The US Census Bureau's survey of income and program participation (SIPP) has, among others, the following variables:

  • epdtbhn: Paid job during the reference period
  • ersnowrk: Main reason for not having a job during the reference period
  • tpmsum*: Earnings from job received in this month
  • eeno*: Across-wave employer index/number Unique job number that will remain the same from wave to wave.

Now, I'm looking at the following individual from the 2008 wave:

br ssuid ersnowrk epdjbthn tpmsum*  eeno* if ssuid == "019925011535"
  • In May 2012, that individual was reported as Unable to find work, epdjbthn == No.
  • In the next month, epdtbhn == Yes: He/she had a job (consistent with ersnowrk == Not In Universe). However, all the employment variables (tpmsum*, eeno*) are also all not in Universe.

This is clearly inconsistent data. The allocation flag for epdjbthn says that the data was not imputed. The allocation flag for the employment variables is meaningless (as they are not in universe). Does this mean I can rely on the person to have found an actual job, even if none of the other employment variables are speaking towards it?

  • $\begingroup$ Classic inconsistency example in survey data. Had many like this when using BHPS for UK. Does Not in Universe also stands for missing data? If you are interested in studying wage related issues, this is not a useful observation, and the most you could do is evaluate if there is some type of missing-data selection, based on non-missing characteristics. The other option is to look at job histories. As your data seems to be a panel, maybe they include such variables. The BHPS at least is full of them, and allow you to evaluate these issues. $\endgroup$ – luchonacho Apr 21 '17 at 8:07
  • $\begingroup$ I've edited the question to make it clear that it relates to the US as SIPP may have other meanings elsewhere (in the UK it stands for Self-Invested Personal Pension). $\endgroup$ – Adam Bailey Apr 21 '17 at 9:05
  • $\begingroup$ @luchonacho I want to measure hiring and firings, and distinguish between a hiring and an exit from the labor force. $\endgroup$ – FooBar Apr 21 '17 at 11:39
  • $\begingroup$ Mmm, then these might be indeed important observations. Do you have job histories? Other variables like firm size, occupation, industry, or something else that indicates whether the person is in fact employed? You can categorise your data by quality, like: full info, partial info, no info, and run your analysis for different subsets. This might be an interest comparison. In any case, I would still provide some basic selection analysis like a Heckman or so, to check for selection bias. $\endgroup$ – luchonacho Apr 21 '17 at 12:28

Your Answer

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

Browse other questions tagged or ask your own question.