xtgee

[Masoumeh Sanagou]

Hi everyone,

I have a continues, all positive outcome (CTDI) and it is skewed. I checked ln(CTDI) which is still not normal. I'm not sure which of the following options is the correct one.
1- xtgee CTDI var1 var2, corr(independent) (I've used this option because I know that GEE approach forgoes the distribution assumption, providing valid inference regardless of the distribution of the data.)

2-xtgee ln_CTDI var1 var2, corr(independent) efom (I ran this model and got very different results from option 1)


3-xtgee CTDI var1 var2, family(gamma) link(log) corr(independent)

Regards,

[Masoumeh Sanagou]

Thanks for pointing out those considerations. I will make those clear and then think about approaches.

[Joseph Coveney]

So you recommend

I don't know what your research objective is, and the approach that I would take to attaining it might be something completely different from the path that you're on.

So, no, I wasn't making recommendations for particular regression models. I was pointing out considerations from among those in your current approach.

Here are a few other considerations that you might want to entertain if you haven't already (they're certainly not clear to me based upon what's gone before in this thread): why you've chosen the covariates that you have, why you are interested in the association between state and CTDI (I would have suspected that radiological health laws and regulations are fairly uniform and uniformly enforced), why you are interested in the association between referring physician's medical specialty and CTDI (that is, if that's what PracticeTypeID refers to—is neurologist versus gastroenterologist intended to be a surrogate for absorbed dose in CT scanning of head versus abdomen?), and why you suppose exogeneity of the referring physician's practice given the fixed effects that you're including in the model.

[Masoumeh Sanagou]

Thanks for the reply.

So you recommend either

regress CTDI  PatientWeight Age_year b1.Sex b2.State b4.PracticeTypeID, vce (cluster PracticeID_code)

or

xtset PracticeID_code
xtgee CTDI  PatientWeight Age_year b1.Sex  b2.State b4.PracticeTypeID    ,  corr(independent) vce(robust)

[Joseph Coveney]

Could you please make it more clear to me that what "your model is misspecified" means and what should I do for that?

Well, look at the -rvfplot-. I'm guessing that the radiologists set the output of the CT scanner (hence the CTDI) based upon some indirect nonlinear function of the patient's body weight (perhaps thoracic or abdominal circumference, if it's not a head scan), and the inclusion of body weight linearly in the model is what's giving rise to the strange appearance.

You can use other regression diagnostic plots to help pin down what's going on, if you're inclined to. But you seem to be using body weight etc. as covariates and interest lies in geography and medical specialty (below). You're not really interested in normality or skew or obtaining a well-fitted explanatory model involving these covariates.

I want to see the effect of the two variables b2.State b4.PracticeTypeID on outcome.

You seem to have a large sample where efficiency wouldn't be an overwhelming consideration and so you might just want to rely upon the -robust- option of -xtgee- with canonical link functions to produce usable asymptotic standard errors of these variables' coefficients even in the presence of model misspecification. In that case, it might not pay to transform CTDI, and you could consider just going with your first model above. You could even consider

regress CTDI b2.State b4.PracticeTypeID, vce(cluster PracticeID_code)

and include the covariates if you think that they will increase efficiency.

[Masoumeh Sanagou]

regress CTDI PatientWeight Age_year b1.Sex i.PracticeID_code
rvfplot, yline(0)

regress ln_CTDI  PatientWeight Age_year b1.Sex i.PracticeID_code
rvfplot, yline(0)

[Masoumeh Sanagou]

I don't know whether adding the practice ID will take care of it or not,

regress CTDI PatientWeight Age_year b1.Sex /* b2.State b4.PracticeTypeID omit these two because of collinearity */ i.PracticeID_code

but your model is misspecified. I recommend that you work on that first, especially if you're planning to use noncanonical link functions, such as your third model above.

I want to see the effect of the two variables b2.State b4.PracticeTypeID on outcome.


Could you please make it more clear to me that what "your model is misspecified" means and what should I do for that?

[Masoumeh Sanagou]

Masoumeh:
why -xtset-ting your data before -regress-?
Besides, if patients are nested within hospitals which, in turn, are nested within states, why not considering -mixed- (I assume that your -depvar- is continuous)?

Sorry that was a typo mistake. The two graphs are for

regress CTDI  PatientWeight Age_year b1.Sex b2.State b4.PracticeTypeID
rvfplot, yline(0

and

regress ln_CTDI PatientWeight Age_year b1.Sex b2.State b4.PracticeTypeID
rvfplot, yline(0)

and yes the depvar is continuous.

[Carlo Lazzaro]

Masoumeh:
why -xtset-ting your data before -regress-?
Besides, if patients are nested within hospitals which, in turn, are nested within states, why not considering -mixed- (I assume that your -depvar- is continuous)?

[Joseph Coveney]

I don't know whether adding the practice ID will take care of it or not,

regress CTDI PatientWeight Age_year b1.Sex /* b2.State b4.PracticeTypeID omit these two because of collinearity */ i.PracticeID_code

but your model is misspecified. I recommend that you work on that first, especially if you're planning to use noncanonical link functions, such as your third model above.

[Masoumeh Sanagou]

The structure of the data: patients go to facilities to do CT scan. There are patient-level characteristics e.g. age, sex, weight and facility-level characteristics e.g. state, practice type(private, public).
each facility has one practice ID.





The panel ID is PracticeID_code

xtset PracticeID_code
regress ln_CTDI PatientWeight Age_year b1.Sex b2.State b4.PracticeTypeID
 rvfplot, yline(0)

[Joseph Coveney]

It looks like you have a bigger problem than skew, but go ahead and do the same with logarithmically transformed response values.

Is State your panel ID?

[Masoumeh Sanagou]

Thanks for the reply.



The attach is residual-versus-fitted plot.

regress CTDI  PatientWeight Age_year b1.Sex b2.State b4.PracticeTypeID
rvfplot, yline(0)

[Joseph Coveney]

If your concern is normality, then what do the residuals of

regress CTDI var1 var2

look like?

I wasn't aware that GEE forgoes distribution assumptions.