You are on page 1of 15

J Cult Econ (2006) 30:141155 DOI 10.

1007/s10824-006-9006-7 ORIGINAL ARTICLE

The demand for books estimated by means of consumer survey data

Vidar Ringstad Knut Lyland

Received: 5 August 2004 / Accepted: 21 January 2006 C Springer Science + Business Media B.V. 2006

Abstract In this analysis we present some results for book demand (schoolbooks are excluded) in Norway, obtained by means of a three-goods model (books, other cultural goods and non-cultural goods) and survey data for more than 18,000 households from the period 19861999. Various methods of estimation are used, and they provide surprisingly unambiguous results. Our hypotheses about the price and income sensitivity of book demand are conrmed. Books turn out to be luxury goods. Our calculations do also suggest that they are quite price sensitive and that they are close substitutes to other cultural goods. The results for socio-demographic variables indicate that access to outlets for books, sex and age matter for book demand. Moreover, we nd that single persons and households with small children, especially those with children less than 7 years, are frequent book-buyers. Keywords Book demand . Income elasticity . Price elasticities . Consumer survey data

1. Introduction Reliable information about the determinants of book demand is important in several contexts. For instance, it is common to assume monopolistic competition for books due to many competing titles and large set up costs, cf. Dixit and Stiglitz (1977). This implies an adjustment of production to a level of demand where the price elasticity is less than 1. If demand analyses conrm that books are price sensitive, they conrm the model to be reasonably accurate as well. If not, they suggest that there could be something wrong with this model specication. The price elasticity of books is also of critical importance for the effects of the modes of public literature policy. Thus price sensitive demand makes subsidies a possible efcient means to increase book consumption. The welfare costs of xed book prices, used in several European countries, are also critically dependent on the price elasticity, cf. Beck (2003) and Ringstad (2004). The impact of income on book demand is also important. The

V. Ringstad K. Lyland Telemark Research Institute, Box 4, 3833 B i Telemark, Norway e-mails:, Springer


J Cult Econ (2006) 30:141155

income sensitivity of books, combined with a positive general trend in the income among consumers, is of decisive importance for the size of the book market. The bigger the income elasticity, the more likely it is for the book market to obtain a bigger share of gross national product over time. Books are also competing with a number of more or less related goods. The effect on the demand for books of price changes on such goods depends on the size of the cross-price elasticity and on how big the price changes actually are. The relative price of likely substitutes such as TV and video has decreased during the last decades. This could have had substantial crowding out effects on the demand for books. The limited number of studies of book demand reects the status of research about the economics of books in general. We have found only a few studies with fairly complete documentation of how the estimates on various parameters of the demand function are obtained. Bittlingmayer (1992) deals with the demand for separate titles on the German book market. The data are obtained from a publisher of a wide range of mostly professional books and the analysis is thus not necessarily representative for books in general.1 The analysis covers the years 19841986, but only a selection of the titles published these years is included. The central explanatory variable is the book price, and a main nding based on the more statistically reliable results, is a price elasticity between 2 and 3. Most of the differences in demand are due to other factors, however. This reects the uncertainty of the demand for books, which implies a corresponding economic risk of book production. Hjorth-Andersen (2000) postulates a relation for aggregate demand in the Danish book market with four explanatory variables: the price of books, aggregate disposable income, number of titles, and a trend measured by the number of years during the period studied, 19731991. This study also concludes that book demand is price sensitive, with an estimated price elasticity of about 1.4. It makes sense that aggregate book demand is less price elastic than the demand for single books, estimated by Bittlingmayer (1992). Increasing the price of a particular title reduces the demand for this title, but increases the demand for other books, leaving total demand for books less affected. Hjorth-Andersen nds that the income elasticity of books in the Danish book market is about 1.8. Books are, thus, luxury goods. The number of titles does not seem to affect demand. On the other hand there is a signicantly autonomous negative trend in book demand, perhaps due to the crowding out effect of new and cheaper close substitutes to books. Prieto-Rodr guez et al. (2004) analyze book demand based on quarterly Spanish consumer survey data of about 3,200 households for the period 198595. They estimate a model consisting of 19 goods of which three are cultural goods. The rst one is an aggregate of cinema, theater, museums and other cultural events (the CTM-good). The second consists of books, magazines and newspapers (the BMN-good). The third good is lm and music on magnetic media (the FMM-good). For the BMN-good they nd an income elasticitity of about 1.4, conrming Hjorth-Andersens ndings. They also obtain estimates of the own price elasticity of 1,65, which is in between the estimates of Hjorth-Andersen (2000) and Bittlingmayer (1992). Prieto-Rodr guez et al. (2004) also estimate cross-price elasticities between the three cultural goods in question, and nd the BMN-good to be complementary to both the CTM- and the FMM-good, contrary to what is usually assumed. Related calculations based on data for the years 19761990 are presented in Fishwick and Fitzsimons (1998). They suggest a price elasticity of about 0.9 and an income elasticity slightly above 1.4. Goolsbee and Chevalier (2002) have carried out an analysis of the demand

