Understanding Society User Support

i now realise there arent any youth longitudinal weights, i do not think id have the capabilities to create my own weight given my time constraints. Could you please suggest another alternative? would i be able to use the cross sectional youth weight? or perhaps the household longitudinal weight as that is the exposure

Thanks very much for this. I looked through the previous queries and it seems that when a custom weight is not created, the longitudinal enumeration weight (psnenus_lw) is usually recommended. Would this be suitable for analyses focused on youth outcomes? I understand that it adjusts for attrition across waves, but I am unsure whether it is designed specifically for youth samples? in practice, do researchers typically use this weight for youth-focused analyses?

Attrition is something i want to account for issue in this study, and I believe psnenus_lw accounts for general dropout over time. However, I am not sure whether it captures the structural missingness in the youth files as participants age out of eligibility?

Finally, if psnenus_lw is used, should it be applied to the final merged dataset after linking the household, adult, and youth files?

Akansha,

Thank you for your question. Ideally you would create your own tailored weight following the training here: https://www.understandingsociety.ac.uk/help/training/creating-tailored-weights/ .

But if you have time constraints, and as you correctly pointed, you could use a longitudinal person weight. I suggest you use o_psnenui_lw in your analysis. Note, the weight comes from wave 15, and note, it is ui and not us (higher sample size).

If you were to create you own weight you could start with m_pwnenui_lw, and predict nonresponse in one regression from wave m to being in your model. Note, you would want to limit in the model to eligible people (3 years of age, I think, who would be 10-13 in wave m, as they still have to be able to participate in wave 15 in youth questionnaire).

Hope this helps,
Olena

Thank you so much for your advice on this. I tried the weight you suggested, but it appears to assign a weight of 0 to a substantial number of participants, which reduces the analytic sample size by nearly half.

I’m wondering whether there is a more general population or cross-sectional weight I could use instead—e.g., a weight that accounts for the sampling design and any over/under-representation of certain groups. Since I have imputed outcomes for participants who were age-eligible for the Wave 15 youth survey, I’m not sure whether a longitudinal weight is still necessary in this context.

Could you advise which population/cross-sectional weight would be most appropriate here, and what adjustment it would be making in practice? I looked through the documentation, but I wasn’t confident I was identifying the best option for this specific scenario. Also, do the population/cross-sectional weights tend to include zero weights as well?

Many thanks again,
Akansha

Akansha,

You could try to start with f_psnenui_li and model nonresponse yourself, if you like. You could either create a tailored weight or model nonresponse through imputation (but do remember to use multiple imputation for correct variance estimation). Note, if you are imputing 50% of responses you should be very clear about this when presenting results. I also would suggest to double check the results with just o_psnenui_lw and with the new nonresponse correction. By expectation the results should be the same, and slightly smaller confidence interval.

I suggest you listen to the first video (10 minutes) in this course:
https://www.understandingsociety.ac.uk/help/training/creating-tailored-weights/

Hope this helps,
Olena

Hi Olena,

Apologies for the confusion, I’ll try to clarify my setup more clearly.

My analysis uses youth Waves 13 and 15 only (with household and adult covariates). I first used multiple imputation by chained equations to impute outcomes at Wave 15 for participants who were present at Wave 13 and age-eligible for the Wave 15 youth survey. I then ran the regression using the o_psnenui_lw longitudinal weight, but noticed that a substantial portion of the sample receives a weight of zero, which effectively reduces the analytic sample size quite a lot.

I wanted to ask what you meant by modelling nonresponse through multiple imputation. Do you mean that this would be an alternative to using the longitudinal weight, or that both approaches should be used together?

Given that I have already imputed Wave 15 outcomes for those age-eligible, I was also wondering whether a more general population or cross-sectional weight might be more appropriate in this situation — for example, one that adjusts for sampling design and representation of different groups, rather than longitudinal follow-up. In other words, I am unsure whether a longitudinal weight is still necessary once the Wave 15 outcomes have been imputed.

Akansha,

Thank you. I understand you use youth children, who are presumably 10-13 in wave 13 and become 12-15 in wave 15, and use youth questionnaire answers in both waves. Your sample size is very small because the analysis is restricted to 3 age year groups, and relatively lower response rate to youth questionnaire. We do not have a specific weight for you, and youth weight provided is not suitable as you are using a longitudinal analysis.

You best approach would be to start with m_psnenui_lw and model nonresponse in your analysis as per tailored weight training, or you could use o_psnenui_lw as a suboptimal weight. To have the highest sample size you could use f_psnenui_li and model nonresponse from wave f. For this you could use either imputation or additional weighting correction. This would be on top of our weight that you will use, as this accounts for only a small (though important) fraction of total nonresponse.

The imputation that you've done is ok, but additional correction would be needed to fill the gap between our enumeration weight (psnen) and the youth questionnaire.

Hope this helps,
Olena

Hi Olena,
Thanks for this, i will attempt to do what you say regarding creating a tailored weight, i am wondering if any resources exist in R as i am not a STATA user and the module seems to provide most guidance in STATA

also would it be sufficient to account for psu and strata? or would this need to be done alongside the weights

i also realised that some psnenui weights are NA (95 to be precise) for my sample, is that typical?

Akansha,

Yes, we have material in R, which we can share. Could you email to "UKHLS User Support" <usersupport@understandingsociety.ac.uk> with the request, please?

I am not sure what you mean by only accounting for psu and strata. In any analysis you should always account for psu, strata and weights. Weights are never optional.

Also, not sure what you mean by NA for psnenui weights. Can you be more specific?

Hope this helps,
Olena