Limited mobility bias correction

Simen Gaure Address: Ragnar Frisch Centre for Economic Research, Oslo, Norway
Date: November 12, 2015
Abstract.

A description of what limited bias correction is. And how to compensate for it.

1. Introduction

We assume we have an OLS model in matrix form

(1) Y=X⁢β+D⁢α+F⁢γ+ϵ

where X is a (n×k)-matrix, D is a (n×g1)-matrix, and F is a (n×g2)-matrix. It is assumed that D and F are matrices of dummy variables and that both g1 and g2 are very large (as in 105−107). The lfe-package estimates α and γ as OLS coefficients. Some applications study the variances var⁡(D⁢α) and var⁡(F⁢γ), as well as the covariance cov⁡(D⁢α,F⁢γ). It was shown in [1] that if one uses the estimates D⁢α^ and F⁢γ^ for this purpose, the resulting variances are positively biased, and the covariance is typically negatively biased. The biases can be large enough to change the sign of the covariance. They also provided explicit formulas for the size of the bias, in terms of the trace of some very large matrices.

In short, the bias for var⁡(D⁢α^) can be computed as

(2) δα=σϵ2⁢tr⁡(A),

where A is a large matrix depending only on X, D and F. The unbiased variance can then be estimated as

(3) var⁡(D⁢α)=var⁡(D⁢α^)−σϵ2⁢tr⁡(A)

In some typical instances, the matrix A will be too large to handle directly, being of size g1. The bias for var⁡(F⁢γ^) and the covariance, are of the same form.

2. How lfe handles this

In [2] a method for estimating δα was outlined. It uses the fact that

(4) tr⁡(A)=E⁡(xt⁢A⁢x),

if x is a vector of independent random variables xi with E⁡(xi)=0 and var⁡(xi)=1. In fact, each xi can optimally be choosen as drawing from {−1,1} with uniform probability. This works because even if A is too large to be handled directly, it is possible to compute A⁢x for a vector x. Thus, lfe’s fevcov routine uses a sample mean to estimate tr⁡(A).

3. Another method

Formula 3 suggests another method for estimating var⁡(D⁢α). We rewrite it as

(5) var⁡(D⁢α^)=var⁡(D⁢α)+σϵ2⁢tr⁡(A)

Now, b=var⁡(D⁢α) and a=tr⁡(A) are constants, so we can write

(6) var⁡(D⁢α^)=a⁢σϵ2+b

If we resample and scale the residuals ϵ^, and do the estimation of α^ over again, we will obtain a series of different observations of formula 6. We may then do an OLS on model 6, and the estimate a^ will be an estimate for tr⁡(A) which can be used in equation 3. The var⁡(D⁢α^) is not normally distributed around its mean, this typically skews b^ more than a^, so a^ is a better estimate. lfe does not implement this method.

References

  • [1] M. Andrews, L. Gill, T. Schank, and R. Upward, High wage workers and low wage firms: negative assortative matching or limited mobility bias?, Journal of the Royal Statistical Society(A) 171(3) (2008), 673–697.
  • [2] S. Gaure, Correlation bias correction in two-way fixed-effects linear regression, Stat 3 (2014), no. 1, 370–390.