1 Due to the economic sensitivity of the information used (prices, gross margins for the book stores and costs) the identity of the publisher is not disclosed.


J Cult Econ (2006) 30:141155


for books on the Internet. Due to differences in approach and methodological framework, the results obtained are not comparable to those referred to here, however. Our analysis is carried out to get a better understanding of how the Norwegian book market works, and it is based on consumer survey and price data collected by Statistics Norway (SN) for the construction of weights in the calculation of the consumer price index. The advantage of these data is the rich and detailed information provided for each household about its income, wealth, location, age and the sex of the main person, and other socio-demographic characteristics. This information is available for 19861999, but unfortunately not for the same households. The combined cross-section/time series nature of the data allows for an exploration of a fairly wide variety of interesting hypotheses about the determinants of the demand for books. A disadvantage is, of course, that these data are not collected for our purpose or for demand studies in general. This raises several problems not usually present in data collected for econometric calculations. Even if there are no costs of collecting these data, there are others connected to their preparation for econometric calculations, as well as to the exploration of their peculiarities, that, if ignored, could bias our results seriously, or even destroy them completely. One problem is that we have data about the consumption of books (as well as other goods) only for just two weeks for each household. This causes a serious problem when the data are used for our purpose, since for most households it is rather incidental if they use money at all on books during a two-week period. Actually, a great majority of them do not, but no outlays on books during the registration period does not necessarily mean that the households concerned do not have any demand for books in a longer time perspective one year for instance. This is something books have in common with a majority of all small goods. Thus there may be several pitfalls in these data, and in spite of their potential information value, the data problems could turn out to be rather severe. Since reliable knowledge about the demand for books is limited, the information value of this source of data is nevertheless worth exploring. The data used are the kind collected by the national statistical bureaus in all industrialised countries. The outcome of our efforts, if reasonably successful, could thus initiate corresponding analyses in other countries. In Section 2 we explain the theoretical and methodological basis of the analysis. The sources of data, the information available and how it is used and how the variables entering the calculations are constructed are presented in Section 3. In Section 4 the results of the analysis are presented and discussed, and Section 5 contains a summary of the analysis and the main ndings obtained.

2. Methodological issues The large number of households reporting no demand for books leads us to confront a related issue as well, namely: Who buys books and what are the characteristics of those households buying books during the registration period? Such issues can properly be dealt with by maximum likelihood estimation of the parameters of binary choice models. The likelihood that a household buys books can be expressed as follows:
R s =

f (s )ds = F ( R),



J Cult Econ (2006) 30:141155

where R is a vector of explanatory variables, is a vector of corresponding parameters, f (s ) is the density function, s is the integration variable and F () is the cumulative distribution function (cdf). In econometric specications of (1), the error term of the postulated relationship is usually assumed to have either a normal or a logistic distribution leading to the probit and the logit method, respectively. We choose the latter alternative which gives: Pr (B = 1) = F ( R) = e R = 1 + e R ( R) and Pr (B = 0) = 1 1 + e R (2)

In our case B = 1 for households reporting outlays on books and B = 0 for those who do not. The explanatory variables, R, are the same as those entering the demand functions, cf. later. The demand function can be derived from the general textbook theory of consumer demand, cf. Deaton and Muellbauer (1993). The household has a certain disposable income, pays certain prices for the various consumer goods, among them books, and tries to allocate income to maximise its utility. Thus, the demand for books and other goods is simultaneously determined. Here we focus both on the partial demand for books, ignoring the demand for other goods, as well as on a complete demand model for all goods. In both cases we use a three-goods specication; books, other cultural goods, and non-cultural goods. Thus the model is quite simple, but it allows us to focus on the good of main interest, books, as well as its sensitivity towards price changes on presumed substitutes. In addition to prices and income, preferences and time-dependent factors are also likely to affect consumer demand. We expect preferences to be related to socio-demographic factors, while we include the unemployment rate and a trend to explore time-dependent factors. Both could be related to the time budget of the consumers. In the three-goods case we then have: X i = f i (p, Y , Z, M) (i = 1, 2, 3) (3)

