LIMDEP 8.0 is a major expansion of previous versions for cross section, panel and time series data analysis. We have added dozens of new estimators and major new innovations in the tools for analyzing panel data - many of these appear in no other software and extend existing results in the econometrics literature. The array of programming and analysis tools has been extended, as well. We have also completely revised the documentation. The new manual, with nearly 2000 pages, contains full reference guides for the program, background econometrics, and sample applications.

With this release, we have greatly expanded NLOGIT, our companion program for discrete choice modeling and FIML estimation. NLOGIT 3.0 includes all of LIMDEP 8.0 plus the estimators for discrete choice models. All the new features described below for LIMDEP 8.0 are in NLOGIT 3.0.

The complete list of new features is described below.

Panel Data Models and Tools

The panel data suite of programs now includes the dynamic linear models, GMM (Arellano/Bond) estimators, fixed and random effects models, random parameters models, latent class models, specialized estimators for sample selection models, stochastic frontier models and much more. LIMDEP's new procedures for analyzing panel data include:

• Dynamic Panel Data Regression Models and the Arellano and Bond Estimator
  The Hausman and Taylor estimator which forms the platform for the Arellano and Bond estimator is now supported for the general linear model. For dynamic models, several formulations of the Arellano/Bond estimator are provided. This is an effects model that contains lagged dependent variables. Instruments may include the full available matrix or different subsets. The stochastic specification may take several forms.
  
• Broad Model Frameworks Supported for Four Panel Data Classes
  Most contemporary programs contain a few panel data estimators, such as fixed and random effects for the linear model, random effects for the probit model, and conditional fixed effects for the Poisson and binary logit model. LIMDEP now contains a range of panel data treatments for nearly all the models in the package, including binary choice, censored data, truncation, count data, survival models, and so on. Most of these do not appear in any other software, and some (the random parameters and latent class models) will be new to the literature.
  The following lists some of the model frameworks for which all of the general classes of estimators listed below are supported:
  • Linear regression
  • Stochastic frontier
  • Binary Choice: probit, logit, complementary log log, Gompertz
  • Ordered probability: probit, logit, complementary log log, Gompertz
  • Loglinear models: exponential, gamma, Weibull, inverse Gaussian
  • Survival: Weibull, exponential, lognormal, loglogistic
  • Bivariate probit: bivariate, partial observability, bivariate probit with sample selection
  • Count data: Poisson, negative binomial, zero inflated Poisson, negative binomial
  • Limited dependent variables: tobit, truncation, Heckman sample selection, grouped data
    
• LIMDEP support fours classes of panel data models:
  
  • Fixed effects: The fixed effects (dummy variable) estimator is generally available in other packages for the linear model (where it is fit by within groups least squares) and the Poisson and binary logit model. The practical reason for the very limited availability of this specification is that only these models allow one to condition out the dummy variables. LIMDEP now allows you to fit true (unconditional) fixed effects models by maximum likelihood - we can estimate up to 20,000 dummy variable coefficients. This means that predictions, marginal effects, and other model extensions can be computed, unlike for the conditional cases. This framework is available for all the models in LIMDEP, including tobit, truncation, Poisson, survival, and dozens of others.
  • Random effects: The random effects model can be implemented more easily than the fixed effects estimator, yet it is still usually available only for a few cases. LIMDEP's new panel data estimators include random effects, fit by quadrature or by maximum simulated likelihood, for all the models in the program, binary choice, censoring, truncation, counts, survival, and so on. We have also modified the random effects estimator for the linear model to allow heteroscedasticity at one or two levels and to simplify the calculation of variance components so that the problem of negative estimated variances is avoided.
  • Random parameters: The random parameters model is called variously 'the hierarchical model,' 'multilevel model,' and 'mixed model.' This can be found in other programs for the linear regression model, probit, logit, and Poisson regression. LIMDEP's new implementation of this framework extends it beyond these to all the binary choice, count data, censoring, truncation, survival, and many other models. This implementation is unique. The random parameters class formulates the included models with the general panel data specification b(i) = b + Dz(i) + Gv(i) where i indexes groups, b(i) is the model parameters, generally including both slopes and ancillary parameters such as scale parameters, z(i) is a vector of covariates, v(i) is a vector of random effects which may have a variety of distributions and be constrained in various ways, or may evolve in an AR(1) fashion, and b, D, and G are structural parameters to be estimated with autocorrelation parameters in the AR(1) case.
  • Latent classes: The latent class model has been applied in the literature and in other software virtually exclusively to the Poisson regression. LIMDEP's implementation allows you to apply this specification to all the models supported by the package, including binary choice, censoring, truncation, the loglinear models, survival models, and so on. The latent class model allows for individual heterogeneity in a panel to be analyzed through discrete variation. The estimator provides estimates of the structural parameters within the classes, the class probabilities, and for each individual, posterior estimates of their own class probabilities and a conditional estimate of the model parameters.
    
