www.gllamm.org

What's New


Since 15 July 2012

Frequently asked questions (FAQs) has been added (work in progress).


Since 22 March 2012

Our book is out!

Rabe-Hesketh, S. and Skrondal, A. (2012). Multilevel and Longitudinal Modeling Using Stata (Third Edition). College Station, TX: Stata Press.

Volume I: Continuous Responses.

Rabe-Hesketh, S. and Skrondal, A. (2012). Multilevel and Longitudinal Modeling Using Stata (Third Edition). College Station, TX: Stata Press.

Volume II: Categorical Responses, Counts, and Survival.

Volume II of the book demonstrates and explains how gllamm (and Stata's official commands where applicable) can be used to estimate models for all sorts of response types: Binary, ordinal, and nominal responses or discrete choice, counts, discrete-time survival or durations, and continuous-time survival

Volume I of the book does not use gllamm. To find out how to fit most of the models from Volume I using gllamm, download:


Since 7 September 2011

By downloading some additional files, gllamm can now be used with Bayesian priors for variance parameters.

There was a bug in gllapred with the ustd option for models with several random effects at the same level. This has been fixed.

Levels of confidence can now be non-integer percentages, e.g., level(99.5)


Since 15 September 2010

gllamm previously did not exponentiate all coefficients in multinomial logit models with the eform option and this has been fixed.

There was a bug when attempting to estimate models with 10 or more random effects or latent variables and this has been fixed. gllamm no longer works in Stata version 6 (Stata 7 was released in December 2000 and the current version is Stata 11.1).

The init option, which sets the random part of the model (apart from the level-1 error, if applicable) to zero, did not behave correctly when there was nothing to estimate, or when it was used together with the from() option. These bugs have been fixed. When the init option is used with the trace option, the "General model information" output now states which parameters will not estimated.


Since 15 September 2008

gllapred with the fsample option now computes empirical Bayes predictions of random effects (u option), posterior mean probabilities (mu option), and posterior means of the linear predictor (linpred option) for observations where the response variable is missing. Predictions of random effects are also given if any of the explanatory variables are missing. See slides, do-file and data for 2008 UK Stata Users Group Meeting.

gllamm previously did not show the estimated threshholds or cut-point parameters for ordinal logit models when the eform option was used. This has been fixed for Stata versions 8, 9, and 10. For earlier versions, display the threshholds after estimation using the command: gllamm (no variables or options).


Since 30 January 2008

Our book is out!

Rabe-Hesketh, S. and Skrondal, A. (2008). Multilevel and Longitudinal Modeling using Stata (Second Edition). College Station, TX: Stata Press.
Datasets for the book


Since 30 December 2007

Review of latent variable modeling was published:

Skrondal, A. and Rabe-Hesketh, S.(2007). Latent variable modelling: A survey. Scandinavian Journal of Statistics 34, 712-745. Local


Since 11 November 2006

gllapred now calculates posterior correlations of random effects or latent variables - see corr option.

gllamm previously did not produce correct robust standard errors with the init and weight() options if there were weights different from 1 at levels 2 or above. This has been fixed.

gllamm previously did not accept the ip(g) option, but now it does.

gllamm previously did not exclude observations from the estimation sample where the variables specified in the expanded() option were missing. Now it does.

The command gllamm, robust previously did not work if the previous gllamm command specified several ordinal links. This has been fixed.


Since 18 July 2006

gllapred produced incorrect predicted marginal probabilities for three-level or higher-level models (with the mu and marg options), and this has been fixed. Thanks to Ann Berrington for pointing this out.


Since 23 May 2006

gllamm previously ignored the exposure() option and now produces a proper error message and stops (offset() should be used instead). Thanks to Dale Needham for pointing this out.

The eval option now works when combined with the pweight() option.

gllamm now prints a warning when non-integer frequency weights are used.


Since 14 April 2006

Stata/MP, a parallel version of Stata that runs simultaneously on several processors, has been released. On a dual core or dual processor machine, gllamm will run up to twice as fast depending on the problem - see Appendix E of the Stata/MP Performance Report for some preliminary results for gllamm.


Since 18 March 2006

The gllamm wrapper ssm has been updated to correct the way standard errors are calculated for some of the parameters.


Since 13 March 2006

A review of gllamm has been written by Leonardo Grilli and Carla Rampichini


Since 27 February 2006

A solutions manual (172 pages) and do-files for all exercises in Multilevel and Longitudinal Modeling Using Stata are now available for instructors. See: http://www.stata.com/bookstore/mlmus.html


Since 8 December 2005

New gllamm wrapper ssm for endogenous switching and sample selection models for binary, count, and ordinal variables. Written by Alfonso Miranda and Sophia Rabe-Hesketh.


Since 19 October 2005

Bugs that occurred in gllamm with the eval and robust options have been fixed.


Since 10 August 2005

Our book is out!

Rabe-Hesketh, S. and Skrondal, A. (2005). Multilevel and Longitudinal Modeling using Stata.. College Station, TX: Stata Press.
Datasets for the book

Since 3 February 2005

Worked example added for Explanatory Item Response Models by De Boeck and Wilson (Eds.)