where X1 is the demand for books and X2 and X3 are the demand for other cultural and non-cultural goods, respectively. p is a vector of prices, Y is income, Z is as a vector of socio-demographic factors, and M is a vector of time-specic factors.2 To be able to carry out calculations, a functional form has to be specied. A linear logarithmic form in prices and income is frequently used in empirical research, and implies constant elasticities. This requires a positive demand for all units, however, a requirement not satised in our data. Here we use a semi-logarithmic specication instead, with the budget shares ( Si ) as the dependent variables. For good i this is dened as Si = piYX i . Adding an error term we then have the partial model for books (i = 1) as:

S1 = 1 +
j =1

1 j ln p j + 1 ln Y + 1 Z + 1 M + u


and corresponding relations for other cultural and non-cultural goods (i = 2, 3). The large number of households, reporting no outlays on books during the registration period, raises other potential serious problems in the estimation of the demand relationships. It is obvious that the distribution of the error term of the postulated demand function must be rather peculiar and OLS will generally provide biased estimates. One way to solve this

These explanatory variables correspond to the R-vector in (1) and (2). Springer

J Cult Econ (2006) 30:141155


problem is to leave out the zero observations. This could lead to estimates even more seriously biased, however, since we then loose the effect of prices and income on the likelihood of buying books which certainly must be an important element in the demand for books (cf. Greene, 2003, Section 22.3). An alternative approach in such situations is to use the Tobit method. This implies the following specication:
S1 = 1 + 3

S1 = S1 =

1 j ln p j + j =1 0 if S1 0, and S1 if S1 > 0.

1 ln Y + 1 Z + 1 M + u (5)

S1 is a latent variable for the budget share for books and u is a stochastic error term, assumed to be normally and independently distributed. To compute the elasticities of interest, we must take into account the fact that the S1 is unobserved. Due to this fact, the marginal effects based on the observed S1 must be corrected by a scale parameter reecting the likelihood of buying books (cf. Greene, 2003, Section 22.3). Thus, for instance, we get the estimated marginal effect of ln Y for the sample mean as:

S1 = 1 ln Y 1 +

1Z + 1 j ln p j + 1 ln Y + 1M j =1 .


is the cdf operator in the normal distribution case and is the standard deviation. Thus, a marginal effect of ln Y consists of two components according to McDonald and Moftt (1980): One due to the conditional mean of the latent variable, S1 , in the non-zero part of the distribution, and one due to the likelihood that the observation will fall in that part of the distribution. The various price and income elasticities depend on the factor shares, and for books in the Tobit case, they are given by: 11 = 1 ()/ S1 1 1k = k ()/ S1 1Y = ()/ S1 + 1 The degree and nature of the censoring in our case, has largely to do with the short registration period. There is a substantial random element in the likelihood of a household buying books during any one two-week period during the year, however. And it turns out that among all households included in the analysis, about 23% only buy books during the registration period. This also implies that no reported demand for books generally understates the demand for books of the households concerned in a wider time perspective. And it implies, correspondingly, that the outlays on books of those households that actually buy books during the registration period, overstates their demand for books in a wider time perspective. This peculiarity in our data cannot be taken care of properly by the Tobit method. Actually, it is likely to make both the Tobit and the OLS methods provide positively biased estimates of the income elasticity.

(k = 2, 3)



J Cult Econ (2006) 30:141155

There seems to be no proper solution for this problem, but it is likely to be substantially reduced by constructing aggregate or annual households. Since there are about 26 two-week periods in a year, this is done by sorting them by income for each year and grouping them into aggregates with 26 in each group.3 This gives us 690 aggregate households among which only 30 have no registered consumption of books, and it underlines the impact of the random element in the reporting of a small good like books during a two week period. While about 77% report no outlays on books during the registration period, less than 5% of the aggregate households do so. The advantage of this procedure is that the objection to the Tobit method, raised above, is now substantially weaker. A disadvantage, however, is that we lose all information about the socio-demographic variables. Nevertheless, we expect this treatment of the data to provide useful results about how serious the errors are that we commit by using the Tobit method on the original data. One problem remains, however: The demand for books and the demand for other goods are determined simultaneously. Our complete demand model consists of (4) and two more relations of the same kind, for other cultural goods and non-cultural goods, respectively. In addition, we impose the following restrictions implied by the theory of consumption:
3 3 3 3

