S3E11: You say tomato, I say tom-ah-to: a (somewhat) head-spinning discussion about interaction analyses
SERious EPIShow Notes
Matt and Hailey take a deep dive into Chapter 26 in Modern Epidemiology, 4th Edition, Analysis of Interaction. This episode needs a content warning- it is among the most advanced and conceptually complex topics we have ever covered on SERious Epi. Interaction occurs when the effect of one exposure on outcome depends in some way on the presence or absence of another exposure. Seems like a simple enough concept, right? However, as you’ll see in this episode, there are many different layers of complexity to consider related to terminology, scale, and interpretation of interaction analyses. A note from Matt and Hailey: since this material is very complex, we reached out to Dr. Jay Kaufman for his perspective on the episode before releasing it. He had some very helpful thoughts, and we would like to share them with you (paraphrasing with his permission): Part of what is confusing about this topic is the terminology differences, with Hailey using terminology (“interaction”) that lines up with that used by VanderWeele, ME4, and the Hernán and Robins textbook chapter and Matt using terminology (“interdependence”) from other articles in the literature, such as Greenland and Poole (1988). When there are joint effects that are exactly multiplicative, or supermultiplicative, you know it’s a causal interaction (i.e., synergistic or biologic interaction) because multiplicativity is necessarily super-additive as long as both exposures meet consistency, exchangeability, and positivity assumptions. However, knowing that joint effects are submultiplicative is not informative about additive interaction or synergism. It is also not possible to make a conclusion about additive interaction when a results section tells you only that in a logistic or Cox regression analysis there is “no significant interaction effect (p<0.05)” as that just tells you an effect is not exactly multiplicative. Multiplicativity has some causal implications because it is super additive as long as the causal assumptions listed above are plausibly satisfied. There are several proposed causal mechanisms that would generate multiplicative joint effects especially from the cancer epidemiology literature (e.g., Koopman 1990). In general, considering interaction on the additive scale is more useful for assessing public health relevance (e.g. Panagiotou and Wacholder 2014). Some of these concepts are difficult to convey in podcast format, so we’re including some helpful resources for anyone interested in learning more about this topic. Thanks again to Dr. Kaufman for helping us put this list together:
Greenland S, Poole C. Invariants and noninvariants in the concept of interdependent effects. Scand J Work Environ Health. 1988 Apr;14(2):125-9. doi: 10.5271/sjweh.1945. PMID: 3387960. VanderWeele TJ. On the distinction between interaction and effect modification. Epidemiology. 2009 Nov;20(6):863-71. doi: 10.1097/EDE.0b013e3181ba333c. VanderWeele TJ. The Interaction Continuum. Epidemiology. 2019 Sep;30(5):648-658. doi: 10.1097/EDE.0000000000001054. PMID: 31205287; PMCID: PMC6677614. Greenland S, Poole C. Invariants and noninvariants in the concept of interdependent effects. Scand J Work Environ Health. 1988 Apr;14(2):125-9. doi: 10.5271/sjweh.1945. PMID: 3387960. Koopman JS, Weed DL. Epigenesis theory: a mathematical model relating causal concepts of pathogenesis in individuals to disease patterns in populations. Am J Epidemiol. 1990 Aug;132(2):366-90. doi: 10.1093/oxfordjournals.aje.a115666. PMID: 2372013. Panagiotou OA, Wacholder S. Invited commentary: How big is that interaction (in my community)--and in which direction? Am J Epidemiol. 2014 Dec 15;180(12):1150-8. PMID: 25395027.
More from SERious EPI
In this season finale, Matt and Hailey reflect on what they learned, and are still wrestling with, after working through Part I of Causal Inference: What If. They discuss how revisiting familiar concepts can reveal new layers of nuance, why the current enthusiasm for target trial emulation shouldn’t be mistaken for a shortcut to causal inference, and the importance of recognizing when the available data or scientific question simply doesn’t support a target trial. The conversation turns to the growing role of AI in epidemiologic analysis and the risks of being able to generate sophisticated analyses without fully understanding what is happening behind the scenes. They also preview the discussion about advanced methods awaiting them in Season 6. And because no season finale would be complete without settling the truly important questions, they finish by deciding which epidemiologic biases best correspond to Friends characters, ice cream flavors, decades, Muppets, and statistical software.
In this episode of SERious Epidemiology, Hailey and Matt are joined by Lucy D’Agostino McGowan to discuss Chapter 10 of Causal Inference: What If? and the problem of random variability. This episode explores the distinction between random error and systematic bias, what confidence intervals and standard errors actually tell us, and why increasing sample size can improve precision without making a biased estimate any closer to the truth. Lucy offers a statistician’s perspective on why random error may actually be the easier problem, and why epidemiologists should be more concerned about where an estimate is centered than about squeezing out a little more efficiency. We also discuss topics like replication, matching versus weighting, baseline adjustment in randomized trials, p-values in Table 1, and the persistent temptation to overinterpret statistical significance.
In this episode of SERious Epidemiology, Hailey and Matt are joined by Dr. Louisa Smith to discuss Chapter 8 of Causal Inference: What If on selection bias. The conversation explores how selection bias can arise through conditioning on a collider, how it differs from confounding, and how loss to follow-up and censoring can introduce bias even in randomized trials. A major theme of this episode is the relationship between selection bias and confounding. We also discuss generalizability, target populations, competing events such as death. Finally, we talk about how challenging it is to address selection bias via analytic techniques including inverse probability weighting. This episode serves as a good reminder that when it comes to selection bias, an ounce of prevention is worth a pound of cure.
In this episode, we talk with Dr. Alexis Reeves about Chapter 9 of Causal Inference: What If, focusing on measurement bias (the bias formerly known as information bias). Measurement bias arises when exposures, outcomes, confounders, or colliders are measured incorrectly. We discuss different types of measurement bias: differential, nondifferential, dependent, and independent, and errors in measurement of continuous variables vs. categorical variables. We follow the structure of the chapter, next discussing DAGs and time-related issues in measurement of variables. We end the episode considering whether accurate measurement should be treated as its own causal identification assumption and by highlighting that mismeasured confounders deserve more attention because adjusting for them may leave residual bias or even worsen bias.
“Confounding, Confounding, Confounding” is like the epidemiologist’s version of “Marcia, Marcia, Marcia” from the Brady Bunch. To discuss Chapter 7 of Causal Inference: What If, we welcome Dr. Ashley Naimi. In this chapter, we discuss confounding as a central problem when estimating causal effects from observational data. The chapter emphasizes that confounding is not just an imbalance in covariates across exposure groups, but a causal problem that depends on the underlying structure of how treatment, outcome, and other variables are related. In this episode, Dr. Naimi helps explain concepts related to confounding, exchangeability, and faithfulness. We (try to) talk through confounding-related DAGs and how they are a useful tool to understand confounding bias. This episode shows why confounding gets so much attention in epidemiology: it is everywhere, often misunderstood, and, like Marcia Brady, it has a way of stealing the spotlight.