15

I'm looking for an R package for estimating the coefficients of logit models with individual fixed-effect (individual intercept) using Chamberlain's 1980 estimator. It is often known as Chamberlain's fixed-effect logit estimator.

It's a classic estimator when dealing with binary outcome panel data (at least in econometrics), but I just don't find anything related to it in the CRAN.

Any clue?

Glen_b
  • 257,508
  • 32
  • 553
  • 939
Kamixave
  • 159
  • 1
  • 1
  • 4

3 Answers3

13

Conditional logistic regression (I assume that this is what you refered to when talking about Chamberlain's estimator) is available through clogit() in the survival package. I also found this page which contains R code to estimate conditional logit parameters. The survey package also includes a lot of wrapper function for GLM and Survival model in the case of complex sampling, but I didn't look at.

Try also to look at logit.mixed in the Zelig package, or directly use the lme4 package which provide methods for mixed-effects models with binomial link (see lmer or glmer).

Did you take a look at Econometrics in R, from Grant V. Farnsworth? It seems to provide a gentle overview of applied econometrics in R (with which I am not familiar).

chl
  • 50,972
  • 18
  • 205
  • 364
  • 1
    Actually, "conditional logit" is a very ambiguous term. It some contexts (mainly when dealing with panel data), it is equivalent to Chamberlain's estimator, but it is very unfrequent. Most of the times, it refers to a cross-sectional model where the outcome variable can take more than 2 values. All your proposals actually refers to packages that consider this last possibility. The same with mixed-logit: it is not a fixed-effect logit. I've already taken a look at Farnsworth's overview, but it is not exhaustive enough to speak about this estimator. Anyway, thank you for your answer ! – Kamixave Aug 26 '10 at 10:11
  • "Conditional logit" does *not* refer to having more than two outcome levels. Some functions might extend it to that situation, but that is not the point. – Aniko Aug 26 '10 at 13:09
  • 1
    Yup, but the conditional logit model can (as I said) take more than 2 values, which differentiates it easily from Chamberlain's model, just like the fact that Chamberlain is designed specifically for panel data. This is thus a relevant information ; precise description of the usual conditional logit isn't (and the description of both would take more than 600 chars). – Kamixave Aug 26 '10 at 16:07
3

You can run the Chamberlains model using glmer. It is basically a RE model but with more variables:

glmer(y~X + Z + (1|subject), data, model=binomial("probit"))
  • X are the variables you consider explain your fixed effect model (a simple case it is the mean of Z)
  • Z are your exogenous variables
  • Subject is the variable where the heterogeneity comes from

I hope this helps.

Isabel
  • 31
  • 1
  • 3
    I think that would restrict heterogeneity to be orthogonal to X and Z while the requested estimator allows for it. – Alex Nov 28 '12 at 13:44
  • No is ramdon effects. It is not chamberlain model which si fixed effects – Ari.stat Nov 10 '21 at 15:01
1

The mclogit package seems to implement conditional logit of the Chamberlain variant.

Julian Schuessler
  • 2,025
  • 11
  • 16