2

I have a balanced panel data set with 5 waves.

My DV is an ordinal outcome (physical activity), with 3 values. I am looking at the treatment effect of a random event (diagnosis of chronic disease) on the outcome over time.

I believe adjusting for baseline will allow me to get more accurate coefficients,as those who were more active before diagnosis will continue to be so.

However I cannot find much information on using baseline variable as a covariate for ordinal data, and wanted to ask about its validity. The coefficients I have obtained for baseline are highly significant (p-value=0.000)

Also, how would I go about incorporating baseline variable and time together, possibly as an interaction term?

user112638
  • 61

1 Answers1

1

Just represent the baseline covatiate in the most flexible way that the sample size allows, perhaps using indicator variables in your example. If it was continuous you could use a spline.

Frank Harrell
  • 91,879
  • 6
  • 178
  • 397
  • Thanks, I've made it a binary variable instead and it seems to be producing more accurate results. – user112638 Aug 31 '16 at 12:01
  • It cannot be more accurate; it can only cause a less good fi with a loss of powert. Never be very economical with the number of parameters assigned to the baseline version of the dependent variable. – Frank Harrell Aug 31 '16 at 21:36