i = 1,
i =1 i =1

i j = 0,
i =1

i = 0,
j =1

i j = 0,

and i j = ji


The rst three restrictions in (8) follow from adding up, i.e. that the share of expenditures must sum to unity. The fourth follows from the fact that the share equations are homogenous of degree zero in the prices, and the fth follows from an assumption of symmetry in the crossprice effects. Adding error terms and a price deator to this system of demand relations, we get a system well adjusted to the Almost Ideal Demand System (AIDS), cf. Deaton and Muellbauer (1980, 1993).4 The system can be estimated by the Seemingly Unrelated Regressions Estimator (SURE).5 The AIDS-calculations yield demand function results for all three goods, but only the results for books will be reported here. The relationships between the parameter estimates of (4) and the various elasticities are somewhat different when using AIDS as compared to the partial estimation case. Using the same notation as in (4) and (7), we now have: 11 = 1 / S1 (1 + ) 1k = k / S1 1Y = / S1 + 1 Sk S1 (k = 2, 3) (9)

3 We then start at the top and thus end up with the households with lowest income. This is done separately for each year, and for most years we end up with a residual number of households less than 26. They are added to the last ordinary aggregate household. Both budget share and income of the aggregate household are calculated as arithmetic means. 4 In the logit and Tobit calculations all prices and income are deated by the consumer price index. In the AIDS calculations nominal prices are used, while income is deated by a Stone index ln P = i Si ln( pi ). This is an approximation to the theoretically correct deator and is frequently used in empirical applications of the AIDS model, cf. Asche and Wessells (1998). 5 Attempts have been made to combine models and methods for censored data like the Tobit method with models and methods for simultaneous equations like AIDS, cf. Amemiya (1979) and Heien and Wessells (1990). There are several unsettled issues in this context, and thus we have not explored their performance.


J Cult Econ (2006) 30:141155


Thus, in this case as well, the elasticities depend on the budget share of books and the crossprice elasticities on the share of the other goods as well. As in the partial estimation case, the AIDS elasticities will be calculated for the empirical means of the budget shares.

3. The data The main source of data is the consumer surveys carried out by Statistics Norway (SN) to obtain weights in the consumer price index. Such surveys have been carried out each year from 1974, but before 1986 the data are not comparable to those collected later. 1999 was the last year for which data were available for this analysis. Each year, SN makes a selection of about 1,500 households. Some decline to take part in the survey, and for various reasons some of the households selected are excluded by the SN. Thus the nal net number is about 1,1001,200. These are supposed to come close to a representative selection of Norwegian households. For the 14 years covered by our study we thus have information for more than 18,000 households. Data are collected both by interviews and by registration schemes. The former is used to collect information about income, wealth, nancial status and transactions as well as sociodemographic characteristics of the households. The registration is made for the outlays of the households on the various consumer goods, using a very detailed specication, all in all for about 900 consumer goods. In addition to the consumer survey data, we use price data collected by SN for consumer price index calculations. These are macro prices, which vary over time only. Thus they are common for all households of the same year. We also use data for unemployment as an indicator of the business cycle. Table 1 gives some main statistical characteristics of the variables used in the analysis.

4. Main ndings In Section 2 we have explained the issues explored in this study as well as the methods used for our explorations. This section is organised in the same way as Section 2. First, we estimate, by means of the logit method, how various characteristics of the households, as well as book prices and other time specic variables, affect the likelihood of buying books. Second, we present results for demand functions for books, both when they are estimated separately, and when they are simultaneously estimated as a part of a complete three-goods system of demand functions. The two sets of calculations are presented in Tables 2 and 3, respectively. For the main explanatory variables, income and prices, all estimates are presented since they are used as the basis for calculation of the various elasticities. For the parameters of the Z- and M-vector variables, only the results for parameters signicantly different from zero at the 10% level are reported. These estimates are marked with a letter, ad, which corresponds to the 10, 5, 1 and 0.1% level of signicance, respectively. From Table 2 we note that the income of the household, as expected, has a substantial and signicantly positive impact on the likelihood of buying books during the two-week registration period. This is not the case for any of the prices, nor the unemployment rate, and there is no signicant autonomous trend in this likelihood. We also note, not surprisingly, that access to books matters. The likelihood of buying books is less in small communities with few bookstores than in cities. Sex also matters,


