I have a question regarding the use of weights when pooling the 9 waves of Understanding Society. At the moment, I simply want to pool all the 9 waves and perform a cross-sectional analysis. Later on, I would like to also use the panel dimension of the data.
I then have two questions about the use of weights:
(1) For the cross-sectional analysis, since I am only using the 9 waves of Understanding Society, can I simply append all the waves and use the associated _xw weights for each wave (namely indinui_xw for waves 6 to 9, indinub_xw for waves 2 to 5, and indinus_xw for wave 1) to create a new weight, say weight_xw? Or would I have to perform any additional transformations? And lastly, if this is true, do I understand correctly that the data should be used in combination with the psu (for example, in Stata: svyset psu [pweight=weight_xw]), and that using the strata information is less important in this case?
(2) When using the panel dimension of the data to perform, for example, a fixed effects analysis, is only the _lw weight in the most recent wave relevant? In other words, I would merge all 9 waves of understanding society using the respondent id, transform the data into long format, and in all analyses use the i_indinui_lw.
Would you say that this approach is correct? Thanks you in advance for you advice!
Updated by Alita Nandi over 3 years ago
- Assignee set to Maria Cotofan
- % Done changed from 0 to 80
(1) Take a look at item 12 in the weighting FAQ which discusses pooling waves for cross-sectional analysis. https://www.understandingsociety.ac.uk/sites/default/files/downloads/documentation/user-guides/mainstage/weighting_faqs.pdf
On behalf of Understanding Society User Support Team
Updated by Maria Cotofan over 3 years ago
Thank you for your quick reply!
Regarding point (1), I was already familiar with item 12 in the FAQ, but I'm not sure it fully answers my question. If I understand correctly, the example discusses how to pool data from one single calendar year using multiple waves ("For example, for a financial year (April to March), months 4 to 15 from wave n can be combined with months 16 to 24 from wave n-1 and months 1-3 from wave n+1. And equivalently for any other period that is a multiple of 12-months.").
However, I would like to be able to pool the data from all the 1 to 8 waves and look at it as a pooled cross-section. I thought this previous discussion was relevant to my question: https://iserswww.essex.ac.uk/support/issues/1257 . The first answer states that "for example observations from wave 1 would have wave 1 weight, and observations from wave 2 would have wave 2 weight and so on. It would be good to scale the weights [...] but this is less important if you are using only UKHLS."
So to sum up, I don't want to pool different waves within the same calendar year, but rather use all available waves/years in the UKHLS and treat them as a pooled cross-section. In that case:
(1) Can I simply use the weight in each wave and create a new weight as suggested here: https://iserswww.essex.ac.uk/support/issues/1257 ?
(2) Is scaling still an issue if I only use UKHLS data and do not include the BHPS?
(3) Would it be problematic if each wave has all the 24 months of data in the pooled cross-section?
I hope I was able to clarify my question and thank you again in advance!
Updated by Olena Kaminska over 3 years ago
Thank you for your questions. I am replying in order:
In your original message:
(1) Yes, your set up is correct - use the most relevant xw weight per wave. Ideally it would be good to scale, but it is less important if you start at wave 1 than if you start in 1991. Also, yes, always correct for clustering (psu) and weighting. Stratification is indeed optional - it is to your advantage and without it your results are more conservative. Your results and conclusions will be correct if you omit stratification though.
(2) The weight you choose depends on what you study rather than the set up of your data. If for example you study between wave change (always using 2 waves for your analysis), you should use weight b_ for waves a-b combination, weight c_ for waves b-c combination etc. This logic extends to any number of wave-combinations you use - always use the lw weight from the last wave for each wave combination.
In your follow up message:
You are right about the relevance of FAQ and #1257 - the latter is more relevant to you.
(2) Don't exclude BHPS when looking at cross-sectional weights in UKHLS - they include BHPS. Just use the data as you described in (1) earlier. Scaling is more important if your analysis starts before wave 1 of UKHLS. Scaling is to compensate for the difference in sample size across waves.
(3) Our data is designed to be used in 24 months chunks - so this makes perfect sense.
Hope this helps,