1
$\begingroup$

I am not sure about the method selection for my data.

Dependent variable: Continuous variable and grouped in ordered categories. As continuous it is not normally distributed but skewed to the right. As categories, they are similar in size in the number of observations except for baseline which is defined based on clinical cut-off.

Explanatory variable: Continuous variable grouped as binary based on clinical cut-off. As continuous it is normally distributed.

Potential risk factors looking into: 2 binary, 1 with three non ordered categories

I looked at potentially using ordered logistic regression, but are now considering a non-parametric method. Was recommended to use the Whitney U test, but not sure what method I can use for multivariate analysis. Considering non-parametric series regression.

What method and test would you use?

Improve this question
$\endgroup$
2
  • $\begingroup$ Welcome to Cross Validated! // 1) I don’t see anything multivariate about your problem. What do you see as multivariate? // 2) What do you want to do with your data? What do you want to learn? // 3) Binning continuous variables is discouraged, as it literally destroys information. $\endgroup$
    – Dave
    Commented Nov 17, 2021 at 10:59
  • $\begingroup$ Thanks, Dave! What I want is to estimate the association between outcome and exposure, and adjust for the 3 confounders (if there is a confounding effect). I want to find a nonparametric method for this analysis. Isn´t it thereby multivariate as there are outcome + exposure + confounders in model? The Binning of the continuous variables was done due to the clinician's instructions, and I am not allowed to change it. $\endgroup$
    – Moa Herrgård
    Commented Nov 17, 2021 at 12:16

1 Answer 1

Reset to default
3
$\begingroup$

This is a multivariable problem, not a multivariate (multiple dependent variable) problem.

If you were to analyze the binned data a semiparametric ordinal regression model such as the proportional odds model would be the logical choice. But do not analyze the binned data. The physician has made a terrible mistake in discarding valuable information in the data by the use of arbitrary binning. It may be worth spending some time demonstrating the information loss, e.g., by comparing the likelihood ratio $\chi^2$ statistic for the whole model with and without binning, using the proportional odds model for both. (Yes the PO model works great for continuous Y just as its special case the Wilcoxon test does). An introduction to the PO model may be found in the Nonparametrics chapter of BBR.

To the bigger question, the role of a statistician is to give the collaborator what she needs, not what she wants. Collaboration is optimal when each collaborator uses their own expertise and there is a logical division of labor. It is a bad idea to be a spectator to bad analytical practice. There is an ethical obligation to not use methods known to be inferior. See https://www.fharrell.com/post/principles.

If you are in a setting where your expertise is not respected it is sometimes best to find a new setting or a new collaborator in an existing place of employment. Good statisticians/data scientists can find jobs almost anywhere nowadays. Sometimes we have to capitalize on this "buyer's market" in order to be in an environment for optimal growth as an analyst.

For information about binning and information loss see https://www.fharrell.com/post/ordinal-info. Another way to demonstrate information loss is to show that a predictor predicts variation of Y even within one of the forced intervals.

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ Regarding the binning, I was about to post a quote of yours, along the lines of, “If you’re afraid of the customer, you’ve already lost.” $\endgroup$
    – Dave
    Commented Nov 17, 2021 at 13:20
  • 1
    $\begingroup$ Thank you Frank for your response! I am new to the profession and was thereby unsure if I was wrong when the clinicians suggested something else. But following your comment, I made the case. Let's hope that the clinicians will listen and adapt. $\endgroup$
    – Moa Herrgård
    Commented Nov 22, 2021 at 18:45

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.