Understanding Society User Support

Tom,

Thank you for your questions.

1. Yes, different weights. If you use multiple waves in your analysis (e.g. comparing wave 4 to 6) you will have to use a longitudinal weight. For pooled analysis xw weight seems to be sufficient. Rescaling is advisable.
2. No bearing on rescaling, but this means you need a longitudinal weight;
3. Really doesn't matter which number you rescale to - you just need the same weighting total for each wave after rescaling to ensure that each wave contributes the same amount of information to the total analysis;
4. No adjustments for calendar years needed. Also, see below for the advice on simpler rescaling.
5. No, weights do no specifically take seasonality into account - they are for the analysis of the whole wave. While nonresponse may differ over seasons, in our weights it is the overall effect that is accounted for. Seasonality, and therefore nonresponse related to it, would be important if you study different months of the year separately. The approach would be similar to the one described for a calendar year in FAQs, and if you are concerned about seasonality of nonresponse and it's influence on your estimates, there would be a possibility to create your own tailored weights to account for this effect specifically. This would not be necessary though if you are planning to use the whole wave in your analysis (instead of splitting it by seasons).

Hope this helps,
Olena

Olena,

Many thanks, this is very helpful. However just three clarifications / further queries ...

First, apologies but I can't seem to see the "advice on simpler rescaling" you mention in your response to Q4 - could you forward/resend?

Second, I think I have been unclear in how I explained my first type of analysis ("Descriptive statistics of changes over time across every other wave, ie comparing Wave 4 vs. 6 vs. 8 vs. 10...;"), hence why you've recommended a longitudinal weight here.

To clarify: I'm carrying out cross-sectional time series analysis, ie creating a cross-section of each wave and seeing how the job quality scores, percentage deprived, etc change in each cross section. I'm not doing any longitudinal analysis: ie what is the score of person X in wave 4, person X in wave 6, etc. The UKHLS user guide seems to advise a cross-sectional _xw weight for my kind of analysis - eg see the Stata code on p.25, which happens to mention the exact weight I use: https://www.understandingsociety.ac.uk/sites/default/files/downloads/documentation/mainstage/user-guides/bhps-harmonised-user-guide.pdf.

Taking all this together, my understanding - based on what you've said - is I should therefore use the same weight for all of my analysis (indinub_xw), but with rescaling for the pooled data.

Finally, you're correct to say that I won't be studying months of the year separately, only wave-by-wave or pooled waves. My question was more about the implications for time series analysis if, say, people in Wave 4 are more likely to be interviewed later in the year than people in Wave 6, etc. My initial analysis suggested that there is quite a bit of wave-by-wave difference in when people are interviewed, and I suspect this will be especially difficult for waves during the pandemic, should I analyse them? However, I'm fine with your conclusion regardless.

Many thanks once again for your advice.

Best wishes,

Tom

Olena Kaminska wrote in #note-1:

Tom,

Thank you for your questions.

1. Yes, different weights. If you use multiple waves in your analysis (e.g. comparing wave 4 to 6) you will have to use a longitudinal weight. For pooled analysis xw weight seems to be sufficient. Rescaling is advisable.
2. No bearing on rescaling, but this means you need a longitudinal weight;
3. Really doesn't matter which number you rescale to - you just need the same weighting total for each wave after rescaling to ensure that each wave contributes the same amount of information to the total analysis;
4. No adjustments for calendar years needed. Also, see below for the advice on simpler rescaling.
5. No, weights do no specifically take seasonality into account - they are for the analysis of the whole wave. While nonresponse may differ over seasons, in our weights it is the overall effect that is accounted for. Seasonality, and therefore nonresponse related to it, would be important if you study different months of the year separately. The approach would be similar to the one described for a calendar year in FAQs, and if you are concerned about seasonality of nonresponse and it's influence on your estimates, there would be a possibility to create your own tailored weights to account for this effect specifically. This would not be necessary though if you are planning to use the whole wave in your analysis (instead of splitting it by seasons).

Hope this helps,
Olena

Tom,

Yes, sounds like cross-sectional weight will be suitable. After wave 6 it will be 'ui' instead of 'ub'.
You don't need rescaling if you are not pooling information into one model (e.g. separate cross-sectional analysis for each wave).

Seasonality should not be a problem with UKHLS - the interviews are carried over 2 years in all seasons. There are sampling months that are issued in each months of this 24 months period. Covid didn't create a problem with this regard, so most definitely use those waves.

Rescaling:
In pooled analysis and sometimes in other types of analysis you may need to apply an additional scaling to our weights. Our weights have a mean of 1 in each wave, which means that if combined in a pooled analysis the waves with smaller sample size will have a smaller contribution in your analysis. This includes BHPS waves and later waves (as sample size decreases with attrition). Ideally, when combining events / states over 30 years (for example) you want each year to have the same importance. To ensure this follow this example to calculate an additional scaling for your weights.
For example, you are looking at job quality and therefore are pooling information from wave 2, 4, 6 & 8 as these are the waves when the questions are asked. Here is how to create a scaled weight for this analysis.

ge weightscaled=0
replace weightscaled=b_indpxub_xw if wave=2

ge ind=1
sum ind [aw=b_indpxub_xw] if wave=2
gen bwtdtot=r(sum_w)
sum ind [aw=d_indpxub_xw] if if wave=4
gen dwtdtot=r(sum_w)
sum ind [aw=f_indpxub_xw] if if wave=6
gen fwtdtot=r(sum_w)
sum ind [aw=h_indpxub_xw] if if wave=8
gen hwtdtot=r(sum_w)

replace weightscaled=d_indpxub_xw*(bwtdtot/dwtdtot) if wave=4
replace weightscaled=f_indpxub_xw*(bwtdtot/fwtdtot) if wave=6
replace weightscaled=h_indpxub_xw*(bwtdtot/hwtdtot) if wave=8

You can double check by looking at the sum of ind with weightscaled for each wave – it should be the same.
sum ind [aw=weightscaled] if wave==2
sum ind [aw=weightscaled] if wave==4
sum ind [aw=weightscaled] if wave==6
sum ind [aw=weightscaled] if wave==8