Вы находитесь на странице: 1из 25

Representation of taste heterogeneity in willingness to pay indicators using parameterization in willingness to pay space

Stefan Mabit Sebastian Caussade Stephane Hess September 15, 2006

Abstract One of the most important uses of discrete choice models is as a mean to estimate willingness to pay (WTP). The paper gives a review of dierent ways to incorporate taste heterogeneity in WTP indicators. Furthermore it investigates the dierence between the two approaches termed modelling in preference space and WTP space. On one hand the conclusion of the paper is that none of the approaches seems to dominate when it comes to model t, hence they should both be within the toolbox of the discrete choice modeller. On the other hand the investigations show that not including correlation between coecients in mixed logit models is a bigger problem for models in preference space. When it comes to the inclusion of explanatory variables the empirical investigation gives evidence to the fact that they should be incorporate with caution to avoid confounding heterogeneity in WTP with heterogeniety in scale.

Introduction and context

Discrete choice (DC) structures belonging to the family of Random Utility Models (RUM) have established themselves as the preferred tool for the analysis of choice behaviour in the area of transport research. They are used

Centre for Trac and Transport, Denmark, e-mail:slm@ctt.dtu.dk Department of Transport Engineering, Ponticia Universidad Catolica de Chile Centre for Tranport Studies, Imperial College, London

across a variety of contexts, ranging from mode choice to destination choice, via the choice of departure time and the choice of route. An ever increasing number of dierent structures are available to modellers, ranging from the basic Multinomial Logit (MNL) model to more advanced structures, such as Mixed Multinomial Logit (MMNL), and other Generalised Extreme Value (GEV) mixture models. While DC models are also used for purposes such as the forecasting of market shares after the implementation of hypothetical policy changes, the main output of studies using DC models is the computation of willingness to pay (WTP) indicators. These WTP indicators give an estimate of the readiness by a respondent to accept an increase in the cost of an alternative in return for an improvement to the alternative along some other dimension, such as travel time. This specic case, namely the trade-o between travel cost and travel time, gives rise to the most commonly used WTP measure, namely the valuation of travel time savings (VTTS). The VTTS is a measure indicating how much a respondent is willing to pay in additional travel cost in return for a reduction of his travel time by a certain measure k , where the VTTS is typically calculated for a value of k equal to one hour. The computation of VTTS measures has been one of the main applications of RUM models, with some recent discussions of the topic including Algers et al. (1998), Hensher (2001a,b,c), Lapparent & de Palma (2002), and Sillano & Ort uzar (2004). The VTTS is used for example for cost benet analysis (CBA) in the context of planning new transport systems, or for pricing. In discrete choice models, the computation of VTTS measures is relatively straightforward, given by the ratio of the partial derivatives of the utility function with respect to travel time and travel cost (i.e. the marginal rate of substitution between travel time and travel cost, at constant utility). Although this is an intuitively plausible approach, it is important to appreciate that the justication for this approach to the valuation of travel time savings rests not on plausibility but rather on a substantial body of microeconomic theory that addresses the issue of how individuals allocate time amongst alternative activities, including travel. Indeed, the topic of time allocation and valuation has been the subject of intense study from a variety of dierent perspectives for several decades (see, among others, Becker 1965, Oort 1969, De Serpa 1971, Evans 1972, Truong & Hensher 1985, Bates 1987 and Jara-Diaz & Guevara 2003). The papers by Jara-Diaz (2000) and Mackie et al. (2001) provide excellent overviews of the development of this literature. While the vast majority of discussions looking at the computation of WTP measures have focussed solely on the case of VTTS, it is important 2

to stress that various other WTP measures are of interest in the context of transport research. As such, policy planners may be interested in travellers willingness to pay for increases in frequency, improvements in on-time performance or reductions in schedule delay. In fact, the term willingness to pay can be rather misleading in this case, as the dimension along which the respondent accepts a decrease in attractiveness need not necessarily be a monetary one. Indeed, one can similarly be interested in respondents willingness to accept increases in travel time in return for improvements along some other dimension. One of the main topics of research in the area of WTP indicators has been the representation of heterogeneity in these indicators, i.e., a variation in their value across respondents. A lot of work has recently been done in this area, and several dierent ways of modelling heterogeneity in WTP indicators have emerged. However, there has not been any investigation in the transportation eld of specication in WTP space. Hence, the aim of this paper is to take another look at the problem of the computation of WTP indicators, with a specic focus on representing the heterogeneity across respondents directly in WTP space as opposed to traditional specications. Aside from oering a review of existing work in this area, the main contribution of the paper comes in several case studies that compare the dierent ways of accounting for heterogeneity in WTP indicators. The remainder of this paper is organised as follows: in the next section 2 we review the literature on how to represent heterogeneity in discrete choice models. In section 3 and section 4 we describe our empirical investigations of how to represent taste heterogeneity in mixed logit models on two dierent data sets. Section 5 summarizes the main points of the paper.

Representation of heterogeneity in WTP indicators

In this section, we look at dierent ways of representing heterogeneity in WTP indicators. While the majority of the discussion focusses on the case of VTTS, the ndings directly extend to other trade-os, be they monetary or not.

2.1

Computation of WTP indicators

Under the assumption that all eects of travel time and travel cost are captured in the observed part of utility V , the VTTS measure is simply

computed as: V /T , V /C (1)

with V giving the observed part of utility, and T and C representing the travel time and travel cost attributes respectively. In the case of xed taste coecients, and with the commonly used linear-in-attributes utility function, this formula reduces to T /C , where T and C are the time and cost coecients, giving the marginal utilities of increases by one unit in travel time and travel cost respectively. Estimates of these marginal utilities are produced by estimating the model on the choice data. Even with the use of non-linear transforms, such as the natural logarithm, the computation remains relatively straightforward, although the actual values of the attributes now enter into the computation of the trade-os.