• Some Specific New Models for Panel Data
  
  • Sample selection: We continue our research on the sample selection model. Several different fixed and random effects and random parameter variants are now supported for Heckman's sample selection model.
  • Stochastic frontiers: LIMDEP contains a major implementation of the stochastic frontier model. This includes not only the full range of cross section variants, but also a variety of random and fixed effects formulations.
• Tools and Refinements
  Since so many of the estimators in the program are now focused on panel data, we have provided a set of tools to work with the data as well.
  • Panel specification: A common format is used throughout the program to specify the panel. This is the count variable which provides the group sizes. A stratification variable, which gives, instead, a group identifier (this need not be a consecutive set of integers) may be used instead. A one line instruction is provided for converting a stratification variable to the group count variable.
  • Automatic creation of the group sizes variable: Use of the linear regression model for panel data automatically computes a stratification variable based on consecutive integers and a group count variable, as well as scalar averages of the group sizes and reciprocals of the group sizes. For more general uses in other models, a single command is provided that will create the group sizes variable for a given sample using your stratification or group identification variable.
  • Reordering a balanced panel: A panel is generally organized so that each block of observations is the set of T(i) observations for individual i. If the data are arranged so that each block is the N(t) observations for period t, then a command is provided to reorder the data to take the more conventional format.
  • Matrix functions: Several new matrix functions are provided specifically for manipulating panel data. The Gsum function creates within group weighted or unweighted sums of a set of variables; Gxbr computes a matrix of group means of a set of variables; Gsiz returns a vector of group sizes; the Gmmw function returns a matrix of weighted moments to be used as the weighting matrix for GMM estimation. One of the more useful functions is Gsum which allows weighted sums where the weights are within group specific. An example is provided by the test for random parameters suggested in Train and McFadden (Journal of Applied Econometrics, October, 2000.) The Gsum function can be combined with the CLOGIT command to carry out this test with a total of only four or five short commands.
  • Bootstrapping panels: The DRAW command will now allow you to obtain bootstrap samples of the groups in a panel. That is, the sample is not individual observations, it is a random sample of groups of observations. Thus, you can bootstrap any of the panel data estimators.
  • Merging data sets: Panel data sets often contain time invariant variables. If these are maintained in a data file that is separate from the time varying data, the data input command may now be specified to merge these two data sets.
  • Increased limits on data set size: In previous versions, the number of groups in the linear panel data model was limited to 20,000. There is no effective limit now on the number of groups. This applies to linear as well as nonlinear models.
  • Double precision: In earlier versions of LIMDEP, the group means for the linear model were stored in single precision. This entailed some loss of accuracy in some problems. They are now stored in double precision, so the level of accuracy is maintained with that of the rest of the calculations.
  • Weighted panels: Panel data estimators may use weights of any sort.
  • Missing data handling: The panel estimators will automatically bypass missing observations and reduce group sizes accordingly.
    

New Estimators and Extensions of Current Estimators
In addition to the array of panel data estimators, we have developed estimators for over 25 new models, including loglinear models, stochastic frontiers, models for count data, binary choice models, time series models, extensions of the linear regression model and more.

