Weighting binary response based on certainty

[Sarah Thorne]

Hi everyone,

If I have a binary response variable (yes/no to a hypothetical referendum on a survey) and included certainty scales, is there any way to estimate a logit model using certainty as a weighting variable? If so, which type of weight would be appropriate? I am only familiar with pweights and fweights, neither of which seems like the right choice in this case.

I have multiple observations per person, so the code is

logit vote tax, vce(cluster ID)

Should I be running a panel data model using xtlogit instead?


Thanks for any advice.

[Mike Lacy]

This sounds like an "importance weight," about which see -help weights-. At least syntactically, it looks to me like using [iweight = certainty] would work. It also occurs to me that using certainty as an explanatory variable (though admittedly a different thing) might be substantively relevant.

Yes, I'd use -xtlogit-, since presumably there's a substantial person-level effect regardless of the effect of your tax or other explanatory variables.

[Sarah Thorne]

Thanks Mike! It occurred to me that when using the panel data model, the weight needs to be the same across all observations of one person, which isn't the case. However, when using the non-panel form of the logit model adding certainty as an importance weight leaves results almost identical results with or without the weights. I'm not quite sure what that exercise tells me other than that importance weights based on certainty scale ratings might not be a great way to adjust for hypothetical bias.