1
$\begingroup$

I have data set with binary outcome, with 5 continuous covariates and 4 discrete covariates.

I am little confused as to how I test for association, for the discrete covariates, I used a chi sq test, testing at 95% if there is any association. However for the continuous covariates I don't know if I should be using two sample t-test or anova test having fitted a linear model OR I should fit a GLM (binomial family) and see if the continuous covariate is significant in determining the outcome.

Please anyone able to explain to me which approach I should take and why it's the correct approach I would greatly appreciate it.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

Once you start talking about 'outcome' variables or dependent variables or the like, I think you are into the modeling world. Chi-square tests for association and does not posit that one variable is dependent. Similarly, the various kinds of correlation are tests of association.

Since your outcome is binary, the usual method is logistic regression

Improve this answer
$\endgroup$
6
  • $\begingroup$ so if i wanted to test association between say outcome (success, failure) and a variable revision (yes,no), isn't chi qs correct test? And if i wanted to test association between outcome (success, failure) and variable time_of_revision (continuous), i should preform logistic regression then use anova or t-test? $\endgroup$
    – khoshnaw
    Commented Aug 29, 2015 at 11:51
  • $\begingroup$ If you want to test association, use a test of association. But then you don't have an outcome variable. If you have an outcome variable, use some form of regression. After you do logistic, you don't need to do ANOVA or t-test (and, in fact, ANOVA is inappropriate if the dependent variable is a dichotomy. $\endgroup$
    – Peter Flom
    Commented Aug 29, 2015 at 11:55
  • 2
    $\begingroup$ In addition to what Peter said, you have to consider whether unadjusted tests are appropriate. Often they are not. To do an adjusted (partial) test one usually needs multivariable models. Think of a test of association (or partial association) as a test of flatness in a regression model. $\endgroup$
    – Frank Harrell
    Commented Aug 29, 2015 at 12:06
  • $\begingroup$ Sorry am little confused, then it is not possible to test association by performing a GLM (using binomial family ) and looking to see if the variable is significant ? $\endgroup$
    – khoshnaw
    Commented Aug 29, 2015 at 12:49
  • 2
    $\begingroup$ To be explicit, a multivariable statistical model is usually the best way to perform a test of association. $\endgroup$
    – Frank Harrell
    Commented Aug 29, 2015 at 14:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.