1
$\begingroup$

I have a microdata recipient survey and a macro (aggregate) donor. How can I fuse the (binary categorical) data? The Statistical Matching techniques/software I'm familiar with are micro-to-micro. Furthermore, a web search revealed nothing (but perhaps my terms were wrong).

One (possibly over-simplified/naive) approach is:

  1. Build a probability distribution of the donor's response in question, broken out into (harmonized) common variables like age, income, etc...

  2. A respondent in the recipient that most closely matches donor profile X (e.g.: Nearest Neighbor) is assigned a "yes" response with probability X (based on #1).

Any pointers, comments on the above, references and software (R is a good one) are greatly appreciated.

$\endgroup$
  • $\begingroup$ I'm interested in what your goal is here. Are you trying to match recipients to donors? Or determine the likelihood that two individuals will be matched given information on each of them? I'm going to guess that you don't have any information on how donor attributes are correlated, if you only have summary statistics. Do you have any other information that might be useful here? $\endgroup$ – Matt Jun 7 '14 at 18:41
  • $\begingroup$ Unfortunately, I only have summary statistics for the donor (hence no info on correlations), but observations for the recipient. I'm trying to enrich the recipient set with some donor variables such that the summary statistics match. Put another way, how can one do statistical matching when the donor is summary? $\endgroup$ – R. Barzell Jun 9 '14 at 11:41
2
+50
$\begingroup$

The most conservative approach here would be to aggregate your micro data into summary statistics, and test for differences between the groups. I think that the reason you can't find anything is because most people in this situation would step back and do macro-macro comparisons. You've lost information from the donor dataset that won't come back, so I'd be very wary of trying to reconstruct the original set from your summaries.

My instinctive first response was to construct a set of pseudo-observations by generating 'samples' from your summary statistics, using correlation information to match the pseudo-observations more closely to the real ones from which you found your summary statistics. Of course, that's not a useful approach here.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks! Although the desired output is micro data, this is used to generate on the fly summary data (by filtering on variables). As such, this can work if I do on-the-fly fusion prior to presenting the summary results. Do you have some links, pointers or suggested terms I can Google? $\endgroup$ – R. Barzell Jun 10 '14 at 14:17
  • $\begingroup$ I'm afraid I don't have any firsthand recommendations for you. On-the-fly data fusion, dynamic data fusion, and, of course, 'data fusion' should help if that's the route you're going to take. Again, I'd like to qualify all this by noting that there isn't anything to fuse on one side, at least not to my eye. Although, if you have information on subsets of your donors (which seems unlikely in your case), that would be closer to your true goal. Best of luck. $\endgroup$ – Matt Jun 10 '14 at 17:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.