What’s New in NLOGIT Version 4.0?
All the new features described for LIMDEP 9.0 are in NLOGIT 4.0. In addition, there are many new features in Version 4.0 of NLOGIT. We have added several enhancements to give you greater flexibility in analyzing different types of data. Many of the features of NLOGIT, existing and new, are designed to let you go beyond just computing coefficients, to analyzing and using your model. We have also added two new models, a generalization of the nested logit model and an equivalent to the random effects model in the panel data framework. The mixed logit (random parameters logit) model is currently the most general and flexible model available for analyzing individual choice. NLOGIT 3.0’s adaptation of the mixed logit model was already the most complete available in any program. In NLOGIT 4.0, we have added several new specifications. Altogether, we have added dozens of features in NLOGIT 4.0, some clearly visible ones such as the new models and some ‘behind the scenes’ that will smooth the operation and help to stabilize the estimation programs. The following will summarize the important new developments.
New Multinomial Choice Models
We have added two major model classes to the package.
Generalized Nested Logit Model
The nested logit model is one of the most popular extensions of the multinomial logit model. One of the shortcomings of the model is the strict requirement that the tree structure allocate each alternative to exactly one branch in the tree. The ‘generalized nested logit model’ allows alternatives to appear, probabilistically, in several branches at once.
Error Components Logit Model
The multinomial logit model has served as the basic platform for discrete choice modeling for decades. Among its restrictive features is its inability to capture individual choice specific variation due to unobserved factors. The error components logit model has emerged as a form that allows this. In a repeated choice (panel data) situation, this will play the role of a type of random effects model.
Model Extensions
Heterogeneity in Variances
Arguably, heterogeneity in the variances of utility functions is as important (more so, according to some) than heterogeneity in the levels. We have added specifications to accommodate heteroscedasticity to the mixed logit model, the covariance heterogeneity (nested logit) model and the heteroscedastic extreme value model.
Multinomial Logit (GME)
Generalized maximum entropy estimation provides a method of calibrating parameters that is, in many cases, more closely aligned with the ‘fit’ of the model to the data than is maximum likelihood. We have added a GME estimator to both the multinomial logit model and the conditional logit model – that is, all forms of the basic MNL model.
Mixed Logit Models
As noted, the mixed logit model represents the frontier in multinomial choice modeling. We have added many new features to NLOGIT’s implementation of this model. A partial list includes:
- Numerous specifications have been added to build realistic, plausible parameter distributions. For example, the Weibull and triangular distributions provide useful alternatives to the lognormal for imposing sign constraints on coefficients. We have also built optional specifications into the definitions of the random parameters to allow variation in the characteristics that appear in the means and standard deviations of different distributions.
- Models for heterogeneity in the variances (heteroscedasticity) of the random coefficients are now provided.
- The error components logit model described above may be layered on top of the mixed logit model.
- In addition to estimation of individual specific expectations of the random parameters, you may now also compute individual specific measures of ‘willingness to pay’ measures, which are computed as ratios of coefficients.
- The mixed logit model may now be fit to ranks data.
Reported Results
The simulator can now be used to compute and report elasticities. (Since the program was already designed to compute the changes in probabilities induced by discrete changes in attributes, arc elasticities are a natural extension of the results.)
Partial effects and elasticities computed with the ; Effects:attribute(alternatives) specification, such as
+---------------------------------------------------+ | Elasticity averaged over observations.| | Attribute is GC in choice AIR | | Effects on probabilities of all choices in model: | | * = Direct Elasticity effect of the attribute. | | Mean St.Dev | | * Choice=AIR -.7156 .4021 | | Choice=TRAIN .2678 .6005 | | Choice=BUS .4876 1.0049 | | Choice=CAR .6092 .9216 | +---------------------------------------------------+
are computed by averaging the person specific values. The value reported is the mean of the sample. NLOGIT now reports the standard deviation of those observations, which will give you an idea of the amount of variation of this result in your sample. The table above (computed for a heteroscedastic extreme value model) gives an example.
The cluster estimator for clustered data sets that has been built into the other estimators in LIMDEP has now been added to the models in NLOGIT. The cluster estimator is a correction to the standard errors of an estimator for assumed panel data effects.
Data Setup and Types
Three major enhancements to the setup of your data for multinomial discrete choice modeling are available in NLOGIT 4.0. Primary data are often erroneously coded or arranged inappropriately for analysis with NLOGIT. You may now request the program to inspect the data, observation by observation, for about 20 different problems that would impede estimation. Some of these are done automatically, while others are available when you request this procedure. For experimental work and model development, NLOGIT will simulate the choice data based on the type 1 extreme value distribution when you supply the utilities. Finally, in some situations, survey responses indicate that individuals have specifically ignored certain attributes in making choices. NLOGIT can automatically accommodate such data during model estimation. (Simply setting the attributes to zero is not the right approach – consider zeroing a price, for example. The model, itself, must be modified.)
The NLOGIT 4.0 Reference Guide
Users of NLOGIT 3.0 should see immediately, we have completely reworked the NLOGIT manual. There are two major extensions here. First, we have included in this manual documentation of the foundational discrete choice models described in detail in the LIMDEP Econometric Modeling Guide, including binary choice and ordered choice models. These are presented here to develop a complete picture of the use of NLOGIT to analyze data on discrete choices. Second, we have added extensive explanatory text and dozens of new examples, with applications, for every technique and model presented. Altogether, this documentation for NLOGIT 4.0 is nearly three times as long as the manual for NLOGIT 3.0.