Understanding Society User Support

Hi, I think I've figured this out, thanks.
I realised I don't need to change the longitudinal weights since all my observations are independent, i.e. separate people.

Dear Spencer,

This should answer your question:
Pooling can be cross-sectional or longitudinal. Theoretically, you will be combining ‘separate samples of events / states’ each of which will have the corresponding weight.
If you are just interested in events (and what happened at the same time / wave) you are looking at pooling cross-sectional information. For this create a new weights variable new_weight, and give it a value of the cross-sectional weight from each wave (e.g. new_weight=a_indinus_xw if wave==1; new_weight=b_indinub_xw if wave==2 etc.)
Alternatively you may be interested in what happens before and / or after a particular event (e.g. studying work pattern for 3 years before birth of a first child and 3 years after for new mothers). In this situation you need to choose a longitudinal weight from the last wave in your analysis for each combination of waves (e.g. for birth at wave 3 where we observe waves 1-6, the weight will be f_indinus_lw; for birth at wave 4 with information in the model of waves 2-7 –it will be g_indinub_lw etc.).

Hope this helps,
Olena

Hi Olena,

Thanks, this is really helpful. This is indeed what I was trying to do:

Alternatively you may be interested in what happens before and / or after a particular event (e.g. studying work pattern for 3 years before birth of a first child and 3 years after for new mothers). In this situation you need to choose a longitudinal weight from the last wave in your analysis for each combination of waves (e.g. for birth at wave 3 where we observe waves 1-6, the weight will be f_indinus_lw; for birth at wave 4 with information in the model of waves 2-7 –it will be g_indinub_lw etc.).

However, why not just use the latest wave for everyone? To use your example - and assuming there are no more groups than what you mention, so overall we have data from wave 1 to wave 7 - why not just use the wave 7 weight for everyone?

If I take your approach, then if I want to look at studying patterns in the third year after the birth of a first child, I would be combining f_indus_lw (the first cohort) and g_indub_lw (for the second cohort), and similarly for the other time periods (2, 1, 0, -1, -2, -3). Wouldn't that be bad practice? Why not just use g_indub_lw for both cohorts?

P.S. - I should have clarified, I am not using wave 1 in my work. I mention this because in your example there was a distinction between indus and indub, but since I am only using waves 2 onwards it would only be indinub, if I've understood correctly.

To make it more concrete, I looking at people who start doing care work in waves 4, 5, and 6. I want to track their earnings over time, in years -2 (two years before starting care work), -1, 0 (year of starting care work), 1, 2, 3, 4, and 5 (five years after starting care work). So, for the cohort starting care work (t = 0) in wave 4, I need at least waves 2-9, for the cohort starting in wave 5 I need 3-10, and for the cohort starting in wave 6 I need 4-11.

So the question is whether I just use k_indinub_lw for all three cohorts (what I've done) or whether I use i_indinub_lw, j_indinub_lw, and k_indinub_lw respectively. I suppose the latter option would increase the sample as it would not be corrected for attrition in years that I don't require, but it would involve combining weights as I allude to above.

Dear Spencer,

Yes, you are completely right. Both approaches will be correct and should give you similar results. The second approach will give you higher sample size, so more statistical power to detect significant effects. In your particular situation the sample size difference is not very large as you use so many waves in your analysis, so you could choose the first option (only one weight), and only switch to the second option if you have a marginally significant result...

Hope this helps,
Olena