• Loglinear Models
  • Beta regression: This is for a random variable whose range is limited to an interval. The two parameters of the distribution are specified to depend on a set of exogenous variables.
  • Loglinear regression models: Exponential, Weibull, gamma, inverse Gauss: These four specifications apply to a nonnegative dependent variable. The first two are special cases of the gamma, whereas the inverse Gauss is not. In all cases, the location, scale and shape parameters of the models are specified as functions of exogenous variables. All four models are also specifically implemented in the package of survival routines.
• Sample Selection Models
  • Fixed effects: This model includes effects in both the selection and regression equation. The model is fit by two step maximum likelihood, with bootstrapping to estimate the standard errors at the second step.
    
  • Random effects: These are modeled both in the selection equation and in the regression. The random effects may also be modeled as dependent on the means of the included variables.
  • Attrition: This is a two period model in which individuals may leave at the end of the first period. The model is fit by maximum likelihood.
  • Random parameters: The random parameters model, which is generally for single equation specifications, is extended to this two equation model. The model is fit by maximum simulated likelihood.
  • MLOGIT selection rule: The multinomial logit model may be used as the selection mechanism. The full set of computations are handled internally. Previously, this was computed using a lengthy set of LIMDEP commands.
  • Alternative selection models: Numerous nonlinear models in LIMDEP can be fit subject to sample selection, including probit, Poisson, and negative binomial. The multivariate probit and ordered probit models are now supported as well.
    
• Stochastic Frontier Models
  
  • Normal-gamma: The normal-gamma model is now supported. The estimator is maximum simulated likelihood.
  • Heteroscedasticity: The model may be fit with either or both of the error components heteroscedastic.
  • Normal-truncated normal: Several forms of this model are now supported.
  • Panel data models: Panel data forms of the stochastic frontier include the random parameters and latent class forms, random effects, and several different forms of the fixed effects model. The panel data forms may be integrated with the truncation and heteroscedasticity models.
• Count Data Models
  
  • Sample selection model (maximum likelihood): This is a full information maximum likelihood estimator for a sample selection model for count data.
  • Underreporting: LIMDEP now includes a maximum likelihood estimator of a model in which the observed counts represent only the reported fraction of the total events which have occurred.
    
  • Gamma: The gamma model is an extension of the Poisson model which allows both underdispersion and overdispersion.
    
  • Hurdle: The hurdle form is an alternative to the zero inflation models. The hurdle form can be fit by single equation limited information methods. The new form is a full information estimator.
    
• Binary Choice Models - Several new functional forms are provided, including:
  • Gompertz
  • Complementary log log
  • Burr (the Scobit model)
  • Nonparametric binary regression (kernel density)
  • Klein & Spady semiparametric estimator
  • Bivariate probit with random effects. Several forms of this model are supported, including one in which the random effect is common to both equations.
    
• Ordered Probability Models: New functional forms supported for this model class are Gompertz and complementary log log. A sample selection model has been developed as well.
• Generalized Maximum Entropy Models: The binomial and multinomial logit models have been extended to support the generalized maximum entropy form. This is a less parametric form of the multinomial choice model than the full MNL model.
  
• Time Series: The ARCH/GARCH/GARCH(m) estimators have been converted to maximum likelihood estimation. Any order of ARCH, GARCH, or GARCH in mean model may be specified.
• Survival: New parametric forms include the inverse Gauss and extensions of the generalized F.
• Scale Heterogeneity: Several models have been added to the set in which the scale parameter (such as the regression variance) may be made a function of exogenous variables. The new forms include all the parametric survival models, the negative binomial regression model and the Heckman sample selection model. These are added to the set which previously included the probit, tobit, binomial logit, truncated regression, and grouped data regression model, in addition to the linear model (Harvey's model).
• Linear Regression Model
  
  • Least absolute deviations is automated as an algorithm.
  • Inequality constrained least squares. The linear least squares estimator may be estimated subject to a set of inequality constraints.
  • For small to moderate sized data sets (up to 100,000 values), the least squares estimator is computed using the QR decomposition. This produces extremely accurate solutions. For example, the NIST Filippelli problem, which most other programs cannot solve at all, is solved with eight or more digits of precision.
  • In the time series cross section model:
    • Panel corrected standard errors for the least squares estimator are computed automatically.
    • Subsets of the observation groups may be specified to be uncorrelated.
      

