I have a processed dataset where patients took a medication and gained a varying amount of weight. In addition, any ICD9 diagnoses the patients have received while being monitored are recorded. I'd like to see if certain diagnoses are correlated with amount of weight gained. What would be a good procedure for testing this?

Of note, each patient has several diagnoses: there are about 1,200 unique diagnoses in total.

Currently, I have calculated the average wt. gain along with the standard deviation, then grouped the patients by standard deviation, say:

(not actual #'s)
Total Patients: 420
Total in Group 1 (>2 Sd weight Loss): 10
Total in Group 2 (>1 Sd weight Loss: 50
Total in Group 3 (<1Sd weight gain or loss): 300
Total in Group 4 (>1 sd weight Gain): 50
Total in Group 5 (>2 sd weight Gain): 10

I was thinking of looking at how many patients have x,y,z (etc.) diagnoses in each group. Say Type II DM:

Total: 200 Pts Have it
In Group 1: 4/10 Pts have it
Group2: 15/50 Pts have it

With this method, would simply computing a Pearson correlation on the absolute count of Diagnoses vs. wt change group be appropriate? Additionally, if by the time I get to Group 2 or Group 4 0 patients carry a particular diagnoses, how will this impact running the analysis this way?

Appreciate any pointers.

  • $\begingroup$ Do you have the actual weight measurements, or is it just those categories? (I hate this medical approach of taking an arbitrary threshold and classify everybody either above or below it; economists, at least, would've broken this down to five quintiles and have the same number of people in each group.) I would think that the s.d. needs to be computed within the diagnosis, rather than across everybody, to make the standardized effect sizes comparisons meaningful. $\endgroup$
    – StasK
    Aug 12, 2011 at 21:51
  • $\begingroup$ Yes, have the original weights; can break them down into as many groups as necessary. $\endgroup$
    – george
    Aug 13, 2011 at 13:45

1 Answer 1


Before proceeding you need to make sure that the subject was free of all the diseases at the time of the initial weight.

If you were just considering one diagnosis you could estimate the probability of the diagnosis as a function of a smooth function of the initial weight and a smooth function of the final weight. It is important not to assume that weight change is an adequate summary of the two weights.

With multiple diagnoses a complex but general solution may possibly be had by turning the problem backwards: predict the optimum linear combination (first canonical variate) of the weight variables and nonlinear transformations of them from 1200 binary dummy variables, using lasso-like penalization of the disease effects.

For a few diagnoses (especially diabetes using glycohemoglobin as a continuous outcome measure) you can turn to NHANES data to model patient disease outcomes or medical test results, starting with a cohort that is disease-free with respect to that disease.

  • $\begingroup$ Why would they need to be free of disease before the initial weight? To be clear, I'm interested in the 1) The disease indication for the medication as affecting the outcome 2) Any co-morbidities that may affect the outcome. Additionally, the intial diagnoses is sometimes incorrect or very general, with the specific, accurate diagnoses coming out further into treatment. $\endgroup$
    – george
    Aug 13, 2011 at 13:48
  • $\begingroup$ A drug may be given because someone has a disease. You can't claim then that the drug caused the disease; it's the other way 'round. Also, some diseases may predispose patients to gain weight regardless of their meds. There may be a way around this but you need to very explicitly lay out all your assumptions, then critically assess them. $\endgroup$ Aug 13, 2011 at 13:51
  • $\begingroup$ I'm sorry, let me clarify - indications for the drug are wide - pt may be given the drug based on diagnosis A,B,C, etc... In addition, they may also have other diagnoses un-related to the one they were given the medication for. I'm interested in both sets. $\endgroup$
    – george
    Aug 13, 2011 at 18:21
  • $\begingroup$ That makes sense, but how do you make sense when you have a mixture of patients who already had the disease before the weight gain, and those who developed the disease after gaining weight? $\endgroup$ Aug 13, 2011 at 19:12

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.