Partially paired data in a before-and-after study

[Cathrine DC]

Hallo

I have little experience with Stata and are now faced with a stastistical/coding problem due the design of my study.

In a hospital setting, I am exploring all contacts to the hospital before-and-after an Intervention was implemented on number of admissions (among other things).
Because data is at contact level, the same patient may enter the before periode more than once and also, the same patients may enter both the before and after period. However the majority of patients do not appear more than once.
In other words, data are partially paired both wthin and between periods - leaving me confused about which tests and methods to use.
Data are non-normalised distrubed.

Is robust standard errors the solution ?

logit prepost admissions, vce(cluster id)

Or do I need to cluster the data into means/median for each patients in the before and after period. This way I only have an overlap of individuel patients between periods (not within periods) of 7,5%.

Your help is much appreciated

Cathrine

[Carlo Lazzaro]

Cathrine:
welcome to this forum.
I would try:

logit i.prepost admissions, vce(cluster id)

As far as the -i.- notation is concerned, please take a look to the really wonderful Stata built-in command -fvvarlist- (that y6ou can access just typing -help fvvarlist- from within Stata).

[Cathrine DC]

Thank you Carlo

I also want compare my before and after period with non-parametric tests using chi2 for categorical variables and Mann Whitney - Wilcoxon test for continuous variables. However, they require independecy also. I far as I can read, the vce(cluster id) can no be used with these test.

Is there another way I do this ?

[Carlo Lazzaro]

Cathrine:
yes, you're correct: vce(cluster id) can no be used with these tests.
To achieve what you have in mind, you may want to interact -prepost- with -admissions-:

logit i.prepost##admissions, vce(cluster id)

Additionally, as far as continuous variables are concerned, you may want to consider a bootstrapped paired -ttest-, as in a sort of mirror study, where each patient acts as her/his own control.

[Cathrine DC]

Thank you very much Carlo!