Since 5 January 2005

There was a bug in gllamm with any or all of the options robust, cluster() and pweight() that caused gllamm to repeatedly print a red error message. This problem has been fixed.


Since 26 October 2004

The GLLAMM manual has been updated.


Since 21 October 2004

gllapred previously didn't work with the mu option for models with more than two levels. Thanks to Phil Schumm for reporting this bug. This has now been fixed.

gllapred with options xb and nooffset did not work correctly and this has been fixed.


Since 6 October 2004

Thanks to Bobby Gutierrez and Stata Corporation, gllamm (updated 6 October 2004) is now faster for certain kinds of models in Stata 8.2 (updated on or after 6 October 2004). Specifically, gllamm is faster for models with composite links and latent class models where probabilities of latent class membership depend on covariates.


Since September 2004

Composite links now available in gllamm, see here for gllamm commands and output for classic blood types example and here for the paper:
Generalized Linear Latent and Mixed Models with Composite Links and Exploded Likelihoods. Proceedings of the 19th International Workshop on Statistical Modeling (Editors A. Biggeri, E. Dreassi, C. Lagazio and M. Marchi), Florence, Italy, pp. 27-39.

gllamm now prints out the iteration log from iteration 0 when adaptive quadrature is used (instead of from iteration 1).

The wrapper cme for covariate measurement error models now collapses the data if several units have the same values on all the relevant variables. As kindly pointed out by Hendrik Juerges, a bug in the previous version prevented this from happening.

Some more worked examples have been added here.


Since May 2004

Our book is out!

Skrondal, A. and Rabe-Hesketh, S. (2004). Generalized Latent Variable Modeling: Multilevel, Longitudinal and Structural Equation Models. Boca Raton, FL: Chapman & Hall/CRC.
Datasets and do-files
Remarks and corrections

Some worked examples and do-file have been added here


Since April 2004

The sample size saved in e(N) is now the number of level-1 units instead of the number of top-level units.

gllamm now produces an error message if an invalid ip() is specified instead of assuming the default.

gllapred now works for mixed responses if some responses are unordered categorical (mlogit link).


Since February 2004

New courses announced

New iterate(#) option in gllamm to run gllamm no more than a given number of iterations even if convergence is not reached.

gllamm now produces a proper error message if any of the variables in the i() option are not numeric.


Since September 2003

New URL for gllamm webpages: http://www.gllamm.org

New wrapper cme for gllamm to estimate generalized linear models with covariate measurement error. See Wrappers.

New nodisplay option to stop gllamm displaying estimates (useful for wrappers}.

Some bugs in gllamm have been fixed. e.g. the cluster() option sometimes did not work.


Since June 2003

New cooksd option for gllapred to compute Cook's distances for top-level units.

New ustd option for gllapred to obtain approximately standardised empirical Bayes predictions of the random effects or latent variables (sampling variance approximated by 'prior variance - posterior variance').


Since January 2003

Thanks to Stata Corporation, gllamm is now considerably faster in Stata 8 (updated on or after 17 January 2003) than in Stata 7.


Since November 2002

New peqs() option for latent class models to allow latent class probabilities to depend on covariates.

New pweight() option to specify sampling weights. This should be used with caution if the sampling weights apply to units at a lower level than the highest level in the multilevel model.

New cluster() option to obtain robust standard errors (sandwich estimator) if the highest level units in the multilevel model are nested in (even higher level) clusters.


Since October 2002

New robust option for gllamm to obtain the sandwich estimator of the covariance matrix of the estimated parameters.

New mu option for gllapred to obtain the expectation of the response (or predicted probabilities for dichotomous, ordinal or nominal responses). The following variants are available:

  • Conditional expectation given specified values of the latent variables. This is useful for looking at 'conditional' effects of covariates for given values of the latent variables.
  • Expectation with respect to the posterior distribution of the latent variables, i.e. conditional on the observed responses. This provides the 'best' predictions for the units.
  • Expectation with respect to the prior distribution of the latent variables. This is useful for looking at the 'marginal' or population average effects of covariates.

New pearson, deviance and anscombe options for gllapred to obtain 'level 1' residuals, either for specified values of the latent variables or their posterior expectation.


Since September 2002

New 'post-estimation' command gllasim for simulating the responses and/or latent variables for the model just estimated in gllamm: See programs

New webpages:

New ethresh() option for gllamm. Like the thresh() option, this option can be used to specify models for the thresholds in cumulative models for ordinal data to relax the parallel regression assumption. However, a different parameterization is used to impose the order restriction on the thresholds. See worked example for King et al. (2002)

Improved behaviour of gllamm with the init (initial values) option. Constraints are for example now permitted.

New from() option for gllapred to make predictions for different parameter values (than those estimated).


Since April 2002

Thanks to Stata Corporation, adaptive quadrature is now considerably faster in Stata 7 if updated since 4 April 2002.

Several options have been added to gllapred including:

  • if and in
  • ll to compute the likelihood contributions of the highest level clusters.
  • fac to compute predictions of the latent variables when these are regressed on observed or other latent variables. The u option in this case provides predictions of the disturbances.

  •