good). The experiments observed that for situations where there is a high reference price and the price of the product is 'high' (that is, in the case of televisions and holidays but not books or chocolates) then transaction value, purchase intentions, acquisition value and quality perceptions are significantly greater than for the same product when the reference price is smaller. The authors suggest that this indicates that high reference prices can lead to consumer detriment by artificially increasing transaction value, acquisition value and quality and this is particularly strong for goods which are a larger portion of a consumers budget (high ticket items).

3.16

OFT1226 | 26

3.17

Alford and Biswas (2002) report that consumers psychological makeup drives the extent to which they are influenced by offers. Specifically the authors look at two psychological features which they call price consciousness and sale proneness. Price consciousness is the degree to which the consumer focuses exclusively on paying a lower price, such that a highly price conscious consumer will have lower perceptions of the offer value when faced with a reference price. In other words they will discount the reference claim to a greater extent. Sale proneness on the other hand, is the degree to which the consumer will respond to an offer because of the form in which the sale price is presented. Consumers that have high sale proneness will have higher perceptions of the offer value. In other words, they will attach less discount to the reference price. In this study the authors did find that participants who were ranked as price conscious placed less weight on the reference than participants who were ranked as low price conscious while participants who were ranked as sale prone put more weight on the reference price. Interpreting these findings requires some care. First, as most studies in this field it is based on a purely hypothetical choice situation, not real (incentivized) choices. Second, the data analysis ignores the problem that 'psychological makeup' is endogenous. Whether a consumer is price conscious or not may be driven by the same factors that also drive how they respond to reference pricing (for example, their degree of impatience). In other words, there is no proof of causality.. A number of studies have investigated if the form of the offer matters. DelVecchio, Krishnan, and Smith (2007) analyse the effect of using cents-off promotions versus percentage-off promotions on participants perceptions of the non-sale price. They ran a number of experimental studies using university students and the goods shampoo and canned drinks (however these experiments were not incentivised). The authors observed that a cents-off frame lead participants to estimate a lower non-sale price as compared to a percentage-off frame, even though both frames resulted in the same savings as compared to the non-sale prices. Therefore, in this study a percentage-off frame leads consumers to believe the savings are greater as compared the use of cents-off promotion.

3.18

OFT1226 | 27

3.19

The Nottingham study (discussed above) also investigated the form of offers. The observations were mixed. Namely, when the reference price was high and the ticket price of the good was high (that is, TVs and holidays but not chocolates and books), then using the recommended retail price as a reference did have an effect on consumers perceived value of the deal (transaction value) and their intention to purchase. However, when considering acquisition value (defined above) using the recommended retail price as a reference had the greatest impact when discounts were low to medium for televisions and high for holidays. Of course, one aspect of promotions and sales is that they simply change the relative prices of goods within one store and a key strategic decision that store owners have to take is which goods, and how many alternatives, they want to offer consumers. A review of these issues can be found in Simonson (1999). One relevant finding Simonson discusses is that there is more switching from low-quality to high-quality options if the high-quality option is advertised than the other way round. Of course, this might be a simple consequence of the relevant cross-price elasticities. There is, not necessarily, anything particularly mysterious about this. The limited literature on reference pricing implies that consumers can be influenced by a reference price, and thereby perceive that the saving on the good may be greater than if they did not have a reference price. It also suggests, however, that consumers are sceptical of reference prices and thereby discount their beliefs about the saving represented by the reference.

3.20

3.21

Time-limited offers
3.22 There is very little published research in this area. Devlin, Ennew, McKechnie, and Smith (2007) analysed time-limited promotions in the context of durables (TVs) and came up with a very clear conclusion: Time-limited price offers do not impact upon consumers perceptions of a goods value, nor do they have an impact on consumers' behaviour in terms of search and buying choice. As such this study suggests that

OFT1226 | 28

marketing managers, policy makers and regulators should not be concerned about this form of pricing. 3.23 Cialdini (2009) draws upon the psychology literature to explain why sellers use time limits when advertising products. Cialdini argues that time limits, which state that the price offered will not be available at a later time, are used by sellers to generate a want for the product even if previously there was none. Cialdini basis his explanation for this behaviour upon the proposed human belief that things which are difficult to get are typically better than those that are easy to get and as such the availability of an item is a 'short-cut' method for determining quality of a good (Lynn, 1989). Further, Cialdini asserts that humans do not like to lose their freedoms, which psychologists term 'reactance theory', and as such whenever humans feel their choices are limited the need to maintain choice, and the goods and services associated with the choices, increases (Brehm and Brehm, 1981). Therefore scarcity invoked through time limits aims to interfere with humans prior access to a good such that the human will react against this loss of freedom by wanting the good more. Cialdini provides anecdotal examples of this effect electrical goods, gym memberships and automobile sales - in his book Influence: Science and Practice. Complex pricing and baiting sales 3.24 We were not able to locate meaningful evidence on the effects of complex '3 for 2' pricing. There is, however, an interesting study on the effects of 'multiple-unit pricing' where offers are explained through the price you get when buying more than one unit ('4 units for 2' instead of '50p per unit'). Wansink, Kent, and Hoch (1998) report results from a field experiment to compare such multiple-unit presentations of offers with straight single-unit offer frames. Utilizing a rather impressive data set from 86 stores that are randomly assigned to one of the two offer frames they find that, for a wide range of products, sales under the multiple-unit frame were, on average, 32 per cent higher than under the single-unit frame. The multiple-unit frames typically used the format 'buy two for x'. The observed effect is not only highly significant but also very large. Rational consumers would, of course, not exhibit this effect

OFT1226 | 29

provided there is no confusion about whether the offer is also effective if you buy a single unit (that is, that the per unit price is not higher if a single unit is bought). In all cases, the offer would have been effective for single units. But the authors cannot rule out that shoppers actually perceived prices to be non-linear (in which case this would provide evidence for complex pricing). If shoppers thought they would have to pay a higher per-unit price for a single unit, then increasing demand when faced with the 'buy two for x' pricing might be perfectly rational. (Notice that, technically speaking, consumers choice sets would be nonconvex.) Unfortunately, the study does not have exit interviews that could have clarified this issue. Without exit interviews it is impossible to understand what is really driving this result; misperception of the nature of pricing or some sort of confusion triggered by the multiple-unit pricing. 3.25 Similarly, we could not locate much empirical work on the effects of traditional forms of baiting. However, Ellison and Ellison (2005) analyse a market characterized by the presence of price search engines. Price search engines are designed to make consumer search and price comparison easier, and as such, reduce friction in the market such that the price of identical goods should be identical. However, retailers may seek to put friction back into the market by making price search more difficult and thereby less of a threat to profitability. One such method is to prevent price comparison sites including the listed price. Ellison and Ellison, using field data from an actual price comparison site in the US observed a more subtle and highly successful version of 'bait and switch' where sellers advertise a low-quality product at a very low price as a bait and then try to convince consumers to switch to better quality more expensive products once they are in their online store (despite the cheap product being available).

Conclusions from the empirical studies

3.26 On balance we find the evidence on price frames rather thin and shaky and believe that this OFT research will make a large contribution to the literature.

OFT1226 | 30

Empirical evidence on price transparency

3.27 There has not been much research on the effects of price transparency for consumer behaviour and market outcomes. There are however a few notable exceptions, most of which focus on the strategic behaviour of firms and the resulting impact on consumers as opposed to specifically testing consumer behaviour. Davis and Holt (1996) test the Diamond paradox that predicts that for homogeneous goods (that is identical goods) the smallest consumer search costs can raise equilibrium prices from the competitive to monopoly levels. This extreme prediction is not borne out in the laboratory data although market prices are increasing in search costs. More support for the Diamond prediction is reported in an experiment by Cason and Friedman (2003). For policy applications this implies that relatively small search costs are less of a worry than theory predicts. Rather, the amount of attention the regulator might want to pay to search costs should be increasing in these search costs. Lynch and Ariely (2000) study the role of price transparency in a field experiment with online wine retailers. They show how lower costs for price searches increases price sensitivity of consumers. However, they also show that this does not necessarily induce ruinous price competition, where firms prefer to exit the market, if it is offset by a simultaneous increase in quality transparency. Therefore, in a policy context, if sellers can differentiate themselves based on quality features/differences, then even if the cost of searching for price information is very low, competitive pressures exerted by consumers do not lead to excessive competition where quality cannot be profitably maintained. In general, there is some evidence that online markets tend to increased price transparency and, thus, price competition which, of course, is precisely the reason why online retailers react to this with more attempts to obfuscate prices (see our discussion of Ellison and Ellison above). The best evidence available so far stems from comparing internet and traditional markets for cars and life insurance. Morton, Zettelmeyer, and Silva-Risso (2001) estimate that online cars are two per cent cheaper than cars bought offline, controlling for the type of car. In a similar vein,

3.28

3.29

OFT1226 | 31

Brown and Goolsbee (2002) demonstrate that life insurance premia fell between eight per cent and 15 per cent through the 1990s in response to the rise of the internet in the consumers homes. 3.30 Notice, however, that there is also clear evidence that the internet does not eliminate price dispersion. While Brown and Goolsbee show that price dispersion in life insurance markets declined as internet use became more common (after an initial increase), Baye, Morgan, and Scholten (2004) show that there is substantial price dispersion in online markets for consumer electronics. They utilize over four million price observations from the price comparison site Shopper.com and find that price dispersion in internet electronic markets depends crucially on the number of firms competing which is in line with standard economic models predicting price dispersion. Baye et al. (2006) study the introduction of the Euro as a natural field experiment. In online markets the common currency makes price comparisons easier and, thus, increases price transparency. The authors show that, quite counter-intuitively, this can increase prices due to strategic responses by retailers. Essentially, the intuition for this follows three steps. First, more transparency implies more competitive pressure for consumers who shop around by comparing prices. Second, more competitive pressure in this market means reduced profits for firms. Third, reduced profits for targeting consumers who shop around increases the incentives to simply rely on 'loyal' consumers and charge higher prices to them. Remarkably, this is borne out in the data (taken from Kelkoo on gaming consoles, computer games, music CDs, PDAs, printers and scanners). After the introduction of the Euro, the authors document a price increase of 3 per cent that can be attributed to the increased transparency (that is, after controlling for cost, demand and market structure effects).

3.31

3.32

Theoretical literature 3.33 With the rise of behavioural economics several researchers in industrial organisation have become interested in modelling markets in which at least some consumers are not perfectly rational. This is still a small but

OFT1226 | 32

rapidly growing literature and there are a few studies that explicitly focus on price framing. 3.34 Armstrong and Zhou (2010) study how time-limited offers might arise endogenously. If firms can track consumers they have an incentive to offer buy-now discounts which, in equilibrium, helps to raise market prices. Gabaix and Laibson (2006) analyse markets where sellers can shroud the prices of expensive add-ons. Such examples include the sale of printers and the add-ons for ink cartridges. Or, consumer credit products and the additional costs for late minimum payments. Consumers that are fully rational (called Bayesian consumers) will always understand in these markets that shrouded prices are likely to be high that is, the devil is always in the detail. Consequently, shrouding would disappear in a purely Bayesian world because competition would drive firms that tricked consumers through the use of shrouding out of the market. But when consumers are not fully rational (non-Bayesian consumers) they do not always understand that shrouded prices are likely to be high. In this case not even perfect competition can eliminate shrouding. This arises because in situations where consumers can be tricked (that is, nonBayesian), then the benefit from training them to recognise that high add-on costs may be shrouded accrues to none of the firms. By informing consumers that high cost elements may be shrouded, firms will loose customers and profits to other firms that do shroud because consumers are not fully rational. On the other side of the coin, sophisticated Bayesian consumers might actually prefer to buy from the shrouding firms. This is because they know that the add-ons are high price but benefit from buying their desired quantity of the cheap nonshrouded component while fully integrating the high add-ons into their decision making. Of course for those consumers that are not sophisticated, they may continue to be tricked by not recognising that high cost elements may be shrouded and thereby not fully integrated the high add-ons into their decision making.

3.35

OFT1226 | 33

3.36

This theoretical work implies that drip pricing with shrouding can indeed lead to reduced outcomes for consumers because they dont realise that the (often) compulsory add-ons are high price. Chioveanu and Zhou (2009) study oligopoly markets (markets in which there are only a small number of firms) where identical sellers of an identical (homogenous) product choose both, prices and price frames. The authors assume that price frames have two effects on consumer behaviour. Consumers may fail to compare prices correctly because of the complexity of a price frame and/or because of the difficulty comparing different frames. Remarkably, these behavioural biases can completely overturn standard intuition about the functioning of markets. Indeed it is shown that an increase in the number of firms can increase industry profits and harm consumers (while standard theory would assume that new entrants reduce industry profits, thereby reducing prices and benefiting consumers). This arises because the framing acts as a form of price differentiation meaning consumers cannot compare prices for identical goods. This work implies that how a price is framed may be important for example, Morwitz et al, (1998) found that how a discount is framed might influence consumers assessment of the partitioned price (above). However, the complexity of the framing may be of more importance when considering its effect on consumer behaviour. For example, multipart tariffs for mobile phone contracts or mortgage contracts with different features can require complex computations in order to directly compare across different offers. In this (latter) situation firms may be able to take advantage of this raising price and increasing profits at a cost to consumers. In related work, Piccione and Spiegler (2009) analyse duopoly markets where consumers make price comparisons only with a certain probability that is assumed to depend on the price frames chosen by firms. Broadly in line with the results by Chioveanu and Zhou, they argue that product differentiation can also be viewed as a means to reduce consumers ability of comparing prices effectively.

3.37

3.38

3.39

OFT1226 | 34

3.40

From a policy point of view, this opens up an entirely new question.: Where do firms engage in product differentiation that is not driven by responses to different tastes of consumers but essentially just designed to make comparison more difficult? In this context, Bakos (2001) makes the interesting observation that in online markets (where, generally, price comparison is easier) there has been proliferation of customization and product differentiation in minute details.

Conclusions of the literature review 3.41 We have undertaken a review of both the marketing and economic literature in regard to price framing. We were surprised to find that the published literature is very limited. Further, while there is some evidence to suggest that how prices are framed impact upon consumer behaviour, this evidence is not very strong. The implications of this for the OFT is that the Market Study into Advertising and Pricing, and this study using controlled economic experiments to test and observe behaviour under different price frames, is, we believe, one of the first to systematically study the impact of price framing on consumer behaviour in markets.

3.42

OFT1226 | 35

THE EXPERIMENT

The experimental environment

4.1 The task is to analyse all of the five pricing practices (plus a baseline with straight unit pricing) in a way that allows the OFT to compare consumer decision making between the different practices and the baseline. This created several constraints when designing the experimental environment.15 Specifically, we need an environment with multiple shops in which multiple units of at least one good can be purchased. Moreover, we need scope for an advertising stage such that the experimental environment needs to mirror not only the shops but also the consumers home where he might be reached by some advertising before actually going to a shop. We opted to include all these facets in the simplest manner. The philosophy behind this is that we want to minimize the noise in decision making that originates from the sheer complexity of the basic environment. The less noise that originates from this baseline complexity, the sharper will be the results of the price-frame comparison. For the implementation of the five different price frames we have opted for 'typical' incarnations as inspired by the web trawl (Annex B). For example, we have opted for two price drips which we found to be common.16 Of course, our comparisons between different price frames are a function of our design choices. For example, complex pricing could be much more complex than '3 for 2' (we could imagine highly nonlinear tariffs) to an extent where consumers completely fail to understand the marginal prices of extra units. Our design choice has

4.2

4.3

15

An experimental environment is the experiment setting. It is the simplified setting which mirrors the important real world features that have a bearing on the policy questions to be tested in the experiment. In the web trawl we did find cases where there were up to 4 compulsory price drips.

16

OFT1226 | 36

been to opt for simple generic versions of all price frames in order to maximize possibilities for comparability. 4.4 As we have opted for a simple basic environment and simple incarnation of the price frames informed by the web trawl. The implications of a simple environment for external validity will be highly asymmetric: If we find that consumer subjects do well in the experiment, it would be hard to extrapolate this finding. However, if we find that they are doing badly, we might feel more assured that in more complicated real-life environments the problems are likely to be at least as severe as identified in the experiment.17 Our basic environment is designed in the following way. There are two shops, both of which sell the good that the consumer wishes to buy. At the start of each round, the consumer is at home. The consumer can go back and forth between his home and the two shops as often as he likes and buy units of the good at the shops (up to four units in total). Unless he receives advertising, he does not know the price of the good at either shop until he visits it. All goods are of the same quality. Again this allows us to focus on the pure effect of the frames. In many real-life markets prices (current or former) may serve as a quality signal which renders the decision problems much more complicated and confounds the issue of price framing.18

4.5

4.6

17

Of course, markets might have some self-healing power. For example, firms could establish a reputation for not using certain confusing practices. We discuss these issues in a section below devoted to the external validity of our study.

18

Some of the literature on reference pricing (called sales in this experiment), for example, Simonson (1999), suggests that the reference price can encourage consumers to switch from low-quality to high-quality options (Chapter 3).

OFT1226 | 37

4.7

In the baseline, each shop draws a random price from a price interval.19 For the purposes of our presentation, let us choose the price interval as [60, 120] (in the experiment the range was sometimes different, but could always be normalized to this interval). Both shops draw their prices from a uniform distribution over this interval. That is, all prices between 60 and 120 are equally likely. Exogenous prices have the crucial advantage of easing comparability between treatments. There are, of course, consequences for the interpretation of our experiment. Clearly, in an environment where firms can adjust prices optimally the effects of any suboptimal consumer behaviour would presumably be more pronounced. In this sense, exogenous prices help us to compare treatments but imply that we will underestimate the consumer detriment (if any) because firms are static and do not respond to consumer behaviour. The consumer has a utility (payoff) function with decreasing marginal utility.20 For the first unit he buys he receives a payoff of 120, for the second unit 80, for the third unit 20, and for the fourth unit 10. So the total payoff for buying one, two, three or four units is 120, 200, 220 or 230 respectively. Notice that with straight unit pricing, as in the baseline, the consumer will never buy more than two units as the marginal utility of the third unit is smaller than the lowest possible price of 60. Notice also that we never allow the purchase of more than four units. Each time the consumer travels to a shop he has to pay some search/travel costs, c. The search costs vary in the experiment, being

4.8

4.9

4.10

19

The price interval is selected before the experiment begins. The subjects playing the role of consumers in the experiment are told what the price interval is at the beginning of the experiment. Marginal utility is the additional satisfaction (payoff) the consumer derives if he buys and consumes another unit of the good in question. Marginal utility decreases as the consumer gets, and consumes, more units of the good.

20

OFT1226 | 38

randomly chosen each round. Specifically, we examine three different levels of c, low (2), medium (6), and high (12). 4.11 The decision process a consumer must therefore undertake is summarised in the box below and a simple schematic diagram of the process is reproduced in the subsequent figure.

Box 2: Consumer decision making process

I can go to shop 1 or shop 2 to buy up to four units of the good available in this time period. Whenever I go to a shop, I incur a travel cost which will be deducted from the points I receive when I purchase a unit of the good. The two shops have different prices. I can only find out what price each shop is selling a unit of the good at by travelling to the shop. At the home screen (see Figure A.1 in the annex for a shot of the home screen), I can see that the travel cost is two per journey in this time period. I will go to a shop. Suppose I choose to go to Shop 1. I see a price of 105 (as in Figure A.2). I know that I receive 120 points from buying 1 unit, for two units I receive 200 points, and for three units I receive 220. So the additional points (my marginal utility) are 80 and 20 for buying the second and third unit. I have already incurred a travel cost of two points. Therefore if I buy one unit at 105, my earnings will be 120 105 2 = 13 points. If I buy two units at this shop, then my earnings will be 200 105 * 2 2 = -12 points. If I buy three units at this shop, then my earnings are 220 3 * 105 2 = -97 points. If instead I choose to go to shop 2 to find out the price of the good there, I will incur another travel cost (for travelling from home to Shop 2). This means my travel costs, for visiting the two shops, will be 2 * 2= 4. To make it worth my while, the price at Shop 2 will need to be lower than at Shop 1. If the price at shop 2 is higher than shop 1, then I may want to come back to shop 1 again. In this instance I will incur a third travel cost meaning my total travel cost will be 2 * 3 = 6. Looking at the price range, the lowest possible price is 60 and the highest is 120. I have observed 105 for a unit at Shop 1. If I choose not to visit Shop 2,

OFT1226 | 39

then it will be best to buy 1 unit at Shop 1 (because I get 13 points, better than -12 or -97). Should I go to Shop 2? Well, the price at Shop 1 is quite high and so it is likely that the price at Shop 2 will be lower. If it is at least two points lower (that is, lower than 103), then I will more than cover the extra travel cost. This is quite likely. Even if it is higher than Shop 1, I can always come back to Shop 1 and have only lost the cost of two extra trips (a cost of four points). So, I will go to Shop 2 and see the price there. Once there, I will do the same calculations as I have just done here to decide the optimal number of units to buy (or whether it is better to go back to Shop 1). Figures A.3, A.4, A.5, and A.6 show the buying and results process. For simplicity, suppose rather than visit Shop 2, I chose to buy 2 units at Shop 1 and then end the round. I must confirm my purchase see Figure A.3. If I confirm my purchase, I successfully buy 2 units (at 105 each) see Figure A.4. If I then go home and click 'Im done' on the home screen to finish the period, I am prompted to confirm my decision Figure A.5. Upon confirmation of my exit, the results are shown see Figure A.6.

