I have a panel survey of self-assessed levels of household resilience in three nearby villages, with observations collected over the course of 10 separate survey waves (observations are collected on a continual basis, with each full wave requiring 6-8 weeks to finish, before the next one then starts up again).

By linking this with observations from a single weather station that is close to all three villages, I can create a dataset that matches household resilience with the weather on the day of each interview.

What I would like to do is examine the how changes in weather within a season, and within a year, affect self-assessed resilience - preferably in a way that allows me to look at each separately.

My first thought is to use a fixed effects model of the sort:

lm(resilience ~ temperature + precipitation + factor(wave) + factor(household), data=df)

Using wave fixed effects means that I have a one observations per household, and allows me to account for time-invariant confounders.

However, while the 6-week period under each wave is useful, it does not overlap with the seasons observed in the surveyed area (of which three are three per year, each 3 months in duration).

Another thought would therefore be to use:

lm(resilience ~ temperature + precipitation + factor(season) + factor(household), data=df)

The same could also be done to look at within-year associations:

lm(resilience ~ temperature + precipitation + factor(year) + factor(household), data=df)

However, in both cases this would mean multiple observations for each time period. I assume this is problematic for inference?

Another option would be to aggregate the data, and average resilience and weather values on the day of interview. But I don't think this would be conceptually valid either.

Any tips would be warmly appreciated!



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.