J Cult Econ (2006) 30:141155 Table 1 Descriptive statistics of the variables included in the analysis (n = 18,103) Variable label Households reporting outlays on books, B = 1. If no outlays, B = 0. (Schoolbooks are excluded). Aggregate households reporting outlays, on books, B = 1. If no outlays, B = 0. (n = 690) Budget share, books (Schoolbooks are excluded) Budget share, other cultural goods Budget share, non-cultural goods Price index, books, deated by the consumer price index Price index, other cultural goods, deated by the consumer price index Price index, non-cultural goods, deated by the consumer price index Total expenditures, NOK 1999. Unemployment in per cent Time trend Share of households in rural communities Share of households in communities with 2001,999 inhabitants Share of households in communities with 2,00019,999 inhabitants Share of households in communities with 20,00099,999 inhabitants Share of households in communities with more than 100,000 inhabitants (basis) Share of single person households Share of households with children, 719 yearsa Share of households with children, 06 years. Households with more than one person but with no children less than 20 years (basis) Share of households with female as main person Age of main person Number of consumer units Number of video and TV sets in the household
a Due

Variable name B

Mean 0.23





S1 S2 S3 p1 p2 p3 Y ZU Zt MB1 MB2 MB3 MB4 MB5

0.0048 0.0217 0.9736 95 86 102 263,874 5.3 0.25 0.16 0.23 0.15 0.21

0.0152 0.0384 0.0420 6 11 2 163,334 2.4

0 0 0.3271 88 71 100 8,146 1.8

0.399 0.673 1.00 103 102 104 2,165,390 8.7


0.15 0.29 0.23 0.33


0.26 46 2.2 1.8 15 0.8 0.9 17 0.6 0.0 94 7.8 9.0

to the preset household categories in the data, it was not possible to separate single parent households with children 06 years old from this category


J Cult Econ (2006) 30:141155 Table 2 Results for the estimated logit models (standard errors in parentheses) Variable ln(Y ) ln( p1 ) ln( p2 ) ln( p3 ) Zu Zt MB1 MB2 MB3 MB4 MH1 MH2 MH3 MAF MK MHA MTV n Log-likelihood Full sample 1.0334d (0.0384) 0.2916 (1.5991) 0.2971 (1.0096) 8.9219 (6.4545) + + 0.1288a (0.0568) + + 0.3051d (0.0597) 0.4344d (0.0667) 0.0967b (0.0465) 0.0080d (0.0016) 18103 9079.47 Aggregated data 3.4517d (0.4715) 6.5074 (20.7902) 5.9598 (12.1586) 30.7257 (76.2574)


690 73.4014

In the calculations households in areas with more than 100,000 inhabitants are used as the basis. Correspondingly, households consisting of more than one person, but with no children less than 20 years old, are used as basis for type of household; a Signicant at 10%-level; b Signicant at 5%-level; c Signicant at 1%-level; d Signicant at 0.1%-level

supporting the notion that women are better customers of booksellers than men. Moreover, there is a tendency of decreased likelihood of buying books by age. The results for type of household are rather surprising. Households with children less than 20 years old are more inclined to buy books than households with more than one adult and no children less than 20 years old (the base-group in the calculations). Couples with small children (less than 7 years old), is actually the type of household with the highest likelihood of buying books. Both their economic budget and their time budget are likely to have negative effects on their demand for books. But these effects, apparently, are more than neutralised by an inclination toward buying children books. This at least, seems to be the more likely explanation, and as we note from Table 2, the corresponding coefcient is highly signicant. In Figure 1 we illustrate the big impact income has on the likelihood of buying books. We note for instance that a household with an income (or rather total consumer expenditure) of NOK 800,000 (measured in 1999-prices) is about ve times more likely to buy books in the two week registration period than a household with an income of NOK 100,000.


