When pooling two different samples from the same population together do have I to adjust my weights?

Question

I have two different data sources (different surveys) from two different years (2010 and 2012) which are on the same population. As both surveys are representative of the population and have questions that are completely the same I am tempted to use both datasets and add the year variable to my unique database to assess the effect of time on my outcome.

Both datasets provide cross sectional weights and PSU. If I want to pool together the two datasources do have I to adjust the weights? If so, how can I do it?

Both surveys give representative estimates of the population and the way the weights were calculated is pretty much the name, but in one case the survey contains responses from around 10000 individuals, in the other case around almost the double. Therefore, the two sample size are quite different.

My question is: when pooling together the two datasets can I use the weights the way they are provided or shall I do some sort of adjustment based on N of both databases?

edit: an example of a typical question I would be interested to look at is whether there has been a decrease/increase in the number of specialist outpatient visits in the same age categories after controlling for socio-economic variables (e.g. education level, job status)

Thanks

Answer 1

This is a neat question, and there has been some methodological work on this.

First, one semi-natural way to incorporate several data sources is with Bayesian methods, where the posterior based on the data from Source 1 can act as the prior for the analysis of the data from Source 2. As an example, Raghu et. al. (2007) combine the data from two health surveys, one areal face-to-face (National Health Interview Survey), and one landline phone (Behavioral Risk Factor Surveilance Survey) using a Bayesian model which additionally correctly for (a known) lower quality of BRFSS frame which did not cover neither the cell-phone-only nor the no-phone-service population.

Second, in some situations where you have the same population and the same variable, you can jointly calibrate the two surveys to ensure that the estimates based on both are identical, and improve the precision of each. Changbao Wu published a few papers on this based on the empirical likelihood calibration idea (this and that), but this can also be achieved by less computationally intensive GREG approach. I vaguely recall a JASA paper from late 1990s to mid 2000s, but I can't seem to locate it; instead, my Google searches stumble upon this and that.