# Analysing data from non-equivalent groups in a quasi-experimental design

I am doing a research on two groups, one experimental and one control, which are not randomly selected. In fact the two groups are intact groups. The purpose is to find the effect of the number of languages that the learners know on their academic achievement. The selected design is Pretest Posttest Nonequivalent Group.

Which statistical method should I use to analyze the data?

-
Can you say more about what you mean when you say the groups are not randomly selected? Also, what do you mean by "intact groups?" –  Macro Jul 9 '12 at 15:10
by intact groups, I mean 'the naturally assembled groups', 'intact classes' and when I say the groups are not randomly selected I mean 'I can not use the process of random selection due to the nature of my study so my samples are not random samples.' –  user12487 Jul 10 '12 at 6:38

It sounds like you want some form of regression, depending on how your dependent variable (academic achievement) is measured. If it's something like grade point average, then linear regression is a good start.

Then the question is how to measure "languages known". Clearly, it's a count variable (that is, it has to be a non-negative integer; in this case, a positive integer) but (depending on where you are doing this) you may have a huge preponderance of 1's.

You will also want to include covariates that relate to achievement. You might want to do some sort of matching, possibly propensity scores.

-
what about the other methods such as ANOVA, t-test or even the Mann-Whitney U test? I have two variables, one independent (the number of the languages the students know) and one dependent variable. I have two groups as I said before. the observations are independent. But I'm not sure about whether the variables are interval or nominal. How to find? thank you –  user12487 Jul 10 '12 at 7:03
ANOVA is regression (they are mathematically equivalent). If you only have two variables, then you can do a t-test. I suggest, though, that you try to get more variables, especially for a non-equivalent design, in order to control for things known to be linked to achievement. –  Peter Flom Jul 10 '12 at 10:24
I, myself, think that t-test is appropriate although I've read the studies that have two two variables and two groups but they use ANOVA and I don't know how. one point that makes it difficult for me to choose t-test is that it is a kind of parametric tests and one assumption for parametric tests is the issue of random selection which my study lacks. am I correct? thanks –  user12487 Jul 10 '12 at 11:33
A t-test is just an ANOVA with only 2 groups. ANOVA is also parametric. The assumption of random selection isn't related to parametric vs. non-parametric; it's related to how generalizable the results are. You need more variables to try to control for the non-random selection. –  Peter Flom Jul 11 '12 at 10:50
even if I try to control the confounding variable such as age, sex, educational level, etc., i should be worry about the non-random selection problem? –  user12487 Jul 11 '12 at 16:03