OFT1226 | 40

Figure 4.1: Consumer decision process in the baseline treatment

The implementation of the pricing practices

4.12 The experiment studies five different pricing practices and compares these practices to a baseline of, straight, per-unit prices. As described previously, in this baseline treatment, consumers see the per-unit price a shop charges once they visit this shop. This price stays constant within one round. That is, if the consumer returns to the shop he will still get the same price. This holds for both shops. Notice that we design the experiment in a way such that both shops always employ the same type of price framing. The setup for the sales frame is almost identical. The selling prices are determined in precisely the same way as for the baseline and the only difference to the baseline is that consumers are additionally informed about a former price. Specifically, they are told that the former price is a price that is chosen randomly between the actual selling price and the

4.13

OFT1226 | 41

maximum possible price. They are also told about the consequent 'discount' that this represents in percentage terms. As such, it should be easy for the subjects to see that the sales information is actually entirely meaningless. Even though subjects are not told how the former price is generated it is apparent that the good has not been sold previously at this price. Of course, it might have been to some other subjects in some earlier round but for the subject faced with the decision at hand this is irrelevant information and should easily be identified as such. The screen shot for the sales frame is presented in figure A.14. 4.14 Figure 4.2 presents a schematic of the consumer decision process in the sales frame.

Figure 4.2: Consumer decision making process in the sales frame

4.15

Drip pricing is also virtually identical to the baseline. Actual selling prices are determined in precisely the same manner. This time, however, consumers learn about the selling price only in drips. Once they visit a shop, they see a base price (with no mention of additional charges).

OFT1226 | 42

Once they decide to buy one or multiple units, they see a first drip and need to click ok to proceed. If they do so, they are informed about a second drip and need to ok this as well. Finally, they see the total price (and its decomposition into base price plus shipping plus handling) and they need to click one more time to confirm their purchase. So the only difference to the baseline is that subjects need to click two more times to learn the actual selling price. As mentioned above, actual selling prices are determined in precisely the same way as in the baseline. The actual selling price is then decomposed by the computer into the base, the first drip and the second drip. The first drip is randomly chosen to be between five per cent and 15 per cent of the selling price, and the second drip is randomly chosen to be between 10 per cent and 20 per cent of the selling price. The base price is the remainder. The screen shots for drip pricing are presented in figures A.8 to A.10. 4.16 Figure 4.3 shows a schematic of the consumer decision making process in the drip pricing frame.

OFT1226 | 43

Figure 4.3: Consumer decision making process in the drip pricing frame

4.17

Under complex pricing ('3 for 2') the unit prices are again determined in the same way but consumers are not charged for a third unit if they wish to buy one. They are informed of this offer once they enter the shop and the offer remains in place for an entire round. Figure A.7 presents a screen shot for complex pricing. Figure 4.4 shows a schematic of the consumer decision making process in the complex pricing frame.

4.18

OFT1226 | 44

Figure 4.4: Consumer decision making process in the complex pricing frame

4.19

In time-limited offers, consumers are told when they enter the first shop that the price they are confronted with now is only on offer at this instance. If consumers come back at a later stage, the price will have changed. Specifically, the shop simply draws a new price from the same distribution, that is, sometimes consumers will actually get a price upon return that is below the time-limited offer price. The other shop engages in the same tactics (that is if the consumer had visited the other shop first, he would also have been confronted with a time-limited offer). However, once the time limit passes both shops offer a price that then remains fixed. This means that the consumer never experiences a timelimited offer at the second shop (simply because the offer there was also just valid for one time period). Figure A.15 shows a screen shot for the time limited offer frame.

OFT1226 | 45

4.20

Figure 4.5 shows a schematic of the consumer decision making process in the time-limited offer frame

Figure 4.5: Consumer decision making process in the time-limited price frame

4.21

Finally, under baiting both stores advertise prices. That is, in contrast to all other treatments, consumers have some price information at their home screen. These prices are under a generic 'while stocks last' qualification that is printed next to the price information that subjects can see at the home screen. Specifically, if the selling price is between 60 and 72, the advertised price is equal to the selling price. If the selling price is greater than 72, then there is a 50 per cent chance that the advertised price will be the selling price and a 50 per cent chance that the advertised price will be some price randomly drawn from the interval [60, 72] (a 'bait'). When a consumer visits their first shop, they will see

OFT1226 | 46

whether the price is real or a bait. The selling price at the other shop is now randomly drawn. If the consumer returns to the home screen, the advertising has gone. Figures A.11 to A.13 show the screen shots for the baiting frame. 4.22 Figure 4.6 presents a schematic of the consumer decision making process in the baiting frame.

Figure 4.6: Consumer decision making process in the baiting price frame

The consumers optimal search strategy in the baseline

4.23 The first decision is to choose one of the two shops to visit first. As there is no history and no information about the two shops, this is inevitably a random (and, hence, meaningless) decision. (Of course, this changes in the presence of some advertising, as in our treatments on

OFT1226 | 47

baiting.) So, without loss of generality, we can call the shop the consumer chooses first 'shop 1'. 4.24 Once the consumer is at shop 1, he can see the unit price that the shop charges in this period. He can then either buy as many units as he desires (up to a maximum of four) or he can also decide not to buy and return to the home screen empty-handed. He can then travel to shop 2 (or indeed go back to shop 1 but that would, of course, not be reasonable as no new information would be revealed), again paying the travel/search costs, c. At shop 2 the same rules apply. That is, the consumer learns the unit price charged at shop 2 and can buy as many units as he desires. Again, he can also return empty-handed and, if he desires, return to shop 1. He can, of course, also return to shop even if he has bought some units already, either at shop 2 or even at shop 1, but, as is easy to see, that would also not be rational. We can analytically derive the optimal consumer search strategy. The optimal search strategy is what a fully rational consumer should do. Here we summarize the solution. The theory is presented in detail in Annex C. Notice first that the marginal utilities and prices used in this experiment mean that it is only ever optimal to buy one or two units in total. The marginal utility of the third unit is always less than the price. At shop 1, there are two cut-off prices, p and p. p is the cut-off for buying one unit or two units and p is the cut-off for buying at shop 1 or searching more by visiting shop 2. If the price at shop 1, p1, is above p, then the consumer will not buy at shop 1, but rather go to shop 2 and see what the price is there. The consumer may return to shop 1 later, depending on the price at shop 2 and the prevailing search costs. If p1 is below p, the consumer will do all their shopping at shop 1. If p1 is below p, the consumer will buy two units at shop 1 and end the round. If p1 is above p he will buy one unit at shop 1 and end the round.

4.25

4.26

4.27

4.28

4.29

OFT1226 | 48

4.30

As described above, in case of a price p1 above p, the consumer will travel to shop 2. There are four different possibilities that can then arise: (i) He can buy two units at shop 2 and end the round. (ii) He can buy one unit at shop 2 and end the round. (iii) He can decide not buy at shop 2 and return to shop 1 to buy two units there. (iv) He can decide not to buy at shop 2, return to shop 1 and buy one unit there. Which of these four options is optimal depends on the search costs and the two prices p1 and p2. The higher the search costs, the higher the cut-off p. Table 4.1 shows the cut-offs for the search costs we have implemented.

4.31

Table 4.1: p cut-offs for the various treatments

Search cost Baseline* Complex Time-limited offer 79.2 85.8 94.8

c=2 c=6 c = 12

74.5 86.3 97.8

74.5 83.9 91.9

Note: The optimal strategies, and therefore the p'' cut-offs for the sales, drip pricing and baiting frames are the same as the baseline (single unit straight pricing). This is explained further in the following sub-section of this chapter.

The consumers optimal search strategy in the price frames

4.32 Two of the five price frames leave the optimal search strategy from the baseline completely untouched because the true unit price does not change and is clearly discernible to the unbiased consumer. In the sales frame, the consumer directly sees the true unit price and just receives some completely irrelevant additional information (the former price)21 and under drip pricing the true price gets clearly revealed, albeit only in drips.

21

In practice, former prices might often serve as a signal of quality. By abstracting from such quality issues we can measure the pure effect of framing a price as a sales price.

OFT1226 | 49

So in both setups, a fully rational consumer will simply adopt the same search strategy as that we have derived as optimal above. 4.33 Optimal search is, however, slightly different under the other three frames. Under complex pricing, however, the derivation of the optimal search strategy is essentially identical. Under '3 for 2' we can revisit our old analysis and simply change the marginal utility of the second unit (80). Specifically, we replace it by the sum of the marginal utility of the second and third unit (80 + 20 = 100). The logic is simple. As the third unit gives you some positive utility you will always take it if you do buy two units. In other words, you will never refuse the offer and check out with just two units. However, you might still sometimes prefer just to buy one unit. This changes the cut-offs. Table 4.1 shows the relevant cut-offs for complex pricing for each search cost level. For the remaining two frames, the structure of the analysis also changes slightly. First, let us consider time-limited offers. If a time-limited offer is not taken and the consumer returns home, both shops draw new prices. Consequently, if a consumer does not take the time-limited offer and returns to his home screen, he faces a situation that is identical to the original straight pricing baseline because, from now on, the shops do simply stick to one price and refrain from any further such offers. This implies that essentially, when the consumer sees the time-limited offer at shop 1 (the shop he visited first), he has the choice between buying now and ending the round or playing the original game. This means the consumer can simply compute the utility he would receive from buying optimally one or two units now (for the time-limited offer price) and compare this with the ex ante expected utility he receives from playing the baseline game, assuming, of course, he plays this optimally. This generates the different cut-offs shown in Table 4.1. Finally, let us consider baiting. Under baiting the consumer needs to take into account that low prices in the range from [60, 72] might turn out to be baits while with higher prices he can rest assured that they will turn out to be true. This solution necessitates knowing the precise rule the firms employ which, in the experiment, is initially not the case but, through repetition, may be able to be learnt approximately. As the

4.34

4.35

OFT1226 | 50

baiting offer is only available initially, the rational consumer essentially has to take just one additional decision that goes beyond the optimal baseline search strategy. At the home screen, he has to decide which shop to visit first! Notice that this is indeed the only frame where the initial choice of shop is meaningful. Once he is at the first shop and the true price gets revealed, the consumer is back to the original problem of the baseline. It is, of course, this elegance of the rational benchmark solution that has inspired the specific design here, that is the decision to have the baits only initially and revert to a new price draw after that (for both firms). 4.36 The optimal choice of shops is not simply determined by the lowest advertised price. As low prices may be baits, but high prices tend to be honest, the consumer needs to work out average expected payoffs based on the offered prices. For prices that look like potential baits, but that are not extremely low, the consumer might actually be better off to seek out a shop that advertises a higher but honest price.

Experimental procedures
4.37 A total of 166 subjects participated in this experiment. Each subject was confronted with the baseline and two of the five frames. Subjects played for thirty rounds, ten for each type of price frame. The sequence in which they faced the different frames was random. This repetition enables subjects to learn about the environment and adapt their behaviour. This enables us to determine whether any potential effects of the price frames can be overcome through experience. In order to enhance attention (and make sure that each round was viewed as a truly new round) we scaled payoffs in four different ways. Specifically, subjects faced four different goods, GREEN, ORANGE, BLUE and RED. Utilities and prices for each good were obtained from the model above through different ways of up scaling (the model above details the utilities and prices for RED). This ensures that the basic problem is always identical, regardless of the specific goods subjects could buy. The search cost was randomly chosen each period from low,

4.38

OFT1226 | 51

medium and high. The actual payoff, price and search costs are given in Table 4.2.

Table 4.2: Parameters for different goods

Product C 0 (search units costs) 1,3,6 1,3,6 2,6,12 0 0 0 1 unit 2 units 100 140 180 3 units 110 170 190 4 units 115 195 190 Price Range 30 to 60 50 to 80 50 to 110 60 to 120

GREEN ORANGE BLUE

60 80 110

RED

2,6,12

120

200

220

230

4.39

There are 10 combinations of two out of five price frames and we implemented all of them. That is, we studied ten different groups of subjects where each group of subjects is characterized by a combination of two price frames subjects are faced with in addition to the baseline that every subject experiences. This allows us both within and between subject comparisons, maximizing the statistical power we get for our data analysis. In addition to the 30 rounds of the experiments, each subject undertook: 1. pre-experiment test to ensure that they understood the experimental instructions 12-question IQ test (incentivised) 15-question personality test22

4.1

2. 3.

22

We found no correlation between personality and behaviour in the experiment.

OFT1226 | 52

4.

feedback questionnaire about the experiment.

The personality test and the feedback questionnaire are reproduced in the annexe. 4.2 Experimental sessions lasted on average 145 minutes and average earnings were approximately 20 which included a 5 show-up fee.

OFT1226 | 53

5
5.1

EXPERIMENTAL RESULTS
In our analysis, we proceed as follows. We first study whether price frames have any effect on consumer welfare. To make treatments comparable we define losses in consumer welfare relative to what consumers could have achieved under optimal behaviour.23 In a second step, we employ econometrics to analyse errors in behaviour, distinguishing errors in search and errors in purchasing. Finally, we will zoom in more closely on search patterns and purchasing behaviour, in order to understand what the root causes of poor performance are and how they relate to known behavioural phenomena.

Consumers
Consumer welfare 5.2 We now turn to the first fundamental question this research addresses: do price frames matter for consumer welfare? Despite our extremely simple environment and a rather smart subject pool, we find that price frames do matter. Since the amount of consumer welfare obtainable under optimal behaviour differs between the different frames, we need to normalize achieved payoffs in an appropriate manner. In order to do this, we take, for each of the 4895 observations we have, the difference between the actual achieved payoff and the payoff that would have resulted from following the optimal decision rule. We call this variable the consumers loss. If a consumer could have achieved a payoff of 87 under optimal behaviour but only achieved a payoff of 69 then his loss is 87 69 = 18. Often we look at the average loss a consumer has made in a particular environment which is simply calculated as the arithmetic mean of all the losses in all rounds.

5.3

23

The optimal strategy is the one a subject that knows the experimental environment and is fully rational would employ.

OFT1226 | 54

5.4

Additionally, we compute a further welfare indicator, the extra loss relative to the baseline loss. This is computed as follows. For each subject we have three average loss variables, one for the baseline, and two for the two price frames encountered.24 The extra loss a subject incurred under a price frame is then simply defined as the difference between the average loss in this price frame and the average loss in the baseline. We can then also compute the extra loss made on average by all subjects under a particular price frame. This difference-in-difference approach controls for both different earning potentials under different frames and subject-specific differences in performance levels. Table 5.1 shows both loss and extra loss for all price frames and the baseline. Ranks are assigned according to extra loss. The magnitude of losses as shown in Table 5.1 refers to the normalized model that we sketched above. One unit of loss translates to 0.02 in each round of the experiment. So, for example, the average loss under time-limited offers of 7.64 translates into a monetary loss of 1.58 over 10 rounds in which time-limited offers are experienced. Given average earnings of 15 on top of the show-up fees over thirty rounds, this shows that the loss is substantial: more than a third of what subjects make on average in ten rounds.

5.5

24

Each subject did three treatments including the baseline within a two-hour session.

OFT1226 | 55

Table 5.1: Welfare losses under the different price frames

Rank 1 2 3 4 5 6 Frame Baseline Complex Sales Baiting Time-ltd offers Drips Loss 5.66 7.47 7.10 5.83 7.64 10.69 Extra Loss 0 0.81** 2.01 2.04** 2.13** 3.23***

Note: Stars indicate significant differences to baseline, ** 5 per cent, *** 1 per cent

5.6

For each price frame, we test whether the average performance is significantly different from the baseline performance. With the exception of the sales frame, differences are indeed significant. The significance levels are 1.8 per cent for complex pricing, 1.6 per cent for time-limited offers, and 2.0 per cent for baiting. For drip pricing the difference in performance even reaches a significance of 0.1 per cent.25 Two main results emerge from Table 5.1. First, price framing is indeed detrimental for consumer welfare and this seems to be the case for all frames, with the potential exception of the sales frame. Second, drip pricing emerges quite clearly as the worst culprit with the biggest average loss and the biggest average extra loss.

5.7

Errors in consumer behaviour

5.8 There are generally two types of errors subjects can make in our experiment: Errors in their search activity and errors in purchasing

25

This means that subjects' extra losses in the price frames were significantly larger (or smaller) than the extra losses incurred by subjects in the baseline.

OFT1226 | 56

behaviour. Errors in search activity occur where a consumer makes more or fewer visits to the shops than is optimal. Specifically, there are two types of search errors. A consumer makes a search error if he buys at the present shop but should optimally have continued his search. Or, vice versa, if he continues his search but should have optimally bought at the present shop. Errors in purchasing behaviour occur where subjects do not buy the optimal amount of the good. For example, they buy one unit when it would be optimal to buy two units given prices and marginal utilities of consumption. Notice that search and purchasing errors can also occur together. For example, when a consumer buys one unit at the first shop while according to the optimal strategy he should have bought two units at the second. 5.9 There are some arguments for why in the context of this study it might be more important to focus on errors rather than overall performance as the performance measures above are sensitive to the precise parameters chosen in the experiment. For example, the losses would have been much bigger if the value of the goods had been higher. In order to get a first grip on errors in decision-making, we examine each of the 4895 rounds played by our 166 subjects. Whenever the observed behaviour in a given round departs from the optimal decision rule, we classify the round as a round with an error. We then regress errors on prices, search costs, the scaling factor and the different price frames (where each frame is captured by a binary dummy with the baseline serving as the reference). We use probit regressions and cluster the standard errors on the subject level in order to account for dependencies resulting from repeated measurement. Table 5.2 shows the estimated marginal effects. (All detailed regression results are contained in the technical appendix where we also include estimations from linear probability models that throughout confirm the robustness of our results.) The way to read Table 5.2 is the following: the estimated coefficients show if subjects in the experiment either searched 'too much' or 'too little' or bought 'too many' units or 'too few' units (overall errors) than was optimal in the relevant price frame as compared to the

5.10

OFT1226 | 57

overall errors that subjects made in the baseline treatment with straight per unit pricing.

Table 5.2: Probit estimation of errors in decision making.

Variable p1 (price at 1st shop visited) p2 (price at 2nd shop visited) c (search cost) Highvalue (the scaling; Green, Blue, Orange or Red goods) d2 (Complex pricing) d3 (Drip pricing) d4 (Baiting) d5 (Sales) d6 (Time limited offers) Coefficient 0.393 Standard error (p-value) 0.070 ***

0.218

0.050 ***

-0.516 0.012

0.214 ** 0.014

0.019 0.143 0.091 0.069 0.185

0.026 0.026 *** 0.026 *** 0.025 ***

0.25 ***

Note: The estimation model is probit regression 1.a. in the annexe.

5.11

The coefficients shown in this table are to be read as the percentage increase in errors if the explanatory variable increases by one unit. With the exception of complex pricing, we observe significantly more erroneous behaviour (relative to the optimal behaviour benchmark) in all price frames, including the sales frame which increases the error rate by almost seven per cent. This higher error rate induced by the simple 'was X is now Y' statement appears striking, the reference price is quite obviously meaningless in the experimental environment.

OFT1226 | 58

5.12

The highest marginal effect on error rates are observed under timelimited offers and drip pricing with 19 per cent and 14 per cent more errors respectively than in the baseline. Baiting generates around nine per cent more errors. With respect to the other variables, we make the following three observations. (i) The 'high value good' variable is not significant. Hence, there are no effects of the scaling and our results are robust to the incentives. In other words, higher stakes do not reduce errors. (ii) Lower prices are associated with lower errors, presumably because with low prices it is easier to detect when purchase of two goods is optimal. Moreover, as we will see in more detail below, consumers exhibit in several treatments a tendency 'to buy anyway' (instead of optimally continuing their search). And this behaviour is obviously only erroneous when prices are high.26 (iii) With higher search costs consumers make fewer errors. Tables 5.3 and 5.4 contain two further regressions that analyse the quality of choices in more detail. Specifically, we distinguish between two kinds of errors, errors in search behaviour and errors in purchasing behaviour. Errors in search occur whenever a consumer continues his search while the optimal strategy prescribes immediate purchase or vice versa. Errors in purchasing behaviour occur when subjects buy a suboptimal number of units. While determining the optimal number of units is a rather simple task (it just requires a very basic understanding of the marginal utility/pay-off table) the decision about search is more demanding. Table 5.3 can be read in the following way: The estimated coefficients show If subjects in the experiment either searched 'too little' or 'too much' than was optimal in the relevant price frame as compared to

5.13

5.14

5.15

26

Notice the difference between price and value. High value refers to a higher multiplier for both, consumer payoff and price. Changing this multiplier leaves the decision-relevant (relative) price completely unaffected.

OFT1226 | 59

search errors that subjects made (searching too little or too much than was optimal) in the baseline treatment.

Table 5.3: Probit estimation of errors in search behaviour

variable p1 (price at 1st shop visited) p2 (price at 2nd shop visited) C (search costs) Highvalue (the scaling; Green, Blue, Orange or Red goods) d2 (Complex pricing) d3 (Drip pricing) d4 (Baiting) d5 (Sales) d6 (Time limited offers) coefficient 0.206 standard error 0.067 ***

0.221

0.042 ***

-0.463 0.007

0.189 ** 0.013

0.018 0.095 0.071 0.047 0.185

0.024 0.027 *** 0.024 *** 0.022 ** 0.026 ***

Note: The estimation model is probit regression 1.b. in the annexe.

5.16

The estimation results for errors in search are almost identical to the estimations for errors in general. This is not surprising given that errors in search are the main source of suboptimal behaviour. Table 5.4 can be read in the following way: The estimated coefficients show if subjects in the experiment either bought 'too many' units or 'too few' units than was optimal in the relevant price frame as compared to purchasing errors that subjects made (buying too few or too many units than was optimal) in the baseline treatment with straight per unit pricing.

