Weighting by reciprocal of group size

[Henry Cust]

Hi,

I have a dataset with multiple responses per group where each group size differs. I wish to run some analysis at the group level, accounting for the different number of observations per group through generating weights, but I am not sure the best way to go about this. In other words, I want the contribution of each observation in a group with 3 members to be 1/3, and the contribution of each observation in a group of 10 members to be 1/10.

Would generating the reciprocal of group size and applying one of the weighting categories be sufficient? If so, which type of weighting would be suitable? Would this have implications for the parameters estimated?

Unfortunately, the data does not lend itself to collapsing by group and restructuring the data to run the analysis.

Any help would be appreciated.
Henry

[David Radwin]

Please see https://www.statalist.org/forums/for...ights-in-stata

[Leonardo Guizzetti]

Just like your question on Reddit, you don't provide sufficient information to give a targeted response. Without understanding the intent of your analysis, and assuming analysis of continuous data using linear regression, one approach could be to analyze average values for each subject and use the number of observations as the analysis weight.

[Henry Cust]

Hi,

Apologies for coming back to this so late.

I have two types of analysis I wish to perform. 1) basic summary statistics - this I would typically do by collapsing into an observation per group using means and summarising no problem, 2) logistic regression - here lies the problem that my outcome variable is binary so collapsing would give proportions hence I want to find a weight that will adjust the contribution of each observation by the number of observations in the group.

Henry