2.2

Heterogeneity in WTP indicators

It should be clear that the assumption of a purely homogeneous population cannot in general be seen to be valid. As such, two individuals will dier in their behaviour due to dierences in priorities as well as sensitivities. In a DC context, these dierences are exhibited in the form of dierent values for the taste coecients that relate attributes values to the utility of an alternative. These dierences will clearly have an eect on the calculation of trade-os, leading to heterogeneity in WTP indicators. Given the importance of accurate estimates of WTP indicators for use in policy analysis, the representation of such heterogeneity is of crucial importance. In this section, we look at the issue of heterogeneous WTP indicators in the context of three discussions, namely the representation of deterministic heterogeneity (Section 2.2.1), the use of continuous or discrete mixtures for the representation of random heterogeneity (Section 2.2.2), and the question of whether to work in preference space or in WTP space (Section 2.2.3). 2.2.1 Deterministic heterogeneity

Approaches accounting for deterministic heterogeneity aim to relate variations in sensitivities, and hence WTP indicators, to socio-demographic information, such as trip purpose or income. In the most basic approach, the population is split into separate mutually exclusive segments, where potentially dierent segmentations are used along dierent dimensions. Separate coecients are then associated with 4

separate population segments. While easy to apply, the main issue with this approach is the requirement to specify appropriate segmentations. If too few segmentations are used, the majority of heterogeneity is left unexplained. On the other hand, if too many segmentations are used, the number of observations per segment can be so low as to lead to unstable estimation results. Finally, with the interest being on the computation of WTP indicators, it should be noted that the use of L segments along one dimension and M segments along the other dimension leads to LM dierent WTP measures, where the number of individuals in the combined segment lm might be very low. A slightly more advanced method consists of using continuous interactions between estimated parameters and socio-demographic attributes. It is clear that such continuous treatments of interactions have advantages in terms of exibility when compared to the more assumption-bound segmentation approaches. As an example, the following approach can be used1 : f (s, x) = x s s
s,x

x,

(2)

where s is the observed value for a given socio-demographic variable, such as income or trip distance, and s is a reference value for this attribute, such as the mean value across a sample population. In this example, the sensitivity to an alternatives attribute x varies with s. The choice of the reference value s is arbitrary, and has no eect on model t, or the estimate for s,x . However, the use of the mean value, s , guarantees that the estimate x gives the sensitivity to x with most precision at the average value of s in the sample population2 . The estimate of s,x gives the elasticity of the sensitivity to x with respect to changes in s; with negative values for s,x , the (absolute) sensitivity decreases with increases in s, with the opposite applying in the case of positive values for s,x . Finally, the rate of the interaction is determined by the absolute value of s,x , where a value of 0 indicates a lack of interaction. At this point, it should be said that a problem with this approach in the present context could be the fact that income information is presented in the form of a set of separate incomeclasses, as opposed to absolute income information, leading to a requirement for using class-midpoints, with the obvious averaging error this involves. Here, it should however be noted that similar averaging error occurs in the
This approach was suggested by Mackie et al. (2003) in the context of the reanalysis of the UK value of time study. ` s,x 2 With s = s , the term s disappears from equation 2. s
1

case where dierent income classes are grouped together in a segmentation approach. 2.2.2 Continuous versus discrete mixtures

While it is clearly preferable, from a computational as well as interpretation point of view, to explain as much of the heterogeneity as possible in a deterministic fashion, there is in general a remaining part of random heterogeneity, due to issues with data quality as well as inherent randomness in choice behaviour. To this extent, it is often necessary to explain at least part of the heterogeneity with the help of random coecients models. Two main approaches are possible, using continuous or discrete mixture models, where the former is used far more frequently than the latter. Continuous mixture models, such as MMNL model, use integration of the closed form discrete choice probabilities over the assumed distribution of the random taste coecients. A detailed presentation of these model structures is given by Train (2003). Aside from heightened computational cost, due to the reliance on simulation approaches (cf. Hess, Train & Polak 2006), the biggest issue that needs to be addressed with the use of continuous mixture models is the choice of mixture distribution. While the discussion here centers on the case of T , similar issues arise for C , or indeed also when estimating the VTTS directly (cf. Section 2.2.3). These distributional assumptions play a crucial role, and have a signicant eect on model interpretation. The issue of the choice of distribution is discussed at length by Hess et al. (2005), such that only a brief summary is given here. Essentially, the problem lies in making a choice of distribution that is consistent with a priori expectations as to the true shape of the distribution, without however making too strong a shape assumption. The vast majority of applications rely exclusively on the use of the Normal distribution, potentially leading to severe issues in interpretation. As such, the Normal distribution is unbounded, giving positive probability to positive as well as negative coecient values, while, from a microeconomic perspective, T is required to be non-positive. In the presence of a negative mean value, with a long tail to the left, the estimation of the mean and standard deviation will, due to the symmetry assumption, assign a potentially signicant mass to the positive space of numbers. This issue does also arise with other symmetrical distribution such as the Logistic or the symmetrical Triangular, but does not exist with the use of distributions allowing for an asymmetrical shape. At this point, it should however also be noted that it is not appropriate to impose an a 6

priori constraint limiting the distribution to the negative space of numbers, as factors such as data issues or model imperfections could indeed mean that positive numbers are possible. With the use of the Normal distribution, it is impossible to know whether the presence of positive coecient estimates is in fact revealed by the model, or is simply a result of the distributional assumptions. As such, the use of exible distributions is most appropriate, allowing for asymmetries as well as the estimation of bounds on the distribution. Here, the risk of values with the wrong sign being caused by the shape of the distribution, as with the Normal, largely disappears, although problems may still occur in the case of a signicant mass at the endpoints, such as in the presence of individuals with zero VTTS. Finally, it should be noted that another possible approach comes with the use of empirical or non-parametric distributions. However, such approaches are only beginning to be exploited in the estimation of MMNL models, and dicult issues of implementation and estimation need to be faced (cf. Fosgerau 2005). An alternative to the use of continuous mixture models comes in the use of discrete mixture approaches. Here, the integration over the distribution (as in the MMNL model) is replaced by a weighted summation over a nite number of dierent support points for the distribution. For a recent discussion of this approach, and a detailed comparison between the results obtained with continuous and discrete mixture models, see Hess, Bierlaire & Polak (2006). Discrete mixture models are not aected by the issue of the shape of the distribution, allowing for symmetrical as well as asymmetrical distribution, as well as multi-modal distributions. Furthermore, they can potentially be exploited to allow for a mass at a specic value, such as a zero VTTS. Another advantage lies in the absence for a need to use simulation, where summation over a nite (and generally low) number of support points is used. However, the estimation of discrete mixture models is often not straightforward, especially when using multiple support points, where issues with clustering can arise. 2.2.3 WTP space versus preference space

The majority of studies looking at the estimation of the VTTS work in a way, such that the VTTS estimate is obtained on the basis of separate travel time and travel cost coecients. While this is a rather straightforward approach, issues arise in the case of random coecients models, and specically, in the presence of a randomly distributed travel cost coecient3 .
In principle similar issues apply to MNL models since VTTS is obtained as the ratio between two asymptotically normal coecients.
3