5.17

OFT1226 | 60

Table 5.4: Probit estimation of errors in purchasing behaviour

variable p1 (price at 1st shop visted) p2 (price at 2nd shop visited) C (search costs) Highvalue ((the scaling; Green, Blue, Orange or Red goods) d2 (Complex pricing) d3 (Drip pricing) d4 (Baiting) d5 (Sales) d6 (Time limited offers) coefficient 0.620 standard error 0.061 ***

-0.156

0.051 ***

0.479 0.026

0.202 ** 0.014

0.013 0.142 0.095 0.060 0.211

0.025 0.028 *** 0.026 *** 0.024 ** 0.025 ***

Note: The estimation model is probit regression 1.c. in the annexe.

5.18

While it is less common that subjects buy the incorrect number of units (we will document this in more detail further below) it does happen and the estimation results in Table 5.4 reveal some clear patterns. Once again, price frames increase error rates (and once again time-limited offers and drip pricing are the worst culprits). But some of the other variables have an effect on purchasing behaviour that is rather different from their effect on search behaviour. Specifically, subjects make more errors in purchasing the right number of units if the price at the second store is comparatively lower; errors increase in search costs; and for high-value goods there are more errors than for low-value goods. All three effects point to a psychological mechanism whereby consumers get overly keen to buy more of a good after being frustrated by high

OFT1226 | 61

prices at the first shop and having spent much on extra search to travel to the second. This justification of earlier incurred costs is an incarnation of the sunk cost fallacy and appears to be the main source of purchasing errors beyond pure cognitive failure to trade off marginal utility and unit price. 5.19 Notice that the regressions suggest a slightly different ranking of the five price frames than our welfare results. There are a number of reasons behind this. First of all, we do not control for the size of errors in the regressions (precisely to offer an alternative view on consumers performance that is less dependent on our parameter choices). Second, we do not account for multiple errors in the regressions. For example, subjects who search excessively (and through these multiple errors lose substantial amounts of money) are classified for the purpose of the regressions in the same way as a subject who makes a single mistake. Notice that we have also estimated linear-probability models as a robustness check. (See the technical appendix for details, linear regression model 1.a. 1.c.). The stability of the results is remarkable: the marginal effects reported above are virtually identical to the parameter estimates obtained through the linear estimations.

5.20

The role of search costs

5.21 As we have seen above search costs have an ambiguous effect on the quality of consumers choice behaviour. Higher search costs reduce, on average, search errors but increase purchasing errors. In order to better understand the effect of search costs we need to investigate their differential impact on behaviour across the different treatments. We investigate whether search costs have treatment effects by rerunning our regressions including interactions between search costs and treatment dummies. (See the technical appendix for all detailed estimation results, estimation models probit 6.a. 6.c.). In the general error regressions we find that search costs as such lose their significance but do show up negatively if interacted with drip pricing (1.11), baiting (-1.32), and time-limited offers (-1.51). In other words, we

5.22

OFT1226 | 62

find that higher search costs reduce errors only under these three price frames and not in any of the others. 5.23 Decomposing errors, we find the following. For search errors, the picture is virtually identical to what we have seen in the general error estimations. Search costs have no effects in the baseline nor under complex pricing or the sales frame. On the other hand, they do reduce search errors under drip pricing, baiting, and time-limited offers. For purchasing errors a different pattern emerges. The search cost variable is significantly positive that is, purchasing errors increase with search costs. However, for drip pricing and time-limited offers this effect is wiped out (which can be seen by comparing the significant coefficients of the relevant interaction terms with the estimated coefficient of the general search cost term). In all, we have to conclude that price frames have a stronger effect on consumer behaviour in markets with lower search costs. In particular, drip pricing, baiting and time-limited offers cause more consumer detriment if search costs are comparatively low. We will revisit the causes for these effects further below when we discuss the behavioural biases that are driving these deviations from optimal consumer choice.

5.24

5.25

The lessons so far

5.26 We have seen very clean evidence now that identifies drip pricing as the price frame that is most detrimental to consumers, both in terms of errors and consequences. Time-limited offers come second (with intermediate welfare losses but even more errors than drip pricing), which is surprising given that the literature so far has either ignored or exonerated time-limited offers. Consumers have the least problems with complex prices and sales frames where the former has no effect on errors and the latter no effects on welfare. They are, however, both not completely unproblematic. In the middle is baiting with systematically more errors and a significant welfare loss.

OFT1226 | 63

5.27

In order to better understand the behavioural forces that are the root cause of inferior consumer decision making stems we will now study search and purchasing behaviour in much finer detail.

Zooming in
5.28 Before we actually turn to a detailed analysis of consumer behaviour under the different price frames, we need carefully to examine the baseline. Table 5.5 shows all observations we have for our subjects at the first shop in the baseline. The rows indicate what would have been optimal, the columns what has actually been chosen. So, the first row ('0') in the table indicates all situations where the optimal strategy prescribes further search (the purchase of zero units at the first shop). The second row contains all cases where consumers should have bought one unit and the third all cases where consumers should have bought two units. The columns indicate the actual number of units purchased. While the table shows that subjects do generally very well in this situation (78.6 per cent of all choices are optimal) it also reveals two interesting asymmetries. First of all, it shows that errors are typically errors in search while purchasing the wrong number of units is much rarer. In fact, 90.5 per cent of all errors are search errors. The second asymmetry occurs within the class of search errors. While subjects do not buy in 86.8 per cent of all cases when it is optimal to continue to search, they buy optimally only in 68.3 per cent of all cases where it is optimal to buy. In other words, there is a clear tendency to oversearch. This is particularly dramatic when subjects should optimally buy just one unit. In this case only 50.8 per cent of choices are correct.

5.29

5.30

OFT1226 | 64

Table 5.5: Optimal vs actual choices at the first shop visited in the baseline (straight per unit pricing)
Actual choice 0 Optimal choice 0 1 2 Total 804 83 119 1006 1 93 99 14 206 2 25 11 410 446 3 1 1 3 5 4 3 1 4 8 Total 926 195 550 1671

5.31

Next we turn to what happens at the second shop (where we only track those subjects who went there optimally).27 The rate of optimal behaviour here is 86.7 per cent even higher than at the first shop. This partly reflects that it is an easier decision (as now all uncertainty has been resolved and all prices are known) but is also due to a selection effect. After all, the table only contains those subjects who have already made one correct decision. The asymmetries also largely disappear. There is no clear pattern of oversearch once consumers have reached the second store.

27

Notice that the number of cases in all tables that show behaviour at the second shop does not always coincide with number of cases in the (0,0) box of the first-shop tables. This is due to the fact that in some instances subjects did not buy at all.

OFT1226 | 65

Table 5.6: Optimal vs actual choices at the second shop in the baseline (straight per unit pricing)
Actual choice 0 Optimal choice 0 1 2 Total 144 24 2 170 1 17 313 18 348 2 2 31 234 267 3 0 3 3 6 4 0 3 3 6 Total 163 374 260 797

5.32

It is interesting to examine the stores sales volume, relative to what they would sell under optimal consumer behaviour. The first shop sold 0.69 units per customer which compares to 0.78 units that are predicted under optimal consumer behaviour.28 In other words, in the baseline the first shop reaches only 88.4 per cent of its sales potential. In contrast, the second store achieves 103.4 per cent of its sales potential. Let us now examine the same two tables for drip pricing. Table 5.7 shows behaviour at the first, and Table 5.8 at the second shop. Comparing Table 5.7 with the baseline reveals how dramatic the effect of drip pricing really is. The rate of optimal decisions falls to 70.9 per cent. But even more strikingly, the phenomenon of relative oversearch that was so prevalent in the baseline not only disappears but is actually replaced by a stark pattern of undersearch. In other words, while many consumers who should have bought at the first store in the baseline decided to check out the second store, this is reversed under drip pricing. Consumers who should continue to search are lured into buying.

5.33

28

Notice that these numbers can be directly computed from the tables showing optimal vs. actual choice.

OFT1226 | 66

In 26.9 per cent of all cases where consumers should not have bought from the first shop they do so now. This compares to just 11.0 per cent in the baseline. 5.34 As a consequence drip pricing means that many more units are sold at store 1. The first store sells more units as predicted, and not fewer as in the baseline. It reaches 111.8 per cent of its sales potential, a relative increase of 26.5 per cent over the baseline. These findings appear all the more remarkable as the difference between the two environments is really rather small. In essence, it just requires two extra clicks to see the true full price under drip pricing. Everything else is completely identical. Moreover, the environment is very simple and the subject population is highly selected and presumably much more capable of sophisticated behaviour than the average consumer.

5.35

Table 5.7: Optimal vs actual choices at the first shop in the drip pricing frame
Actual choice 0 Optimal choice 0 1 2 Total 274 25 39 338 1 68 39 7 114 2 28 7 137 172 3 0 0 0 0 4 5 0 6 11 Total 375 71 189 635

5.36

Turning to Table 5.8 we also find that consumers performance does not much improve at the second shop with just 72.6 per cent of optimal choices (compared to 86.7 per cent in the baseline). However, the second store does not benefit to the same extent as the first store, even conditional on being visited. In fact, it only reaches 97.2 per cent of its sales potential.

OFT1226 | 67

Table 5.8: Optimal vs actual choices at the second shop in the drip pricing frame
Actual choice 0 Optimal choice 0 1 2 Total 40 25 5 70 1 6 73 11 90 2 2 21 83 106 3 0 1 2 3 4 0 0 1 1 Total 48 120 102 270

5.37

Tables 5.9 and 5.10 show the same data for consumer behaviour at the first and second shop for time-limited offers. Table 5.9 for the first shop looks remarkably similar to Table 5.7 for the first shop under drip pricing. Again the phenomenon of oversearch is reversed and consumers tend to buy when optimal behaviour suggests they should not. However, the achieved sales volume is slightly smaller than under drip pricing (albeit still much better than under the baseline). With time-limited offers for the first store reaches 101.2 per cent of its sales potential. While consumer behaviour at the first store under time-limited offers is qualitatively similar to that under drip pricing, there is a marked difference between Tables 5.10 and 5.8 which show consumer behaviour at the second shop. While the second shop could not benefit from drip pricing (if anything consumers reversed to oversearch when facing drips at the second shop), it does benefit substantially from the time-limited offers consumers faced at the first shop. It (the second shop) reaches an enormous 140.4 per cent of its sales potential. From that it becomes apparent that consumers fear to face substantially higher prices upon return to the first store. In that sense, time-limited offers do the trick, they do confuse consumers substantially. But,

5.38

OFT1226 | 68

somewhat ironically, the competition benefits more from this than the shop first visited by the consumer.

Table 5.9: Optimal vs actual choices at the first shop in the time limited offers frame
Actual choice 0 Optimal choice 0 1 2 Total 297 22 28 347 1 69 32 15 116 2 9 7 178 194 3 0 0 1 1 4 1 0 0 1 Total 376 61 222 659

Table 5.10: Optimal vs actual choices at the second shop in the time limited offers frame
Actual choice 0 Optimal choice 0 1 2 Total 117 5 1 123 1 70 24 7 101 2 3 5 60 68 3 0 1 0 1 4 0 0 0 0 Total 190 35 68 293

5.39

Tables 5.11 and 5.12 show the same data for baiting, Tables 5.13 and 5.14 for complex pricing, and Tables 5.15 and 5.16 for the sales frame.

OFT1226 | 69

5.40

Just as in drip pricing and time-limited offers, a baiting strategy eradicates oversearch and generates a pattern of undersearch with many consumers (even) buying two units at the first shop where a continued search would have been optimal. There is no indication of consumers 'punishing' sellers that lured them into their shops with low offers that then turn out to be unavailable. In contrast to time-limited offers, there are no particular effects of baiting on consumer behaviour at the second shop. Generally, consumers do very well at the second shop with the vast majority of observations in Table 5.12 on the main diagonal (meaning actual choice was optimal). Baiting is the only treatment where the initial choice of shops is nonrandom. It is worthy to note that subjects in the experiment invariably chose the shop with the lower advertised price even in those cases when the optimal strategy suggests to choose the more expensive store.29 Table 5.13 and 5.14 for complex ('3 for 2') prices looks, of course, considerable different from the others as it is never optimal to buy two units (when the third is effectively free). Once again, there is undersearch instead of oversearch with more than 20 per cent of consumers buying the 3-for-2 offer at the first shop when, in fact, they should continue their search. A particularly interesting case arises at the second store where it can be optimal to buy only a single item of the good on offer. This is psychologically demanding as it requires that consumers decide on purpose to buy the product at a higher per-unit price than the price that is on offer. Due to the parameter constellations this case does not arise very often: There are only 41 observations. In 23 cases consumers do

5.41

5.42

5.43

5.44

29

Of course, it is not easy to learn this peculiarity in the experiment. Remember that subjects are not told the precise baiting strategy that shops employ and even if they did know it, it would require some non-trivial computations to determine the optimal rule for choosing among the two shops.

OFT1226 | 70

optimally buy one unit and there is no discernible pattern in deviations. Not buying is almost as frequent as buying the offer. 5.45 Finally, let us investigate the sales frame, in many ways the weakest treatment as it should be easy for subjects to understand that the former price that is mentioned is entirely meaningless. While the welfare losses that subjects incurred under the sales frame were statistically not significant we did find significantly higher error rates in our econometric estimations. As Tables 5.15 and 5.16 show these errors mainly arise at the first shop where only around 60 per cent of all choices are optimal. Interestingly, even the sales frame destroys the strong pattern of oversearch that we detected in the baseline. However, this is not reversed into an undersearch pattern. Rather over- and undersearch are roughly equally frequent.

Table 5.11: Optimal vs actual choices at the first shop visited under baiting
Actual choice 0 Optimal choice 0 1 2 Total 187 21 18 226 1 84 40 17 141 2 36 17 234 287 3 2 0 1 3 4 0 0 3 3 Total 309 78 273 660

OFT1226 | 71

Table 5.12: Optimal vs actual choices at the second shop visited under baiting
Actual choice 0 Optimal choice 0 1 2 Total 28 7 1 36 1 4 62 9 75 2 0 2 64 66 3 0 0 1 1 4 0 0 2 2 Total 32 71 77 180

Table 5.13: Optimal vs actual choices at the first shop under complex pricing ('3 for 2')
Actual choice 0 Optimal choice 0 3 347 34 1 23 1 2 1 6 3 102 214 4 0 2 Total 473 257

Total

381

24

316

730

OFT1226 | 72

Table 5.14: Optimal vs actual choices at the second shop under complex pricing ('3 for 2')
Actual choice 0 Optimal choice 0 1 3 Total 52 8 8 68 1 4 23 13 40 2 0 0 5 5 3 6 10 214 230 4 0 0 2 2 Total 62 41 242 345

Table 5.15: Optimal vs actual choices at the first shop visited under a sales frame
Actual choice 0 Optimal choice 0 1 2 Total 300 30 53 383 1 56 43 12 111 2 10 2 152 164 3 0 0 0 0 4 0 1 1 2 Total 366 76 218 660

OFT1226 | 73

Table 5.16: Optimal vs actual choices at the second shop visited under the sales frame
Actual choice 0 Optimal choice 0 1 2 Total 42 16 1 59 1 9 106 11 126 2 0 13 95 108 3 0 1 2 3 4 1 1 0 2 Total 52 137 109 298

Table 5.17: Achieved sales potential

Frame per cent sales potential shop 1 88.4 122.9 87.3 117.9 101.2 111.8 per cent sales potential shop 2 103.4 97.5 98.3 96.9 140.4 97.2

Baseline Complex Sales Baiting Time-ltd offers Drip pricing

5.46

Table 5.17 compares all treatments by showing the achieved sales potentials for the two shops. Notice that this table is not equivalent to actual total sales as it does not include sales through return visits or sales to consumers who strayed from the optimal strategy at a previous

OFT1226 | 74

stage. Therefore, the table is more indicative of consumer behaviour than of total profitability for firms which will analyse separately below. 5.47 The table reveals one more stark effect, the large sales volume at shop 1 under complex pricing. This is simply due to the fact that 21.6 per cent of those customers that should continue to search actually buy three units at the first store. Of course, these numbers cannot directly be compared to those under the other price frames as effective prices are much lower when '3 for 2' offers are purchased. The overall picture from this more detailed analysis is that price frames are effective because they mainly change behaviour at the first shop. To some extent they reduce inefficient oversearch. However, in particular under drip pricing, and as we can see from Table 5.11 also under baiting, the picture actually reverses and consumers buy too often at the first shop. Relative to the baseline this implies even worse outcomes for consumers. Given the general tendency of experimental subjects to explore environments they are faced with rather more fully than they may in the field, our results presumably underestimate the consumer detriment caused by these practices simply because an experiment like this will always bias subjects towards some oversearch. And clearly, with less search, the baseline would perform better and the different price frames would perform worse.

5.48

5.49

Learning and IQ
5.50 Given that the decision environment in this experiment is not completely trivial and given that subjects take the same kind of decisions repeatedly, there is, just as in real life, ample scope for learning. In this sub-section we discuss whether there is any evidence for learning and whether learning is different across the different price frames. In a first step we run the regressions on errors in subjects behaviour (errors are documented above in Tables 5.2, 5.3, and 5.4) again, this time controlling for a linear time trend (see the technical annexe for precise details, estimation model probit 2.a. 2.c.). Crucially, the

5.51

OFT1226 | 75

estimates on the coefficients reported above turn out to be robust to the inclusion of such a time trend. 5.52 The linear time trend variable turns out to be highly significant and important in terms of its impact. The estimate is that per period the error rate falls by 0.58 percentage points. This may initially look like a small number but remember that there are thirty periods altogether. A linear approximation would, thus, suggest that subjects error rate is more than 15 percentage points lower at the end of the experiment than at the beginning. Allowing for non-linear effects (through a quadratic term) suggests that learning is levelling out towards the end of the experiment. Comparing the size of the learning effects with the overall error rates that we have examined above we can conclude that learning substantially reduces but does not eliminate erroneous behaviour. Decomposing the errors into search and purchasing errors, similar pictures emerge although the speed of learning is slower when it comes to search errors. These are estimated to decline by 0.22 percentage points per period while purchasing errors decline by 0.53 percentage points. In a further set of estimations we also add results from an aptitude (IQ) test that we conducted at the end of the experiment. The aptitude score is between zero and 12 and the marginal effect of one extra point is estimated to be 2.06 percentage points for overall errors, 0.88 percentage points for search errors, and 2.02 percentage points for purchasing errors in the expected direction. All three coefficients are highly significant. In a further stage we add interaction terms between price frames and the linear time trend in order to examine whether learning speeds differ between different time trends. For the overall error rate we find no differences in learning speed between the baseline and the different price frames with the exception of time-limited offers. For time-limited offers learning is significantly slower (and very significantly so). In fact, the estimate is so big that it completely wipes out the general learning trend. On balance, we find that there is no learning at all under time-limited

5.53

5.54

5.55

OFT1226 | 76

offers. Consumers appear to believe that the first shop will always charge a higher price upon return and therefore typically do not return. Hence, they can never learn that their belief is wrong. To what extent this holds in real-life markets will crucially depend on the precise informational structure. For example, for shops located along a high street, consumers might notice if time-limited offers advertised in a window display turn out to be repeated over time or are replaced by even lower prices. In other markets where such information does not come for free it would be similarly hard to learn as in our experiment. 5.56 The same findings hold for the decomposed search and purchasing errors: There is no differential speed of learning between the baseline and the other price frames with the exception of time-limited offers where there is simply no learning at all. Neither search, nor purchasing errors are reduced over time in the presence of time-limited offers. In fact, search errors are even slightly increasing over time for time-limited offers. Clearly, this examination of learning renders our finding that consumers have trouble with time-limited offers substantially more worrying. Not only is it difficult for subjects to optimally adjust their search strategy in the presence of time-limited offers they also are not able to improve their performance over time. As we have evidence on declining error rates that are attributed to learning we can also examine whether consumer welfare increases over time. For that purpose we conduct simple non-parametric tests to compare subject welfare in the first half of the experiment with subject welfare in the second half of the experiment. (All test statistics are to be found in the technical appendix.) Welfare is found to be significantly increasing in the second half of the experiment under all price frames (including the baseline) with one exception: time-limited offers. This is shown in the following table we report welfare loss in the first and second half of the experiment sessions by price frame.

