You know Frank I agree completely with your response and have said the same thing to colleagues who have misguidedly promoted use of log-linear risk or (worse) linear risk models. In fact I’ve been advocating our shared view on that since the 1970s (although, as I cited earlier, I have encountered exceptions in pair-matched cohorts in which log-linear risk models outperformed logistic models for a common outcome).
So perhaps you can imagine why after over 40 years I get exasperated when someone comes along and ignores all nuances and most literature on the topic to make vastly oversimplified and harmful claims, e.g., that we should be using ORs to summarize studies, ignoring all OR problems with a completely botched understanding of heterogeneity and noncollapsibility problems. This kind of toy one-method-fits-all approach to stats is the bane of science. It is as if someone doing a shift on a medical helpline advised every caller complaining of a bad headache to take aspirin (which would help some and kill others).
Allow me to explain the origin of my thinking: In the mid-1970s dissertation data involved a very common outcome (c-section, >50% in some groups but with huge variation) and many discrete clinical covariates in a large and very complete cohort database, which I analyzed using the then-new ML loglinear-count model software based on the then-state-of-the-art methods in Bishop Fienberg & Holland (1975) (as you know, loglinear count models include logistic models as a special case and log-odds/log-ORs are their natural parameters). The overall summary OR for the treatment of central interest, electronic fetal monitoring, was about 1 (null as can be).
But that concealed completely the most striking data feature: The titanic (and stable) OR variation across covariates, as expected from clinical input that some effects could reverse direction across other covariates, even when the “background risk” wasn’t varying much. Obstetricians could expect this because they were the mechanism of variation and could explain how they chose the outcome (c-section or vaginal delivery) based on the covariate values. Furthermore, one could see the extreme variation in effects across studies, as expected from the varying policy, practice and case mix across hospitals. [What a contrast to cancer epidemiology, where you can’t ask the body why it it chose to develop a tumor or not based on its exposures.]
So the lesson I learned from the start is that the model better be as rich as sustainable to allow for OR variation: Just because ORs vary less than RRs or RDs is no excuse for pretending they are anything like constant or transportable across studies or settings. And I learned that using p > 0.05 from a homogeneity test to ignore heterogeneity was inviting disaster, since the power of those tests for important variation is pathetic (in the same data one could also see that using p > 0.05 to ignore a confounder was biasing, especially since by definition of a confounder the target parameter is an adjusted one). From that experience I further learned the importance of reading the contextual literature, talking to the clinicians, and using that information to properly graph out the causal orderings of the variables under analysis.
I summarized those lessons in early articles such as
Greenland S. Limitations of the logistic analysis of epidemiologic data. Am J Epidemiol 1979;110:693-698
Greenland S, Neutra RR. Control of confounding in the assessment of medical technology. Int J Epidemiol 1980;9:361–367
Greenland S. Tests for interaction in epidemiologic studies: a review and a study of power. Stat Med 1983;2:243–251
Those are dated, but still they promoted basic ideas of statistical modeling as smoothing and causal modeling as a crucial input for statistical model selection.
Of course in light of much data experience afterward along with extensive refinements of causal modeling in the 1980s (notably by Robins, Pearl, Rosenbaum etc.) my conceptual thinking evolved quite a bit. And subsequent computing advances allowed practical implementation of hierarchical/multilevel models (empirical-Bayes, semi-Bayes, penalized regression) to expand models while controlling estimation stability, as I reviewed in later articles including
Greenland S. Multilevel modeling and model averaging. Scand J Work Environ Health 1999;25 (suppl 4):43–48
Greenland S. When should epidemiologic regressions use random coefficients? Biometrics 2000;56:915–921
Greenland S. Principles of multilevel modelling. Int J Epidemiol 2000;29:158–167
Greenland S. Smoothing observational data: a philosophy and implementation for the health sciences. Int Stat Rev 2006;74:31–46
Back in the 20th century computing was costly (every analysis for my dissertation required tedious card punching and overnight mainframe runs using data on tape), while journals were expensive laboriously typeset physical items that had to impose strict word limits. These economic problems led to a focus on simple summarization and often to a distortive compression of results. Computing advances since then make it easy for reports to go beyond oversimplified data summaries (like ORs from exponentiated model coefficients) and computer typesetting has slashed article production costs; yet narrow presentation limits are still imposed. Nonetheless, online supplements allow thorough presentation not only of data but also of causal narratives about its generation and how those were accounted for in the statistical models. I suggest that will usually be a better approach for advising practice than compressing variation into one number with some interval around it that captures nothing but some hypothetical “random error”.