Here, the distribution of the VTTS can often not be obtained directly, and simulation needs to be used. Further complications arise in the case where both coecients follow a random distribution, where correlation between the coecients needs to be taken into account. One of the main issues in the simulation of the ratio arises in the case where the distribution allows for extreme values (and potentially positive values), leading to very large or very small values for the ratio, such that special care is required in the analysis of the results (cf. Hess et al. 2005). In this context, it may often be preferable to work in (WTP) space as opposed to preference space. Here we use the vocabulary introduced by Train & Weeks (2005). With this approach, the distribution of the VTTS is estimated directly from the data, as opposed to being based on the ratio of two separately estimated coecients. Specically, let: Ui = + T T + C C + . . . . This can easily be rewritten as: Ui = + C V T + C C + . . . ,
T is an estimate of the VTTS that is obtained directly from the where V = C data. This approach is no dierent from working in preference space when all coecients are non-random except for a possible inuence on condence intervals of the estimates. However, when working with random coecients, the approach has potential advantages, as issues with the simulation of the ratio do no longer arise. However, several questions need to be addressed, notably in terms of the potential impact of the choice of distribution for C on the results for V , where the correlation between the two parameters also needs to be taken into account.

Danish data

The empirical investigation of this paper will focus on the questions concerning WTP space. We investigate this on two dierent data sets, one collected in Denmark and the other collected in Chile. For both data sets estimation is done in Biogeme, see Bierlaire (2003).

3.1

Data

The rst data set is taken from the Danish 2004 VTTS study, DATIV. The sample used here consists of 477 car commuters. 8

Table 1: Models spec/model basic corr. ev. ev. corr. ev. II mnl yes no yes no yes mix yes yes no no no mixv yes yes no no no log yes yes yes yes yes logv yes yes yes yes yes loginvv yes yes yes yes yes sbv yes yes yes yes no mixlog yes yes yes yes yes mixlogv yes yes yes yes yes

Individuals in the sample each answered 9 binary SP choices with reference to a recent car trip. Each SP choice had 3 attributes: cost, free ow time, and congested time. The congested time had a xed ratio to the total time. Since this was explicit to the respondents the SP can be view as described by cost, time and congestion percentage, where the congestion percentage enter as explanatory variable describing the individual. For each individual 8 of the SP choices came from a balanced design pivoted around the reference trip values. The 9th choice had a dominated alternative i.e. one of the alternatives was faster and cheaper. The 477 respondents all chose the fast and cheap alternative which conrms that they understand the SP task. The dominated SP choice is not included in the present analysis hence the sample has 3816 observations. Besides the percentage of congestion the data included many background variables. From these we will only use income, xinc and travel time of reference trip, tref . As they are often seen to be the two most inuential explanatory variables.

3.2

Models

On the sample we investigated dierent specications with respect to mixing distribution, correlation, and explanatory variables in both preference space and VTTS space. Table 1 gives an overview of the models tested. The rst row of models are all basic models with either the specication in preference space Ui = T Ti + C Ci + i , or in VTTS space Ui = V T T S s Ti + s Ci + i , 9 (4) (3)

or in inverse VTTS space


1 Ui = s Ti + s V T T S Ci + i .

(5)

For the basic models the two distributions are assumed independent. The mix models have both coecients normally distributed. The three log models have both coecients follow a log normal distribution. In the Sbvtts model the scale is log normal and the VTTS follows an Sb distribution. The nal model has a normal VTTS coecient and a log normal scale. In the corresponding models in in the second row the correlation between the two distributions is estimated. Both of the above model types are enlarged to include explanatory variables income and tref . This is done using the multiplicative specication from equation 2 in the time coecient for the mnl and preference space models. xinc inc Ti + C Ci . Ui = T xinc In the vtts space models the explanatory variables multiply the VTTS coefcient. The models with two normal distributions are not elaborated since these models have sign problems that will be discussed in the result section. The last row of models change the parameterizations by explanatory variables. Here the coecient not parameterized above is parameterized instead.

3.3

Estimation and results

All the result on models without explanatory variables are reported in Table 2a-c. As a base model we estimated an MNL model. It was tested against an MNL including serial correlation for each individual. The serial correlation was insignicant which seems reasonable for an unlabeled route choice. It is seen that the alternative specic constant is signicant. The signs of all coecients in the MNL are as expected and all coecients are signicant. First we compare the models without explanatory variables. The rst models in preference space either have two normal distributions, two log normal distributions or one of each. They all clearly outperform the MNL model with all the random coecients having signicant variance and the model with two log normals having the highest log likelihood value. The signs on the normal distributions in the 3 specications have the expected negative 10

sign. For the log normal distributions a shift was tested but in all cases it was insignicant and left out to keep the VTTS distribution as simple as possible. The three models above were next estimated allowing for correlation between the coecients. For all models the correlation turned out signicant and positive. The improvements in log likelihood for the models with one or two log normals were very large. Corresponding to these models we estimated 3 specications in wtp space and a fourth with log normal scale and wtp following an Sb distribution. For the specication with normal distributions the signs were as expected. For the log normal distributions and the Sb distribution shifts were found to be insignicant and hence left out of the nal models. The model with Sb distribution had problems converging and only did so when the parameter 0 was xed based on prior estimation. The models were then estimated allowing for correlation. For all specications the correlations were signicant. The correlation is found as the product of tc and cost . Hence for all models in preference space it is positive. For the models in WTP space the results are mixed with a positive correlation for the model with two normals while for the remaining it was negative. Last we also estimated a model in inverse wtp space for the specication with two log normals. This is reasonable only because the inverse for a log normal distribution is log normal. This specication actually outperforms the other specication based on log likelihood. The model was tested with a shift for the inverse VTTS distribution, which turned out insignicantly. When allowing for correlation the correlation turned out insignicant, but the shift became signicant. Next we discuss models with explanatory variables on income and reference travel time. The results can be seen in Table 3a-c. In all models the signs are the same that is VTTS rises with income and travel time. Furthermore the income eect is best modelled with a dummy making VTTS higher for low or unknown income than the general income trend. An argument for this is that unknown income can mean any income group. For the low income group some of the people might not be aected by the low income because of living in a couple. Once again all the mixed models clearly outperform the MNL 11

with the inverse vtts space specication obtaining the highest log likelihood. The last models test an opposite specication of the explanatory variables. These models highlights the problem of heteroscedasticity mixing with the eect of explanatory variables.