5.57

5.58

5.59

OFT1226 | 77

Table 5.18: Welfare loss across first and second half of the experiment sessions
Frame Baseline Complex Sales Baiting Time-ltd offers Drip pricing first half second half Average 5.66 7.47 7.10 5.83 7.64 10.69

8.65 9.87 10.64 8.03 8.72 14.98

2.56 5.07 3.56 3.63 6.46 6.04

Behavioural forces at work

5.60 We conclude our empirical analysis of consumer behaviour by reflecting on what the data tell us about behavioural forces at work. We briefly discuss each of the five pricing practices under investigation. Drip pricing: As we have seen drip pricing turns consumers who tend to search too much into consumers who tend search too little. This effect is particularly striking as in our experiment the drips are revealed through just two mouse clicks and consumers can see the total price very clearly before they make their final purchasing decision. The objective costs of going through the drips are very close to zero (just a few seconds that pass for the two clicks). Accordingly, it is completely implausible to attribute the change in behaviour to increased costs of search and sunk costs. In principle, consumers might rationally decide to accept higher prices after being led through complicated drips if they were to expect equally costly practices elsewhere. Accepting a higher price at the first outlet would after all avoid the costs of clicking through a labyrinth at a competing outlet. Notice that this is an entirely rational response and has nothing to do with the so-called sunk cost fallacy where consumers

5.61

OFT1226 | 78

ignore that they cannot recover the costs of (failed) activities and, consequently, may 'throw good money after bad'. Here the extra search costs are so tiny that any explanation along the sunk costs line is simply not justified. Rather the data suggest that consumers who see a low base price and do not yet know that the effective price will go up through 'shipping and handling' charges experience an increase in their willingness to pay for the good which is in line with loss aversion and the so-called endowment effect.30 Consumers who decide to buy the product at the low price experience a shift in their reference point as they already imagine departing with the good. Changing the initial decision, that is, giving up the good that is already in the virtual basket would be perceived as a loss. This loss can be avoided by purchasing the product despite an increased price. 5.62 Time-limited offers: Time-limited offers eliminate oversearch at the first store but have an even more dramatic effect on the sales at the second store. Consumers who have rejected the time-limited offer at the first store simply tend not to return to it even if the price at the second store is comparatively high. The underlying problem in consumer behaviour is obvious: As the consumers believe the store (that is, as they erroneously believe that prices will go up) they have a) a tendency to buy more now at the first store, and b) if they dont buy at the first store, they buy at the second shop if this appears at all profitable. This false preconception can then never be revised, simply because consumers do not learn that time-limited offers are not real offers with true discounts but that prices can go up and down once they expire. Consumers simply do not understand the real mechanism because of naive beliefs, confusion and too little exploration. In all, this effect appears to be due to purely cognitive problems. Baiting: As under drip pricing consumers buy too much at the first shop under baiting while their behaviour at the second shop is largely optimal. As baiting generates expectations about good deals, it appears once

5.63

30

This is similar to the findings of previous researchers such as Morwitz, Greenleaf and Johnson (1998), and Hossain and Morgan (2006) discussed in chapter 3.

OFT1226 | 79

again that it is loss aversion and the endowment effect that is at work. Consumers pick a store from which they expect a good deal and this act in itself raises their willingness to pay for the good in comparison to the baseline. In contrast, to standard experiments on the endowment effect it is, once again, anticipated or imagined ownership that causes the positive shift in consumers valuation of the good. 5.64 Complex pricing: This frame is different from the other price frames in that it actually lowers the prices. Controlling for that, we find that consumers make slightly higher welfare losses under complex pricing than under the baseline. Error rates are not significantly higher than in the baseline but the type of error is very different. Instead of oversearch consumers undersearch and buy too often at the first shop. Notice that they do not simply buy too many units of the good (which would follow from an endowment-effect like shift in valuations) but that they do buy the offer. This suggests that the offer has an attraction beyond the mere reduced price or that consumers are cognitively limited. Sales frame: Finally, we find that somewhat surprisingly even the meaningless by-line 'was X' where X is a higher price than the current price eradicated oversearch. However, the sale frame is not able to trigger loss aversion or the endowment effect to the point where consumers start to undersearch. Rather it generates evenly distributed errors suggesting mainly cognitive problems with processing information that optimally should be discarded. In light of the behavioural forces that we have identified it is worthwhile to revisit the issue of how search costs impact on consumer choice. As we have seen earlier, search errors are falling in search costs under drip pricing, baiting and time-limited offers precisely those treatments where we now conclude that search errors are driven by loss aversion. So why and how do search costs matter in these treatment? The nexus is simple. Loss aversion increases consumers willingness to pay at the first shop (once they imagine themselves as owner of the good). Now we simply need to observe that the larger the search costs, the more likely it is that the decision to buy at a comparatively high price is actually optimal! In other words, in environments with large search costs

5.65

5.66

OFT1226 | 80

it matters less if consumers want to buy straight away because they would experience not buying as a loss. 5.67 We have also argued that purchasing errors are either due to cognitive failure or to the sunk cost fallacy and we have seen earlier that they increase with higher search costs under all treatments with the exception of time-limited offers and drip pricing (but including the baseline). The basic intuition for the impact of search costs on the sunk costs fallacy is obvious. The higher the search costs consumers have spent, the bigger their urge to justify them through purchasing too many units. It is less clear why this interacts with the different treatments. One possibility is that, insofar purchasing errors occur after prolonged search, that those who are prone to loss aversion are less likely to make them (simply because they tend to buy straight away at the first shop). Furthermore, if these biases are correlated (as recent research by Burks et al. (2008) suggests) then this implies that under those treatments that trigger loss aversion there will be fewer biased consumers who do long searches. In other words, the heterogeneous treatment effect on purchasing errors is quite plausibly simply due to a selection effect. This would then suggest that the sunk cost fallacy occurs indeed independent of price framing.

Feelings
5.68 At the end of each session, and before the subjects left the laboratory, they were asked to complete a qualitative questionnaire about their experiences with the price frames in the experiment.31 We now turn to the individual price frames and subjects' reported experiences. In the drip pricing frame subjects reported annoyance and irritation because the additional charges were not shown with the base price.

5.69

5.70

31

The questionnaire was optional, and was not incentivised. Further, as previously mentioned this is controlled laboratory experiment and as such we used university students. Therefore, this survey is not designed to be representative of the general population.

OFT1226 | 81

Subjects often reported feeling disappointment because they felt they were receiving a good deal when they saw the base price, and only after they had decided to buy the good were they then told that they would incur additional charges. Subjects reported that they still bought the good even after they were informed of the additional costs, but felt cheated/annoyed because their profit on the units were reduced. 5.71 In the sales frame, 59 per cent of responses we received reported that the 'special offer - discounted price' had no influence on their decision making; subjects continued to make their decision based on the post sale price. However, 41 per cent of subjects reported that the special offer did influence their decision making.32 These subjects felt they were 'getting a good deal', 'saving money', 'receiving a discount' and therefore were more tempted to buy the product. In the complex pricing frame, subjects reported that they recognised that the offer may not always be the best choice for them. Subjects report that they would buy only one unit if the return to them was better than buying two and getting the third unit for free. However, 32 per cent of respondents to this question reported that always chose the offer. Baiting, subjects' reported that the frame enticed them to go to the shop with the lower offer first and to do it quickly (click quickly) so as not to miss out. Subjects reported anger, frustration and disappointment if the offer price had changed from the advertised price when they got to the store. However, they may buy anyway. Time limited offers, subjects reported that this frame 'compelled [them] to buy', or it was 'something not to be missed'. Subjects reported they found it strange that the price upon return to the shop may be lower than when they first visited. And, some subjects reported that the special offer enticed them to buy without searching further.

5.72

5.73

5.74

32

This is in line with our results which find that the essentially meaningless sales frame in this experiment does encourage subjects to reduce oversearch.

OFT1226 | 82

5.75

If we consider the subjects' experience in the experiment session as whole we find the following qualitative observations. In regard to annoying aspects: he additional costs in the drip pricing frame were by far the most reported annoyance to subjects. subjects reported annoyance if the baiting price was not available once they travelled to the shop. subjects were annoyed if the price at the second shop visited was higher than the first visited shop, and annoyance arose because subjects wanted to know the prices before choosing to travel.

5.76

5.77

When asked about enjoyable aspects subjects reported the following: making a profit was, not surprisingly, the most enjoyable aspect reported by subjects subjects enjoyed the choice aspect of the experiment, namely the decision to continue to search or not similarities with real shopping experiences 'felt like high street shopping with the travel costs', 'shipping and handling is like online shopping', 'makes me realise the calculations I do subconsciously when shopping'.

5.78

Finally we asked subjects to report on how they thought the experiment mirrored actual shopping experiences. Subjects reported the following: many subjects reported that the price frames mirrored well those they have experienced when shopping subjects reported that they encountered the same trade-off in real life as to whether to keep shopping around for a better deal or to simply accept a price that is reasonable (to them) in the first shop they visit

OFT1226 | 83

there were different feelings about the travel costs. Some subjects reported that the travel costs reflected the real costs incurred to search or not, but others felt that the travel costs were too high because it is only 'their time' and 'going for a walk' which they must invest to go to another shop, while others reported that they always check prices online which is quick and easy before choosing to buy anything the experiment did not capture the emotional aspect of shopping, for example, 'if I love it I will buy it no matter the price or whether I need it' quality differences between the same type of product was not included which is sometimes important when comparing between shops in the real world.

Sellers and total welfare

5.79 In our analysis we have so far very much focussed on the effect of price frames on consumer behaviour and welfare although we had some results that pointed towards the effect on firms and, generally, the two are of course closely linked. While shops were completely computerized it is still interesting to ask how their performance is affected by price frames. Price frames that perform well for the shops are not necessarily those that are detrimental for consumers. Some price frames could hurt both buyers and sellers. Of course, one would expect that in a market environment mainly those price frames are used that actually improve sellers performance. Table 5.18 shows average number of units sold by the two shops under the different price frames as well as average turnover. The table also contains units sold and turnover for the entire industry. These industry performance indicators are perhaps the most important ones as, a specific shop, is in our experiment equally likely to become the first or the second shop.

5.80

5.81

OFT1226 | 84

Table 5.18: Average number of units sold by the two shops

Frame Units shop 1 Units shop 2 All units (Industry indicator) Baseline Complex Sales Baiting Time-ltd Drip pricing .95 1.67 .95 1.23 1.02 .99 .63 1.06 .64 .38 .57 .60 1.58 2.73 1.59 1.61 1.59 1.59 73 133 73 94 79 77 52 88 52 30 42 47 Sales shop 1 Sales shop 2 All sales (Industry indicator)

125 221 125 124 121 124

5.82

The table shows impressively that, contrary to what we might have expected, the shops gain nothing from employing the different price frames. In fact, under time-limited offers, which is one of the two most detrimental practices for consumers, shops experience even reduced sales. Sales for complex pricing are, of course, substantially higher but whether this would translate into higher profits depends entirely on unit costs which remain unmodelled. If unit costs are equal to the average market price (as suggested by a Bertrand model) then complex pricing would not be profitable. If there is a higher average mark-up (say if costs were equal to the lowest price ever charged) then complex pricing would be profitable for firms. While the other price frames have essentially no effect on industry performance there are dramatic shifts in the distribution of sales, generally to the advantage of the shop that is first visited (which is in line with what we have said above about the reversal from oversearch to

5.83

5.84

OFT1226 | 85

undersearch). While in our experiment, the sequence in which shops are visited is essentially random (with the exception of baiting), firms can in many markets influence the order in which consumers search, for example through advertising or sponsored links in search engines. Our results suggest that the incentives to engage in such activities are substantially increased through elaborate price framing. We would thus predict that, across different markets, the occurrence of price framing is positively correlated with the amounts firms spend on enticing consumers to search their offers first. This is a prediction that could be easily tested using data from auctions for sponsored search. 5.85 Of course, such activities to change consumers search order would reduce total welfare even further. In our setup, price framing reduces total welfare already in any case as it harms consumers and does not benefit firms. However, contests for being searched first would render the picture even bleaker.

Summary
5.86 In Table 5.19 we attempt a summary of our main findings, indicating for each price frame how it impinges on welfare and errors, how it affects stores; which kind of behavioural bias (if any) can be identified to be driving the data; and whether learning helps consumers to overcome or reduce the effects of the bias.

OFT1226 | 86

Table 5.19: Summary of main findings

Significant welfare losses Drip pricing Large Significantly more errors first shop benefits Behavioural biases Learning helps

Yes, substantially

Yes

Endowment Yes effect/loss aversion/maybe sunk cost fallacy Cognitive errors/maybe sunk cost fallacy Endowment effect/loss aversion/sunk cost fallacy Cognitive errors/sunk cost fallacy Cognitive errors/sunk cost fallacy No

Time-ltd offers

Medium

Yes

Yes

Baiting

Medium

Yes

Yes, strongly

Yes

Complex pricing

Small

No

Yes

Yes

Sales frame

No

Yes

No

Yes

OFT1226 | 87

EXTERNAL VALIDITY AND IMPLICATIONS FOR NONLABORATORY MARKETS

External validity for studies of this nature is typically highly asymmetric. Both simplification and stylization of the decision problems and the highly selected subject pool imply that we are more likely to observe good or even optimal performance in the laboratory than in the field under real-life conditions. This implies that whenever we find close to perfect performance external validity is severely limited. If student subjects do well in a simple task it is hard to conclude from that the general population would do well in a more complicated task. However, the other way round things look much brighter. If a highly selected student sample does badly in a simple decision environment that also offers scope for repetition and learning, it would be very surprising if the general population did much better in more complicated situations. Given this asymmetry, the design choices we took were risky. We designed a very simple search environment that was, given our requirement (multiple units, two stages) pretty much minimal. It would have been difficult to come up with an even simpler environment. Similarly, the implementation of the price frames was typically simple and subjects could experience them repeatedly in almost identical manner. The risk of these design choices was that we might have found close to optimal performance in all treatments in which case we would have learned very little from this study. The alternative strategy to design a more complicated environment would have entailed different risks. In more complicated environments decision errors and noise will invariably go up and accordingly it will be more difficult to detect differences between treatments for given sample sizes. However, as we have documented we have observed substantial difference between treatments with surprisingly poor performance in some. The results on drip pricing stand out. Being 'just two clicks away' from the baseline its effects on search patterns and performance are dramatic. Outside the laboratory where drip mechanics are more

6.1

6.2

6.3

6.4

OFT1226 | 88

elaborate and it is more time consuming to reach a stage where full prices are clearly visible it is likely that the effects that stem from loss aversion or the endowment effect will be even stronger. On top of that, more elaborate drips will also increase the true costs of searching for the price which will enhance the effects from loss aversion. Similarly, we have not tried to optimize the drip sizes and it would be very bewildering if we had, by accident, stumbled across the most effective drips. Once again, this suggests that, if anything, we might still underestimate the true consumer detriment resulting from drip pricing. 6.5 Given our basic design choices and selection of subjects the observation holds, of course, more generally: there is a built-in tendency to underestimate consumer detriment for all price frames. This also implies that the sales frames which receives almost a clean bill of health in this study could potentially be more harmful than we detect. As soon as we are in environment where former prices contain some hard information there is much more scope for consumers to process this information in a less than adequate way. On the other hand, references to former (higher) prices could also serve as a useful signal of quality. In this context it is important to notice that, with the exception of complex (3 for 2) pricing, we isolate the pure effect of price frames and not of offers. Of course, real offers with lower prices might benefit consumers even if their presentation confuses them or triggers some behavioural biases. The net effect of lower prices and the adverse consequences we measure here might still be positive. What this study shows, however, is that also real offers could benefit consumers more if presented in a straight way as in our baseline treatment. In our experiment, both price frames and prices themselves are exogenously fixed while in real markets they are, of course, chosen. Given our consumer data it appears clear that certain price frames will allow firms to charge higher prices (in particular those that trigger loss aversion and the endowment effect as these effects are akin to increased willingness to pay or an outward shift of the demand curve). Consumers would then suffer doubly, from their direct negative consequences we measure in this experiment and from the higher prices.

6.6

6.7

OFT1226 | 89

6.8

One aspect of endogenous choice of price frames that we have not studied at all is that sellers in the same market might choose different price frames which makes price comparisons much harder. As Chioveanu and Zhou (2009) show these effects can even overturn standard intuition on how the number of firms in a market relates to consumer welfare. With added confusion from a greater variety of price frames, consumers might actually suffer from the entry of additional sellers. On the other hand firms may elect to not use price frames that annoy customers. Firms may seek to establish a reputation for not using annoying practices, such a drip pricing. While we can neither validate nor reject these theoretical predictions we can say a little about how our findings on relevant behavioural biases would impact on markets. Clearly, the strongest force that causes consumer detriment in our experiment is the endowment effect or loss aversion. Consumers imagination of owning a good shifts their willingness to pay. We observe strong evidence on this in both drip pricing and baiting. In the field there will be many other practices of hot selling that play on these effects. If the consumer tries out a product in a shop it will give him some objective information about how the product handles but it also makes envisaging ownership easier and what we have seen here is that envisaging ownership is all that is needed to increase willingness to pay. There are, of course, many institutional and physical details that will matter for the effect of these practices in non-laboratory markets. For example, it might be easier to encourage the imagination of ownership for some goods than for others. By thinking about the product characteristics that make imagination of ownership easier, we could then derive comparative static predictions about in which markets we would expect to observe certain price frames more frequently. Similarly, there might be particular characteristics of sellers that tinge the decision problems in real-life markets. For example, for closing-down sales (where, say, the consumer can see that a building is about to be

6.9

6.10

6.11

6.12

OFT1226 | 90

torn down) the time-limitedness of offers might be more credible than for other 'mid season' sales. 6.13 For one important aspect of real-life markets we do have some indication in our data, the role of search costs. Our analysis suggests that the detrimental effects of loss aversion increase with lower search costs. This is intuitive: Buying too early too often, tends to coincide with optimal behaviour when search costs are high. On the other hand, we have found that the sunk cost fallacy (that drives some of the purchasing errors) increases with search costs. Again this seems plausible for real-world applications. The more time and money consumers have spent on search, the more desperate they might be to justify these high search costs through making (too many) purchases. In our experiment, this effect is much smaller than the effect of search errors which can be traced to loss aversion. However, this may well be a consequence of our parameter choice. For example, in markets with very high search costs (for example, because of location in just one city in a larger geographical area, say, some sort of fair) it is plausible that consumers might be very frustrated to leave empty-handed and would thus be tempted to buy more than they would have bought had the market taken place at their doorstep. Summarising, let us stress however again that we have good reason to believe in the general external validity of our results that these practices do cause consumer detriment and that what we identify in the lab is probably rather the tip of the ice berg as there are many aspects of real-life markets that will accentuate the problems we document here.33

6.14

33

Although inevitably there are also likely to be some factors in real life which will mitigate concerns about practices (even when frames are false offers) such as the desire for firms to build reputation, and consumers to learn about honest firms, as discussed in the next section.

OFT1226 | 91

7
7.1

POLICY RECOMMENDATIONS
From a policy point of view, this experimental study has two broad implications. We identify price frames that clearly cause consumer detriment. As they do so in a comparatively simple environment and with a comparatively sophisticated subject pool, it appears clear that these practices are also harmful outside the laboratory. Our findings suggest that firms will be tempted to confuse consumers through drips, baits or time-limited offers. Clearly, the occurrence of any of these three practices which do not represent genuine offers might provide reason to worry. While drip pricing has previously been known to be problematic, timelimited offers have never before been identified as a source of consumer confusion and detriment. As we have shown, consumer errors under time-limited offers are particularly severe in that they are not reduced through learning (at least in this setting). Thus, even in markets for goods that are purchased frequently, one would expect that time-limited offers are used and indeed problematic. Consequently, one may be particularly worried about drip pricing in markets for not particularly frequently purchased goods or time-limited offers in markets for goods that are purchased with high frequency.34 We also have clear indication that in this experimental setting these effects are particularly pronounced in environments with low search costs for consumers. However, this result would require further analysis before clear policy conclusions could be drawn from it. Not least because, with higher search costs consumers become prone to the sunk cost fallacy and this effect could become much stronger if search costs are substantially higher.

7.2

7.3

7.4

7.5

34

Clearly, as a seller one would rather employ frames that are robust to learning for frequently purchased items.

OFT1226 | 92

7.6

If the main effect of price frames is that they shift demand from one firm to another, enforcement (which helps to promote a level playing field) may be welcomed not only by consumers but also by firms. But this needs more research into environments with endogenous prices. This could be done in the same experimental framework. If consumers are annoyed by some price frames, there is some scope for the self-healing powers of the market. Firms may gain a reputation for not using such practices. This is more likely to work for practices that are indeed perceived as an annoyance such as drip pricing. In contrast, time-limited offers might, in fact, be perceived as something positive by consumers who will not even be aware of their struggle with understanding the true nature of such offers. Accordingly, one would have less hope that the market can overcome the occurrence of such positively perceived practices. These issues could be investigated in a similar experimental study where firms are no longer simple static computers.

7.7

OFT1226 | 93

SCREEN SHOTS FROM THE EXPERIMENT

Figure A.1: Home Screen in baseline

OFT1226 | 94

Figure A.2: Shop 1 screen in baseline

OFT1226 | 95

Figure A.3: Purchase confirmation dialogue in baseline

OFT1226 | 96

Figure A. 4: Purchase notification

OFT1226 | 97

Figure A.5: Confirm exit from period

OFT1226 | 98

Figure A.6: Results screen

OFT1226 | 99

Figure A.7: Shop 1 screen in Complex pricing

OFT1226 | 100

Figure A.8: Drip 1 screen

OFT1226 | 101

Figure A.9: Drip 2 screen

OFT1226 | 102

Figure A.10: Confirmation in drip pricing

OFT1226 | 103

Figure A.11: Home screen in baiting

OFT1226 | 104

Figure A.12: Shop 2 screen in baiting (price different from advertised)

OFT1226 | 105

Figure A.13: Home screen in baiting after returning home

OFT1226 | 106

Figure A.14: Shop 1 screen in sales price frame

OFT1226 | 107

Figure A.15: Shop 1 screen in Time limited offer

OFT1226 | 108

B
B.1

WEB TRAWL OF PRICING PRACTICES

Here we provide details of the web trawl. The purpose of the trawl was to collect information on the practices used by different firms in the UK. The web trawl was completed between the 28 September and the 9 October 2009. This trawl was completed as background information to guide the design of the experiment treatments. Industries chosen for the web trawl are illustrative only; the experiment is concerned with the impact of pricing frames in general and not just for these industries listed. The web trawl includes the following industries: airfares package holidays accommodation and hotels theatre tickets electronic goods (for example, audiovisual and CD players) computers and computer equipment, and furniture.

