weighting BHPS sample correctly in longitudinal analysis
Hi, I want to make sure I am using the BHPS longitudinal weights correctly in my analysis. I am analysing respondents to the self-completion interview in 4 waves of data: BHPS waves N, O, & R, and USOC Wave B, using both random effects models (using xtmixed in Stata) and fixed effects models (using xtreg, fe). My analyses are restricted to English 18+ respondents only.
As there is no BHPS longitudinal self-completion weight, I have been using the BHPS 2001 longitudinal main interview weight (b_indin01_lw) as the best substitute. For example:
xtreg y x1 x2 [pweight=b_indin01_lw], fe
Am I OK with this, or would another approach give me less biased estimates?
#1 Updated by Redmine Admin over 6 years ago
- % Done changed from 0 to 50
If you are interested in descriptive statistics and your analysis
sample is UKHLS Wave 2 BHPS sample, then using the weights
b_indin01_lw, produces unbiased estimates for the 2001 UK (and
England) population that has survived until 2010. These weights
correct for unequal selection probability (particularly because of the
regional boost samples) and attrition from 2001 to 2010. As you are
using only the England sample, unequal selection probabilities is not
a problem, but attrition is. As you rightly pointed out these weights
do not correct for the additional self-completion non-response. One
option is to estimate a self-completion non-response model yourself
and then multiply the existing weight with the inverse of the response
But in case of multivariate analysis, there is another problem. The
answer to the question, whether using weights in multivariate analysis
produces consistent estimates of the coefficients is not quite
straight-forward. Here is a good reference. "What are we weighting
for?" Solon, Haider and Wooldridge, NBER working paper 18859.