3.4

VTTS calculation

The parameter estimates are asymptotically normal. Based on this we use the approach discussed in Hensher & Greene (2003) to simulate the distribution of VTTS. From the simulated distribution it is then possible to calculate the mean as well as quantiles of interest. For each of the models we calculate the mean, the standard deviation and the 95% quantile. The mean values range from 5 to 19 $ pr. hour if we neglect the strange behavior of model 16. In 3 cases the mean is the ratio of two normal distributions and we see that only in the case of the MNL model this does not cause problems. Two immediate observations are that log normal specication gives rise to higher VTTS and that estimation in WTP space lowers the VTTS. In preference space correlation is seen to lower VTTS and in WTP space the VTTS rises in the models allowing for correlation. Since the true VTTS is unknown it is dicult to assess which distribution/space gives rise to the best mean VTTS but it seems reasonable to expect that the true mean has the preference space estimate as upper bound and the WTP space estimates as lower bound. One useful reference could be that the mean net hourly wage in the full sample is around 18 $/hr.

4
4.1

Chilean data
Data

For the second application we use the SP data collected by Caussade et al. (2005) on car route choices. Here we use 2166 observations out of the 8020 in the full data set. These are the observations where individuals make route choices based on ve attributes: Free ow time (FFT), Moderate congestion time (MCT), Heavy congestion time (HCT), Trip time variability (TTV), and Travel cost (TC). One of the alternatives was an original RP choice. The purpose of the trips is mainly leisure. The relevancy of varying TC comes from the fact that car travelers pay tolls when driving on highways 12

in Chile. Even though it will not be of importance in this study it is worthwhile mentioning that the sample is far from being a random sample of the Chilean population. The reason why this is not important is that we are not interested in the level of WTP but in the way it changes depending on estimation in preference and WTP space.

4.2

Estimation and results

To have a base model we rst estimated an MNL model. The preferred model was estimated taking serial correlation into account. The parameters have the expected signs and are all signicant. The derived WTPs are reasonable in the sense that WTP rises with congestion. The MMNL models are all estimated with normal distributions only or log normals only. The rst type of model has uncorrelated coecients in preference space. These models are what is most common in research and application. For both distributions we see that they outperform the MNL model and that all the mixing parameters are signicant except for the mixing of TTV in the log normal case. The signs in the normal case are negative as expected. In the log normal specication we have entered shifts to test the sign constraint implied by a log normal coecient. Here we see that the shift is only signicant for FFT. This should not be seen as an implication of the existence of negative WTP merely as an indication that the log normal is an approximation to the true distribution and that it approximates it better over the whole range of WTP by being shifted to the left. Both of the above specications were then estimated allowing for correlation. For both distributions the model with correlation outperform the model without. For the normal specication the signs and magnitudes remain the same, but all variances become less signicant which seem reasonable. The correlations are all signicant except for FFT-MCT and MCT-TC. All the negative correlations contain HCT. The fact that most correlations are positive contrary to the intuition that individuals with lower marginal utility of income have higher marginal utility of time supports that the data has a random scale since the scale would induces positive correlation at least for the log normal specication.

13

4.3

WTP calculations

Here we calculate the mean, standard deviation and 95% quantile as described in section 3.4. Again we observe the problems that arise when the VTTS is the ratio between two normal distributions. The values for log normal models are very high. This could be due to poor identication of the mixing distribution. For both distributions the values make more sense for the WTP models. One the other hand it is a problem that MCT has a lower value than FFT for the models in WTP space.

Summary and conclusions

This paper gives a review of how to incorporate taste heterogeneity in willingness to pay indicators. Furthermore it investigates on real data sets the dierence between the two approaches termed modelling in preference space and WTP space. The conclusions of the paper is that there is non of the approaches that seems to dominate when it comes to model t but that not including correlation between variables seem like a bigger problem for models in preference space. This underlines that it is very probable that most of the correlation comes from a random scale when individuals evaluate SP choices. When it comes to the inclusion of explanatory variables the empirical investigation of the paper gives evidence to the fact that they should be incorporate as in WTP space models to avoid mixing eects of scale. Based on the review and the empirical investigation of the paper three future directions of research that deserve attention are: non parametric methods to aid in the guessing of appropriate distributions, semi parametric methods and diagnostic tests to assess whether the chosen distributions are correct, and nally the question of which explanatory variables to use on one hand to avoid the problem of omitted variables and on the other to avoid endogenous explantory variables.

Acknowledgment

This work was initiated during the rst workshop on Applications of Discrete Choice Models organised at EPFL in Lausanne, Switzerland, September 2005.

14

References
Algers, S., Bergstroem, P., Dahlberg, M. & Lindqvist Dillen, J. (1998), Mixed Logit Estimation of the Value of Travel Time, Working Paper, Department of Economics, Uppsala University, Uppsala, Sweden. Bates, J. (1987), Measuring travel time values with a discrete choice model: a note, The Economic Journal 97, 493498. Becker, S. G. (1965), A theory of the allocation of time, Economic Journal 75(299), 493517. Bierlaire, M. (2003), BIOGEME: a free package for the estimation of discrete choice models, Proceedings of the 3rd Swiss Transport Research Conference, Monte Verit` a, Ascona, Switzerland. Caussade, S., Ort uzar, J. de D., Rizzi, L. I. & Hensher, D. (2005), Assessing the inuence of design dimensions on stated choice experiment estimates, Transportation Research Part B: Methodological 39(7), 621640. De Serpa, A. C. (1971), A theory of the economics of time, The Economic Journal 81(324), 828846. Evans, A. (1972), On the theory of the valuation and allocation of time, Scottish Journal of Political Economy, February pp. 117. Fosgerau, M. (2005), Investigating the distribution of the value of travel time savings, Transportation Research Part B: Methodological, forthcoming. Hensher, D. (2001a), Measurement of the valuation of travel time savings, Journal of Transport Economics and Policy 35(1), 7198. Hensher, D. (2001b), The sensitivity of the valuation of travel time savings to the specication of unobserved eects, Transportation Research Part E: Logistics and Transportation Review 37(2-3), 129142. Hensher, D. (2001c), The valuation of commuter travel time savings for car drivers: Evaluating alternative model specications, Transportation 28, 101118. Hensher, D. & Greene, W. (2003), The Mixed Logit Model: The State of Practice, Transportation 30(2), 133176.