B.2

B.3

B.4

A summary of the pricing frames used by different firms is presented in the tables below.

OFT1226 | 109

Table B.1: Drip pricing

Industry and Firm
Airfares Carrier 1

Number of compulsory drips

Headline price does not include taxes/fees Online check-in (5.00 pp) Handling charge for payment (5.00 pp, per flight), not shown until last page.

Number of optional drips

Priority boarding (3 pp) Bags in hold (10 first bag, 20 second and third each , max 3) Add bulk sports item, sports equipment, musical instrument, ski (30) Add baby equipment (10) Request SMS confirmation (1.00)

Shrouded drips

Other details/feat ures

List of all fees and charges available using link on booking page. No (clear) mention of additional cost of food/drink on board.

Handling charge not seen until last page.

3 compulsory drips

Carrier 2

Headline price includes government taxes and airport fees. Handling charge for booking 2.95 Credit card fee 2.5 per cent of total cost or 4.00 whichever is lowest.

2 compulsory drips

Up to 5 optional drips Priority boarding (8.00 pp) Bags in hold (8.00 per bag up to max 8 with total weight 20kg) Extra kilos in hold (21 per 3 kilo lots up to 30kg max) Add sports equipment (18.50 per unit of equipment) Carbon offset programme depending on distance Up to 5 optional drips Carbon offset depending on distance. 1 optional drip

Found no shrouding

1 bag per person is the default. Travel insurance included is the default. No mention (clear) of additional cost of food/drink on board.

Carrier 3

Price includes taxes fees and charges (which can be seen separately at bottom of screen) Price includes handling fee (but not applicable from UK). Price includes 1 hold bag up to 23kg. Credit card payment method 4.50

Found no shrouding

Uses selling feature that all charges are included in price, including food on board. List of any fees and charges available on booking page

OFT1226 | 110

Industry and Firm

Number of compulsory drips

1 compulsory drip (credit card) Headline price includes taxes fees and charges which can be seen separately. Drip for credit charge fee. 1 compulsory drip Headline price does not include taxes and charges. Check-in fees: online no bags 0, online 1+bags 30, airport 0 bags 3.00, airport 1+ bags 6.00 Handling fees for debit card (3.5 per cent), credit card fees (3.5 per cent + 2.25 per cent), PayPal (3.49 + 1.5 per cent) 3 compulsory drips (unless check in online with 0 bags)

Number of optional drips

Shrouded drips

Other details/feat ures

Carrier 3

No optional drips but travel insurance is available if wanted. 1 optional drip

Found no shrouding

Carrier 4

Bags in hold (9.99 per bag max of 3 max weight 22kg) Seat selection: standard 6.99, extra leg room 15.99. Add sports equipment 20 up to 20kg Travel insurance 6.50 pp. Meals onboard (8.00)

Found no shrouding

List of fees and charges available on booking page.

5 optional drips

Carrier 5

Headline price does not include taxes and fees.

Payment by credit card 4.00 pp No charge for debit card

No shrouding found

List of fees and charge on booking page

1 compulsory drip 1 optional drip Package holidays Company 1

Credit card charge

1 compulsory drip

Offset emissions (4.41) Book a car (7684.86) Airport parking (30.95) Travel insurance (21.68-43.36) Credit card fee 5 optional drips Upgrading possibility of accommodation

No shrouding found

Company 2

Fuel supplement (20) TOD charges

No shrouding found

First price includes optional

OFT1226 | 111

Industry and Firm

Number of compulsory drips

(30)

Number of optional drips

(84-238) World Care Fund donation (2 per person) Insurance (19.98-37.98) Credit card charge if dont pay by debit card

Shrouded drips

Other details/feat ures

donation Coach transfer is free and optional Online saving possibility (386) No flight meals Holiday+fligh t party price includes taxes, online discount, donation to World Care Fund and return airport transfer Price changes due to the system updates (increases and decreases)

2 compulsory drips

4 optional drips Company 3 No compulsory drips Junior suite (98) No flight meals included Insurance and excess waiver Credit card charge if dont pay by debit card Family holiday choice shows price only per person at first

Company 4

System updates (price differs from the one shown at first )

1 compulsory drip

3 optional drips Accommodation options half board (98) Resort transfer (20 per person) Insurance (19.99 per person) 3 optional drips Price of the holiday depends on the flight time one selects (varies up to 30) In flight meal (15) Extra luggage (5) Resort transfer (20) Seats together Mid-hl (12) 5 optional drips Transfer (227361) Car hire (140-

Credit card charge not shown

Company 5

Adult supplement (105) Fuel supplement (10) Late booking fee (15)

Credit card charge not shown

Allowed 15kg luggage Online booking discount 10

3 compulsory drips

Company 6

Airport taxes Fuel supplements

Credit card charge not shown

A notice that it is a fair price that

OFT1226 | 112

Industry and Firm

Number of compulsory drips

1 piece of hold luggage Full financial security of ABTA and ATOL membership (no prices shown for the services)

Number of optional drips

285) Car parking (5888)

Shrouded drips

Other details/feat ures

includes all UK airport taxes, fuel supplements, 1 piece of hold luggage per passenger, and full financial security of both ABTA and ATOL Membership

3 optional drips

4 compulsory drips

Theatre tickets Provider 1

Booking fee (4.95)

No optional drips

No shrouding found

1 compulsory drip Provider 2 Service charge (5.30) Order processing fee (jncluded) Delivery charge for customers who live outside of Republic of Ireland, Northern Ireland and Great Britain (3) No optional drips No shrouding found

On the first page it says 'from 10', actual price 49.50 Notes that all prices exclude fees and charges

Provider 3

3 compulsory drips Booking fee (3.50) Transaction fee (2.25) 2 compulsory drips Commission fee (30.70) Delivery charge depending the country 2 compulsory drips Delivery charge (7.50-40) Booking fee (13.35) IVA Tax (2.14) 3 compulsory drips Delivery (1)

No optional drips

No shrouding found

It is actually said in the first place what are the extra costs The commission and delivery fees is listed alongside the ticket price at the outset All the extra costs are listed alongside at the outset The delivery

Provider 4

No optional drips

No shrouding found

Provider 5

No optional drips

No shrouding found

Provider 6

No optional drips

No shrouding

OFT1226 | 113

Industry and Firm

Number of compulsory drips

1 compulsory drip

Number of optional drips

Shrouded drips
found

Other details/feat ures

cost is listed alongside at the outlet The booking cost is listed alongside at the outlet. Delivery cost is not shown not until last page

Provider 7

Booking fee (1.95) Delivery cost (1.79-1.95)

No optional drips

No shrouding found

1 compulsory drip

Furniture Company 1

Delivery fee (35) 1 compulsory drip

No optional drips

No shrouding found

Company 2

Delivery (5.808.70)

2 year insurance (40-99.99)

No shrouding found No shrouding found No shrouding found

Company 3

1 compulsory drip Delivery fee (5.80) 1 compulsory drip Delivery costs (5-20 or free if purchase over 400) 1 compulsory drip Delivery fee (25) 1 compulsory drip

1 optional drip No optional drips

Delivery cost is being mentioned on the left corner of the page, not easily noticeable Chance to collect purchased items at store

Company 4

No optional drips

Company 5

No optional drips

No shrouding found No shrouding found Free delivery statement on first page of purchase

Company 6 No compulsory drips Computers Company 1 No compulsory drips

Pre-paid returns (4.95)

Accessories like speakers, number pads etc. (4.9999.95)

No shrouding found

Free delivery shown at the check out

Company 2

Up to 8 optional drips 'Other customer

No shrouding

Subtotal

OFT1226 | 114

Industry and Firm

Number of compulsory drips

No compulsory drips

Number of optional drips

also bought' variation of products offered (16-27) Up to 7 optional drips Offers to supplement the product bought (14-69) 'essential' accessories (4.99/6.86) Up to 7 optional drips (possibility to view even more) No optional drips No optional drips

Shrouded drips
found

Other details/feat ures

including VAT Free delivery

Company 3

Norton antivirus was added to the basket even when to remove the tick (29)

No shrouding found

No delivery charge

1 compulsory tick

Company 4 Company 5

No compulsory drips Delivery charge (3.40)

No shrouding found No shrouding found

VAT included

Company 6

1 compulsory drip Delivery charge (5.65) 1 compulsory drip

No optional drips

No shrouding found

Price with and without Vat is shown

Audivisual & DVD Company 1

Company 2

Delivery charge (international fee 6.48$) 1 compulsory drip Delivery charge (6.99-15.99) 1 compulsory drip No compulsory drips

No optional drips

No shrouding found

Free delivery with Super Saver Shipping

Accessories

No shrouding found

Company 3

N number of optional drips Accessories, products and services Standard (free) or premium delivery N number of drips Extra warranty (40) Cable offer (29) Delivery choice (next day, Saturday) up to

No shrouding found

Standard delivery free

Company 4

No compulsory drips

No shrouding found

5 days standard delivery free

OFT1226 | 115

Industry and Firm

Number of compulsory drips

Number of optional drips

9.95

Shrouded drips

Other details/feat ures

Company 5

Delivery charge (10)

3 optional drips No optional drips

No shrouding found

1 compulsory drip Company 6 No compulsory drips No optional drips No shrouding found

Price including and excluding VAT shown on the first page Free delivery after 3 days

Table B.2: Sales

Industry and Firm
Airfares Carrier 1 Carrier 2 Carrier 3 Carrier 3 Carrier 4 Carrier 5 Package holidays Company 1 Company 2

Existence of reference pricing

None none none none yes none

Example of reference pricing

none none none none 15 per cent off across Spain; Winter Sun Offer 10 per cent none

Shrouded references
No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

Other details/feat ures

None

Yes Yes

Hotel sale extended up to 50 per cent off Save 10 per cent guaranteed on all holidays departing in November ends soon; massive winter clearance 10 per cent guaranteed savings in November and December None Summer 2010: guaranteed 10 per cent off Prices from only ...

No shrouding found No shrouding found

To find special offers and deals, one has to click on the section before

Company 3 Company 4 Company 5

None Yes Yes

No shrouding found No shrouding found No shrouding

OFT1226 | 116

Industry and Firm

Company 6 Theatre tickets Provider 1 Provider 2 Provider 3 Provider 4 Provider 5 Provider 6 Provider 7 Furniture Company 1 Company 2

Existence of reference pricing

Yes

Example of reference pricing

(Mexico, Madeira etc) Bargains (Cyprus, Turkey etc) Up to 50 per cent off theatre; none none none none Half price and discount theatre tickets 'Concert tickets on sale'

Shrouded references
found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found Not sure if the 'original' price is accurate Not sure if the 'original' price is accurate No shrouding found No shrouding found

Other details/feat ures

Yes none none none none Yes yes

none yes

none Save at least 400 today (shows the 'original' price and the amount one saves) Save 40 on a dishwasher ('original' price and final price shown) 25 per cent off selected Bosch power tools; up to price selected furniture; Save 1/3 off on all modular and freestanding furniture; save 15 per cent all sofas; save 40 per cent piano wall hung fires none Autumn sale now on, prices reduced now up to 50 per cent; clearance sale no up to 70 per cent

Company 3

yes

Company 4

yes

Company 5

yes

Company 6 Company 7

none yes

No shrouding found No shrouding found

Computers Company 1 yes 1TB hard drive only 23.99; Acer 5536 Laptop save 150; massive stock clearance 140, 30 etc No shrouding found

OFT1226 | 117

Industry and Firm

Company 2 Company 3

Existence of reference pricing

none yes

Example of reference pricing

none Various products 'x, sale save over x, was x', save up to 200 on large kitchen appliances SAVE 45 was 244.95 inc vat 199.00 inc vat none none

Shrouded references
No shrouding found Not sure if the 'original' price is accurate No shrouding found No shrouding found No shrouding found

Other details/feat ures

Company 4 Company 5 Company 6 Audivisual & DVD Company 1

yes none none

yes

DVDs up to 80 per cent off;

No shrouding found

Info not on the front page, one has to click on the selection first

Company 2 Company 3 Company 4 Company 5 Company 6

yes none none none none

Denon DM37DAB save $50 199; none none none none

No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

OFT1226 | 118

Table B.3: BAITING SALES

Industry and Firm
Airfares Carrier 1 Carrier 2

Existence of baiting sales

yes Yes

Example of baiting sales

Homepage: Beds from 8; shuttle direct from 7 Munich from 29.99; ski flights to Geneva from 23.99 etc Edinburgh escape from 185 per person; Hong Kong city breaks from 579 per person etc Hotels all over the world (LA, Jamaica etc) from 43 (60, 100 etc); Caribbean holidays from 515 Winter sun getaways from 29.99 one way including taxes Cairo and Tel Aviv from $558

Shrouded baiting
No shrouding found No shrouding found No shrouding found No shrouding found

Other details/feat ures

Carrier 3:

yes

Carrier 3

Yes

Carrier 4: Carrier 5: Package holidays Company 1

Yes Yes

No shrouding found No shrouding found

Yes

Company 2

Yes

Hotels (London, Paris, Barcelona etc) from 35, Costa del Sol from 214 etc All inclusive holidays from only 267

No shrouding found No shrouding found To find special offers and deals, one has to click on the section before

Company 3:

yes

Company 4

Yes

Company 5 Company 6 Theatre tickets

none none

Winter escapes from only 99 pp; summer 2010 holidays from 140 Last minute holidays: from 149 pp; summer 2010 cruises from only 419 pp none none

No shrouding found No shrouding found No shrouding found No shrouding found

OFT1226 | 119

Industry and Firm

Provider 1 Provider 2 Provider 3 Provider 4 Provider 5 Provider 6 Provider 7 Furniture Company 1 Company 2 Company 3

Existence of baiting sales

Yes none none none none yes none

Example of baiting sales

Priscilla queen of the desert from 20 none none none none Blood Brothers discount tickets from 20 none

Shrouded baiting
No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

Other details/feat ures

none none yes

none none 1000s DVDs & CDs from 2.95; gift experiences from only 5 Up to price on selected furniture Save 1/3 off on all modular and freestanding furniture none Autumn sale now on, prices reduced now up to 50 per cent; clearance sale no up to 70 per cent

Company 4 Company 5 Company 6 Company 7

yes yes none yes

No shrouding found No shrouding found No shrouding found No shrouding found

Computers Company 1 Company 2 yes yes Great range of laptops from 249 Inspiron 546 feat. 6GB RAM and large hard drive from 299; inspiron 15 with 500GB hard drive from 349 Save up to half price on sat navs; save up to 1/3 on Panasonic phones none none Samsung Netbook range from 249.99, laptops No shrouding found No shrouding found

Company 3

yes

No shrouding found No shrouding found No shrouding found No shrouding found

To see the price one has to click to go further

Company 4 Company 5 Company 6

none none yes

OFT1226 | 120

Industry and Firm

Audivisual & DVD

Existence of baiting sales

yes

Example of baiting sales

from under 300 5MP3 albums at 5 songs from 29p; autumn offers: CDs from 2.98

Shrouded baiting
No shrouding found

Other details/feat ures

Info not on the front page, one has to click on the selection first

Company 1 Company 2 Company 3

none yes yes

none Your dream installation fully installed for only 1999.95 incredible! Free delivery on all orders LG KC550 Pay as You Go Mobile Phone 49.99 free delivery

No shrouding found No shrouding found No shrouding found Not sure how much does one save, since there is no 'before' price

Company 4 Company 5 Company 6

none none none

none none none

No shrouding found No shrouding found No shrouding found

OFT1226 | 121

Table B.4: TIME LIMITED PRICE OFFERS

Industry and Firm
Airfares Carrier 1 Carrier 2

Existence of time limited offers

Yes Yes

Example of time limited offers

Book until midnight 8.10.09 from $1 Book a hotel by 9 October 9 (the same offer as '5 for the price of 1') none Asia and Africa: book by 9th October; last minute deals all over the world None Earn 30 by joining Company Rewards before 31 December

Shrouded offers
No shrouding found No shrouding found

Other details/feat ures

Carrier 3 Carrier 4 Carrier 5 Carrier 6 Package holidays Company 1 Company 2

None Yes None Yes

No shrouding found No shrouding found No shrouding found No shrouding found

Yes Yes

20 meal and show deals for 20 for a limited time period Save 10 per cent guaranteed on all holidays departing in November ends soon; massive winter clearance 10 per cent guaranteed savings in November and December none Hurry escape before the deals do! none none

No shrouding found No shrouding found To find special offers and deals, one has to click on the section before

Company 3 Company 4 Company 5 Company 6 Theatre tickets Provider 1 Provider 2 Provider 3 Provider 4 Provider 5

none Yes none none

No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding

Yes none none none none

Chicago theatre tickets 48 hr sale up to 50 per cent off none none none none

OFT1226 | 122

Industry and Firm

Provider 6 Provider 7 Furniture Company 1 Company 2 Company 3 Company 4

Existence of time limited offers

none none

Example of time limited offers

none none

Shrouded offers
found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

Other details/feat ures

none none none yes

none none none Guaranteed kitchen installations before Christmas when you order before 11 October none none none

Company 5 Company 6 Company 7 Computers Company 1 Company 2

none none none

none yes

none Save up to 327 offers end 14.10.09; 14 days of deals, offer ends 14.10.09 none none none none

No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

Company 3 Company 4 Company 5 Company 6 Audivisual & DVD Company 1 Company 2 Company 3 Company 4 Company 5 Company 6

none none none none

none none none none none none

none none none none none none

No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

OFT1226 | 123

Table B.5: COMPLEX PRICING

Industry and Firm
Airfares Carrier 1 Carrier 2 Carrier 3 Carrier 4 Carrier 5 Carrier 6 Package holidays Company 1 Company 2 Company 3 Company 4 Company 5 Company 6 Theatre tickets Provider 1 Provider 2 Provider 3 Provider 4 Provider 5 Provider 6 Provider 7 Furniture Company 1 Company 2 Company 3 Company 4

Existence of complex pricing

none Yes none none Yes none

Example of complex pricings

none Hotel offer: 5 for the price of 1 none none 1000s 3 for 2 hotel rooms none

Shrouded complex pricing

No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

Other details/feat ures

None

none none none none none none

none none none none none none

No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No t sure how

none none none none none none none

none none none none none none none

none none none yes

none none none price kitchen units

OFT1226 | 124

Industry and Firm

Existence of complex pricing

Example of complex pricings

when you buy 3 or more

Shrouded complex pricing

much is the units price No shrouding found No shrouding found No shrouding found

Other details/feat ures

Company 5 Company 6 Company 7 Computers Company 1 Company 2 Company 3 Company 4 Company 5 Company 6 Audivisual & DVD Company 1

yes none none

Buy one, get one free (different products) none none

none none none none

none none none none

none

none

No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

yes

Buy one get one free on Disney DVD & Blue Ray

No shrouding found

Info not on the front page, one has to click on the selection first

Company 2 Company 3 Company 4 Company 5 Company 6

yes none none none none

Buy all 3 together for only 379.95 none none none none

No shrouding found No shrouding found No shrouding found No shrouding found No shrouding found

OFT1226 | 125

THEORY: OPTIMAL SEARCH STRATEGY

OFT1226 | 126

OFT1226 | 127

OFT1226 | 128

OFT1226 | 129

OFT1226 | 130

OFT1226 | 131

REGRESSION DETAILS
Variables used in regressions: Dependent variables:
label var err_v 'Indicator variable for error in the number of visits made'. These are search errors. label var err_n 'Indicator variable for error in the number units bought'. These are purchasing errors. label var err 'Indicator variable for any type of error '. These are overall errors.

Independent variables:
subject: Subject number price_strategy: Pricing strategy (1=baseline, 2=complex, 3=drip, 4=bait, 5=reference, 6=time limited) p1: Price at first shop visited p2: Price at second shop visited c: Search cost highvaluegood: Dummy for whether the good was high value aptitude: Aptitude (0 = worst, 12 = best) d2: Dummy for complex pricing d3: Dummy for drip pricing d4: Dummy for baiting d5: Dummy for reference pricing (sales) d6: Dummy for time limited offers pd2: Interaction between period and d2 pd3: Interaction between period and d3 pd4: Interaction between period and d4 pd5: Interaction between period and d5 pd6: Interaction between period and d6 cd2: Interaction between c and d2 cd3: Interaction between c and d3

OFT1226 | 132

cd4: Interaction between c and d4 cd5: Interaction between c and d5 cd6: Interaction between c and d6

Regressions included:
For each specification, there are three probit regressions using the three different dependent variables. The specifications are as follows: 1. 'base' specification using independent variables: p1 p2 highvaluegood c d2 d3 d4 d5 d6 2. 'base' plus period using independent variables: p1 p2 highvaluegood c d2 d3 d4 d5 d6 period 3. 'base' plus aptitude using independent variables: p1 p2 highvaluegood c d2 d3 d4 d5 d6 aptitude 4. 'base' plus aptitude and period using independent variables: p1 p2 highvaluegood c d2 d3 d4 d5 d6 aptitude period 5. 'base' plus aptitude, period and interaction between period and treatment dummy using independent variables: p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude pd2 pd3 pd4 pd5 pd6 6. 'base' plus aptitude, period and interaction between cost and treatment dummy using independent variables: p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude cd2 cd3 cd4 cd5 pd6

Following this, and included as a robustness check, we perform exactly the same sequence of 18 regressions using linear probability model (that is, using the reg command rather than probit).

