I have a really basic question about weights though I want to make sure that I'm doing my analysis correctly. Most of my analysis involve the sample of respondents from the last wave (h_). However, sometimes I take into account the responses that those individuals interviewed in wave h_ gave in previous waves. At the moment I am using the cross-sectional weight from the last wave h_, but I'm not sure if this is correct. Should I use a longitudinal weight if I use their responses in previous waves? (even if I only consider those individuals inteviewed in the last wave)
Regarding the issue of weighting, I think I underestimated the complexity of it in the UKHLS. Do you have any slides from a past course about this topic that could be shared online?
#1 Updated by Olena Kaminska 7 months ago
If in your model you only use information from wave h then you can use xw weight. But if you use any information from other waves you need to select longitudinal weight. I suggest that you don't select respondents based on their responses - the weight would do the job for you correctly.
You may find this short video about our weights useful:
#4 Updated by Marina Fernandez Reino 7 months ago
In my model I use respondents from wave 7 only, but I consider the answers that wave 7 respondents gave in previous waves. In particular, I am looking at one of the questions in the harassment module: 'In the last 12 months, have you felt unsafe in any of these places?'. By consider the answers that wave 7 respondents gave in previous waves, I somehow expand the question so I can identify whether any of wave 7 respondents have reported feeling unsafe since 2009 - when the survey started-. That is, I am not considering any respondents from previous waves that have not participated in wave 7 nor I am doing any longitudinal analysis. I think I should use a cross-sectional weight, but I'm not entirely sure.
#5 Updated by Olena Kaminska 7 months ago
Thank you for the detailed explanation - this always clarifies the question. Yes, you definitely have to use longitudinal weights. It doesn't matter where you get respondents from - what matters is which waves you get their responses from. You use responses from previous waves and wave 7, so your analysis should use wave 7 longitudinal weights.
There is one potential exception - if you don't care whether you have information from previous waves. For example if for people who respondent in previous waves you classify them, but for people who don't have any information from previous waves r have partial information you either impute this information or you put them into a separate 'missing' category. If this is the case you could use wave 7 cross-sectional weight. In this situation basically you would be dealing with nonresponse in an alternative way.
Thank you again for clarifying your question,