J Cult Econ (2006) 30:141155

Table 3 Results for the partial and system calculations (standard errors in parentheses). Tobit, total sample, reduced model 1.9367d (0.0763) 1.3663 (3.6579) 0.9230 (2.2948) 8.7202 (14.615) + + AIDS, total sample, reduced model 0.1532d (0.0176) 0.8262 (0.8281) 0.6678 (0.5768) 0.1585 (0.7021) +

Variable ln(Y ) ln( p1 ) ln( p2 ) ln( p3 ) Zu Zt MB1 M B2 MB3 MB4 MH1 MH2 MH3 MAF MK MHA MTV n Loglikelihood R 2 adj ()

Tobit, total sample, full model 1.8435d (0.0880) 1.8061 (3.6629) 1.2305 (2.3017) 11.1139 (14.649) + + 0.3271b (0.1287) 0.2395a (0.1390) + 0.4088b (0.1645) 0.4916c (0.1371) 0.7183d (0.1534) 0.2848c (0.1055) 0.0166d (0.0036) 18103 17079.14

Tobit, aggregated data 0.1680d (0.0196) 0.7125 (0.9949) 0.7693 (0.5792) 0.0147 (1.1055) +

AIDS, total sample, full model 0.1775d (0.0216) 0.9044 (0.8271) 0.8317 (0.5768) 0.07280 (0.7018) + 0.0627a (0.0345) 0.0729a (0.0379) 0.1280d (0.0416) + + 0.0735c (0.0283) 0.0020c (0.0009)

AIDS, aggregated data 0.1457d (0.0186) 0.8458 (0.8373) 0.7434 (0.5838) 0.1025 (0.7158) +

18103 17142.69

690 201.51




0.008 0.20413 0.20635 0.93344



note in Table 2

a Signicant at 10%-level; b Signicant at 5%-level; c Signicant at 1%-level; d Signicant at 0.1%-level


J Cult Econ (2006) 30:141155

1 0.9 0.8
Likelihood of buying books during a year


0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 50 150 250 350

Likelihood of buying books during a twoweek period








Income in thousands of NOK measured in 1999-prices.

Fig. 1 The likelihood of buying books, by income, during a two-week period and during one year

The estimated relationship between income and the likelihood of buying books for an aggregate household (cf. Section 2) is presented in Figure 1 as well. We note that there is a striking difference between this relationship and the corresponding relationship estimated for the original households. For households with an income of NOK 100,000, only about 10% buy books during a two-week period. On an annual basis the likelihood is, as we see, about nine times higher or about 90%. This likelihood increases fast with income and is close to 100% when the income is more than NOK 200,000. The partial calculations are based on the Tobit method. Moreover, we have two sets of data: The original collection of 18,103 households and the 690 aggregate households obtained as explained in Section 2. The results are presented in the left part of Table 3. The corresponding income and price elasticities are found in Table 4.6 We note from Table 3 that the t of the relations, measured by log likelihood in the Tobit case and R2 adjusted in the AIDS case, is extremely poor, leaving most of the differences in demand between the households unexplained. This is quite common for demand analyses based on micro data, cf. the comments on Bittlingmayer (1992) in Section 1 above. We also note from Table 3 that by constructing aggregate households, much of the noise in the data is eliminated, improving the t considerabely. Income turns out to have a signicantly positive impact on the demand for books according to the three sets of partial results. The estimate of the income elasticity obtained by means of the aggregated data is, however, lower than the one obtained on the basis of the total sample, indicating that the Tobit method applied to the original data tends to overestimate the income effect as suggested in Section 2.

The Tobit method does not provide the marginal effects of prices and income needed to calculate elasticities. They are obtained by scaling the parameter estimates by (), presented at the bottom of Table 3, cf. also expression (6) and (7). Thus we obtain the marginal effect on S1 of lnY from the Tobit results for the total sample, full model, as = 1.84350 0.20413 = 0.3763. Springer


J Cult Econ (2006) 30:141155 Table 4 Price and income elasticities for books based on demand function estimation Cross-price elasticities Income Direct price Other cultural Non-cultural elasticity(1Y ) elasticity(11 ) goods(12 ) goods(13 ) Tobit, all households, full model Tobit, all households, reduced model Tobit, aggregated data AIDS, all households, full model AIDS, all households, reduced model AIDS, aggregated data 1.79 1.84 1.33 1.37 1.32 1.30 1.77 1.59 2.40 2.91 2.74 2.77 0.53 0.40 1.52 1.74 1.40 1.55 4.78 3.78 0.03 0.21 0.02 0.08

