Fixed effect model for hospital and patietn level data (really need some help)

[Anqi Zhang]

Hi,

I am exploring a dataset of patient data from 20 hospitals for 3 years. Among which 9 hospitals are in the treatment group, and 11 hospitals are in the control group.

The dataset is pooled cross-sectional at the patient level, but panel at the hospital level. I have two goals:

1. Find out if a hospital-level intervention has an impact on patients' medical costs. I want to fit a fixed-effect model, controlling for hospital and year fixed effects.

First, following https://www.statalist.org/forums/for...in-differences I tried:

reg cost intervention i.hospital_id i.year_id

But the treatment effect is not significant, even I include more patient-level control variables.

Second, I tried:

xtset year_id
xtreg cost intervention i.hospital_id, fe

The treatment effect is still not significant. Am I doing it right? Are these methods correct for my goal #1?

2. Find out if a hospital-level intervention has an impact on the hospitals' admission rates.

In this case, should I aggregate the patient data to the hospital level? How could I do it? Should I take the averages of patient data in a specific year and a specific hospital?

I only have 20 observations at the hospital level, is this too small? Which method could I use if I cannot acquire more observations? Which if I can increase the sample size?


Thank you very much for your reply and favor.

[Noah Mkasanga]

Anqi,

I can see you are more interested to see your results being statistically significant. However, I have read several threads saying that if there is no significance, interpret the results based the magnitude and sings of the coefficients. Assuming that the model is correctly specified.

[Carlo Lazzaro]

Anqi:
if you have a panel dataset at hospital level and you want to go -fe-, your first code should have been:

reg cost intervention i.hospital_id i.year_id, vce(cluster idcode)

as your obsrcations at panel level are not independent.

2) as far as -xtreg.fe- is concerned, why not coding:

xtset hospitalid year

In addition (and for bothe your codes) if -intervention- is a categorical variable, why not coding it as:

i.intervention

3) set aside the irrelevance of statistical significance when compared to 95% CI, 20 observations per panel are really too few.

[Anqi Zhang]

2) as far as -xtreg.fe- is concerned, why not coding:

xtset hospitalid year

In addition (and for bothe your codes) if -intervention- is a categorical variable, why not coding it as:


3) set aside the irrelevance of statistical significance when compared to 95% CI, 20 observations per panel are really too few.

I will need to collapse the pooled cross-sectional patient data to panel data to call xtset hospitalid year. I will try to increase the sample size and categorize the intervention.

Thank you very much, Carlo.