Outgoing group rotation sample longitudinal analysis

I would like to understand the share of recent union members who are newly retired for 2000-2023.

Because my year 1 population uses the union variable, but my year 2 population variable includes nonparticipation, should I use earnwt or wtfinl?

Because of attrition, I suspect I should use a longitudinal weight. I think I can use LNKFW1YWT if the sample is adjusted for wtfinl. But if the sample is adjusted for earnwt, I think I would need to create my own longitudinal weights. Is this correct?

Or, is it possible to do a monthly analysis using MISH 4 and MISH 5 waves, finding the share of new retirees in MISH 5 among people who were in unions 8 months prior (MISH 4)? And this could use LNKFWMIS45WT?

Also, once I have my longitudinal weights, should I use year 1 weights or year 2 weights?

As of now, I keep only mish 4 and 8.

I keep only cpsidvs that are present for both mish 4 and mish 8.

I reshape the data from long to wide.

I keep observations where there are union members in mish 4.

collapse (mean) newret [pw=lnkfw1ywt4], by(year8)

Where newret are people who are retired in mish 8 and employed in mish 4.

Thank you!

There is no official guidance from the Census Bureau or BLS on how to apply weights to this type of analysis. I would recommend creating your own longitudinal weight that is based on EARNWT.

In general, users should apply the most restrictive weight to their analysis. In this case, EARNWT is more restrictive, since it adjusts for the probability of remaining in the CPS long enough to be part of the Earner Study, and the probability of meeting the criteria of the Earner Study (age and employment status). EARNWT should be used with any analysis that uses variables from the Earner Study.

You are correct that EARNWT should also be adjusted for the probability of being observed in the CPS at multiple points a year apart. This is necessary because some respondents drop out of the CPS due to moving, nonresponse, or other reasons between the two observation points you are using. You can modify EARNWT to be a longitudinal weight by raking EARNWT by running a version of the do file linked on this page on linking and the CPS. In the Stata script, you will need to replace WTFINL with EARNWT.

Using year one versus year two weights is a question about whether you prefer a forward-looking versus backward-looking weight. A forward-looking weight should be used if you are interested in individuals observed in time x+1 and leveraging characteristics from time x. A backward-looking weight should be used if you are interested in individuals observed in time x and leveraging characteristics from time x+1. This IPUMS forum post from a few years ago on longitudinal weights in the CPS, particularly the responses from Grace Cooper, may be useful.

I am not aware of anyone using, testing, or verifying this method, but after speaking with my colleagues on the IPUMS User Support team, this is the solution we recommend starting with.

Great, thank you! This has been so helpful.

As a follow-up, the do-file is for a consecutive month longitudinal analysis.

For my purposes of only looking at two observations 12 months apart per individual (4th and 8th MIS): would I input an extract for every month between Jan 2000 and Dec 2023, restrict to if MISH==4 or ==8, and follow the same do-file? (With changing wtfinl to earnwt, of course). And then, during my analysis, divide the weights by 12 to transition from monthly to annual?

Or, would significant changes need to be made to the do-file to an annual panel?
Another way to phrase the question might be: do the LNKFW1MWT and LNKFW1YWT do-files differ significantly, apart from dividing by 12?

Thank you again for all your help.

That is a good question. The do file linked allows you to create a longitudinal version of EARNWT for observations one month apart. Since you are linking observations that are 12 months apart, you could adapt the sample code a bit. This is a bit beyond our expertise in weighting and we do not have sample code to do exactly what you are hoping to, so I’m sorry that I cannot provide very specific guidance. But essentially, you would make a flag to identify the people in MIS (month in sample) 8 and in MIS 4 and the people in MIS 8 but not MIS 4. Then you would inflate EARNWT for the people in MIS 8 and MIS 4 to the sum of EARNWT for everyone in MIS 8.