2
$\begingroup$

I am writing a research protocol which requires an ordinal logistic regression to be done (Kleinbaum and Klein, 2002). my outcome is type of neonatal admission with three levels, namely, NICU admission, High Risk admission, and Well Baby admission. My my predictors are number of antenatal care visits, age of gestation, birth weight, and two interaction terms number of visit x age of gestation and number of visit x birth weight.

From the literature that I searched, researchers used Hosmer-Lemeshow to test the model fit of logistic regression models but I wonder if it is also appropriate for ordered/ordinal logistic with a 3 level outcome? I will be using stata to run my tests.

Improve this question
$\endgroup$

2 Answers 2

Reset to default
3
$\begingroup$

Spend your effort in well-specifying the model. And note that Hosmer-Lemeshow is considered to be obsolete and does not directly apply to ordinal regression. Besides carefully specifying the right-hand-side of the model (relaxing linearity assumptions through regression splines, pre-specifying interactions, etc.), assess the proportional odds model using tools such as partial residual plots with two cutoffs on $Y$. Details are in my book Regression Modeling Strategies and my RMS course notes - see https://hbiostat.org/rms and https://hbiostat.org/bib/po.html

Improve this answer
$\endgroup$
0
0
$\begingroup$

Yes, the Hosmer-Lemeshow test does extend to ordinal models. stata example of such implementation is discussed here.

For R users, there is a recent implementation of the ordinal Hosmer-Lemeshow test in the gofcat package. The hosmerlem() function, in particular, returns results similar to those implemented in stata.

Another R implementation is from the generalhoslem package, however, with results different from stata implementation. The author's note in the package documentation says, "Finally, it has been observed that the results from this implementation of the binary and ordinal Hosmer-Lemeshow tests and the Lipsitz test are slightly different from the Stata implementations. It is not yet clear why this is but is under investigation."

Improve this answer
$\endgroup$
1
  • 2
    $\begingroup$ I highly recommend against the use of the Hosmer-Lemeshow test. Any statistical test that requires binning has ultimately been abandoned. Use smooth continuous calibration curves corrected for overfitting among other methods, as detailed in hbiostat.org/rms $\endgroup$
    – Frank Harrell
    Commented Apr 16, 2022 at 11:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.