15

Hess, S., Bierlaire, M. & Polak, J. W. (2005), Estimation of value of traveltime savings using mixed logit models, Transportation Research Part A: Policy and Practice 39(2-3), 221236. Hess, S., Bierlaire, M. & Polak, J. W. (2006), A systematic comparison of continuous and discrete mixture models, paper presented at the 11th International Conference of Travel Behaviour Research, Kyoto, Japan. Hess, S., Train, K. & Polak, J. W. (2006), On the use of a Modied Latin Hypercube Sampling (MLHS) method in the estimation of a Mixed Logit model for vehicle choice, Transportation Research Part B: Methodological 40(2), 147163. Jara-Diaz, S. R. (2000), Allocation and valuation of travel time savings, in D. Hensher & K. Button, eds, Handbooks in Transport, Vol. 1: Transport Modelling, Pergamon Press, Oxford, pp. 303319. Jara-Diaz, S. R. & Guevara, C. A. (2003), Behind the subjective value of travel time savings - the perception of work, leisure, and travel from a joint mode choice activity model, Journal of Transport Economics and Policy 37, 2946. Lapparent, M. D. & de Palma, A. (2002), Non-linearities in the valuations of time estimates, Proceedings of the PTRC Annual Meeting, 2002, PTRC, London. Mackie, P. J., Jara-Diaz, S. R. & Fowkes, A. S. (2001), The value of travel time savings in evaluation, Transportation Research Part E: Logistics and Transportation Review 37(2-3), 91106. Mackie, P., Wardman, M., Fowkes, A. S., Whelan, G., Nellthorp, J. & Bates, J. (2003), Values of travel time savings in the UK, Report to Department of Transport, Institute for Transport studies, University of Leeds and John Bates Services, Leeds and Abingdon. Oort, O. (1969), The evaluation of travelling time, Journal of Transport Economics and Policy 3, 279286. Sillano, M. & Ort uzar, J. de D. (2004), Willingness-to-pay estimation from mixed logit models: some new evidence, Environment & Planning A 37(3), 525550. Train, K. (2003), Discrete Choice Methods with Simulation, Cambridge University Press, Cambridge, MA. 16

Train, K. & Weeks, M. (2005), Discrete Choice Models in Preference Space and Willingness-to-Pay Space, Springer, pp. 116 in R. Scarpa and Alberini(eds) Application of simulation methods in environmental and resource economics. Truong, P. & Hensher, D. (1985), Measurement of travel time savings and opportunity cost from a discrete choice model, The Economic Journal 95, 438451.

17

TABLE 2a Model: MNL N. Of draws N. of par.: N. of obs.: N.of ind.: Null LL: Final LL: Ll ratio test: Rho^2: Adj rho^2: Final grad.: Var-covar:

3 ML - norm 500 6 3816 477 -2645.05 -2189.04 912.013 0.1724 0.170132 0.369839 hessian Value t-test 0.172 3.962 -0.816 -3.892 -0.420 -9.956 -0.201 -8.654 0.302 9.979 -0.278 -10.428

4 3816 3816 -2645.05 -2494.47 301.151 0.056927 0.055415 0.015469 hessian

4 ML - norm (w. corr.) 500 7 3816 477 -2645.05 -2185.52 919.069 0.173734 0.171087 0.371625 hessian Value t-test 0.176 3.981 -0.855 -6.736 -0.442 -9.569 -0.219 -8.112 0.341 9.282 -0.314 -11.810 0.107 4.674

5 ML - norm (vtts sp.) 500 6 3816 477 -2645.05 -2182.27 925.569 0.174962 0.172694 0.659231 hessian Value t-test 0.174 3.945 2.312 4.466 -0.448 -9.681 0.556 8.353 0.314 8.688 -0.956 -12.083

6 ML - norm (vtts sp. w. corr ) 500 7 3816 477 -2645.05 -2180.72 928.669 0.175549 0.172902 4.84738 hessian Value t-test 0.175 3.965 2.375 4.764 -0.449 -9.957 0.706 10.514 0.306 9.183 -0.888 -13.054 0.345 6.384

Coefficients Value t-test 0.104 3.113 ASC_1 -0.265 -5.218 BETA_cng -0.103 -8.834 BETA_cost -0.086 -9.148 BETA_time SIGMA_cost SIGMA_time rho_time_cost BETA_0 BETA_1 At zero congestion E(VTTS) $/hr std. dev. 95% fractile std. dev. of mean

8.32 1.35 10.71 0.07

11.01 7019.90 44.82 405.29

1.80 1116.50 32.15 64.30

5.46 9.49 21.10 0.54

6.94 9.45 22.53 0.56

TABLE 2b 7 Model: ML - log N. Of draws N. of par.: N. of obs.: N.of ind.: Null LL: Final LL: Ll ratio test: Rho^2: Adj rho^2: Final grad.: Var-covar: 500 6 3816 477 -2645.05 -2147.16 995.785 0.188236 0.185967 0.014883 hessian

8 ML - log (w. corr.) 500 7 3816 477 -2645.05 -2093.61 1102.88 0.20848 0.205833 1.41786 hessian Value t-test 0.199 4.181 1.914 26.686 -0.704 -5.975 -1.166 -11.027 -2.102 -12.865 -0.846 -42.762 -1.302 -8.125

9 ML - log (vtts sp.) 500 6 3816 477 -2645.05 -2112.11 1065.88 0.201485 0.199217 1.60251 hessian Value t-test 0.197 4.159 2.897 318.386 -0.677 -4.645 -0.920 -148.379 2.513 11.055 -1.250 -139.866

10 ML - log (vtts sp. w. corr ) 500 7 3816 477 -2645.05 -2107.25 1075.61 0.203325 0.200678 2.36007 hessian Value t-test 0.198 4.169 2.600 31.998 -0.701 -4.968 -0.732 -24.852 2.482 10.199 -1.169 -249.803 -0.183 -26.847

11 ML - log (inv. vtts sp.) 500 6 3816 477 -2645.05 -2094.05 1102 0.208314 0.206045 0.914216 hessian Value 0.198 2.414 0.595 -1.274 1.205

12 ML - log (inv. vtts sp. w. corr ) 500 8 3816 477 -2645.05 -2093.53 1103.04 0.20851 0.205485 1.79354 hessian

Coefficients Value t-test 0.177 4.013 ASC_1 2.193 5.169 BETA_cng -0.981 -9.011 BETA_cost -1.629 -13.046 BETA_time -1.258 -13.006 SIGMA_cost SIGMA_time rho_time_co st BETA_0 BETA_1 At zero congestion E(VTTS) $/hr std. dev. 95% fractile std. dev. of mean 16.99 54.61 64.67 3.16 -0.884 -9.750

t-test Value t-test 4.146 0.198 4.143 32.491 2.283 56.227 31.026 0.448 10.398 -11.611 -1.225 -10.986 62.323 1.284 26.106 -1.721 0.130 0.109 -9.038 0.717 2.366

