Using fixed effects at the household and intra household levels

[Shailaja Tiwari]

I am working with a panel data set. The data are collected at three points of time 2011, 2015 and 2018 from the same 5000 households
the data contains information on food consumption by the different household members (intra household) and also on the total income of the household. My objective is to estimate the determinants of household income and intra household consumption.

i want to run two seperate regressions. In the first regression i use xtreg with Fe and my dependent variable is household income. In this regression i am including household and time fixed effects.

In my second regression my dependent variable is food consumed by different household members. This is intra household information. The data for this regression is panel set at the intra household level (i.e. individual within the household). I want to run an xtreg with Fe.
My question is that in the intra household regression do i include individual-within-household fixed effects or household fixed effects. Also should the standard errors be clustered at the household level or intra household level.
Thanks

[Carlo Lazzaro]

Shailaja:
1) I would add -i.timevar- as a predictor in both your regressions;
2) I would cluster the standard errors at household level (1st regression) and individual level (2nd regression). As far as your 2nd regression is concerned, you may want to consider multi-way-clustering, that is allowed by the community-conytributed module -reghdfe-;
3) in your 2nd regression I would consider individual fixed effect;
4) in your 2nd regression, individuals are actually nested within family. Therefore, you may want to consider -mixed- (that implies -re- specification, though).

[Shailaja Tiwari]

Carlo Lazzaro thank you for the suggestions. I will try these suggestions.

[Jeff Wooldridge]

I would use individual FEs when analyzing the individual data but cluster at the household level because the households were sampled — the households then form the clusters. I wouldn’t bother with mixed models unless you use a correlated random effects approach.

[Shailaja Tiwari]

Jeff Wooldridge thanks for the reply. I think i get the rationale behind clustering at the household level and controlling for individual effects.
Including Individual fixed effects will eliminate any time invariant individual level characteristic that drives consumption.
Clustering at the household level will ensure accounting for the sample design.