As I understand it there are three potential objectives for multivariable models in biology/medicine. Explanatory (causal), predictive and descriptive. In my field the objective of risk factor studies is rarely explicitly stated. If it is predictive then I would think a proper approach to predictive models with validation and calibration is required. If explanatory, then a prespecified model based upon a discussion of probable confounders and mediators is appropriate. The selection of variables based on data dredging as many of these studies seem to be doing can create a model but what is the utility of such a model? It hasn’t helped the reader in the prediction of an outcome nor has it helped in determining policy or treatment options that will have an impact on the outcome. It seems to me many of this studies suffer from ambiguity of purpose.
Among these for me (a layperson), the relationship of owner financial conflict of interest on the safety of the animals.
This nice article just published in BMJ is relevant to this thread:
https://bmjmedicine.bmj.com/content/4/1/e001375
The “factors associated with” design was mainstream by the 1970s15 16 and has become increasingly common. In 2024, more than 4000 articles with the phrase “factors associated with” in the title were added to the PubMed database, a 10-fold increase from 2004.
Lewer D, Brothers T, O’Nions E, Pickavance J. Factors associated with: problems of using exploratory multivariable regression to identify causal risk factors. BMJ Medicine . 2025;4:e001375. https://doi.org/10.1136/bmjmed-2025-001375
Great to know about. Risk factor analysis even has great difficulties correctly and stably finding only non-causal associations.
Someone mentioned to me that epidemiologists still consider stepwise regression a valid method. Under what circumstances , if any, would this be true?
@Ewout_Steyerberg has shown that if you use backwards stepdown with \alpha \geq 0.5 you do little damage to inference. Still there are better ways to accomplish the goals.