-1.769 -10.306

12.90 25.99 44.50 1.52

8.60 17.25 31.03 0.99

9.60 17.33 33.62 1.05

11.28 21.56 39.93 1.23

14.66 33.30 53.22 1.95

TABLE 2c 13 Model: ML - Sb (vtts sp.) 500 N. Of draws 6 N. of par.: 3816 N. of obs.: 477 N.of ind.: Null LL: -2645.05 Final LL: -2111.61 1066.88 Ll ratio test: Rho^2: 0.201675 Adj rho^2: 0.199406 3.66402 Final grad.: Var-covar: hessian Coefficients Value t-test 0.198 4.171 ASC_1 3.623 89.761 BETA_cng -0.640 -4.128 BETA_cost -2.923 -1760.553 BETA_time -2.477 -9.620 SIGMA_cost -1.545 -333.655 SIGMA_time rho_time_cost BETA_0 7.000 BETA_1 At zero congestion E(VTTS) $/hr 7.26 std. dev. 9.52 95% fractile 28.24 std. dev. of mean 0.54

14 ML - Sb (vtts sp. w. corr ) 500 7 3816 477 -2645.05 -2107.83 1074.43 0.203102 0.200456 6.25139 hessian Value 0.193 3.585 -0.687 -2.658 -2.430 -1.457 0.152 7.000 t-test 4.067 16.522 -4.979 -10.941 -10.292 -14.502 2.191

15 ML - mixlog 500 6 3816 477 -2645.05 -2157.56 974.984 0.184304 0.182035 0.0177412 hessian Value t-test 0.177 4.051 0.600 5.075 -1.033 -9.712 0.254 9.638 -1.248 -15.012 0.195 8.706

16 ML - mixlog (corr ) 500 7 3816 477 -2645.05 -2115.66 1058.78 0.200143 0.197497 0.933118 hessian Value t-test 0.196 4.239 0.583 4.142 -0.901 -8.567 0.433 11.486 -1.837 -21.832 0.171 7.197 -0.309 -11.509

17 ML - mixlog (vtts sp. ) 500 6 3816 477 -2645.05 -2158.87 972.361 0.183808 0.181539 0.266678 hessian Value t-test 0.185 4.024 2.590 14.784 -0.959 -8.282 0.580 24.515 -1.767 -9.193 -0.972 -20.126

18 ML - mixlog (vtts sp.w. corr ) 500 7 3816 477 -2645.05 -2156.37 977.363 0.184753 0.182107 5.37607 hessian Value 0.185 2.598 -0.953 0.680 -1.757 -0.924 0.108 t-test 4.019 7.110 -6.952 20.692 -9.103 -8.151 8.017

8.53 10.52 31.25 0.63

15.54 39.04 62.93 2.24

-17.49 384.03 39.05 22.51

5.70 9.61 21.56 0.55

6.69 9.25 21.90 0.56

TABLE 3a Model: MNL N. Of draws N. of est. Par.: N. of obs.: N.of ind.: Null LL: Final LL: Ll ratio test: Rho^2: Adj rho^2: Final grad.: Var-covar:

19

20 ML - log 500 9 3816 477 -2645.05 -2108.02 1074.06 0.203032 0.199629 0.715932 hessian Value t-test 0.179 4.039 1.834 5.275 -1.134 -11.553 0.263 1.972 0.700 3.988 -1.609 -13.710 0.723 9.490 -0.990 -14.389 -0.976 -16.401

7 3816 3816 -2645.05 -2399.95 490.201 0.092664 0.090017 0.014824 hessian

21 ML - log (w. corr.) 500 11 3816 477 -2645.05 -2071.01 1148.08 0.217025 0.212866 4.74089 hessian Value t-test 0.198 4.181 1.504 8.024 -0.846 -7.340 0.327 3.024 0.189 2.414 -1.048 -9.432 0.482 18.149 -1.871 -10.767 -0.765 -14.999 -1.073 -6.918 -0.031 -2.186

22 ML - log (vtts sp.) 500 9 3816 477 -2645.05 -2094.92 1100.26 0.207985 0.204582 5.14627 BHHH Value t-test 0.197 4.415 2.257 5.460 -0.667 -4.396 0.647 5.591 0.011 0.038 -0.736 -9.332 0.378 5.562 2.507 10.670 -1.196 -18.448

23 ML - log (vtts sp. w. corr ) 500 10 3816 477 -2645.05 -2094.85 1100.4 0.208012 0.204231 6.30058 hessian Value t-test 0.197 4.150 2.258 27.043 -0.663 -4.627 0.647 27.165 0.015 0.322 -0.727 -9.747 0.377 49.049 2.508 9.856 -1.195 -91.882 -0.005 -0.156

Coefficients Value t-test 0.114 3.338 ASC_1 -0.298 -6.364 BETA_cng -0.131 -13.023 BETA_cost 0.290 2.578 BETA_inc 0.373 2.038 BETA_nlinc -0.084 -8.591 BETA_time 0.586 10.018 BETA_tt SIGMA_cost SIGMA_time rho_time_cost BETA_0 BETA_1 At zero cong, log(inc), log(tt) E(VTTS) $/hr 6.35 std 0.9 95% fractile 7.88 std. dev. of mean 0.048

16.13 38.13 60.13 2.25

11.45 32.73 42.2 1.91

9.73 17.74 34.28 1.03

9.82 18.21 34.44 1.09

TABLE 3b 24 Model: ML - log (inv vtts sp.) 500 N. Of draws 9 N. of est. Par.: 3816 N. of obs.: 477 N.of ind.: Null LL: -2645.05 -2068.5 Final LL: 1153.1 Ll ratio test: Rho^2: 0.217973 Adj rho^2: 0.214571 Final grad.: 4.61841 Var-covar: hessian Coefficients Value t-test 0.199 4.172 ASC_1 1.805 18.803 BETA_cng 0.375 27.846 BETA_cost 0.268 20.415 BETA_inc 0.342 17.711 BETA_nlinc -1.203 -10.751 BETA_time 0.522 23.697 BETA_tt 1.078 41.411 SIGMA_cost -1.649 -9.531 SIGMA_time rho_time_cost BETA_0 BETA_1 At zero cong, log(inc), log(tt) E(VTTS) $/hr 12.15 std 18.77 95% fractile 40.32 std. dev. of 1.07

