If the genotype and the phase (pre and post) could be treated as factors and the non-parametric data can be treated as response variables, then you can estimate a model using a binomial distribution or poisson distribution in a repeated measures analysis model.

Simpler approachs based in contingency tables and chi-square tests and wilcoxon signed rank test may be appropriate, too.

Thanks Mr Marcelo. I think the question was not appropriately explained. The question is

"126 patients are given lipid lowering therapy and the lipids decreased significantly after treatment. Now these 126 patients have three genotypes (some have AA, some have aA and rest have aa). I need to test the association of genotypes (AA, aA and aa) with the baseline lipid values, percentage change and post treatment values. I mean whether a particular genotype has high baseline values. This can be done easily if the lipid data is parametric but here the lipid data distribution is non-parametric. Which test is to be applied?. Moreover, non-parametric tests only reveal Z score with positive and negative ranks (not mean percentage decrease), so how to test whether a particular genotype has a better response.

thanks

What is the scale of measurement of the lipid values? Nominal, ordinal, intervalar or ratio? I believe it is, at least, ordinal (high, medium, lower). In this case you may convert these values to 3, 2 and 1 and then you can use a Kruskal-Wallis non-parametric test. If the lipids are measured as a number (intervalar or ratio scale), then you can use a non-parametric test, but we prefer to try a linear model before. Through the linear model you can compare means in ANOVA, but to do this, the residuals must be adherent to the gaussian distribution. Generalized linear mixed models can be useful to treat data adherent to a distributions traditionally analysed by non-parametric tests like Poisson distribution or binomial distribution.

Thanks again Dr Marcelo. Yes the data is collected continuous for example 200, 215, 236 etc. I tried anova and glm univariate but again the problem is nonguassian data distribution. Generalised mixed models says "unable to compute" . I may be putting data in wrong manner.

I wish you to have a look on data if you have some time plz. Actually the original paper is from Brazil itself. Plz inbox me your email. Thanks again