We note from Table 4 that the Tobit results yield even bigger income elasticities than AIDS. As pointed out in Section 2, the Tobit estimates could be positively biased, but there is no reason to believe they should be more biased than the AIDS estimates. Moreover, the causes of these biases should be almost eliminated by the aggregation we have carried out, and we see from Table 4 that the Tobit results for the income elasticity based on these data, are almost exactly the same as those obtained by AIDS on the original collection of households. Thus the income elasticity for books seems to be about 1.3, conrming the ndings of HjorthAndersen (2000), Fishwick and Fitzimons (1998) and Prieto-Rodr guez et al. (2004) that the demand for books is fairly income sensitive. All sets of results imply a direct price elasticity of books well below 1, but we note that the estimated standard errors are quite large. In this respect the results are rather uncertain. On the other hand, the corresponding elasticities are surprisingly unambiguous, and thus they seem to be quite robust to the different econometric specications and datasets employed. They all suggest that book demand is quite price sensitive, conrming the results obtained previously by Bittlingmayer (1992), Hjorth-Andersen (2000), and Prieto-Rodr guez et al. (2004). The calculations also suggest what is usually and reasonably assumed, that books and other cultural goods are substitutes. The level of signicance is poor for these results as well, but like the direct price elasticity results, they are quite unambiguous and robust. This result is contrary to the results in Prieto-Rodr guez et al. (2004). They use an aggregate which, in addition to books, includes magazines and newspapers, however, cf. Section 1. Their results are thus not directly comparable to ours. A basic weakness of the partial results for the various elasticities is an inconsistency due to the fact that the restriction on the sum of the elasticities, implied by the theory of consumption, 1Y + 11 + 12 + 13 = 0, (10)

is ignored. We note from the Tobit results in Table 4 that (10) is far from fullled in our calculations. In the AIDS calculations, the restriction in (10) holds, as long as the restrictions in (8) are fullled. It is interesting to note that AIDS provides more resonable results, with

J Cult Econ (2006) 30:141155


books and non-cultural goods largely independent in demand. It is also interesting to note that system estimation by AIDS has small effects on the estimates of the other elasticities. The Tobit and AIDS results largely conrm those obtained for the logit method concerning the impact of accessibility of books and the differences between the various types of households. In this context there are a couple of aspects that are not so evident from the logit results. From the Tobit results single-person households turn out to be quite important as book buyers. The logit results suggest that they are more inclined to buy books than are other households with only grown-up members. The Tobit and AIDS results also conrm the logit results concerning the sex and the age effect on the demand for books. In this context as well, the effects of the other variables explored; the trend, the unemployment rate, the number of consumption units in the household, and the number of TV and video sets, are small and quite insignicant in statistical terms.

5. Summary and main conclusions In this analysis we estimate the effects of factors explaining the likelihood of buying books, as well as the demand for books, by means of data for more than 18,000 Norwegian households from the period 19861999. In addition to the demand for books and other goods, we have information about the income of the households and various socio-demographic characteristics. Among the latter we utilise data about the size of the local community, number of members of the household and their age, the number of consumer units, the sex and the age of the main person of the household, and the number of TV and video sets. In addition, we use macro data for the unemployment rate and price indices for books, other cultural goods, non-cultural goods and total consumption. In our calculations we also include a trend to catch any systematic changes in the demand for books not embodied in the effects of the other variables. We have encountered several methodological challenges in this analysis. A main one has to do with the fact that the registration period is quite short, namely two weeks. Thus a majority of the households, or about 77%, do not have any demand for books at all according to their reporting. This is obviously misleading over a longer time perspective. On an annual basis for instance, the share of book buyers is likely to be substantially larger. This is conrmed by the results obtained on the basis of aggregate households. These are obtained by grouping households with approximately the same income for each year, with 26 in each group (since a year consists of about 26 two-week periods). More than 95% of the aggregate households then turn out to buy books. There are, not unexpectedly, big differences in the likelihood of buying books by income. On the basis of logit results, the estimated likelihood is about ve times higher for a household with an annual income of NOK 800,000, as compared to one with an annual income of about NOK 100,000. For the former it is about 50% and about 10% for the latter. The logit results for the aggregate households conrm that these calculations seriously underestimate the likelihood of buying books over a longer time perspective. For a household with an annual income of NOK 100,000 there is a likelihood of buying books during one year of about 90%, or about nine times higher than the likelihood of book buying during a randomly selected two-week period. For households with an annual income of NOK 200,000 or more, the likelihood of buying books during one year is close to 100%. Demand functions for books are estimated by means of two methods: the Tobit method and the AIDS method. The latter, in contrast to the former, takes into consideration the simultaneous nature of the demand for books and other goods. Both methods could lead to


