Project

General

Profile

Support #1721

Random effects analysis and weights (no time*person interaction)

Added by Richard Belcher almost 2 years ago. Updated over 1 year ago.

Status:
Resolved
Priority:
Normal
Category:
Weights
Start date:
07/05/2022
% Done:

100%


Description

Dear Support Team,

I am completing an analysis using Understanding Society Waves 1-9, where I would like to generalize findings to the UK adult population, and to do so I can see weighting is important. I am not looking at changes over time, but would like to use the the repeated responses to the same question an individual can give. Therefore, my analysis will be a multi-level (mixed effects) model, with errors clustered around each person. So each case (row) in the data would be an individual wave response (including proxy responses), with a person indicator column (amongst variables of interest, some household variables). This person indicator would be used as a random effect. I don't want to use a balanced panel as i lose a lot of samples. I am not looking at changes in each year (so wouldn't have a time*person interaction in the random effect or elsewhere). Similar questions were asked previously, but I think there was confusion about whether time AND person were the random effect in interaction https://iserredex.essex.ac.uk/support/issues/1257

My questions:

1. As I am not interested in changes over time (or modelling time*person interactions, just a random intercept or slope for each person), would extracting a cross-sectional weight from each of the waves and matching it to the relevant wave be appropriate? E.g. apply the 'a_indpxus_xw' weight for all responses in wave 1.
2. If my first question is correct, do I need to create my own longitudinal weights? Otherwise the majority of respondents are removed. If so do you have any resources that I can read to set this up.
3. My last question is regarding the specifics of what the US sample is technically representative of, and how the weights relate to this. Am I right in understanding that the weights aim to get back to the original population sample intended in Wave 1 (2009), (excluding the oversampling of ethnic minorities in the boost sample) where the sampling was based on geographic location and socioeconomic status (as described here: https://www.understandingsociety.ac.uk/sites/default/files/downloads/working-papers/2009-01.pdf). This is through adjustment for attrition etc. from the originally intended sample. If not what are the specifics of the population that the sample is getting back to? (e.g. is the weighting based on the proportion of ethnicity from a census year etc.). Apologies if this information is available I could not find it in any documentation.

Thank you for your help,

All the best,

Richard

Also available in: Atom PDF