New Model Features and Model Extensions
The list of supported models has expanded, but to allow even greater flexibility, we have also provided tools for users to develop their own estimators, including GMM estimation, bootstrapping, optimization tools such as simulation estimation tools, Halton sequences for simulation, kernel functions and density estimators, and many new matrix, transformation and calculator functions.

• Estimation Procedures: Many of the program revisions and enhancements will make it easier to write your own estimators and computation routines for models that you design.
  • Constrained maximum likelihood estimation:Nearly all maximum likelihood estimators in the program may now be computed subject to linear restrictions. This will simplify likelihood ratio and Lagrange multiplier tests.
  • Optimization - MAXIMIZE, MINIMIZE, NLSQ:
    
    • Integration: Many applications involve integration in the likelihood function or GMM criterion. When closed forms do not exist, it is necessary to use quadrature or simulation. Both of these are now supported in these procedures. Some specific integrals are now supported as well.
    • Quadrature: Maximands may now involve Hermite or Laguerre quadrature for up to five functions.
    • Simulation: Maximands may include simulation integration of up to five dimensions. This feature may be used to construct estimators based on the maximum simulated likelihood procedure.
    • Multivariate normal: Maximands may include integration of the normal density for one, two, or K dimensions. The last of these uses the GHK simulator.
    • Closed forms: Maximands may include complete or incomplete beta or gamma integrals.
    • Function specification: Functions to be maximized may now include linear (b'x), bilinear (b'Ac) and quadratic (b'Ab) forms.
• GMM estimation: The new GMME command provides a means of formally setting up a GMM estimation problem. The command provides the definitions of the moment equation and the name of the optimal weighting matrix to be used. The estimation may be done in two steps, with the first step obtaining an inefficient but consistent estimator, and computing the appropriate weighting matrix and the second step doing full GMM estimation with the optimal weighting matrix.
• Linear programming: MAXIMIZE or MINIMIZE may define a linear programming problem to be solved using the standard simplex method. The new matrix function, Gmmw may be used in some cases to compute this optimal weighting matrix.
• Bootstrapping:
  • DRAW command: This command is used to obtain bootstrap samples, with or without replacement. The command has been extended to allow you to draw bootstrap samples from a panel, so you can bootstrap familiar panel data estimators. The procedure is no more complex than sampling individual observations. You simply indicate the number of groups to be drawn and specify the usual group size indicator to allow LIMDEP to discern how the panel is arranged in the sample.
  • Bootstrapping in procedures: Procedures have been modified so that you can bootstrap any estimator or any other computation done with LIMDEP using a data set. This includes models contained in the program, models that you define, or the result of any matrix or scalar computation.
• Two step estimation:
  • Murphy and Topel correction: An increasingly common application involves two step maximum likelihood estimation in which a predicted value is computed using the model estimated at the first step, then the model estimated at the second step includes the prediction from the first. In general, it is necessary to use a correction such as Murphy-Topel at the second step to obtain appropriate standard errors. The LIMDEP manual (still) contains numerous applications to demonstrate the technique using rather long LIMDEP command sets. This computation has been automated for many of the models in LIMDEP 8.0. At the first step, you will save the prediction and use LIMDEP's ;Hold specification to keep the model available. The second step need only indicate, with a ;2Step = prediction specification, that this is a two step estimator. All of the necessary computations needed to obtain the Murphy-Topel correction are done automatically. This procedure is used for several models including combinations of the tobit, probit, logit, linear regression and Poisson regression.
  • Sample selection models: In the various first step estimators for SELECT, ;Hold now keeps the covariance matrix so it need not be recomputed. This avoids some problems created previously by missing data.
• Robust Covariance Matrices: Various forms of robust covariance matrices are now provided for most of the estimators.
  • Linear models
    • The White estimator and three variants suggested for improved small properties are supported for the linear model.
    • White and Newey-West estimators are provided for the fixed effects linear model.
    • White and Newey-West estimators are provided for the random effects linear model.
  • Nonlinear models
    • The cluster estimator is supported for nearly all models in LIMDEP including the linear regression model.
    • The White estimator is provided for nonlinear least squares.
    • 'Sandwich' style robust covariance matrices are provided for probit, logit, and Poisson regression models.
      
