Fixed versus random effects when migration is high

[Fathima Salih]

Hi,

I'd be really grateful for insights on the following issue.

Say I want to estimate the effects of the experience of a flood on health spending for boys versus girls aged 1 to 9. I have health spending data at the individual level for years 2011, 2014, 2016 and 2018. Around 50% of the sample experienced flooding in 2014 and 2018. I initially thought a fixed effects panel estimation as follows will be best:

xtset person_id year

xtreg health_exp i.flood##1.male age, fe vce(cluster person_id)

health_exp and age are continuous variable, flood and male are dummies. The coefficient on the interaction will be the parameter of interest.

Upon further thinking, however, I feel the model is not correct as health spending will only occur if the child was ill. So I need a hurdle model for panel data such as xtdhreg that estimates a random effects hurdle model.

1. What is the rational for the model to be RE? Does it suit the original question I seem to ask?
2. Around 60% of the households migrated between rounds but are still in the survey. Does this make RE more suitable than FE as there may be between panel effects that need to be captured as well not just within effects?

I have read a bit too much and am confused and shall appreciate some insights.

Thank you.

[Maxence Morlet]

Are you interested in the coefficient on “male” alone? i.e. just this coefficient by itself, the baseline effect, and not only the interaction term? If yes, then that excludes FE unfortunately.

If households migrated but you follow them before and after the migration, that variation is captured by the within estimator.

[Carlo Lazzaro]

Fathima:
as an aside to Maxence's helpful reply, are you talking about health care spending for acute or chronic paediatric diseases? My guess is that your concern makes sense only if you refer to chronic diseases.
If that were the case, why not considering a diff in diff regression (households with healthy children=control group; households with ill children=treatmen group); before the flood=0; after the flood=1.
In addition, I'm not clear whether you're dealing with panel or survey data.

[Fathima Salih]

Maxence Morlet thank you. I am interested in "male" alone as well as the interaction term. So this means I need to adopt a RE? It turns out that xtdhreg is not suitable as a central feature of the panel-hurdle model is that the first hurdle has only one outcome per subject, and that outcome applies to all observations for that subject. But that will not work for my sample as being 'unwell' - the first hurdle is not always 0 or always 1 for a child across the round. I still feel the RE is not the best estimator as around 25% of the panel in any given year has 0 health spending.

@carlolazzaro thank you. my interest is any illness that would lead a household to incur health costs (medicine, traditional healing, hospitalisation etc). I have panel data. My research question is whether flood exposure correlates with differential health health spending of boys versus girls.

Is a random effects estimation (with no correction for the fact that around 25% of the sample have 0 health spending) appropriate?

[Carlo Lazzaro]

Fathima:
thanks for clarifying.
Therefore DIS is not the way to go there.
I agree that the -re- estimator is probably correct.
As far as the health spending=0 for the 25% of the sample triggers (or not) the need for a hurdle model, take a look at the literature in your research field.