OFT1226 | 133

Probit regressions
1. 'Base' regressions
a. Reported in Table 5.2: Probit estimation of overall errors in decision

making
dprobit err p1 p2 highvaluegood c d2 d3 d4 d5 d6, cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(9) = 132.78 Prob > chi2 = 0.0000 Log pseudolikelihood = -3227.2947 Pseudo R2 = 0.0270

(Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .39767 .0688223 5.84 0.000 .746815 .262781 .532559 p2 | .2200164 .0498589 4.41 0.000 .745319 .122295 .317738 highva~d*| .0111229 .0139256 0.80 0.424 .503689 -.016171 .038417 c | -.5291529 .2172676 -2.44 0.014 .055726 -.95499 -.103316 d2*| .0188239 .0252819 0.75 0.455 .145563 -.030728 .068376 d3*| .1544238 .0253296 6.17 0.000 .12662 .104779 .204069 d4*| .0788943 .0262817 3.03 0.002 .131605 .027383 .130405 d5*| .066988 .0254915 2.67 0.007 .131605 .017026 .116951 d6*| .1731629 .0257119 6.76 0.000 .131406 .122768 .223557 ---------+-------------------------------------------------------------------obs. P | .374676 pred. P | .3713474 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 134

b. Reported in Table 5.3: Probit estimation of errors in search behaviour

dprobit err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6, cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(9) = 95.85 Prob > chi2 = 0.0000 Log pseudolikelihood = -2909.8074 Pseudo R2 = 0.0220 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .2234849 .0672924 3.35 0.001 .746815 .091594 .355376 p2 | .2162868 .0414883 5.19 0.000 .745319 .134971 .297602 highva~d*| .0045715 .0128177 0.36 0.721 .503689 -.020551 .029694 c | -.5120419 .1901925 -2.70 0.007 .055726 -.884812 -.139271 d2*| .0244149 .024193 1.02 0.306 .145563 -.023003 .071832 d3*| .0979155 .0259545 3.92 0.000 .12662 .047046 .148785 d4*| .0648617 .02431 2.74 0.006 .131605 .017215 .112508 d5*| .0463191 .0223851 2.13 0.033 .131605 .002445 .090193 d6*| .178912 .0259026 7.24 0.000 .131406 .128144 .22968 ---------+-------------------------------------------------------------------obs. P | .2803589 pred. P | .2762831 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 135

c. Reported in Table 5.4: Probit estimation of errors in purchasing behaviour

dprobit err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6, cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(9) = 264.01 Prob > chi2 = 0.0000 Log pseudolikelihood = -2858.4259 Pseudo R2 = 0.0525 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .622151 .0597762 11.16 0.000 .746815 .504992 .73931 p2 | -.1465648 .0504143 -2.90 0.004 .745319 -.245375 -.047755 highva~d*| .0261268 .0135208 1.94 0.053 .503689 -.000373 .052627 c | .4850114 .205159 2.34 0.019 .055726 .082907 .887116 d2*| .0153017 .0241285 0.64 0.523 .145563 -.031989 .062593 d3*| .1551735 .0274357 5.99 0.000 .12662 .1014 .208947 d4*| .080523 .0264786 3.14 0.002 .131605 .028626 .13242 d5*| .0571337 .0255033 2.32 0.020 .131605 .007148 .107119 d6*| .1975555 .0259052 7.90 0.000 .131406 .146782 .248329 ---------+-------------------------------------------------------------------obs. P | .289332 pred. P | .2779313 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 136

2. Base regressions including period (time trend) to account for learning a. Reported in paragraph 5.52 (overall errors)
.
dprobit err p1 p2 highvaluegood c d2 d3 d4 d5 d6 period , cluster(subject )

Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(10) = 177.45 Prob > chi2 = 0.0000 Log pseudolikelihood = -3201.4808 Pseudo R2 = 0.0348 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .4028919 .0692408 5.88 0.000 .746815 .267182 .538601 p2 | .213841 .0497624 4.29 0.000 .745319 .116308 .311373 highva~d*| .0093243 .0140595 0.66 0.507 .503689 -.018232 .03688 c | -.5504067 .2193025 -2.52 0.012 .055726 -.980232 -.120582 d2*| .0201932 .0255385 0.79 0.427 .145563 -.029861 .070248 d3*| .1549183 .0256867 6.11 0.000 .12662 .104573 .205263 d4*| .0809213 .0265381 3.08 0.002 .131605 .028908 .132935 d5*| .0681888 .0257969 2.69 0.007 .131605 .017628 .11875 d6*| .172756 .0258861 6.71 0.000 .131406 .12202 .223492 period | -.0057894 .0008282 -6.98 0.000 15.2881 -.007413 -.004166 ---------+-------------------------------------------------------------------obs. P | .374676 pred. P | .3701252 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

b. Reported in paragraph 5.53 (search errors)

dprobit err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 period , cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(10) = 105.55 Prob > chi2 = 0.0000 Log pseudolikelihood = -2905.4031 Pseudo R2 = 0.0235 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .2258963 .0670752 3.40 0.001 .746815 .094431 .357361 p2 | .2135218 .0414073 5.14 0.000 .745319 .132365 .294679 highva~d*| .0039001 .0128066 0.30 0.761 .503689 -.0212 .029001 c | -.5200878 .1900081 -2.75 0.006 .055726 -.892497 -.147679 d2*| .0250345 .0242547 1.05 0.295 .145563 -.022504 .072573 d3*| .0976357 .0260292 3.90 0.000 .12662 .046619 .148652 d4*| .0655612 .0243515 2.77 0.006 .131605 .017833 .113289 d5*| .0465966 .022437 2.14 0.033 .131605 .002621 .090572 d6*| .178545 .0259155 7.23 0.000 .131406 .127751 .229339 period | -.0022006 .0007915 -2.76 0.006 15.2881 -.003752 -.000649 ---------+-------------------------------------------------------------------obs. P | .2803589 pred. P | .275887 (at x-bar)

OFT1226 | 137

-----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

c. Reported in paragraph 5.53 (purchasing errors)

dprobit err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 period , cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(10) = 309.22 Prob > chi2 = 0.0000 Log pseudolikelihood = -2832.7844 Pseudo R2 = 0.0610 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .6286775 .0604273 11.19 0.000 .746815 .510242 .747113 p2 | -.1552339 .0502741 -3.09 0.002 .745319 -.253769 -.056698 highva~d*| .0245179 .0135889 1.81 0.070 .503689 -.002116 .051152 c | .471413 .2066389 2.26 0.024 .055726 .066408 .876418 d2*| .0173205 .0244133 0.72 0.474 .145563 -.030529 .06517 d3*| .1553828 .0278635 5.91 0.000 .12662 .100771 .209994 d4*| .0825454 .0268375 3.18 0.001 .131605 .029945 .135146 d5*| .0582946 .0257837 2.34 0.019 .131605 .007759 .10883 d6*| .19809 .0260925 7.87 0.000 .131406 .14695 .24923 period | -.0053769 .0008142 -6.67 0.000 15.2881 -.006973 -.003781 ---------+-------------------------------------------------------------------obs. P | .289332 pred. P | .2757784 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 138

3. Base regressions including aptitude (IQ test) no time trend a. Aptitude score and overall errors (no time trend): This is not reported in the main document but provided for additional information
. dprobit err p1 p2 highvaluegood c d2 d3 d4 d5 d6 aptitude , cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(10) = 136.47 Prob > chi2 = 0.0000 Log pseudolikelihood = -3204.5041 Pseudo R2 = 0.0339 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .3925986 .0693896 5.73 0.000 .746815 .256597 .5286 p2 | .2152091 .049812 4.32 0.000 .745319 .117579 .312839 highva~d*| .0103742 .0138171 0.75 0.453 .503689 -.016707 .037455 c | -.5364496 .2173673 -2.48 0.013 .055726 -.962482 -.110417 d2*| .0232376 .0255684 0.91 0.361 .145563 -.026875 .073351 d3*| .1478671 .0244522 6.11 0.000 .12662 .099942 .195792 d4*| .0838768 .0259858 3.26 0.001 .131605 .032946 .134808 d5*| .0629328 .0250123 2.56 0.010 .131605 .01391 .111956 d6*| .1773022 .0253245 7.04 0.000 .131406 .127667 .226937 aptitude | -.020245 .0060059 -3.40 0.001 9.61416 -.032016 -.008474 ---------+-------------------------------------------------------------------obs. P | .374676 pred. P | .3705894 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

b. Aptitude score and search errors (no time trend): This is not reported in the main document but provided for additional information
dprobit err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 aptitude , cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(10) = 100.17 Prob > chi2 = 0.0000 Log pseudolikelihood = -2904.7271 Pseudo R2 = 0.0237 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .220039 .0672009 3.30 0.001 .746815 .088328 .35175 p2 | .2141076 .0416203 5.13 0.000 .745319 .132533 .295682 highva~d*| .0042712 .01278 0.33 0.738 .503689 -.020777 .029319 c | -.5128878 .1901901 -2.70 0.007 .055726 -.885654 -.140122 d2*| .0260137 .0241937 1.09 0.275 .145563 -.021405 .073432 d3*| .0940223 .0253924 3.84 0.000 .12662 .044254 .14379 d4*| .0669203 .024382 2.83 0.005 .131605 .019132 .114708 d5*| .0443031 .0222778 2.04 0.041 .131605 .000639 .087967 d6*| .1806035 .0258304 7.34 0.000 .131406 .129977 .23123 aptitude | -.0087638 .0043133 -2.04 0.041 9.61416 -.017218 -.00031 ---------+-------------------------------------------------------------------obs. P | .2803589

OFT1226 | 139

pred. P | .2759511 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

c. Aptitude score and purchasing errors (no time trend): This is not reported in the main document but provided for additional information
. dprobit err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 aptitude , cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(10) = 280.93 Prob > chi2 = 0.0000 Log pseudolikelihood = -2831.8944 Pseudo R2 = 0.0613 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .6192829 .0606703 11.05 0.000 .746815 .500371 .738195 p2 | -.1545655 .0504147 -3.07 0.002 .745319 -.253377 -.055754 highva~d*| .0256831 .0135018 1.91 0.056 .503689 -.00078 .052146 c | .4813031 .2043267 2.33 0.020 .055726 .08083 .881776 d2*| .0196412 .0243649 0.81 0.416 .145563 -.028113 .067395 d3*| .1475948 .0268628 5.79 0.000 .12662 .094945 .200245 d4*| .0861477 .0258976 3.45 0.001 .131605 .035389 .136906 d5*| .0536591 .0249353 2.23 0.026 .131605 .004787 .102531 d6*| .202571 .0254268 8.29 0.000 .131406 .152735 .252407 aptitude | -.0199857 .0059569 -3.40 0.001 9.61416 -.031661 -.00831 ---------+-------------------------------------------------------------------obs. P | .289332 pred. P | .2762123 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 140

4. Base regressions including aptitude (IQ test) with time trend a. Reported in paragraph 5.54 Aptitude score and overall errors (with time trend)
dprobit err p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude , cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(11) = 179.25 Prob > chi2 = 0.0000 Log pseudolikelihood = -3178.0582 Pseudo R2 = 0.0419 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .3979432 .0699006 5.76 0.000 .746815 .260941 .534946 p2 | .2088008 .0496715 4.19 0.000 .745319 .111446 .306155 highva~d*| .0086067 .0139609 0.62 0.537 .503689 -.018756 .03597 c | -.5592682 .2193316 -2.56 0.010 .055726 -.98915 -.129386 d2*| .0247506 .0258238 0.96 0.335 .145563 -.025863 .075364 d3*| .1483134 .0248002 6.05 0.000 .12662 .099706 .196921 d4*| .0861345 .0262493 3.32 0.001 .131605 .034687 .137582 d5*| .064239 .0253229 2.58 0.010 .131605 .014607 .113871 d6*| .1769457 .0254744 6.98 0.000 .131406 .127017 .226875 period | -.0058771 .0008369 -7.01 0.000 15.2881 -.007517 -.004237 aptitude | -.0205834 .0060692 -3.42 0.001 9.61416 -.032479 -.008688 ---------+-------------------------------------------------------------------obs. P | .374676 pred. P | .3693388 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 141

b. Reported in paragraph 5.54 Aptitude score and search errors (with time trend)
dprobit err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude , cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(11) = 111.85 Prob > chi2 = 0.0000 Log pseudolikelihood = -2900.2538 Pseudo R2 = 0.0252 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .2224458 .0669998 3.35 0.001 .746815 .091129 .353763 p2 | .2113218 .0415244 5.07 0.000 .745319 .129935 .292708 highva~d*| .0036092 .0127683 0.28 0.777 .503689 -.021416 .028635 c | -.5211159 .1899626 -2.75 0.006 .055726 -.893436 -.148796 d2*| .0266448 .024253 1.11 0.265 .145563 -.02089 .07418 d3*| .0937191 .0254602 3.82 0.000 .12662 .043818 .14362 d4*| .0676542 .0244233 2.85 0.004 .131605 .019785 .115523 d5*| .044614 .0223347 2.05 0.040 .131605 .000839 .088389 d6*| .1802525 .0258386 7.32 0.000 .131406 .12961 .230895 period | -.0022186 .0007919 -2.78 0.005 15.2881 -.003771 -.000667 aptitude | -.0088248 .0043276 -2.05 0.040 9.61416 -.017307 -.000343 ---------+-------------------------------------------------------------------obs. P | .2803589 pred. P | .2755517 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 142

c. Reported in paragraph 5.54 Aptitude score and purchasing errors (with time trend)
dprobit err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude , cluster(subject ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(11) = 318.94 Prob > chi2 = 0.0000 Log pseudolikelihood = -2805.6533 Pseudo R2 = 0.0700 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .6257942 .0614191 11.06 0.000 .746815 .505415 .746173 p2 | -.1636083 .0502934 -3.27 0.001 .745319 -.262182 -.065035 highva~d*| .0240438 .0135759 1.78 0.076 .503689 -.002565 .050652 c | .4663176 .2057147 2.24 0.025 .055726 .063124 .869511 d2*| .0218483 .0246595 0.90 0.371 .145563 -.026484 .07018 d3*| .1477884 .027274 5.72 0.000 .12662 .094332 .201244 d4*| .0884172 .0262874 3.49 0.000 .131605 .036895 .13994 d5*| .0549567 .0252275 2.26 0.024 .131605 .005512 .104402 d6*| .2031956 .0255897 8.28 0.000 .131406 .153041 .25335 period | -.0054489 .0008251 -6.65 0.000 15.2881 -.007066 -.003832 aptitude | -.0202333 .00601 -3.42 0.001 9.61416 -.032013 -.008454 ---------+-------------------------------------------------------------------obs. P | .289332 pred. P | .274009 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 143

5. Base regressions including aptitude (IQ test) and interaction between price frames and time trend

a. Reported in paragraph 5.55 Interaction terms between price frames and time trend (overall errors)
. dprobit err p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude pd2 pd3 pd4 pd5 pd6 , cluster(subject > ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(16) = 188.84 Prob > chi2 = 0.0000 Log pseudolikelihood = -3171.6004 Pseudo R2 = 0.0438 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .3988921 .0700601 5.77 0.000 .746815 .261577 .536207 p2 | .208588 .0495926 4.20 0.000 .745319 .111388 .305788 highva~d*| .0076987 .0140764 0.55 0.584 .503689 -.019891 .035288 c | -.5503808 .2183509 -2.53 0.011 .055726 -.978341 -.122421 d2*| .0374593 .0472915 0.80 0.424 .145563 -.05523 .130149 d3*| .195949 .0456059 4.33 0.000 .12662 .106563 .285335 d4*| .1134488 .0473532 2.43 0.015 .131605 .020638 .206259 d5*| .0946806 .0449673 2.14 0.032 .131605 .006546 .182815 d6*| .0693022 .0447374 1.57 0.116 .131406 -.018381 .156986 period | -.0057981 .0014002 -4.16 0.000 15.2881 -.008542 -.003054 aptitude | -.0206289 .0060978 -3.41 0.001 9.61416 -.03258 -.008677 pd2 | -.0008274 .0027187 -0.30 0.761 2.25464 -.006156 .004501 pd3 | -.0030261 .0025407 -1.19 0.233 1.91366 -.008006 .001954 pd4 | -.0017216 .0025692 -0.67 0.503 2.04686 -.006757 .003314 pd5 | -.0019452 .002564 -0.76 0.449 2.03968 -.006971 .00308 pd6 | .0069187 .0024703 2.81 0.005 1.96191 .002077 .01176 ---------+-------------------------------------------------------------------obs. P | .374676 pred. P | .3691438 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 144

b. Reported in paragraph 5.56 Interaction terms between price frames and time trend (search errors)
. dprobit err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude pd2 pd3 pd4 pd5 pd6 , cluster(subject >t) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(16) = 120.26 Prob > chi2 = 0.0000 Log pseudolikelihood = -2896.6522 Pseudo R2 = 0.0265 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .2223229 .0670359 3.35 0.001 .746815 .090935 .353711 p2 | .2112912 .04139 5.09 0.000 .745319 .130168 .292414 highva~d*| .0032695 .0128579 0.25 0.799 .503689 -.021931 .028471 c | -.5155089 .1899918 -2.72 0.007 .055726 -.887886 -.143132 d2*| .0031518 .0437461 0.07 0.942 .145563 -.082589 .088893 d3*| .1213581 .0474757 2.69 0.007 .12662 .028307 .214409 d4*| .0758175 .0470318 1.67 0.095 .131605 -.016363 .167998 d5*| .0608236 .047497 1.32 0.186 .131605 -.032269 .153916 d6*| .0976527 .0431666 2.36 0.018 .131406 .013048 .182258 period | -.0027117 .0014193 -1.90 0.057 15.2881 -.005493 .00007 aptitude | -.0088159 .0043357 -2.04 0.041 9.61416 -.017314 -.000318 pd2 | .0015213 .002584 0.59 0.556 2.25464 -.003543 .006586 pd3 | -.0016748 .002524 -0.66 0.507 1.91366 -.006622 .003272 pd4 | -.000486 .0024458 -0.20 0.842 2.04686 -.00528 .004308 pd5 | -.0010077 .0026198 -0.38 0.701 2.03968 -.006142 .004127 pd6 | .0048932 .0022014 2.22 0.027 1.96191 .000579 .009208 ---------+-------------------------------------------------------------------obs. P | .2803589 pred. P | .2753392 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 145

c. Reported in paragraph 5.56 Interaction terms between price frames and time trend (purchasing errors)
dprobit err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude pd2 pd3 pd4 pd5 pd6 , cluster(subjec >t) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(16) = 345.89 Prob > chi2 = 0.0000 Log pseudolikelihood = -2799.2758 Pseudo R2 = 0.0721 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .6283741 .0615721 11.10 0.000 .746815 .507695 .749053 p2 | -.1646668 .0500515 -3.30 0.001 .745319 -.262766 -.066568 highva~d*| .0234938 .0136437 1.73 0.084 .503689 -.003247 .050235 c | .4765438 .2048793 2.30 0.022 .055726 .074988 .8781 d2*| .0112102 .0452586 0.25 0.803 .145563 -.077495 .099915 d3*| .1843825 .0484549 4.03 0.000 .12662 .089413 .279352 d4*| .1108439 .0445429 2.60 0.009 .131605 .023541 .198146 d5*| .0523911 .0460994 1.17 0.243 .131605 -.037962 .142744 d6*| .0879082 .0463086 1.97 0.049 .131406 -.002855 .178671 period | -.0060872 .0013844 -4.45 0.000 15.2881 -.008801 -.003374 aptitude | -.0202804 .0060356 -3.41 0.001 9.61416 -.03211 -.008451 pd2 | .0007137 .0027253 0.26 0.793 2.25464 -.004628 .006055 pd3 | -.0021542 .0023679 -0.91 0.363 1.91366 -.006795 .002487 pd4 | -.0013412 .0025483 -0.53 0.599 2.04686 -.006336 .003653 pd5 | .0001732 .0024392 0.07 0.943 2.03968 -.004607 .004954 pd6 | .0068452 .0023735 2.90 0.004 1.96191 .002193 .011497 ---------+-------------------------------------------------------------------obs. P | .289332 pred. P | .2734712 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 146

6. Base regressions including interaction between search costs and treatment a. Reported in paragraph 5.22 Interaction terms between search costs and treatment (overall errors)
. dprobit err p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude cd2 cd3 cd4 cd5 cd6 , cluster(subject >)

Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(16) = 190.86 Prob > chi2 = 0.0000 Log pseudolikelihood = -3174.0818 Pseudo R2 = 0.0431 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .3962835 .0699802 5.74 0.000 .746815 .259125 .533442 p2 | .2097658 .0497116 4.21 0.000 .745319 .112333 .307199 highva~d*| .0079886 .0139478 0.57 0.567 .503689 -.019348 .035326 c | .1238768 .3822887 0.32 0.746 .055726 -.625395 .873149 d2*| .0528838 .0453782 1.18 0.239 .145563 -.036056 .141823 d3*| .2126438 .0428422 4.96 0.000 .12662 .128675 .296613 d4*| .1618088 .0435639 3.75 0.000 .131605 .076425 .247193 d5*| .0998216 .0451835 2.25 0.024 .131605 .011264 .18838 d6*| .2643349 .0506616 5.12 0.000 .131406 .16504 .36363 period | -.0058292 .0008321 -6.99 0.000 15.2881 -.00746 -.004198 aptitude | -.0209384 .0060683 -3.48 0.001 9.61416 -.032832 -.009045 cd2 | -.4891798 .6655608 -0.74 0.462 .008049 -1.79366 .815295 cd3 | -1.114248 .5958725 -1.87 0.062 .007022 -2.28214 .053641 cd4 | -1.322431 .6199777 -2.13 0.033 .007185 -2.53756 -.107297 cd5 | -.6181273 .6063966 -1.02 0.308 .007444 -1.80664 .570388 cd6 | -1.509331 .7061585 -2.14 0.032 .007424 -2.89338 -.125286 ---------+-------------------------------------------------------------------obs. P | .374676 pred. P | .3693055 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 147

b. Reported in paragraph 5.23 Interaction terms between search costs and treatment (search errors)

dprobit err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude cd2 cd3 cd4 cd5 cd6 , cluster(subject > t ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(16) = 129.53 Prob > chi2 = 0.0000 Log pseudolikelihood = -2888.82 Pseudo R2 = 0.0291 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .2206325 .0670083 3.33 0.001 .746815 .089299 .351966 p2 | .2126269 .0413668 5.13 0.000 .745319 .13155 .293704 highva~d*| .0027607 .012686 0.22 0.828 .503689 -.022103 .027625 c | .3459962 .3405603 1.01 0.310 .055726 -.32149 1.01348 d2*| .0432098 .0416566 1.06 0.288 .145563 -.038436 .124855 d3*| .1835859 .0450089 4.28 0.000 .12662 .09537 .271802 d4*| .198842 .0428474 4.88 0.000 .131605 .114863 .282821 d5*| .0668572 .0406978 1.70 0.088 .131605 -.012909 .146624 d6*| .3145075 .050635 6.39 0.000 .131406 .215265 .41375 period | -.0021322 .0007885 -2.69 0.007 15.2881 -.003678 -.000587 aptitude | -.0093272 .0043372 -2.16 0.031 9.61416 -.017828 -.000827 cd2 | -.2692348 .6201286 -0.43 0.664 .008049 -1.48466 .946195 cd3 | -1.442658 .5538867 -2.59 0.010 .007022 -2.52826 -.35706 cd4 | -2.168836 .6172764 -3.51 0.000 .007185 -3.37868 -.958996 cd5 | -.376066 .5389976 -0.70 0.485 .007444 -1.43248 .68035 cd6 | -2.079233 .6060126 -3.44 0.001 .007424 -3.267 -.89147 ---------+-------------------------------------------------------------------obs. P | .2803589 pred. P | .2750024 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 148

c. Reported in paragraph 5.24 Interaction terms between search costs and treatment (purchasing errors)
dprobit err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude cd2 cd3 cd4 cd5 cd6 , cluster(subject > t ) Probit regression, reporting marginal effects Number of obs = 5015 Wald chi2(16) = 332.42 Prob > chi2 = 0.0000 Log pseudolikelihood = -2802.5502 Pseudo R2 = 0.0710 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | dF/dx Std. Err. z P>|z| x-bar [ 95 per cent C.I. ] ---------+-------------------------------------------------------------------p1 | .6250536 .0614056 11.06 0.000 .746815 .504701 .745406 p2 | -.1625912 .050294 -3.24 0.001 .745319 -.261166 -.064017 highva~d*| .0236248 .0135655 1.75 0.081 .503689 -.002963 .050213 c | 1.048708 .3492857 2.97 0.003 .055726 .364121 1.7333 d2*| .0599929 .0433454 1.42 0.155 .145563 -.024962 .144948 d3*| .2158993 .0470321 4.85 0.000 .12662 .123718 .308081 d4*| .1118733 .044836 2.61 0.009 .131605 .023996 .19975 d5*| .1059318 .0472655 2.35 0.019 .131605 .013293 .198571 d6*| .2872089 .0499563 5.96 0.000 .131406 .189296 .385121 period | -.0054247 .0008206 -6.66 0.000 15.2881 -.007033 -.003816 aptitude | -.0204506 .0060048 -3.46 0.001 9.61416 -.03222 -.008681 cd2 | -.6355408 .5631849 -1.13 0.259 .008049 -1.73936 .468281 cd3 | -1.05095 .5401403 -1.95 0.052 .007022 -2.10961 .007706 cd4 | -.3616208 .5810001 -0.62 0.534 .007185 -1.50036 .777118 cd5 | -.8203628 .5571246 -1.47 0.141 .007444 -1.91231 .271581 cd6 | -1.271652 .6221084 -2.04 0.041 .007424 -2.49096 -.052342 ---------+-------------------------------------------------------------------obs. P | .289332 pred. P | .2736706 (at x-bar) -----------------------------------------------------------------------------(*) dF/dx is for discrete change of dummy variable from 0 to 1 z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 149

Non-parametric test consumer welfare (paragraph 5.58)

by price_strategy : signrank diff_u =diff_u2 -----------------------------------------------------------------------------> price_strategy = 1 Wilcoxon signed-rank test sign | obs sum ranks expected -------------+--------------------------------positive | 56 4182 7245 negative | 105 10308 7245 zero | 9 45 45 -------------+--------------------------------all | 170 14535 14535 unadjusted variance 413036.25 adjustment for ties -2.13 adjustment for zeros -71.25 ---------adjusted variance 412962.88 Ho: diff_u = diff_u2 z = -4.766 Prob > |z| = 0.0000 -----------------------------------------------------------------------------> price_strategy = 2 Wilcoxon signed-rank test sign | obs sum ranks expected -------------+--------------------------------positive | 26 799.5 1350 negative | 46 1900.5 1350 zero | 1 1 1 -------------+--------------------------------all | 73 2701 2701 unadjusted variance 33087.25 adjustment for ties -0.25 adjustment for zeros -0.25 ---------adjusted variance 33086.75

OFT1226 | 150

Linear regressions performed to check robustness

1. 'Base' regressions a. Linear estimation of overall errors in decision making
reg err p1 p2 highvaluegood c d2 d3 d4 d5 d6, cluster(subject ) Linear regression Number of obs = 5015 F( 9, 169) = 15.83 Prob > F = 0.0000 R-squared = 0.0354 Root MSE = .47586 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .3843403 .0651998 5.89 0.000 .2556293 .5130513 p2 | .2177066 .0486591 4.47 0.000 .1216487 .3137646 highvalueg~d | .0110253 .0135952 0.81 0.419 -.015813 .0378635 c | -.5025095 .2105306 -2.39 0.018 -.918118 -.086901 d2 | .0194859 .0237183 0.82 0.412 -.0273363 .0663082 d3 | .149702 .0245467 6.10 0.000 .1012443 .1981597 d4 | .0767684 .024637 3.12 0.002 .0281326 .1254042 d5 | .0638656 .0242424 2.63 0.009 .0160086 .1117226 d6 | .1684747 .0246646 6.83 0.000 .1197843 .217165 _cons | -.1146047 .0653518 -1.75 0.081 -.2436158 .0144064 ------------------------------------------------------------------------------

b. Linear estimation of errors in search behaviour

reg err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6, cluster(subject ) Linear regression Number of obs = 5015 F( 9, 169) = 10.98 Prob > F = 0.0000 R-squared = 0.0268 Root MSE = .44356 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .2190497 .0634417 3.45 0.001 .0938095 .3442899 p2 | .2189598 .0410745 5.33 0.000 .1378747 .3000449 highvalueg~d | .0049569 .012704 0.39 0.697 -.0201221 .0300359 c | -.4940624 .1845707 -2.68 0.008 -.8584236 -.1297013 d2 | .0237349 .0220786 1.08 0.284 -.0198503 .0673202 d3 | .0923835 .0239344 3.86 0.000 .0451346 .1396325 d4 | .0631902 .0221352 2.85 0.005 .0194931 .1068873 d5 | .0431921 .0204709 2.11 0.036 .0027806 .0836036 d6 | .1738838 .0243593 7.14 0.000 .1257961 .2219715 _cons | -.0733924 .0573267 -1.28 0.202 -.1865611 .0397763

OFT1226 | 151

c. Linear estimation of errors in purchasing behaviour

reg err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6, cluster(subject ) Linear regression Number of obs = 5015 F( 9, 169) = 29.92 Prob > F = 0.0000 R-squared = 0.0611 Root MSE = .43983 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .5882927 .055456 10.61 0.000 .4788169 .6977685 p2 | -.128347 .0488204 -2.63 0.009 -.2247233 -.0319706 highvalueg~d | .0252613 .0131397 1.92 0.056 -.0006779 .0512004 c | .4790518 .1984353 2.41 0.017 .0873206 .870783 d2 | .0149644 .0216724 0.69 0.491 -.027819 .0577479 d3 | .1437444 .0257071 5.59 0.000 .0929959 .1944928 d4 | .0738574 .0232609 3.18 0.002 .027938 .1197768 d5 | .0509222 .0228032 2.23 0.027 .0059064 .095938 d6 | .1847231 .0235767 7.83 0.000 .1381803 .2312659 _cons | -.1548482 .0584492 -2.65 0.009 -.2702329 -.0394635

2. Base regressions including period (time trend) to account for learning a. Linear estimation (overall errors)
reg err p1 p2 highvaluegood c d2 d3 d4 d5 d6 period , cluster(subject ) Linear regression Number of obs = 5015 F( 10, 169) = 19.71 Prob > F = 0.0000 R-squared = 0.0452 Root MSE = .47349 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .3849301 .0652101 5.90 0.000 .2561987 .5136615 p2 | .2109145 .0483101 4.37 0.000 .1155455 .3062835 highvalueg~d | .0091038 .0136333 0.67 0.505 -.0178097 .0360173 c | -.5144887 .2108786 -2.44 0.016 -.9307843 -.0981932 d2 | .02087 .0237399 0.88 0.381 -.025995 .067735 d3 | .148945 .0246001 6.05 0.000 .100382 .197508 d4 | .0785549 .0246309 3.19 0.002 .029931 .1271788 d5 | .0652308 .0242331 2.69 0.008 .0173922 .1130694 d6 | .1668361 .0247087 6.75 0.000 .1180587 .2156135 period | -.0055898 .0008006 -6.98 0.000 -.0071704 -.0040093 _cons | -.0231943 .066322 -0.35 0.727 -.1541206 .1077319 -------------------------------------------------------------------------------

OFT1226 | 152

b. Linear estimation (search errors)

reg err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 period , cluster(subject ) Linear regression Number of obs = 5015 F( 10, 169) = 11.04 Prob > F = 0.0000 R-squared = 0.0284 Root MSE = .44324 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .2192702 .0633432 3.46 0.001 .0942244 .3443161 p2 | .2164203 .0410221 5.28 0.000 .1354386 .2974021 highvalueg~d | .0042385 .0126904 0.33 0.739 -.0208136 .0292906 c | -.4985413 .1844228 -2.70 0.008 -.8626104 -.1344723 d2 | .0242524 .0220776 1.10 0.274 -.019331 .0678358 d3 | .0921005 .0239483 3.85 0.000 .0448242 .1393769 d4 | .0638581 .0221411 2.88 0.004 .0201494 .1075669 d5 | .0437025 .0204785 2.13 0.034 .0032758 .0841292 d6 | .1732711 .0243627 7.11 0.000 .1251767 .2213656 period | -.00209 .000786 -2.66 0.009 -.0036416 -.0005384 _cons | -.0392153 .0602411 -0.65 0.516 -.1581372 .0797066

c. Linear estimation (purchasing errors)

reg err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 period , cluster(subject ) Linear regression Number of obs = 5015 F( 10, 169) = 31.88 Prob > F = 0.0000 R-squared = 0.0702 Root MSE = .43772 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .588828 .0557822 10.56 0.000 .4787083 .6989477 p2 | -.1345112 .0484762 -2.77 0.006 -.2302081 -.0388143 highvalueg~d | .0235174 .013098 1.80 0.074 -.0023393 .0493741 c | .4681801 .1986009 2.36 0.020 .0761219 .8602382 d2 | .0162205 .021713 0.75 0.456 -.0266431 .0590842 d3 | .1430574 .0257707 5.55 0.000 .0921834 .1939314 d4 | .0754787 .0232998 3.24 0.001 .0294826 .1214748 d5 | .0521612 .0228149 2.29 0.023 .0071224 .0972 d6 | .183236 .0236425 7.75 0.000 .1365633 .2299087 period | -.0050731 .0007818 -6.49 0.000 -.0066163 -.0035298 _cons | -.0718888 .0590577 -1.22 0.225 -.1884746 .0446971

OFT1226 | 153

3. Base regressions including aptitude (IQ test) no time trend a. Linear Estimation: Aptitude score and overall errors (no time trend). This is not reported in the main document but provided for additional information
reg err p1 p2 highvaluegood c d2 d3 d4 d5 d6 aptitude , cluster(subject ) Linear regression Number of obs = 5015 F( 10, 169) = 14.94 Prob > F = 0.0000 R-squared = 0.0444 Root MSE = .4737 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .3767392 .0651553 5.78 0.000 .2481162 .5053623 p2 | .2113991 .0482665 4.38 0.000 .1161162 .3066819 highvalueg~d | .0103371 .0134045 0.77 0.442 -.0161246 .0367989 c | -.5034809 .2090071 -2.41 0.017 -.9160819 -.0908799 d2 | .0230307 .0237248 0.97 0.333 -.0238044 .0698659 d3 | .1422743 .0233686 6.09 0.000 .0961423 .1884063 d4 | .0804488 .0242342 3.32 0.001 .0326082 .1282895 d5 | .05954 .0236212 2.52 0.013 .0129095 .1061706 d6 | .1712666 .0241892 7.08 0.000 .1235146 .2190186 aptitude | -.0201029 .0059812 -3.36 0.001 -.0319104 -.0082954 _cons | .0895888 .0870977 1.03 0.305 -.0823507 .2615283

OFT1226 | 154

b. Linear Estimation: Aptitude score and search errors (no time trend). This is not reported in the main document but provided for additional information
reg err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 aptitude , cluster(subject ) Linear regression Number of obs = 5015 F( 10, 169) = 10.28 Prob > F = 0.0000 R-squared = 0.0288 Root MSE = .44314 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .2156845 .0632746 3.41 0.001 .090774 .3405949 p2 | .2161673 .0412416 5.24 0.000 .1347523 .2975823 highvalueg~d | .0046523 .0126463 0.37 0.713 -.0203129 .0296174 c | -.4944925 .1844263 -2.68 0.008 -.8585685 -.1304165 d2 | .0253043 .0219837 1.15 0.251 -.0180936 .0687023 d3 | .0890951 .0234131 3.81 0.000 .0428752 .1353149 d4 | .0648196 .0221579 2.93 0.004 .0210776 .1085616 d5 | .041277 .0203521 2.03 0.044 .0010998 .0814542 d6 | .1751199 .0242684 7.22 0.000 .1272116 .2230281 aptitude | -.0089002 .0044762 -1.99 0.048 -.0177367 -.0000637 _cons | .0170106 .0735236 0.23 0.817 -.1281324 .1621535 ------------------------------------------------------------------------------

c. Linear Estimation: Aptitude score and purchasing errors (no time trend). This is not reported in the main document but provided for additional information
reg err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 aptitude , cluster(subject ) Linear regression Number of obs = 5015 F( 10, 169) = 28.69 Prob > F = 0.0000 R-squared = 0.0713 Root MSE = .43746 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .5806488 .0555212 10.46 0.000 .4710444 .6902532 p2 | -.1346901 .0486209 -2.77 0.006 -.2306727 -.0387076 highvalueg~d | .0245693 .0130293 1.89 0.061 -.0011518 .0502904 c | .478075 .1969125 2.43 0.016 .08935 .8667999 d2 | .0185292 .0216601 0.86 0.394 -.0242301 .0612885 d3 | .1362748 .0247686 5.50 0.000 .0873792 .1851704 d4 | .0775586 .0226555 3.42 0.001 .0328344 .1222828 d5 | .0465722 .0222881 2.09 0.038 .0025733 .0905712 d6 | .1875307 .023013 8.15 0.000 .1421009 .2329606 aptitude | -.0202163 .0061826 -3.27 0.001 -.0324214 -.0080112 _cons | .0504971 .0847426 0.60 0.552 -.1167933 .2177876

OFT1226 | 155

4. Base regressions including aptitude (IQ test) with time trend a. Linear Estimation: Aptitude score and overall errors (with time trend)
reg err p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude , cluster(subject ) Linear regression Number of obs = 5015 F( 11, 169) = 18.46 Prob > F = 0.0000 R-squared = 0.0543 Root MSE = .47128 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .3772766 .0651867 5.79 0.000 .2485916 .5059617 p2 | .2045141 .0479022 4.27 0.000 .1099504 .2990777 highvalueg~d | .0083975 .0134437 0.62 0.533 -.0181416 .0349367 c | -.5155485 .2092547 -2.46 0.015 -.9286384 -.1024587 d2 | .0244505 .0237431 1.03 0.305 -.0224207 .0713217 d3 | .1414571 .0234118 6.04 0.000 .0952399 .1876744 d4 | .0822747 .0242268 3.40 0.001 .0344487 .1301008 d5 | .0608824 .0236056 2.58 0.011 .0142825 .1074823 d6 | .1696376 .0242211 7.00 0.000 .1218228 .2174524 period | -.0056277 .0008027 -7.01 0.000 -.0072123 -.0040431 aptitude | -.020252 .0059918 -3.38 0.001 -.0320804 -.0084237 _cons | .1831338 .0874546 2.09 0.038 .0104895 .355778 ------------------------------------------------------------------------------

b. Linear Estimation: Aptitude score and search errors (with time trend)
reg err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude , cluster(subject ) Linear regression Number of obs = 5015 F( 11, 169) = 10.58 Prob > F = 0.0000 R-squared = 0.0304 Root MSE = .44282 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .2158856 .0631799 3.42 0.001 .0911621 .3406092 p2 | .2135899 .0411839 5.19 0.000 .1322887 .2948911 highvalueg~d | .0039262 .0126325 0.31 0.756 -.0210117 .0288641 c | -.49901 .1842637 -2.71 0.007 -.862765 -.135255 d2 | .0258358 .0219822 1.18 0.242 -.0175594 .069231 d3 | .0887892 .0234236 3.79 0.000 .0425486 .1350298 d4 | .0655032 .0221631 2.96 0.004 .0217509 .1092554 d5 | .0417795 .0203588 2.05 0.042 .0015893 .0819698 d6 | .17451 .0242712 7.19 0.000 .1265962 .2224239 period | -.0021067 .0007857 -2.68 0.008 -.0036577 -.0005558 aptitude | -.008956 .0044778 -2.00 0.047 -.0177956 -.0001164 _cons | .0520288 .0746798 0.70 0.487 -.0953967 .1994543 ------------------------------------------------------------------------------

OFT1226 | 156

c. Linear estimation: Aptitude score and purchasing errors (with time trend)
reg err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude , cluster(subject ) Linear regression Number of obs = 5015 F( 11, 169) = 29.81 Prob > F = 0.0000 R-squared = 0.0807 Root MSE = .4353 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .5811369 .055869 10.40 0.000 .4708458 .6914279 p2 | -.1409431 .0482658 -2.92 0.004 -.2362246 -.0456617 highvalueg~d | .0228077 .0129886 1.76 0.081 -.0028332 .0484486 c | .467115 .1969811 2.37 0.019 .0782545 .8559755 d2 | .0198186 .021695 0.91 0.362 -.0230095 .0626468 d3 | .1355326 .0248259 5.46 0.000 .0865237 .1845415 d4 | .0792169 .022696 3.49 0.001 .0344128 .124021 d5 | .0477914 .0222951 2.14 0.033 .0037787 .0918041 d6 | .1860513 .0230631 8.07 0.000 .1405224 .2315801 period | -.0051111 .0007851 -6.51 0.000 -.006661 -.0035613 aptitude | -.0203517 .0061947 -3.29 0.001 -.0325806 -.0081229 _cons | .1354551 .0857042 1.58 0.116 -.0337335 .3046438 ------------------------------------------------------------------------------

OFT1226 | 157

5. Base regressions including aptitude (IQ test) and interaction between price frames and time trend a. Liner estimation: Interaction terms between price frames and time trend (overall errors)
reg err p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude pd2 pd3 pd4 pd5 pd6 , cluster(subject ) Linear regression Number of obs = 5015 F( 16, 169) = 13.46 Prob > F = 0.0000 R-squared = 0.0568 Root MSE = .47089 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .3771078 .0652229 5.78 0.000 .2483512 .5058644 p2 | .2040824 .0477327 4.28 0.000 .1098532 .2983116 highvalueg~d | .0076402 .0135429 0.56 0.573 -.0190948 .0343751 c | -.5054014 .2078975 -2.43 0.016 -.915812 -.0949909 d2 | .037421 .0449334 0.83 0.406 -.051282 .126124 d3 | .1953679 .0438268 4.46 0.000 .1088495 .2818863 d4 | .1115643 .0451669 2.47 0.015 .0224003 .2007283 d5 | .0951622 .043486 2.19 0.030 .0093166 .1810079 d6 | .072565 .043916 1.65 0.100 -.0141297 .1592596 period | -.0053366 .0012772 -4.18 0.000 -.0078579 -.0028153 aptitude | -.0202384 .0059997 -3.37 0.001 -.0320824 -.0083944 pd2 | -.0008417 .0024809 -0.34 0.735 -.0057393 .004056 pd3 | -.0035654 .0024701 -1.44 0.151 -.0084416 .0013108 pd4 | -.0018882 .0024134 -0.78 0.435 -.0066524 .0028761 pd5 | -.0022197 .0024425 -0.91 0.365 -.0070415 .0026021 pd6 | .0065071 .0024687 2.64 0.009 .0016336 .0113806 _cons | .1788373 .089109 2.01 0.046 .0029272 .3547475 ------------------------------------------------------------------------------

