I have a population dataset in whom the outcomes are 1. presence of coronary disease 2. Adverse cardiac events. In univariate analysis, the usual suspects such as LDL, smoking and family history have been consistently non-significant. There are other factors that have shown significance. Now in a multi-variate analysis, should I be including the common risk factors or only those that have shown significance in a univariate analysis? I have submitted to journals previously with using only those with significant p values in the multivariate analysis and have been accepted but during my viva, the examiner was insistent that comparing results based on p values was inappropriate.
Secondly, the examiner also commented on using odds ratio as effect measure and how it is vulnerable to bias due its non-collapsibility. Would greatly appreciate any help in resolving these issues.
In a multivariate analysis, it is generally recommended to include all relevant variables, even if they are not significant in a univariate analysis. This is because the relationship between the outcome and a predictor variable can be influenced by other variables in the model. Including all relevant variables can help control for confounding and improve the accuracy of the model.
Regarding the use of odds ratios as an effect measure, it is true that odds ratios are non-collapsible. This means that the odds ratio for a particular relationship can change when other variables are included in the model, even if those variables are not confounders. This can make interpretation of odds ratios more difficult. However, odds ratios are still commonly used as an effect measure in many fields.
This is a complex question. In theory one should adjust for all independent variables that are considered to be potentially confounding between the independent and dependent variable to be evaluated.
In relation to the Odds ratio, a multivariate regression analysis can be performed by expressing the result in terms of coefficients instead of in terms of Odds ratio, as an alternative.
Hello Shreenidhi,
Here's a link which offers more focused advice for addressing the noncollapsibility concern when comparing unadjusted (single IV) logistic models with adjusted (multiple IV) logistic models:
Article Noncollapsibility and its role in quantifying confounding bi...
Good luck with your work.
- Badges
- Science topic