non-parametric ANCOVA?


New Member
Dear all,
I would appreciate any help on a much debated issue of non-parametric ANCOVA.

I compare 2 gourps (group is a factor, categorical), y is post measure, and x is pre mesure (covariate). The question is if the effect of the treatement is the same on both groups.

Are there non-parametric alternatives, in R software for example?
(I cheked sm package and fANCOVA package but they do not seem to solve my problem.)

Many thanks


New Member
Dear Bugman,

thanks for your reply. My problem with "sm" is two-fold:

- first it seems that in "sm" "non-parametric" refers not to the distributional assumption but to the form of the relationship between y and a continuous covariate, i.e. a smoothing issue. This is the discussion I had here:

- second, even if the first point turns out to be non-limiting for my case (where distributional assumptions are not met), it is uncler to me how to interpret the results from "sm"
E.g. with sm.ancova, for var1 to var4, post as response, pre as co-variate, group as factor, I get:
var1 p-value: non significant for "equality model" and non significant for "parallel model",
var2 p-value: significant for "equality model" and non significant for "parallel model",
var3 p-value: non significant for "equality model" and significant for "parallel model",
var4 p-value: significant for "equality model" and significant for "parallel model".

how to interpret it?

Many thanks!


New Member
But why exactely do you think you need a "nonparametric" analysis?

With kind regards

sorry, I did not mention from the beginning: my 2 groups are very small (10 and 20 subjects), and I have a big number of variables to be compared so that it is not realistic to meet the distributional assumptions for parametric ANCOVA, for all of variables. The variables are numeric values (measures) or ratios of numeric measures. Thanks!


TS Contributor
Have you considered analysing difference scores using U-tests?
Or are the baseline scores to different between groups?

With kind regards



New Member
As far as I understand change score are second best option when comparing two groups - the methods to "adjust" for baseline are much discussed but not clear (for me).
And yes, the baselines are very different, ranging, for different variables, from 0.06 to 1.5 for the ratio (pre in A / pre in B). Even conceptually, I have to take the group as factor, one group is trained people, the other not...
very best.