• Estimators: The following lists extensions and revisions of certain existing estimation procedures.
  • Descriptive statistics procedures: These are new procedures in the descriptive statistics package:
    • Kernel density estimators
    • Box and Whisker plots
    • Burg Estimator for PACF (This produces greater accuracy than Yule-Walker.)
    • Spectral density estimator
    • Phillips-Perron test for a unit root
      
  • Modifications of the descriptive statistics programs:
    • The standard histogram has been replaced with a more detailed figure.
    • Scatter plots now allow specification of a label for the vertical axis.
    • Scatter plots can be done with a stratification to put several kinds of dots in one scatter plot.
    • You can now fix the number of digits reported in descriptive statistics output.
  • Tobit and censored data:
    • Fin and Schmidt's LM test for the tobit specification is computed automatically.
    • The tobit random effects model now allows a variable lower or upper limit value.
    • The various forms of grouped data models now allow the limits to be variables and to vary across observations.
  • Multinomial logit:
    • Models may now contain up to 300 parameters.
    • The parameter vector is reconfigured after estimation and saved as a parameter matrix. This simplifies subsequent computation of probabilities, derivatives, effects, and so on.
    • The full set of probabilities may be saved as a set of variables. In earlier versions, it was necessary to program this explicitly with LIMDEP commands.
    • Marginal effects with standard errors are now computed internally and displayed with the standard output.
  • Bivariate probit model:
    • The estimator saves a parameter matrix with two columns. This is constructed to make it easier to compute probabilities and to display model results.
    • The program will now save the bivariate normal cdf and density evaluated at the data and estimated parameters, if requested.
    • Marginal effects for dummy variables are computed using the change in the probabilities.
  • Seemingly unrelated regression models:
    • Output has been reformatted for easier reading.
    • SURE/MLE has been expanded to 150 parameters and 20 equations.
    • SURE/3SLS now allows weights.
    • SURE creates a parameter matrix which can be used with a namelist to create fitted values easily.
  • Qualitative response models:
    • Marginal effects for dummy variables in the probit, bivariate probit and logit models are now computed using the change in the probabilities rather than approximately using derivatives. Dummy variables are located internally.
    • An array of goodness of fit measures is provided for binary choice models, including Hosmer and Lemeshow's chi-squared statistic.
      
    • You may specify an alternative threshold for predictions from binary choice models. The previous setting was 0.5 for the probability.
    • A full set of marginal effects, including standard errors, is produced for the ordered probability models.
    • The probit model with heteroscedasticity now allows choice based sampling.
      

Program Refinements
Naturally, we took the opportunity to improve the program operation for such features as data input, data set sizes and computational algorithms. A number of new functions have been added to CALCULATE, MATRIX and CREATE. Some of these are listed below. Other refinements relate to definition of namelists and computational algorithms.

• CALCULATE:
  • New functions:
    • Hlt returns a Halton draw from the sequence for the base specified.
    • Sgn returns the sign of a computation result that is in a scalar.
    • Bvn returns the bivariate normal cdf.
    • Bvd returns the bivariate normal density.
    • Bv1 returns the partial derivative of Bvn with respect to the first x.
    • Bv2 returns the partial derivative of Bvn with respect to the second x.
    • Ngi returns the number of groups in the specified panel.
    • Rck returns Kendall's rank correlation coefficient.
    • Cnc returns the coefficient of concordance for a set of ranks.
    • Gmn returns the geometric mean of a variable.
      
  • Moments of subsamples: Functions Xbr, Sum, Sdv and Var will be computed for the subset of observations for which a companion variable is nonzero. Thus, Xbr(income) is the mean of income in the current sample while Xbr(income,female) is the mean of the observations for which female is nonzero. In most cases, the second variable will be a dummy variable that indicates the subsample, but it can be any variable whose nonzero values will indicate inclusion in the sum.
    
