Why is gender and sex critical when doing causal inference, and how critical they are?

Why is gender and sex critical when doing causal inference, and how critical they are? Keling Wang | 王可翎 Department of Epidemiology, Erasmus University Medical Center Rotterdam When gender/sex is a confounder People `adjust for gender` in causal studies everywhere, but what is it? ! Write here: what is the role of sex/gender in your study? Lemme know! When sex/gender defines target population Women vs. people with ovaries: an ovarian cancer screening trial When gender/sex is an intervention Effect of `gender` on influencer marketing, a gender-related context Checklist: considerations of sex and gender in causal study? Source: Effect of Screening on Ovarian Cancer Mortality. J Am Med Assoc 2011;305(22):2295-303. 10.1001/jama.2011.766 The target population is defined as ‘women aged 55-74’ without clear definition. The recruitment includes “males and females”. But, all individuals with at least one ovary should be included as target population3. Gender and sex are multidimensional constructs!1 When talking about sex...  legal indicator assigned at birth?  hormone milieu?  reproductive organs or genitals?  sexual development history? When talking about gender...  self gender identity?  gender expression w.r.t. gender norms?  meta-perceived gender?  cultural and contextual dependence? Do they matter, and how much do they matter? TL;DR: YES! And ‘how much’ will depend on your causal question, and the roles of sex-gender characteristics in that context. Be careful. Information provided in this (opinion) poster:  Magnitude of the biases when ignoring the multidimensionality or minor categories of sex or gender-related concepts when doing causal inference.  Contextual suggestions if sex and gender are critical for a causal question. But we often...  Mess one with another or only capture the wrong dimension.  Use “1/0” binarization or collapse “Others”. Other references: 1. Am J Epidemiol 2023;192(1):122-32. 10.1093/aje/kwac173. 2. J Causal Inference 2023; 1(1):107-34. 10.1515/jci-20120004. 3. J Clin Endocrinol Metab 2008;93(4):1408-11. 10.1210/jc.2007-2808. 4. Biometrics 2019;75(2):685-94. 10.1111/ biom.13009. 5. Chin Sociol Rev 2014;47(1):3-29. 10.2753/CSA2162-0555470101.2014.11082908. 6. PLOS Med 2017;14 (12):e1002479. 10.1371/journal.pmed.1002479. When gender/sex affects transportability A paper introducing an excellent method but ignoring minor categories 2 3 4 The PLCO trial estimates the average treatment effect (ATE) of an ovarian cancer screening strategy on overall and cancerspecific survival in women aged 55-74. It randomized ~78k women. What happened in the original study? What is the sex/gender concern? What went wrong: misalignment 1) Cis women and post-transition trans women were selected and randomized; trans men and people of atypical sexual development were completely ignored. 2) Among ~78k women in the trial, ~9.5k had no ovaries or fallopian tubes. 3) This leads to misalignment between (theoretically) trial-eligible and (actual) trialinvited population4. Which aspect is critical? It is the organ status that matters, not the administrative sex indicator or the gender identity. Illustration of the issue S, S’: selection nodes DSD: difference in sexual development; GI, gender identity; Bias from simulation study What happened in the original study? What went wrong: minor category ignored Illustration of the issue Bias from simulation study 1 Methods: data and simulation set-ups TL;DR: We grab top papers where sex and gender play a role, simulate their causal structures, and illustrate the bias they had per different bias parameters. Data: Manually selected research articles 1) that claim to, intend to, or implicitly aim to assess the causal effect; 2) where an aspect or some dimensions of sex and/or gender have a causal role (target population, intervention, confounder, effect measure modifier, …); 3) where sex/gender is mis-formulated or mis-measured in the study. 0 Illustration of the causal structure: 1) we use directed acyclic graphs (DAG) or selection diagrams2 based on DAGs. 2) we follow Pearl’s structural causal model approach 3) we try to ‘recover’ the causal structure using information from the original studies; in case it cannot be recovered / not reported, we will make additional assumptions. Simulation of the bias: 1) we use R `simDAG` package to simulate the population’s attributes and the causal structure. These attributes are recovered from the original reports. 