1
$\begingroup$

I am using school-level data to compare mean scores for two groups (Group A schools and Group B schools), as measured by a series of indicators: 1) grade-level exam percentile, 2) percent of male students who meet standards on exam, 3) percent of male students who exceed standards on exam, 4) percent of female students who meet standards on exam, 5) percent of female students who exceed standards on exam.

I know these indicators are correlated as they all are based upon upon exam score. A fellow grad student suggested a series of t-tests between school types A and B for each indicator, but I am concerned about an increase in Type 1 error. Then another student suggested I use z-test, but because I am dealing with school-level data this did not make sense to me as it doesn't matter how many students are in each group because I am comparing schools, not student-level data. I think I should be running a MANOVA, but am genuinely not sure if that is the right way to go about this.

EDIT --- extra info from comment

I have a sample size of 60 schools, only school-level data (no student-level data). I have 20 schools in group A and 40 comparison schools in group B.

$\endgroup$
  • 5
    $\begingroup$ Generally politeness is good - and comments are viewed in a slightly different way to actual questions and answers (partly because of differences in how they're picked up in search engines). I wanted to leave an explanation of why I edited your post, but didn't want to touch the misspelt MANOVA in case you actually were referring to something else (I'm always wary of "correcting" acronyms, in case it's just one I'm unfamiliar with). Hopefully someone will be able to answer your Q shortly: editing your question is a good way to give it attention, particularly if you clarify or add example data. $\endgroup$ – Silverfish Apr 28 '15 at 21:35
  • 1
    $\begingroup$ What is the size of your data? Also do you have student-level results for each school? The methods may depend on these factors. $\endgroup$ – rnso Apr 29 '15 at 4:04
  • $\begingroup$ I have a sample size of 60 schools, only school-level data (no student-level data). I have 20 schools in group A and 40 comparison schools in group B. $\endgroup$ – Torifire Apr 30 '15 at 16:02
0
$\begingroup$

I would suggest using Principal Components Analysis (PCA). Given a set of highly correlated features, it may work to use PCA to output a orthogonal (uncorrelated) set of transformed features.

$\endgroup$
  • 3
    $\begingroup$ This is a nice idea but it wouldn't really solve the problem. $\endgroup$ – shadowtalker Apr 28 '15 at 23:13

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.