  • Graphing and tabulating discrete probability distributions: The following functions are primarily for teaching statistics and econometrics. Each returns a table of probabilities and a figure that shows the shape of the indicated discrete distribution.
    
    • Tbb(pi,n) for binomial
      
    • Tbp(lambda) for Poisson
    • Tbg(pi) for geometric
    • Tbn(pi,n) for negative binomial
    • Tbh(p,m,n) for hypergeometric
• MATRIX
  • Matrix results:
    • Ladb(x,y) returns the least absolute deviations coefficient vector.
    • Svdx(x) returns a singular value decomposition.
    • Mvec(a,r,c) rearranges a matrix to be a vector or vice versa.
    • Iden(k) returns an identity matrix.
    • Sign(a) returns a matrix of signs, element by element, of matrix a.
  • Matrix raised to a power: Several forms of this operation are provided, varying by whether the matrix and/or the exponent are scalars or matrices.
  • Syntax for summations: The construction 1'X will sum the rows of the matrix X. The number '1' is interpreted as a column of ones in a matrix multiplication.
  • MATRIX save files: Functions Mput and Mget store and retrieve matrix save files similar to SAVE and LOAD, but only for the matrix work area.
    
  • Sums for panel data: The matrix function Gsum (namelist [, variables not in a list], weight, index) (the additional variables are optional) is a general function that transforms the data set into a matrix of weighted sums. The resulting matrix has number of rows equal to the number of groups in the index set. The number of columns is the number of variables in the namelist and the additional variables listed. The variables in the namelist are weighted by the weighting variable as they are summed while the variables in the list of additional variables, if any are specified, are simply summed, but not weighted. This function is useful for creating estimators. For example, it can be used to extend the Hausman and Taylor random effects estimator.
    
• CREATE
  • Halton draws: The function Hlt(base) creates a column of Halton draws using the specified base, and number of values equal to the sample length.
  • Expanding a qualitative variable: The function Expand(variable) creates a complete set of dummy variables for the discrete variable specified.
  • Logged values in logical expressions: If(...) expressions may now involve lagged values. Indexes for lags may be specific numbers or named scalars (with leading minus signs if needed).
  • Ranks: CREATE now contains a Rnk(variable) function which sorts the variable and assigns to the new variable the rank of the old one in the sorted list.
  • Kernel weights: CREATE contains a Krn(variable,kernel,bandwidth) function which computes for the specified variable a set of kernel weights or kernel density estimators.
• Namelists
  Three new functions are provided for combining namelists x1 and x2:
  
  NAMELIST ; Newlist = Or(x1,x2) $
  
  creates a namelist of all variables in the union of x1 and x2 - that is, variables that are in either or both x1 and x2.
  
  NAMELIST ; Newlist = And(x1, x2) $
  
  creates a namelist of all variables in the intersection of x1 and x2 - that is, only variables that are in both x1 and x2.
  
  NAMELIST ; Newlist = Xor(x1, x2) $
  
  creates a namelist of all variables in the exclusive union of x1 and x2 - that is, variables that are in either x1 or x2 but not in both.
  
• Data Input and Entry
  • The limit on variables in an active data set has been raised from 200 to 900 variables.
  • Observation labels may be input when data are read from a spreadsheet. This variable would be the first column in the spreadsheet, and would give names to observations. The names are used in data listings.
  • Excel files may now use the current (Excel 2000, 2002) format.
  • There is a default ASCII file format, rectangular with a row of names at the top that may be invoked simply by providing the file name. Nvar and Nrec as well as the names may be omitted from the command.
  • File formats DIF and CSV have been added to the supported set. (You probably will not need these.)
    
  • Append may now be used with spreadsheet files.
    
  • Qualitative variables indicated by an alphabetic code may be recoded to a set of appropriate numeric values at the time they are read.
    
  • Data lines in an ASCII data file may now be up to 1024 bytes wide.
    
  • Lists of observations on a set of variables may be sorted based on a specified key variable.
    
