Weights for longitudinal analysis
I am trying to use Understanding Society to estimate how changes in workplace pension saving decisions are associated with changes in other individual circumstances (e.g. moving house, changes in marital status, arrival/departure of children etc.). I'm not sure exactly what waves to use for my analysis, and I was hoping you could help.
The pension saving variables I am interested in are in the work conditions module. This is only asked every other wave. For this reason, to calculate changes in pension saving, I am planning on comparing wave t answer with wave t-2 answer. Then I would measure changes in individual circumstances by comparing the same two waves (I don't think I can just use the annualeventhistory module as this only asks about changes over the course of one year, while I am interested in changes over the course of two years). I will then want to look at both these changes each wave (e.g. how many people joined a pension between waves 2 and 4, between wave 4 and 6 etc.), but also looked at this pooled (e.g. over all even waves 2 to 10, what proportion of people who were not in a pension in wave t-2, had become member of a pension by wave t, and how is this associated with e.g. having a child between t-2 and t).
I am a bit unsure which weights to use for this analysis. I have had a look at the FAQs on the website, and I think I should be using longitudinal weights. I think one option would be to use _lw weights given in Understanding Society. For example, when looking at changes between waves 2 and 4, or 4 and 6, use d_indinub_lw and f_indinub_lw respectively, and then when looking at changes between 6 and 8 or 8 and 10, use h_indinui_lw and j_indinui_lw respectively. Would this be correct?
I suppose one issue with this is that e.g. when using f_indinub_lw for waves 4-6, this only gives weight to people who were in all weights 2 to 6 (if I understand correctly). But I am only interested in waves 4 and 6 in this case. Would I have to use my own weights if I wanted to do this more accurately for extra precision? If so, how would I go about that - do I start with wave 4 cross-sectional weights and then estimate a logit, where the outcome is being in wave 6, and adjust the wave 4 weights with the estimated logit probabilities?
Please let me know if anything is unclear!
Many thanks in advance for your help,