25 ML - log (inv vtts sp. w. corr ) 500 11 3816 477 -2645.05 -2066.89 1156.31 0.218581 0.214422 1.07703 hessian Value t-test 0.198 4.174 1.745 12.809 0.276 4.286 0.452 10.381 0.556 7.042 -1.262 -11.065 0.639 11.433 1.206 15.391 -1.638 -9.874 -0.267 -1.923 0.156 2.381

26 ML - sb (vtts sp. ) 400 10 3816 477 -2645.05 -2100.47 1089.17 0.205888 0.202107 6.29842 hessian Value t-test 0.199 4.190 2.706 68.057 -0.604 -4.152 0.160 22.469 0.068 0.575 -2.683 -152.580 0.254 37.938 -2.478 -10.236 -1.422 -125.125 0.023 6.930 5.109

27 ML - sb (vtts sp.w corr ) 500 10 3816 477 -2645.05 -2091.63 1106.84 0.209228 0.205448 5.96373 hessian Value t-test 0.200 4.196 2.752 75.404 -0.612 -4.251 0.404 55.838 0.070 6.446 -2.516 -69.494 0.346 60.048 -2.412 -11.035 -1.349 -220.180 0.082 6.077 7.000

28 ML - mixlog 500 9 3816 477 -2645.05 -2154.75 980.596 0.185364 0.181962 0.012359 hessian Value t-test 0.177 4.046 0.650 4.402 -1.028 -9.587 0.218 1.723 -0.215 -1.336 0.251 8.669 -0.102 -1.541 -1.291 -13.943 0.179 8.211

10.12 13.76 30.64 0.84

8.15 9.43 28.78 0.57

8.7 9.7 29.66 0.59

16.15 41.96 64.79 2.42

TABLE 3c Model: N. Of draws (MLHS) N. of est. Par.: N. of obs.: N.of ind.: Null LL: Final LL: Ll ratio test: Rho^2: Adj rho^2: Final grad.: Var-covar: Coefficients ASC_1 BETA_cng BETA_cost BETA_inc BETA_nlinc BETA_time BETA_tt SIGMA_cost SIGMA_time rho_time_cost BETA_0 BETA_1 At zero congestion, log(inc), log(tt) E(VTTS) $/hr std 95% fractile std. dev. of mean

29 ML - mixlog (corr. ) 500 10 3816 477 -2645.05 -2110.18 1069.74 0.202215 0.198435 1.62239 hessian Value t-test 0.196 4.231 0.531 -0.918 0.243 -0.050 0.447 0.013 -1.881 0.163 -0.343 4.013 -8.932 4.001 -0.526 11.251 0.338 -18.716 6.806 -8.218

30 ML - mixlog (vtts sp) 500 9 3816 477 -2645.05 -2138.4 1013.29 0.191546 0.188143 1.02916 hessian Value t-test 0.186 4.056 1.967 -0.931 0.241 0.107 0.693 0.426 -1.724 -0.887 28.950 -7.653 2.488 2.493 18.100 6.962 -7.616 -27.011

31 ML - mixlog (vtts sp corr. ) 500 10 3816 477 -2645.05 -2138.17 1013.76 0.191633 0.187852 12.1373 hessian Value 0.187 1.967 -0.931 0.238 0.109 0.712 0.427 -1.725 -0.890 0.010 t-test 4.050 22.749 -7.706 2.938 0.998 12.102 11.476 -8.128 -63.159 0.466

-28.98 544.3 37.06 31.79

6.81 8.77 21.28 0.5

7 8.8 21.51 0.53

TABLE 4 Model: MNL N. Of draws N. of est. Par.: N. of obs.: N.of ind.: Null LL: Final LL: Ll ratio test: Rho^2: Adj rho^2: Final grad.: Var-covar:

32 log

33 logvtts

34

35 loginvvtts 500 9 3816 477 -2645.05 -2076.85 1136.39 0.214815 0.211412 0.28903 hessian

36 mixlog 500 9 3816 477 -2645.05 -2124.24 1041.61 0.196898 0.193496 0.011302 hessian

37 mixlogvtts 500 9 3816 477 -2645.05 -2135.14 1019.82 0.192779 0.189376 2.22981 BHHH

7 3816 3816 -2645.05 -2458.89 372.315 0.0703795 0.0677331 0.0161071 hessian

500 9 3816 477 -2645.05 -2144.07 1001.95 0.189401 0.185998 0.0594964 hessian

500 9 3816 477 -2645.05 -2087.55 1115 0.210771 0.207368 1.82554 BHHH

Coefficients Value t-test Value t-test Value t-test Value t-test Value t-test Value t-test 0.107 3.197 0.177 3.989 0.190 4.270 0.194 4.106 0.180 4.100 0.181 4.166 ASC_1 -0.277 -5.631 2.157 3.814 2.521 5.082 2.375 7.941 0.484 3.744 2.481 6.702 BETA_cng -0.109 -9.684 -0.985 -9.133 -0.904 -6.689 0.573 6.734 -1.134 -11.169 -1.126 -10.749 BETA_cost 0.346 2.405 0.356 2.187 0.161 0.470 0.394 1.478 -0.345 -1.812 0.178 0.663 BETA_inc 0.133 0.863 -0.296 -0.485 -1.780 -3.465 -1.402 -2.951 -0.672 -2.607 -1.186 -3.133 BETA_nlinc -0.069 -6.826 -1.603 -12.457 -0.935 -10.550 -1.371 -11.109 0.276 9.065 0.575 8.965 BETA_time 0.356 4.903 -0.077 -0.849 -1.062 -6.487 -0.646 -5.153 -0.642 -6.964 -0.818 -6.543 BETA_tt -1.287 -12.641 1.918 10.047 1.180 21.572 -1.068 -13.806 -1.345 -9.288 SIGMA_cost -0.854 -10.135 -1.232 -15.831 -1.420 -7.372 0.196 6.884 -0.932 -19.478 SIGMA_time At zero cong, log(inc), log(tt) E(VTTS) $/hr 6.30 16.13 9.73 12.15 16.15 6.81 std 95% fractile std. dev. of mean 1.15 8.29 0.06 38.13 60.13 2.25 17.74 34.28 1.03 18.77 40.32 1.07 41.96 64.79 2.42 8.77 21.28 0.50

