I am investigating the level of association among different categorical variables. Chi-square analysis shows that some variables have significant association, while these variables fell short of significance in binary logistic regression model.
How shall I present the data? I mean can I still use Chi-square findings as an evidence of significant association? or shall I present findings from both analysis?
With thanks.
There is insufficient information to answer. With one dichotomous X and one dichotomous Y variable chi-square and logistic regression are more-or-less equivalent (one gives you a Pearson chi-square test and the other the likelihood chi-square test). With reasonable sample sizes they'll be almost identical and similar even with smaller sample sizes.
As you combine multiple X variables in logistic regression the most likely factor influencing differences is collinearity between X variables.
You can present both results. But to me the results of logistic model is more important.
When you perform logistic regression, please add those that are not significant by a Chi-square test but you believe they can have effect on your dependent variable. A common mistake is just choosing those variables that are significant after a Chi-square test and omitting those that are not significant.
Hi Abdullah,
I side with Thom here. For bivariate analysis (1 outcome and 1 exposure), chi-square and logistic regressions should be more or less equivalent, worst case scenario at least similar. Unless you are following an extremely rigid interpretation of the p-values (chi-square returned 0.04999, thus being "significant" while logistic regression returned 0.0500001, thus being "not significant"), which I would argue against, I believe either there is some mistake on your coding or you are running a multivariate logistic regression (1 outcome, more than one predictor). On a multivariate logistic regressions you expect your associations (and significance) to change due to control for confounding. As usual, more information is needed. Try discussing with your supervisor, look for help in forums that deal with that (if you are using R, try looking at Stackoverflow, if using Stata, try looking at Statalist) and look for books that discuss this topic (statistics, biostatistics, epidemiology...).
Best;
Thank you very much for kind replies.
@Igor Paplosky, I'm running multivariate analysis (1 outcome and multi predictors). The problem I face is, in Chi-square test I get some variables significantly associated with the outcome, while in binary logistic regression they lose their significance.
Hi Abdullah, this is a not a problem, this is one of the objectives someone can have when running a multivariate regression - control for other variables to reduce confounding. Imagine a situation when someone is studying the relationship between smoking and BMI. By running bivariate (outcome is BMI, exposure is smoking) analysis, either using chi-square or a bivariate regression - the results should be almost the same, one might find that smoking is associated to BMI. However, people who smoke more often than not also consume alcohol, practice less exercise and consume low-quality food, all of which are associated both to smoking and to BMI. The risk of a higher BMI found in smokers is, if the analysis is run without controlling for those confounders, likely overestimated. In this situation, if someone tried to show results of the simpler bivariate analysis (regardless of the p-values found), it would readily criticized. One way out in this hypothetical example would be to run a multivariate regression (to try to control for confounding), and then report the findings of the multivariate analysis.
I think the root of the question here is not the reporting of the findings, but a deeper one on why you are running a multivariate analysis to begin with. There might be very good reasons to do so, but I'm not sure you understood them, and this leads to doubts in interpreting and reporting your findings.
I would suggest to you to discuss the issue with your supervisor and to read a chapter on confounding on a epidemiology book, it might help. There are also several links on the internet that explain the issue. To cite a few:
http://sphweb.bumc.bu.edu/otlt/MPH-Modules/BS/BS704_Multivariable/BS704_Multivariable_print.html
https://stats.idre.ucla.edu/stata/dae/logistic-regression/
Wikipedia has a link on confounding:
https://en.wikipedia.org/wiki/Confounding
Best;