Project

General

Profile

Support #1590

Creating a weight variable for fixed effects analysis

Added by Karen Mak about 2 months ago. Updated 25 days ago.

Status:
Feedback
Priority:
High
Category:
Weights
Start date:
10/08/2021
% Done:

90%


Description

Dear UKHLS support,

Hope you are well.

May I ask for your advice on whether you think weighting should be applied in a fixed-effects analysis?

I would like to use data from Waves 2, 4, 6, 8, and 10. From my understanding, if we use the longitudinal weight provided in the Wave 10 data (j_indscub_lw), it will only retain the sample who completed waves from Waves 2 to 10, which could lead to a substantial drop in the sample size.

I have read the paper "Weighting and Sample Representation: Frequently Asked Questions", which was really useful and clear, and learnt that we can derive our own weight variable. However, I am not sure how to do so with a longitudinal dataset (in a long Stata format).

I would be very grateful to have your advice on (i) whether or not weighting is necessary for a fixed-effects model, and (ii) if so, what would be the best way to derive a weight variable for this study.

Best wishes,
Karen

#1

Updated by Understanding Society User Support Team about 2 months ago

  • Status changed from New to Feedback
  • % Done changed from 0 to 50
  • Private changed from Yes to No

(1) This is an analysis question. This is a good discussion about why and when we should weight. http://jhr.uwpress.org/content/50/2/301.refs
(2) Please email us at and we will send you the guidance for producing your own weights.

#2

Updated by Olena Kaminska about 2 months ago

Karen,

Yes, weighting is necessary in fixed effect models.

Look for similar questions about pooled analysis on this forum - they may answer your question.

If you need more help, please give us details on what you want to estimate (e.g. people or events etc.)

Thanks,
Olena

#3

Updated by Karen Mak about 2 months ago

Hi Olena,

Thanks so much for your prompt reply and for your help.

I tried to look for similar questions on this forum, but couldn't seem to find any solutions.

I am interested in how changes in volunteering behaviours are associated with changes in wellbeing. Questions on volunteering were asked in alternative waves (i.e. Waves 2, 4, 6, 8 & 10). I have now merged/append all relevant waves into one dataset in a long format.

(1) I'd like to ask whether it would be more appropriate to use the wave 10 longitudinal weight (j_indscub_lw), or to create a specific weight for the analysis?

(2) If creating a new weight is more appropriate here, what would be the steps to create one in a long data format? Do I need to reshape it to wide format -> then create a binary variable "response" (1=completed all 5 waves, 0=only completed Wave 2) -> run a logistic regression predicting "response" using predictors (e.g. age, gender, martial status) on a condition that participants not known to have died/emigrated and that participants have a Wave 2 weight value greater than 0 -> then generate a new weight = gen weightW25 = (1/p)*b_indscus_lw ?

Thank you and best wishes,
Karen

#4

Updated by Olena Kaminska about 2 months ago

Karen,

There is no particular reason to create your own weight.
To help you select a weight, please explain how you analyse the data: are you looking at a change between 2 waves, or 3 waves at a time?

Thanks,
Olena

#5

Updated by Karen Mak about 2 months ago

Hi Olena,

Thank you for your reply.

I am looking at a change between 5 waves (Waves 2, 4, 6, 8, and 10) at a time using fixed effects modelling. Is it correct to choose j_indscub_lw as the weight for my analysis?

Best wishes,
Karen

#6

Updated by Olena Kaminska about 2 months ago

Karen,

Do you mean you study change between wave 2 and wave 10?
Or do you mean you study change between wave 2 and 4, and separately between 4 and 6 etc.
What is your outcome variable?

Thanks,
Olena

#7

Updated by Karen Mak about 2 months ago

Hi Olena,

So sorry for the confusion. We are interested in studying changes between waves 2 and 4, and separately between 4 and 6 etc. The outcome variables are GHQ12 and SF12.

Best wishes,
Karen

#8

Updated by Olena Kaminska about 2 months ago

Karen,

In this situation use longitudinal weight for wave 4 for the set of waves 2 to 4 outcome, lw weight for wave 6 for the set of waves 4 to 6 outcome etc. In a long format you will create a new weight variable and give it respective values.

Hope this helps,
Olena

#9

Updated by Karen Mak about 2 months ago

Hi Olena,

Thanks so much for your reply and incredible help!

I think I might have misunderstood your question earlier. Just wanted to confirm with you that if I wanted to explore the changes from Wave 2 to 10, I should be using Wave 10 lw weight?

Best wishes,
Karen

#10

Updated by Olena Kaminska 26 days ago

Karen,

Yes, that's correct.

Olena

#11

Updated by Understanding Society User Support Team 25 days ago

  • % Done changed from 50 to 90

Also available in: Atom PDF