Skip to main content Skip to navigation

David Firth, software: brlr


As of version 0.9-1 (31 May 2008), the brlr package is deprecated: it is superseded by more efficient and more comprehensive facilities in the brglm package (by Ioannis Kosmidis). Archived copies of the brlr package sources remain available at

brlr is a package for R, to fit logistic regression models to grouped or individual-level binary response data by a maximum penalized likelihood method. The penalty function applied is the Jeffreys invariant prior. The resulting estimates have asymptotic bias of order n^{-2} rather than the usual n^{-1}, and are guaranteed finite (as are the estimated standard errors associated with them). For the theory behind this see

  • Firth, D. (1993) Bias reduction of maximum likelihood estimates. Biometrika 80, 27--38.
  • Firth, D. (1992) Bias reduction, the Jeffreys prior and GLIM. In Advances in GLIM and Statistical Modelling, Eds. L Fahrmeir, B J Francis, R Gilchrist and G Tutz, pp91--100. New York: Springer.
  • Kosmidis, I (2007) Bias Reduction in Exponential Family Nonlinear Models. PhD Thesis, University of Warwick.

The brlr software is made available free, under GPL version 2. It comes with ABSOLUTELY NO WARRANTY.

Package contents

The package sources for brlr are available at CRAN sites for installation into R in the standard way.

Version history

0.9-1: brlr is now deprecated, in favour of brglm — see the note at the top of this page (2008-05-31)
0.8-8: fixed a bug which caused non-convergence when non-unit weights were used together with br = TRUE (2006-03-22)
0.8-7: fixed a clumsy bug introduced at 0.8-6, whereby response types other than a 2-column matrix resulted in an error (2006-01-10)
0.8-6: improved the behaviour and speed of the fitting algorithm (2006-01-05)
0.8-5: removed predict.brlr as it is now redundant (and it had a bug); the method used is now predict.glm (2005-10-29)
0.8-3: fixed a bug in the handling of the contrasts argument (2005-10-03)
0.8-2: improved the handling of missing values by add1.brlr, and the handling of zero-valued weights by brlr (2005-08-05)
0.8-1: fixed a bug introduced at version 0.8, whereby single-column binary data no longer worked as a way of specifying the response variable. (2003-12-10)
0.8: brlr now accepts a binomial glm as its "formula" argument; "add1" and "drop1" methods provided; various minor bugs fixed. (2003-11-18)
0.7-3: minor administrative changes only (2003-10-14)
0.7-2: added components to brlr objects for improved compatibility with glm methods (2003-09-15)
0.7-1: minor update to documentation (2003-02-10)
0.7: improvements to numerical robustness (2003-02-04)
0.6: added some components to brlr objects for compatibility with glm and lm methods (2002-06-10)
0.5: first public release (2002-05-14)

Related links: