Understanding Society User Support

Nico,

I think the time variable might cause you problems. The weight should be fine for a genuine cross-sectional analysis, but introducing a time dimension changes things a bit. Broadly speaking, the samples issued/interviewed in each calendar year 2011 to 2014 are equivalent, but years 2009, 2010 and 2015 are rather different due to the fact that the "year 1" and "year 2" samples have different components.

In 2009 there is only a "year 1" sample, with the result that Northern Ireland will be over-represented and Bangladeshis will be under-represented (and some other rather minor differences);
In 2010 there will be an over-representation of Scotland and Wales, as the wave 1 weights do not reflect the inclusion of the BHPS sample (as this joined the study only at wave 2);
In 2015 there is only a "year 2" sample, so no Northern Ireland at all, under-representation of Scotland and Wales (as they are boosted in the BHPS sample, which is entirely in year 1), and over-representation of Bangladeshis.

I wonder if you might do some kind of sensitivity analysis by restricting your analysis to 2011-14 and comparing with the unrestricted results? Or maybe restrict to England only, where the issues are more minor?

HTH,

Peter

First, no! You should still use weights, even if restricting the analysis to England only. There are big differences in selection probabilities between ethnic groups and in response probabilities between various socio-demographic groups.

Second, hmm, I realise now that my suggestion is not so simple to implement! The weighting problem that you had will only go away if you restrict the sample to people in households that were sampled in England. This means selecting based on GOR variable from BHPS wave 1 in the case of BHPS sample, and from Understanding Society wave 1 in the case of GPS and EMBS samples.

A simpler approach would be to select based on current wave GOR. This will not remove the sample imbalance problem, but should reduce it to a minimum, so may be good enough.

Nico,

The non-response pattern in your analysis data will be very different from that in a truly longitudinal (balanced panel) sample, so longitudinal weights would not be appropriate. The weight you have created correctly reflects your data structure (6 cross-sections). For example, many of the cases in your analysis will have longitudinal weights of zero.

Peter