First - very new to Statistics; about half through a basic biostats book and an R book.

I have a set of data where I"m trying to see if there is a correlation between a medication and weight gain. The dataset:

  1. ~2200 patients
  2. A list of their primary diagnoses, weight, date's of visit, and whether they were on medication at that time.
  3. Each patient may be on or off the medication at various points, though I have limited the data set to patients receiving at least 90 days of medication.
  4. Different patients are on / off the medication for varying # of days. Also, each patient has recorded visit for a varying amount of total days (though at least 90 based on 90 day min. medication limit from #2).

General Outlay

My main goals are:

  1. Is this medication correlated with weight gain?
  2. If so / not - does it depend on a certain threshold length of constant use?
  3. (Eventually) do either of the above two correlations change based on the patients diagnoses (Some pt's will be grouped together)

My question is - really, where should I begin? What types of analyses should I be doing? I'm willing to put in the reading / work to figure it out; but I'm not quite sure where to look. As stated in intro, I have - some - experience with R, and am expecting to carry out my in depth analysis in that.


So far, I have been looking at Average of change in weight / days on medication; so basically giving each patient a Wt. Change per day of med. Im planning on adding wt. change per day while off medication and all patient variables, but this illuminates what I need something more advanced for --

How do I account for trends occurring independent of medication. A thought - say a patient started a diet, and has been steadily losing 1lb per week prior to medicaiton. If that medication is started, it is possible the overall trend will overshadow the medication effect. In this case - is there a practical way to account for long term trends, without examining each patients data individually?


Welcome to the wonderful world of Statistics. Despite a Masters degree in Medical Statistics, it never ceases to amaze me that the one thing I am learning over and over again is how little I actually understand about this fascinating modality!!

For what it is worth from a non-practising and not very experienced statistician, you need to look at a time-series data analysis with a time varying covariate. The Cox Proportional Hazards model can be used for this data with relative ease.

Check out this paper on the CRAN website: Cox Proportional Hazards Regression (J Fox, 2002)

My two cents - hope it's useful.

| cite | improve this answer | |
  • 1
    $\begingroup$ What would be the "hazard" in this example? The outcome of interest stated by the OP is "weight gain", and hence a statistical technique to evaluate the time to an event does not make any sense. $\endgroup$ – Andy W Jul 20 '11 at 12:46

You have a panel (2200 patiens x 90 days). But it could be complicated to organize it that way. Suggestion as a starting point: (1) Compare two states, i.e. "before" and "after" treatment-period (90 days), and look at changes in weight (x) and medication (y) during the 90-days period. (2) Run a regression (ordinary least squares) between y = f(x) with and without a constant. The constant absorbes autonomous effects from other factors than weight loss. You can also add other, relevant variables in the model.


| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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