I am analyzing national data from the High School Longitudinal Study of 2009, including multiple subsamples selected for specific reasons. I am trying to determine if the normalization of weights for the subsamples effectively adjusts the weight to be nationally representative of the subsample. I would appreciate guidance or resources on this topic.
Marisol -
I'm not familiar with this survey, but if you are talking about the population for a one year or whatever single period slice of a continuing survey, and if the weights are survey weights used for each stratum or subpopulation category, then it should work as follows: Say the subpopulation or stratum for a given category, h, has Nh members, and there is a random (or close to random) sample of size nh, then the weight for that group would be Nh/nh. Or, there could be unequal weighting by size within the group, or if you have auxiliary data, then a calibration weight could be used which is part survey weight and could be part regression weight. There are a lot of possibilities.
Perhaps you could request documentation from the agency which does the survey. That might explain their weights.
(Perhaps you knew all of this and I may have misunderstood your question.)
It sounds like you might be post-stratifying, which means setting up Nh/nh for new categories, after-the-fact, or perhaps you mean something structurally similar where you are rearranging groups. But just be sure that you are not using weights for data in one group and pretending it is for another. Survey weights are supposed to be for a given group and selected randomly (or close to randomly), within that group. You can use models which rely on regressor data on the entire population, and then randomization is not so important, but you have to look at contingencies.
Cheers - Jim
Here is a link to the official document that discusses weighting in this study:
https://nces.ed.gov/training/datauser/HSLS_04.html
- Badges
- Science topic
- Similar topics
- Philosophy
- Philosophy Of Science
- Reasoning