so I'm doing a linear regression with three sorts of religion variables. Two of those religion variables are based on the same question in the questionnaire. People are asked what's their religion and if they go to church much or abide by the religious norms.

So I made a variable of religious identification with the categories 'christian', 'muslim' and 'not religious'. Then I made a variable of religious practice with the categories 'strict', 'not-strict, but religious' and 'not-religious'.

I made them into dummy variables for the regression analysis. So I have 'dummy_christian' and 'dummy_muslim' with the 'not-religious' category as the reference. Then I have 'dummy_strict' and 'dummy_not-strict' with the 'not-religious' as the reference.

When I put the dummy variables of the religious identification variable in the regression analysis, they show significant scores with my dependent variable. The same with religious practice. The problem is that when I put all my religion dummy variables in the regression analysis, I don't have any significant results. And I think it's because there's a big overlap (and a big correlation) between the two not-religious categories.

Is this possible? And can I fix it?

  • $\begingroup$ It's possible. How many data do you have? In essence you have 5 categories (NR, sC, nsC, sM, nsM), how many people are there in each category? $\endgroup$ – gung - Reinstate Monica May 31 '18 at 17:17
  • 1
    $\begingroup$ There are 467 non-religious respondents, 377 strict christians, 191 non strict christians, 342 strict muslims and 281 non strict muslims $\endgroup$ – Elle May 31 '18 at 17:24
  • $\begingroup$ What's the dependent variable? $\endgroup$ – AdamO May 31 '18 at 17:32

The problem you are facing is that you are chasing statistical significance. This is bad statistical practice as it compromises the scientific integrity of an analysis.

It's good that you're describing discrepant findings. I would do the same as you assemble your report. I would then appeal to multiple aspects of the design to try and reconcile findings.

In general, multivariate models rarely agree with the sequence of univariate models describing model findings. This is because of a loss of precision, or possibly because those various factors are confounders, or mediators. Confounding and mediation can both result in multicollinear features. Those incidental findings can be interesting for generating hypotheses in future investigations.

| cite | improve this answer | |

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.