Paul D. Allison "Fixed Effects Regression Models"

Fixed Effects Regression Models (Quantitative Applications in the Social Sciences)

Fixed Effects Regression Models (Quantitative Applications in the Social Sciences)

unobserved heterogeneityを完全にコントロールしてバイアスのない推定値を得られるのがfixed effectsの強みということで、経済学者がrandom effectsよりもfixed effectsを好んでいるように見えるのもこの理由からなのだろう。
しかし、社会学者の関心はしばしばtime-invariantな変数に向いているわけで、本書で紹介されるhybrid methodのように、どのようにお互いの長所をとってゆくのかということが重要になってくるということか。



Regardless of which computational method is used, the fixed effects method effectively controls for all time-invariant predictors, both measured and unmeasured. This is its principal attraction as compared with random effects methods. A key assumption of the fixed effects method, however is that the time-invariant predictors must have the same effects at all occasions. Variables whose effects are not constant across occasions must be explicitly included in the model.

For response variables with more than two categories, conditional maximum likelihood estimation for fixed effects logistic regression is generally unavailable in commercial software. Instead, the best available approach at present for both ordinal and nominal response variables is to use the hybrid method with robust standard errors to correct for dependence.

Fixed effects regression analysis of event history data typically requires that each individual has multiple, repeated events. As we saw with logistic regression, the use of dummy variables to estimate the fixed effects usually leads to biased coefficient estimates for the other variables. This incidental parameters problem can be avoided by using Cox regression with stratification to eliminate the fixed effects from the partial likelihood, a method that is computationally efficient even for large numbers of strata. Under most conditions, the method of stratification produces approximately unbiased estimates.