# Using activity level weights for multi-variate analysis with ATUS WB module.

Hi,

This is a follow up to a previous question I asked about using activity-level weights with the ATUS wellbeing module (here). Thank you for your answer then. I am still confused about on a couple of points.

(1) In your response you advised multiplying the awbwt value by the amount of time in the activity of interest (in my case childcare). But in the codebook it seems to say that the activity-level weight should be multiplied by the amount of time spent in ALL activities eligbile for the well-being module. I would appreciate any clarification you could offer on this.

(2) From what I can tell, the formulas for estimating affect during a particular activity that are laid out in the handbook are aimed at producing estimates of average affect during a particular activity (e.g. in the handbook average pain while at work). But they don’t seem to be designed for use in multi-variate analysis, (e.g. what characteristics predict experiencing more or less pain while at work). Is this interpretation correct? What weights should I be using in a multi-variate analysis to predict affect during a particlar activity? Perhaps this is when I need to multiply by the amount of time in the activity of interest?

Thank you for your help. It is very much appreciated.

For (1), you are correct. AWBWT should be multiplied by the time spent on all activities eligible for the Well-Being module. For (2), if you are using one observation per Well-Being module respondent, along with their average affect during a certain activity, then you should use the WBWT weight for your multivariate analysis.

To summarize, use the AWBWT method described in the previous answer, modified as mentioned in (1), to generate your average affect. Then, use the WBWT weight for your multivariate analysis of this average affect.

Hope this helps.

Thank you very much, Tim. I am using all Well-Being activities that are coded as childcare. This means that some respondents contribute more than one observation to the sample. I’m clustering to account for that. Can I still use the person-level weights in this situation?

Thank you.

Since it sounds like your multivariate analysis is at the level of the activity, then you should use the AWBWT weight instead of the WBWT weight. Note that these are the AWBWT values found in the ATUS-X data, not the adjusted values you used to calculate average affect.

Hope this helps.

Tim,

Thank you again for your help. I hate to belabour the point but I want to be sure that I’m doing everything correctly. There are substantial differences in the results of my analyses depending on whether I use the AWBWT or the WBWT so it matters which I choose.

I’m doing a multivariate analysis predicting affect levels during certain activities (in particular childcare). I want to determine how activity-level variables - the location or timing of the activity, who was present during the activity, and so on - are related to affect. I have searched for other papers using the affect measures in this way but I have not found any that apply activity-level weights to their multivariate analyses. Do I need to use the activity-level weights for this analysis?

Thank you again.

AWBWT takes WBWT and makes two adjustments (see Appendix B). First, it accounts for the probability that this activity was selected for the affect questions instead of another of the respondent’s eligible activities. Second, it accounts for the length of time of the activity, relative to the time spent on other eligible activities. Since the ATUS weights are probability weights (i.e. represent the inverse probability of a case being selected into the sample), using WBWT for your analysis would erroneously assume that all ATUS respondents had the same number of eligible activities and that each eligible activity was performed for an equal amount of time.

To put this another way, AWBWT is “weakening” the impact of activities on your results that had a low probability of being selected and/or were of a relatively longer duration. Presumably, the second adjustment is due to the fact that a well-being measure will be less representative of any minute of the activity, the longer this activity persists.

Finally, it is extremely important to specify in your statistical package that AWBWT is a probability weight.

Hope this helps.