We are trying to run a clinical prediction model (based on regression analysis) to predict the effect of some dietary and non-dietary factors on mortality from a specific disease in a cohort database. For defining predictors, we did a comprehensive systematic literature review. It has been suggested that we should use the results of this review to draw a directed acyclic graph (DAG) to find the major predictors, confounding variables, and collinearity of predictors.
Now, I have some major questions:
1) What is the truest and systematic path to define the major predictors, confounding variables, and collinearity of predictors in this study (clinical prediction models based on regression analysis)?
2) Is it necessary to draw DAG to define the major predictors, confounding variables, and collinearity of predictors?
3) Can we use the Structural Equation Model (SEM) to define the major predictors, confounding variables, and collinearity of predictors in our dataset? And then use its results to run the model directly? If yes, is it true to look for confounding variables within our dataset instead of literature to run the prediction model?
I would suggest that for such a clinical medical study, it would be sufficient to assess from the literature (and/ or your own experience) the putative causal factors and potential confounders and then to explore in your regression model collinearities and actual confounding.
The final model would include the best outcome correlated or most accessible predictor among the collinear variable groups . Actual and potential confounders should be included in a prediction model because exposure distributions in other populations may not be the same as in your cohort.
DAG and SEM are more used in social scence where variables are created from questionnaires and psycho social constructs are used, often involving factor analysis techniques.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size