TABLE 5 MNL Number of observations Number of individuals Number of coefficients Null log-likelihood Init log-likelihood Final log-likelihood Likelihood ratio test Rho-square Adjusted rho-square Final gradient norm Variance-covariance 2152 103 7 -2392.9 -2392.9 -2050.3 685.2 0.143 0.140 0.973 Hessian MMNL PS Normal (no correlation) 2152 103 11 -2392.9 -1656.5 -1605.9 1574.1 0.329 0.324 2.734 Hessian MMNL PS Normal (with correlation) 2152 103 20 -2392.9 -1634.1 -1579.6 1626.6 0.340 0.332 0.304 Hessian MMNL WTP Normal (no correlation) 2152 103 11 -2392.9 -1609.5 -1601.9 1582.1 0.331 0.326 0.006 Hessian MMNL WTP Normal (with correlation) 2152 103 20 -2392.9 -1623.7 -1588.2 1609.5 0.336 0.328 0.000 Hessian MMNL PS LogN (no correlation) 2152 103 12 -2392.9 -1558.2 -1540.6 1704.6 0.356 0.351 0.001 Hessian MMNL PS LogN (with correlation) 2152 103 21 -2392.9 -1548.4 -1515.0 1755.9 0.367 0.358 0.002 Hessian MMNL WTP LogN (no correlation) 2152 103 12 -2392.9 -1581.5 -1581.5 1622.9 0.339 0.334 0.210 Hessian

Coefficients Cost sd_cost shift_cost Freeflow time sd_freeflow shift_freeflow_time Heavy congestion time sd_heavy shift_heavy_congested_time Moderate congestion time sd_moderate shift_moderate_congested_time Trip time variability sd_variability shift_trip_time_variability RP Sigma Lambda Cost

value -0.00009 ---------0.00790 ---------0.03142 ---------0.01335 ---------0.00719 --------0.20769 0.960 ------

t-ratio -8.8 ---------8.4 ---------8.3 ---------5.6 ---------4.7 --------1.7 11.4 ------

value -0.00035 0.00103 -----0.04878 0.08135 -----0.11979 0.11860 -----0.06136 0.09083 -----0.02765 0.05529 -----0.00897 -----------

t-ratio -12.5 12.5 -----14.5 13.3 -----13.3 11.2 -----10.6 12.6 -----7.3 11.1 -----0.1 -----------

value -0.00076 0.00084 -----0.03707 0.08684 -----0.15275 0.09463 -----0.06525 0.08533 -----0.01871 0.06445 -----0.08005 -----------

t-ratio -12.1 5.5 -----10.8 7.0 -----13.3 4.1 -----8.0 5.0 -----3.3 4.3 -----0.5 -----------

value t-ratio -0.00078 -9.7 0.00057 9.8 --------116.070 11.3 119.427 10.9 --------323.044 15.0 286.690 12.9 --------111.647 7.0 111.169 9.4 --------52.985 7.5 76.643 10.3 ---------0.12950 -0.9 ---------------------

value t-ratio -0.00088 -8.6 0.00082 4.0 --------105.681 12.5 137.842 6.0 --------256.704 11.2 289.752 6.9 --------71.982 6.4 117.491 4.4 --------39.443 5.8 89.907 4.0 ---------0.09402 -0.6 ---------------------

value t-ratio -8.03447 -53.1 1.96087 13.9 ---------2.33708 -13.7 0.74949 8.2 -0.04995 -4.0 -2.16965 -17.0 1.14106 16.5 ---------3.82707 -17.3 1.89032 10.6 ---------4.54176 -13.1 1.58425 6.6 ---------0.05926 -0.4 ---------------------

value t-ratio -7.82391 -51.1 3.08879 15.4 ---------2.08091 -14.7 1.09817 5.9 -0.04152 -4.3 -2.00350 -18.4 1.42087 13.5 ---------3.04196 -15.2 2.03318 11.7 ---------4.56039 -13.4 2.35866 11.4 ---------0.06020 -0.4 ---------------------

value t-ratio -6.87971 -51.2 1.82155 16.0 --------4.24812 83.1 1.69202 30.0 --------5.04224 67.0 1.36068 17.1 --------4.32020 62.1 1.28191 15.2 --------3.87372 13.7 0.99350 6.1 -48.10172 -4.2 0.00164 0.0 ---------------------

Value of Time [US$/hr] (*) value VOT_Freeflow $ 9.7 VOT_Moderate $ 16.4 VOT_Heavy $ 38.6 VOT_Uncertainty $ 8.8 (*) 1US$ = $550 (**) standard deviation of simulated mean Covariances Freeflow - Cost Freeflow - Heavy Freeflow - Moderate Freeflow - Variability Heavy - Cost Heavy - Variability Moderate - Cost Moderate - Heavy Moderate - Variability value ----------------------------------------------

std** 0.09 0.35 0.2 0.12

value -$ 74.4 -$ 84.4 -$ 157.5 -$ 90.8

std** 1995 2374 5021 10595

value -----------------

std** -----------------

value $ 12.7 $ 12.2 $ 35.2 $ 5.8

std** 0.8 0.7 1.9 0.5

value -----------------

std** -----------------

value $ 204.4 $ 316.9 $ 553.6 $ 109.0

std** 146.2 331.5 444.5 124.2

value -----------------

std** -----------------

value $ 31.9 $ 18.8 $ 42.8 $ 3.9

std** 6.8 2.3 5.7 0.8

t-ratio ----------------------------------------------

value ----------------------------------------------

t-ratio ----------------------------------------------

value 0.00001 -0.00238 0.00069 0.00135 -0.00005 -0.00094 0.00000 0.00311 0.00292

t-ratio 2.8 -2.1 0.6 2.3 -4.8 -2.1 0.1 3.9 4.7

value ----------------------------------------------

t-ratio ----------------------------------------------

value t-ratio -0.03733 -7.3 8223.965 2.4 4886.024 2.7 205.299 0.3 -0.01502 -1.8 7632.448 3.7 -0.03903 -4.2 6538.102 3.5 1888.070 2.5

value ----------------------------------------------

t-ratio ----------------------------------------------

value t-ratio 1.2099 10.0 0.4496 3.9 1.2944 8.1 0.3320 2.6 -2.1383 -9.6 2.2688 10.2 1.6722 6.5 2.0395 12.3 1.6416 5.8

value ----------------------------------------------

t-ratio ----------------------------------------------

Вам также может понравиться