2) When not reported, we used information from related articles or national statistics as a general estimation of the gender-sex related attributes. 3) Estimates of the causal effect according to different bias parameters will be displayed, and compared with the original estimates. DOMAIN CONSIDERATIONS Causal question What causal question do I try to answer? Do I target any population beyond binary sex/gender categories? Causal roles Which causal roles do sex and gender have in my study? Context Which aspect/dimension of sex and gender is the most important in my study? Is it enough to consider only one dimension? Measurement How should sex/gender be correctly defined and measured in this context? What are the ideal measuring tools and the measurements? Data availability and feasibility Do my data or tools at hand suffice to precisely measure this construct? What additional materials do I need? What to do in case it’s impossible to measure or approximate this construct? Analysis How should I take the complexity of sex/gender when modelling it in statistical analysis? If not possible to do things right: what is the impact on vulnerable populations? Source: Gender effects in influencer marketing. Int J Advert 2021;41(1):128–49.10.1080/02650487.2021.1997455 A 2×2 design experiment estimates the effect of the influencers’ gender on some marketing outcomes, and sees whether the participants’ gender modifies this effect. G: gender KOL: influencer Resp: respondent GE: gender expression Cat: category Sim: perceived similarity Intr: personal interaction Y: outcome Source: Transportability Without Positivity: A Synthesis of Statistical and Simulation Modeling. Epidemiology 2024;35(1):23-31. 10.1097/EDE.0000000000001677 Source: Educational mismatch and mental health: Evidence from China. Soc Sci Med 2024;356:117140. 10.1016/j.socscimed.2024.117140 G: gender Edu: education Req: job education requirement MisM: educationjob mismatch Inc: income D: depression L: confounders i.e. age, marriage, occupation, etc. S: selection node S=0: source; S=1: target G: gender A: treatment (electronic kit vs. walk-in invite) Y: outcome (6-mo STD screening participation) Illustration of the issue What happened in the original study?  This methodological paper examines a strategy that allows one to transport the effect when the ‘positivity over selection’ is violated, with parametric assumptions.  In the case study, they want to know the effect of an electronically ordered test kit vs. walk-in invite on STD screening participation in a clinic C (target) regardless of gender.  What they have in hand, is a trial with only men (source)6. The age also differs. What went wrong: minor categories ignored The original trial divided participants in ‘male, female, transgender’ and had only 4 ‘trans people’. The minor categories are ignored in the source and are still not considered in the target. When handling the positivity issue during this transporting, G is considered a binary variable throughout the article. What happened in the original study? What went wrong: mismeasurements Bias from simulation study  An observational study estimates the effect of “educational mismatch” (whether one’s education level is higher than the job’s requirement) on depression, and if this is mediated by personal income.  It targets an ATT and uses the CFPS data6.  Gender is considered in the original article as a confounder that acts as a common cause of both ‘mismatch’ and ‘depression’. POSSIBLE CAUSAL ROLES OF SEX/GENDER Target population Definition Operationalization Intervention Confounder Effect measure modifier Mediator? arguable* And maybe multiple roles? 1) The influencer’s gender is set by gender norms (e.g. muscles); only two options are provided. 2) Transgender and non-binary respondents are all excluded from the study and are ignored. 3) The minor categories ignored here are typically “influential points’ with high leverage and act differently per category (e.g. fashion, lifestyle). Which aspect is critical? The ‘perceived gender roles’ is critical; minor categories cannot be simply ignored.  The CFPS uses household book registration information for demographic characteristics and uses binarized ‘gender’ and ‘sex’ interchangeably.  The administrative indicators are used in analysis, while it should be ‘gender’ instead. Which aspect is critical? ‘Perceived gender roles’ and ‘whether the perceived gender role differs from assigned sex at birth’ are both important. Bias from simulation study You will see this box if I am still working on it You will see this box if I am still working on it You will see this box if I am still working on it