How to apply the multiple weights in regression model (Health and Retirement Study)?

[Raoping Tu]

Dear all,

Recently I use data from Health and Retirement Study to do the regression analysis, there were three weights in relation to my variables:1. respondent weight for the physical measures subsample; 2.weight:respondent level 3.respondent weight for the biomarker subsample. How to apply those three weights in one regression model?


Look forward to your reply!
Thank you


Best regards,
Raoping

[Steve Samuels]

You can't use them all in one analysis. For analysis of physical measures, use only the physical measures subsample with the corresponding weight. For analysis of biomarkers, use only the biomarker subsample and the biomarker weight. So ,you will need three svyset statements

1. Before analyzing variables measured on everybody, svyset with the final weight

2. Before analyzing physical measure variable: keep only that subsample and svyset with pweight = respondent weight for that subsample

3. Before analyzing biomarkers, keep only that subsample and and svyset with the respondent weight for that subsample.

I don't know what you mean by "weight: respondent level" Is there a "level" subsample? If so, treat it like the other subsamples.

Weights for these subsample analyses are larger than the overall study weights, because the subsample still has to represent the entire population. If, for example, the hormone subsample is 20% of the overall sample, the hormone subsample weights will be about five times the size of the overall study weight.

It would have been helpful if you had provided a link to the study you mention. Statalisters come from many countries and disciplines. "Health and Retirement Study" will mean nothing to many, including, in this instance, me.

[Raoping Tu]

You can't use them all in one analysis. For analysis of physical measures, use only the physical measures subsample with the corresponding weight. For analysis of biomarkers, use only the biomarker subsample and the biomarker weight. So ,you will need three svyset statements

1. Before analyzing variables measured on everybody, svyset with the final weight

2. Before analyzing physical measure variable: keep only that subsample and svyset with pweight = respondent weight for that subsample

3. Before analyzing biomarkers, keep only that subsample and and svyset with the respondent weight for that subsample.

I don't know what you mean by "weight: respondent level" Is there a "level" subsample? If so, treat it like the other subsamples.

Weights for these subsample analyses are larger than the overall study weights, because the subsample still has to represent the entire population. If, for example, the hormone subsample is 20% of the overall sample, the hormone subsample weights will be about five times the size of the overall study weight.

It would have been helpful if you had provided a link to the study you mention. Statalisters come from many countries and disciplines. "Health and Retirement Study" will mean nothing to many, including, in this instance, me.

Dear Steve Samuels，

Thank you for your kindly reply! Health and Retirement Study (HRS):

https://hrs.isr.umich.edu/about?_ga=...119.1531229365. HRS collected blood-based biomarkers on half the sample in 2006, the other half of the sample provided biomarker data in 2008. The first group was asked for blood samples again in 2010 and the second group gave repeat samples in 2012. Now I would like to examine the association between biomarker and self-rated health, physical measure is my covariate.

"weight: respondent level" Is there a "level" of whole sample.


Regards,
Raoping

[Steve Samuels]

Thanks for the link. According to this document the physical measure and hormone data should have been collected on the same subsample, those getting the Enhanced Face to Face Interview (pp 18-19). It appears that some people didn't get both the physical and hormone measures, leading to different weights (pp 26-28).

For your analysis of biomarkers, use the biomarker sample weights on the biomarker subsamples as I said above. People in these subsamples may not have all biomarkers and some may have no physical measures.

Good luck!

[Raoping Tu]

Thanks!

Steve Samuels

Best,
Raoping