  • WRITE now supports the CSV format, so data exchange with spreadsheet programs such as Excel is simple and seamless.
• Computation Algorithms
  • The algorithm in the Cox proportional hazards model has been replaced, with extremely large increases in speed. A proportional hazard model fit on a 600 Mhz machine with 10,000 observations will take approximately 0.5 seconds.
  • The algorithms for the sample selection and bivariate probit models have been modified to improve handling of the bounded range of the correlation coefficient.
  • A new random number generator has been installed. The period is now 2132 draws, which should be sufficient for any simulation problem imaginable.
  • Marginal effects for specific models may be computed at specific data vectors in addition to the data means.
    

NLOGIT Version 3.0
NLOGIT 3.0 provides the widest and deepest array of tools available anywhere for analysis of discrete choice, including (heterogeneous) nested logit models, heteroscedastic extreme value, multinomial probit, mixed logit and panel (repeated measures) estimators for mixed logit, random utility, multinomial probit, and latent class models. NLOGIT is the only program available that supports mixing stated and revealed choice data sets. NLOGIT 3.0 also introduces a simulation package that allows you to analyze alternative scenarios in the context of any estimated discrete choice model with any data set, whether used in estimation or as hold out data for examining model cross validity.

There are dozens of new features in Version 3.0 of NLOGIT. The two major groups of enhancements are the model simulator which allows you to simulate any model with the data that you have used to estimate it or any other sample, and a large number of new model specifications. There are also numerous small refinements that make it easier for you to use the program.

The complete list of new features is described below.

• Model Simulator: A major addition to NLOGIT is the model simulator. After estimation of a model, you can simulate the probabilities computed by the model using the same or a different data set. The simulation can restrict the choice set or use the original one. Scenarios in the simulations involve changing attributes and recomputing probabilities and sample shares to examine the effect of the change on aggregate. The simulator may be used with any model.
• New Model Specifications : Several new extensions of the basic multinomial logit model have been added in Version 3.0. The random parameters logit (RPL) model includes many variations that embody the current state of the art for this specification. The multinomial probit (MNP) model has likewise been expanded and can now be applied to a quite large estimation problem. We have also added a new application, the latent class model. This model, with the RPL and MNP models provide a wide range of model formulations that can all be used with panel data and mixtures of stated and revealed preference data.
  • Multinomial logit model:
    Algorithm: Newton's method will almost always converge quickly when fitting this model. But, certain data sets are such that this algorithm performs poorly. The BFGS method, which takes longer, but is more stable, may now be specified instead. Time differences are noticeable, but small.
    Ties in ranked data: The Beggs, Cardell and Hausman estimator for ranked data now allows a partial ranking. A subset of the ranks may be ties (for last place). The model then applies to the choices ranked better than this group.
    
  • Heteroscedastic extreme value: Specification of the HEV model can fix or change the standard deviations for the underlying random variables.
  • Multinomial-multiperiod probit model: The MNP model now allows panel data. This is the 'multinomial-multiperiod' model in the literature. Specification may be random effects or AR(1).
  • Random parameters logit model:.
    Panel Data: All forms of the RPL model may now be fit with panel data. The random component of the parameter vector may be treated as a time invariant random effect or it may be assumed to evolve in an AR(1) process.
    
    Simulations:
    
    • Simulations may be based on Halton draws rather than on random draws. This greatly speeds up convergence.
    • The estimator now allows more distributions in simulations, including normal, lognormal, tent, and uniform.
    • The standard deviations of the simulation draws may be constrained for purposes of specification or hypothesis testing.
      
      Observation Specific Tastes: ;Par with this model saves the person specific expected coefficient vectors.
      
      
  • Nested logit model: Several different forms of the nested logit model may be specified that impose restrictions on the inclusive value coefficients to enforce the model of utility maximization.
    
  • Latent class model: The latent class model has been extended to the multinomial logit model. This model provides a discrete alternative to the continuous parameter variation that is embodied in the random parameters model. The latent class model is also extended to panel data.
    
• Restricted Choice Sets
  The choice set may be restricted before estimation begins. The model is fit to the data on individuals who chose one of the alternatives in the restricted choice set. This is the same as the IIA test, but a bit simpler to set up and use, and does not entail a testing procedure.