J Cult Econ (2006) 30:141155

biased estimates due to the large number of households reporting no demand for books during the two-week registration period. Thus both methods, for various reasons, may yield biased estimates, but less so when using aggregate data. Despite this background the differences between the results obtained by the various methods and data are surprisingly small. This is the case both for the various price and income elasiticities of interest here and for the effects of the socio-demographic variables. The calculations yield an income elasticity signicantly above 1, which implies that book demand is rather income sensitive. They also suggest that book demand is quite price sensitive, and that books and other cultural goods are substitutes. The stability of the results for the price elasticities is particularly surprising since none of the corresponding coefcients are statistically signicant at any reasonable level of signicance. There are also several interesting differences between households with different sociodemographic characteristics. The more interesting because they are the less obvious are the differences between the various types of households in this respect. Single person households are more inclined to buy books than those with more than one adult and no children less than 20 years old. This is also reected in the demand function results. Households with children 719 years old are also more inclined to buy books than nonsingle households with only grown-ups. Households with small children (less than 7 years old) are, however, the type of household with the highest likelihood of buying books. Both their economic budget and their time budget are likely to have negative effects in this context. But these effects, apparently, are more than neutralised by an inclination toward buying children books. All in all we nd the results of this analysis quite intriguing. The results are fairly reasonable and those obtained for the various demand elasticities are surprisingly unambiguous. We do not know, however, whether or not these characteristics of our main results are incidental. Thus it would be of substantial interest to obtain more results for other cultural goods like newspapers, performing arts demand, cinema, TV-equipment/subscription, video and other AV-products. The core element of our approach is to focus on the good of main interest, the group of goods to which it belongs and the rest. This model for small goods does not overstretch the information about prices in the same way as a ten-goods model say, would do. Naturally, more analyses could turn out to be disappointing. On the other hand they may turn out to support the impression from the present study that such data are a unique source of information, in particular for the demand of small goods like books.
Acknowledgements This research was supported by The Norwegian Research Council. An earlier version of this paper was presented at the A.E.A International Conference on The Econometrics of Cultural Goods, Padova, April 2004. Two anonymous referees are gratefully acknowledged for valuable comments and suggestions on an earlier draft.

Amemiya, T. (1979). The estimation of a simultaneous-equation Tobit model. International Economic Review, 20, 169181. Asche, F., & Wessells, C.R. (1998). On price indicies in the almost ideal demand system. American Journal of Agricultural Economics, 79, 11821185. Beck, J. (2003). Monopoly vs. Oligopoly on the debate on resale price maintenance for books: a preliminary empirical result. Diplomarbeit, Humboldt-Universit at zu Berlin. Bittlingmayer, G. (1992). The elasticity of demand for books, resale price maintenance and the Lerner Index. Journal of Institutional and Theoretical Economics, 48, 588606. Springer

J Cult Econ (2006) 30:141155


Deaton, A.S., & Muellbauer, J. (1980) An almost ideal demand system. The American Economic Review, 70, 312326. Deaton, A.S., & Muellbauer, J. (1993) Economics and consumer behavior. Cambridge University Press, Cambridge, UK. Dixit, A.K., & Stiglitz, J.E. (1977) Monopolistic competition and optimal product diversity. The American Economic Review, 67, 297308. Fishwick, F., & Fitzsimons, S. (1998) Report into the effects of the abandonment of the net book agreement. Book Trust, London. Goolsbee, A., & Chevalier, J. (2003) Price competition online: Amazon Versus Barnes and Noble. Quantitative Marketing and Economics, 1(2), 203222. Greene, W.H. (2003) Econometric Analysis. Macmillan Publishing Company, New York, 5th Edition. Heien, D., & Wessells, C.R. (1990) Demand systems estimation with microdata: a censored regression approach. Journal of Business and Statistics, 8, 365371. Hjorth-Andersen, C. (2000) A model of the Danish book market. Journal of Cultural Economics, 24, 2743. McDonald, J.F., & Moftt, R.A. (1980) The uses of Tobit Analysis. Review of Economics and Statistics, 62, 318321. Prieto-Rodr guez, J., Romero-Jord an, D., & Sanz-Sanz, J.F. (2004) Is a tax cut on cultural goods consumption actually desirable? A microsimulation analysis applied to Spain. Paper Presented at the AEA Conference on the Econometrics for Cultural Goods, Padova, April, 2223. Ringstad, V. (2004) On the cultural blessings of xed book prices: facts or ction? International Journal of Cultural Policy, 10, 351365.