OFT1226 | 158

b. Linear estimation: Interaction terms between price frames and time trend (search errors)
. reg err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude pd2 pd3 pd4 pd5 pd6 , cluster(subject ) Linear regression Number of obs = 5015 F( 16, 169) = 7.83 Prob > F = 0.0000 R-squared = 0.0319 Root MSE = .4427 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .2152265 .0632046 3.41 0.001 .0904544 .3399986 p2 | .2137568 .0410381 5.21 0.000 .1327435 .29477 highvalueg~d | .0036083 .0127172 0.28 0.777 -.0214968 .0287133 c | -.493282 .1840931 -2.68 0.008 -.8567003 -.1298638 d2 | .0042466 .0416249 0.10 0.919 -.0779252 .0864184 d3 | .1223118 .0447887 2.73 0.007 .0338945 .2107291 d4 | .0760956 .04367 1.74 0.083 -.0101134 .1623047 d5 | .0591564 .0450419 1.31 0.191 -.0297608 .1480737 d6 | .1004524 .0421358 2.38 0.018 .0172722 .1836326 period | -.0024335 .0012911 -1.88 0.061 -.0049823 .0001153 aptitude | -.0089475 .0044824 -2.00 0.048 -.0177963 -.0000988 pd2 | .0014008 .0023995 0.58 0.560 -.0033361 .0061377 pd3 | -.00222 .0025422 -0.87 0.384 -.0072386 .0027987 pd4 | -.0006743 .0023445 -0.29 0.774 -.0053025 .0039539 pd5 | -.0011161 .0025334 -0.44 0.660 -.0061172 .003885 pd6 | .0049538 .0023613 2.10 0.037 .0002925 .0096152 _cons | .057128 .0766873 0.74 0.457 -.0942604 .2085165

OFT1226 | 159

c. Linear estimation: Interaction terms between price frames and time trend (purchasing errors)
reg err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude pd2 pd3 pd4 pd5 pd6 , cluster(subject ) Linear regression Number of obs = 5015 F( 16, 169) = 22.24 Prob > F = 0.0000 R-squared = 0.0829 Root MSE = .43498 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .5816763 .0559091 10.40 0.000 .471306 .6920465 p2 | -.1404681 .0479951 -2.93 0.004 -.2352153 -.0457209 highvalueg~d | .0221594 .0130499 1.70 0.091 -.0036024 .0479212 c | .477269 .1960103 2.43 0.016 .090325 .8642129 d2 | .0178651 .0419834 0.43 0.671 -.0650144 .1007446 d3 | .1875255 .0457952 4.09 0.000 .0971212 .2779297 d4 | .1082362 .0403947 2.68 0.008 .028493 .1879795 d5 | .0548557 .0444816 1.23 0.219 -.0329555 .1426668 d6 | .0968941 .044949 2.16 0.033 .0081602 .1856279 period | -.0051491 .0011782 -4.37 0.000 -.0074751 -.0028231 aptitude | -.0203441 .0062037 -3.28 0.001 -.0325908 -.0080974 pd2 | .0001282 .002341 0.05 0.956 -.0044931 .0047496 pd3 | -.0034397 .0023807 -1.44 0.150 -.0081394 .00126 pd4 | -.0018611 .0022774 -0.82 0.415 -.0063569 .0026347 pd5 | -.0004564 .0023133 -0.20 0.844 -.0050231 .0041102 pd6 | .005971 .0024994 2.39 0.018 .0010368 .0109051 _cons | .1349531 .0880065 1.53 0.127 -.0387806 .3086868

OFT1226 | 160

6. Base regressions including interaction between search costs and treatment a. Linear estimation: Interaction terms between search costs and treatment (overall errors). Referred to in paragraph 5.2.
reg err p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude cd2 cd3 cd4 cd5 cd6 , cluster(subject ) Linear regression Number of obs = 5015 F( 16, 169) = 13.49 Prob > F = 0.0000 R-squared = 0.0560 Root MSE = .4711 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .3761938 .0651845 5.77 0.000 .2475131 .5048745 p2 | .2051198 .0480376 4.27 0.000 .1102887 .2999508 highvalueg~d | .0079281 .013412 0.59 0.555 -.0185486 .0344049 c | .1429074 .3526965 0.41 0.686 -.5533509 .8391657 d2 | .0482192 .0418448 1.15 0.251 -.0343865 .130825 d3 | .2049976 .0411469 4.98 0.000 .1237695 .2862257 d4 | .1535119 .0402235 3.82 0.000 .0741067 .232917 d5 | .0930471 .0422811 2.20 0.029 .0095801 .1765142 d6 | .2563992 .0500176 5.13 0.000 .1576594 .355139 period | -.0055708 .0007983 -6.98 0.000 -.0071467 -.0039949 aptitude | -.0206072 .0059922 -3.44 0.001 -.0324365 -.0087779 cd2 | -.422031 .6216148 -0.68 0.498 -1.649161 .8050991 cd3 | -1.143424 .5824744 -1.96 0.051 -2.293287 .0064391 cd4 | -1.289201 .5862457 -2.20 0.029 -2.446509 -.1318926 cd5 | -.5789729 .5792171 -1.00 0.319 -1.722406 .5644598 cd6 | -1.54217 .7042568 -2.19 0.030 -2.932444 -.1518962 _cons | .1495132 .0884576 1.69 0.093 -.025111 .3241374 ------------------------------------------------------------------------------

OFT1226 | 161

b. Linear estimation: Interaction terms between search costs and treatment (search errors)
reg err_v p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude cd2 cd3 cd4 cd5 cd6 , cluster(subject ) Linear regression Number of obs = 5015 F( 16, 169) = 8.46 Prob > F = 0.0000 R-squared = 0.0354 Root MSE = .4419 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_v | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .2149247 .0630822 3.41 0.001 .0903942 .3394553 p2 | .2148667 .0410987 5.23 0.000 .1337337 .2959997 highvalueg~d | .0035045 .0125274 0.28 0.780 -.0212258 .0282347 c | .3256471 .3103176 1.05 0.295 -.2869511 .9382453 d2 | .0368799 .037047 1.00 0.321 -.0362545 .1100144 d3 | .1698329 .0402472 4.22 0.000 .0903809 .2492849 d4 | .1810451 .0372284 4.86 0.000 .1075526 .2545377 d5 | .0577631 .0359549 1.61 0.110 -.0132155 .1287417 d6 | .3015525 .0486231 6.20 0.000 .2055656 .3975394 period | -.0020197 .0007816 -2.58 0.011 -.0035626 -.0004769 aptitude | -.0094729 .0044791 -2.11 0.036 -.018315 -.0006307 cd2 | -.189733 .582171 -0.33 0.745 -1.338997 .9595309 cd3 | -1.458678 .5479735 -2.66 0.009 -2.540433 -.3769234 cd4 | -2.096243 .573799 -3.65 0.000 -3.228979 -.9635056 cd5 | -.2955704 .5094517 -0.58 0.563 -1.301279 .7101384 cd6 | -2.256467 .6486409 -3.48 0.001 -3.536949 -.9759846 _cons | .0096037 .0766216 0.13 0.900 -.141655 .1608623 ------------------------------------------------------------------------------ z and P>|z| correspond to the test of the underlying coefficient being 0

OFT1226 | 162

c. Linear estimation: Interaction terms between search costs and treatment (purchasing errors)
reg err_n p1 p2 highvaluegood c d2 d3 d4 d5 d6 period aptitude cd2 cd3 cd4 cd5 cd6 , cluster(subject ) Linear regression Number of obs = 5015 F( 16, 169) = 21.21 Prob > F = 0.0000 R-squared = 0.0817 Root MSE = .43528 (Std. Err. adjusted for 170 clusters in subject) -----------------------------------------------------------------------------| Robust err_n | Coef. Std. Err. t P>|t| [95 per cent Conf. Interval] -------------+---------------------------------------------------------------p1 | .5803583 .0558194 10.40 0.000 .4701653 .6905513 p2 | -.1404074 .048253 -2.91 0.004 -.2356637 -.045151 highvalueg~d | .0224876 .0129686 1.73 0.085 -.0031136 .0480889 c | .9580039 .3050051 3.14 0.002 .3558932 1.560115 d2 | .0508532 .0353083 1.44 0.152 -.018849 .1205553 d3 | .1862716 .0413694 4.50 0.000 .1046043 .2679389 d4 | .0960869 .036041 2.67 0.008 .0249383 .1672354 d5 | .0883287 .03928 2.25 0.026 .010786 .1658714 d6 | .2523042 .0448162 5.63 0.000 .1638325 .3407759 period | -.0050886 .0007814 -6.51 0.000 -.0066311 -.0035461 aptitude | -.0205489 .0061923 -3.32 0.001 -.0327731 -.0083247 cd2 | -.5555199 .5116064 -1.09 0.279 -1.565482 .4544425 cd3 | -.9127406 .5368703 -1.70 0.091 -1.972577 .1470954 cd4 | -.2974583 .5430896 -0.55 0.585 -1.369572 .7746551 cd5 | -.7239608 .5178312 -1.40 0.164 -1.746212 .29829 cd6 | -1.177779 .6535352 -1.80 0.073 -2.467923 .1123647 _cons | .1099432 .0858383 1.28 0.202 -.0595103 .2793967 ------------------------------------------------------------------------------

OFT1226 | 163

PERSONALITY TEST
The personality survey consisted of the following statements. Subjects had to choose how strongly they agreed or disagreed with them. I am somebody who: 1. Is often worried. 2. Is communicative and talkative. 3. Works systematically and thoroughly. 4. Has original ideas. 5. Is sometimes a little tough on others. 6. Tends to be lazy. 7. Becomes nervous easily. 8. Is sociable and can let their hair down. 9. Values artistic experiences. 10. Can forgive others. 11. Is reserved. 12. Has a vivid imagination. 13. Treats others respectfully and politely. 14. Is relaxed and can cope with stress easily. 15. Solves tasks efficiently and effectively.

Developed by Schupp and Gerlitz , DIW German Institute for Economic Research, 2005.

OFT1226 | 164

FEEDBACK QUESTIONNAIRE

Here are the questions that we asked the subject in the questionnaire. Note that not all subjects received all parts of question 3. 1. 2. 3. Did you feel the instructions were clear? What was your strategy during the experiment (in terms of shops to visit, where to purchase and number of units to buy)? Do you have any comments on the shopping experience where you encountered a: a. price advertised on the home screen that sometimes was not available at the shop? b. price that only existed for the first visit to the shop? c. 3 for 2 offer d. price that was advertised at the shop as reduced from a higher price e. situation where you needed to add shipping and handling after choosing how many units to buy 4. 5. 6. 7. 8. 9. Were there any aspects of the shopping experience that you enjoyed? Please explain what and why. Were there any aspects of the shopping experience that you found annoying? Please explain what and why. How did the experiment, pricing practices you encountered and your behaviour mirror/reflect your experience of shopping? How old are you in years? Which gender are you? What subject, if any, are you studying?

10. Any other comments.

OFT1226 | 165

EXPERIMENT INSTRUCTIONS EXPERIMENTAL INSTRUCTIONS

Welcome to our experiment! In the course of this experiment you can earn a substantial amount of money. The precise amount will depend on your choices and some luck. We kindly ask you to remain silent throughout the entire experiment. Do not attempt to communicate with your neighbours and do not try to look at their screens. If you have any questions, please, raise your hand and we will come and answer it in private. This experimental session will consist of several on-screen stages: 1. 2. 3. 4. 5. Quiz (to ensure you understand the instructions) Experiment (described in the instructions below) Multiple choice quiz Questionnaire about yourself Feedback questionnaire about the experiment

Your payment for this session will consist of the amounts you earn in Stages 2, 3 together with the 5 show up fee. We will pay you in cash. You will need to sign a receipt, which we will supply. In this experiment, we simulate the idea of shopping for different products. The experiment will have 30 periods and in each period, you can buy zero or more units of a product that is available to buy at two different shops (the same product is available at both shops). By buying units of the product, you will get some points. At the end of the experiment, the total sum of all the points you earned over the 30 periods will be converted into pounds, at a rate of 1 for 50 points.

OFT1226 | 166

Let us describe the experiment. You can buy four different products. They are called GREEN, ORANGE, BLUE, and RED. For each of these products, you can see in the attached table (on page 3) how many points you will get depending on the number of units you buy in a period. For example, if you buy a total of two units of GREEN in a period you get 100 points. Or for a total of three units of RED you get 220 points. The prices for these products will also vary from period to period and shop to shop. The table shows you, for each good, the price range in points you can usually expect. This is the price is per unit. The price will be randomly chosen each period from the price range and the price for each shop will be determined separately. At the beginning of each period, you will see your home screen. This will inform you of which product is available to buy in this period (only one product will be sold in any one period and it is chosen at random) so that you know which row of the table is relevant. The product available will also be identified by the colour of the screen surround. On your home screen, there are two buttons, representing the two shops where you can buy the product. They will be labelled 'Go To Shop 1' and 'Go To Shop 2'. There is also a button labelled 'Im done' which ends the period and brings up the results. If you want to go to a shop to buy something (or simply to check the price), you can do so by clicking on this shops button. Every time you visit a shop, you will incur some costs (think of time it takes to get there, petrol, etc). The cost (denoted in points) is displayed on the home screen (the cost is the cost to make one return trip to a shop). Once you are at the shop, you will see the price of the good that is available to buy in the period and, if you want to buy, you can enter the number of units you wish to buy (the maximum total number of units you can buy in any one period is 4). A box will appear where you need to confirm your purchase. Once your purchase is confirmed (or if you choose to cancel), you will be returned to the Shop screen (not to your home screen). If you make a purchase, a box will briefly appear informing you of the number of units you have successfully bought.

OFT1226 | 167

If you dont want to buy anything (because you find the prices too high or you first want to check out the other shop) or once you have finished shopping, you can click the Go home button to return to the home screen. You can go back and forth between your home screen and the shops as often as you like but notice that you are charged for every trip you make. Once you are done with your shopping for a period, you can click the Im done button at the bottom right corner of your home screen and proceed to the results screen, where you are informed of your decisions and earnings in the period. Your earnings in each period depend on the number of units of the product you bought as shown in the attached table. From these points, we take away your travel costs and the amount of points you spent on buying the product. For example, suppose in a period the GREEN product is available to buy and that travel costs are 3 and you take the following actions: 1. Go to shop 2 and observe a price of 48 per unit for GREEN. 2. Go back to the home screen. 3. Go to shop 1 and observe a price of 31 per unit for GREEN. 4. Buy 3 units of GREEN at shop 1. [You are still at shop 1 after the purchase] 5. Go to home screen. 6. Click 'Im done' Then your earnings for the period would be: Points for buying 3 units of GREEN: 110 Minus travel costs (two visits to the shops): 2 * 3 = 6 Minus cost of buying 3 units in shop 1: 3 * 31 = 93 Your earnings for the period would be: 110 6 93 = 11 points

OFT1226 | 168

The prices shown above are purely for example and the strategy of the shopper in this case may or may not be good. The real prices for GREEN that you will encounter will be randomly drawn each period between 30 and 60. Before the experiment, there will be a short quiz on the screen to check that you understand these instructions. You will need to get all the answers right before you can proceed to the experiment, but you can have a second, third, etc. go if you get an answer wrong. After the experiment, there will be another on-screen quiz consisting of 12 multiple choice questions. Instructions will appear on the screen. If you get more than 10 correct, we will pay you an extra 2 or if you get more than eight correct we will pay an extra 1. You have a total of eight minutes to complete the quiz it will time out after this point. Following this quiz, there will be two short questionnaires where we ask you to give us your thoughts on the experiment, your strategy and about yourself. At the end of the experiment, you will be asked to fill in a receipt. We will call you up to the front, one by one, to pay you in cash. After this, you are free to go. Product GREEN ORANGE BLUE RED 0 units 1 unit 0 0 0 0 60 80 110 120 2 units 3 units 4 units 100 140 180 200 110 170 190 220 115 195 190 230 Price Range 30 to 60 50 to 80 50 to 110 60 to 120

OFT1226 | 169

REFERENCES

Armstrong.M and Zou.J (2010) Conditioning prices on search behaviour, University of Munich Library, MPRA Paper. Biswas, A, Wilson, EJ and Licata, JW (1993) Reference pricing studies in marketing: a synthesis of research results, Journal of Business Research, 27, 239256. Alford.B.L., and A Biswas (2002) The effects of discount level, price consciousness and sale proneness on consumers' price perception and behavioral intention, Journal of Business Research, 55, 775-783. Bakos, Y. (2001) The Emerging Landscape for Retail E-commerce Journal of Economic Perspectives, Vol. 15, No.1 pp 69-80, 2001 Baye, M., Morgan, J., and Scholten, P. Price Dispersion in the Small and in the Large: Evidence from an Internet Price Comparison Site, Journal of Industrial Economics 52 (4), pp. 463-496, 2004 Baye, M.. Baye.M.R., Gatti. R.R and Morgan.J (2006) Did the Euro Foster Online Price Competition? Evidence from an International Price Comparison Site, Economic Inquiry, 44(2), 265-279. Blair, EA and Landon, EL, (1981) The effects of reference prices in retail advertisements. Journal of Marketing, 45, 6169. Brehm, S. S. & Brehm, J. W. (1981). Psychological reactance, New York: Academic Press. Brown, J. and Goolsbee, A. (2002) Does the Internet Make Markets More Competitive? Journal of Political Economy, 110 (3), pp.481-507., 2002 Brown.J., Hossain.T., and Morgan.J., (2007) Shrouded Attributes and Information Suppression: Evidence from the Field. Mimeo, University of California (faculty.haas.berkeley.edu/rjmorgan/shrouding.pdf).

OFT1226 | 170

Burks, SV, Carpenter, JP, Gotte, L, and Rustichini, A (2008) Cognitive skills explain economic preferences, strategic behaviour, and job attachment, IZA Discussion Paper No. 3609. Cialdini. R. (2009) Influence: Science and Practice, 5th Edition, Boston: Pearson. Cason.T. and Friedman.D., (2003) Buyer search and price dispersion: a laboratory study, Journal of Economic Theory Chandrashekaran, R. & Grewal, D. (2006). Anchoring effects of advertised reference price and sale price: The moderating role of saving presentation format. Journal of Business Research, 59, 10631071 Charness.G. and Chen.K. (2001). Minimum advertised price policy rules and retailer behaviour: An experiment by Hewlett-Packard. Chioveanu, I, and Zhou, J (2009) Price competition with consumer confusion, MPRA Paper No. 17430. http://mpra.ub.uni-muechen.de/17340 Davis.D and Holt,C (1996) Consumer search costs and market performance, Economic Inquiry, 34, 133-51. DellaVigna, S (2009) Psychology and economics: Evidence from the field, Journal of Economic Literature 47, 315-372. DellaVigna, S, and Malmendier, U (2006) Paying not to go to the gym, American Economic Review 96, 694-719. DelVecchio,D, Krishnan,HS and Smith, DC (2007), Cents or Percent? The Effects of Promotion Framing on Price Expectations and Choice, Journal of Marketing, 71 (3), 158-170. Devlin.J, Ennew.C, McKechnie.S, and Smith.A, (2007) A study of time limited price promotions Journal of Product Brand and Management, 280 285. Dickson, PR, and Sawyer, AG (1990) The price knowledge and search of supermarket shoppers, Journal of Marketing 54, 42-53.

OFT1226 | 171

Ellison, G, and Ellison, SF (2005) Search, obfuscation, and price elasticities on the internet, mimeo, February 2005. Gerstner, E, and Hess, JD (1990) Can bait and switch benefit consumers?, Marketing Science 9, 114-124. Gabaix, X, and Laibson, D (2006) Shrouded Attributes, Consumer Myopia, and Information Suppression in Competitive Markets, Quarterly Journal of Economics, 121, 505-540. Hertwig, R, and Ortmann, A (2010) Experimental practices in economics: A methodological challenge for psychologists?, Behavioural and Brain Sciences 24, 383-403. Hess, JD, and Gerstner, E (1998) Yes, 'bait and switch' really benefits consumers, Marketing Science 17, 283-289. Hossain.T and Morgan.J (2006) Plus Shipping and Handling: Revenue (Non) equivalence in Field Experiment on eBay, Advances in Economic Analysis and Policy, 6(2), 1-26. Lynch.J. and Ariely.D. (2000) Wine online: Search cost affect, competition on price, quality and distribution, Marketing Science, vol.19(1) Lynn, M. (1989). Scarcity effect on value: Mediated by assumed expensiveness. Journal of Economic Psychology, 10, 257-274. Morton, FS, Zettelmayer, F, and Silva-Risso, J (2001) Internet car retailing, Journal of Industrial Economics 49, 501-519. Morwitz, VG, Greenleaf, EA, and Johnson, EJ (1998) Divide and prosper: Consumers reactions to partitioned prices, Journal of Marketing Research 35, 453-463 Ofcom (2010) Using experiments in consumer research, available at: www.ofcom.org.uk/research/telecoms/reports/experiments/ OFT (2009) Road testing consumer remedies, available at: www.oft.gov.uk/shared_oft/economic_research/oft1099.pdf

OFT1226 | 172

Piccione.M. and Spiegler.R. (2009) Farming Competition, available at http://else.econ.ucl.ac.uk/newweb/displayProfile.php?key=54 Simonson, I. (1999). The Effect of Product Assortment on Buyer Preferences. Journal of Retailing, 75, 347370 Wansink B., Kent, R.J., and Hoch S.J. (1998). An Anchoring and Adjustment Model of Purchase Quantity Decisions. Journal of Marketing Research, 65, 7181 Wilkie, WL, Mela, CF, and Gundlach, GT (1998) Does 'bait and switch' really benefit consumers? Marketing Science 17, 273-282.

OFT1226 | 173