Causal Inference with Missing Exposures and Missing Outcomes
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Causal Inference with Missing Exposures and Missing Outcomes".
Jane: Here's the gist of the paper. Missing data are ubiquitous in public health research When estimating causal effects, there are well-established methods to address bias to due missing outcomes The authors demonstrate how this framework can be extended to missing exposures.
Tom: They further extend this framework to incorporate missingness on the baseline outcome, which induces missingness on the population of interest (e.g., persons at-risk).
Jane: To do so, they highlight Counterfactual Strata Effects, a general class of causal estimands where the focus population is subject to missingness and/or impacted by the exposure.
Tom: Framework Extension The standard Causal Roadmap framework can be extended to address multi-source missingness in real-world (i.e., observational) settings.
Paper summary: Tom: This paper, "Causal Inference with Missing Exposures and Missing Outcomes," points out that missing data is everywhere in public health research. They show a framework to address bias when you're missing both the cause and the effect of something.
Jane: They extend methods already used for missing outcomes—where you don't know the final result—to include when the exposure itself is missing, which happens often in real-world studies.
Lu: The core idea they highlight is these Counterfactual Strata Effects, which is a general class of causal estimands where the focus population might be missing or directly affected by the exposure.
Meng: So what's the big claim here? They show how this framework can be applied to real data, specifically looking at alcohol use and incident tuberculosis infection in rural Uganda.
Tom: The paper uses this real-data example to investigate the effect of alcohol consumption on incident TB, highlighting that they use TMLE with Super Learner for their estimation.
Jane: That specific finding shows the method works in practice by giving a concrete number, but we have to remember they also admit their study design had some limitations.
Lu: They specifically note that their test focused on observational studies and didn't use multiple imputation, which is a very common way to fill in missing values when you're dealing with data.
Meng: So the simulation study confirms the theory works well, but they acknowledge that it doesn't solve all real-world data messiness by itself.
Tom: It’s a strong theoretical extension of how we define causal questions when data is messy, so now we need to see how robust these estimators actually are when you run them against messy production data.
Conclusion: Tom: So, looking at "Causal Inference with Missing Exposures and Missing Outcomes," the authors show us a way to find cause and effect even when you're missing data on both what caused it and what happened to it.
Jane: They take established ways of handling missing outcomes—like when you don't know the final result—and they stretch that method out to include missing exposures too, which is a major headache in real health research.
Meng: For someone shipping models, this means they give you a statistical way to define the effect even when there's that missing piece of information in your data stream.
Tom: They tested this with alcohol use and tuberculosis infection in Uganda, finding a specific risk increase of one point four nine for the risk ratio.
Jane: That specific result shows the method works in practice, giving us a concrete number for an intervention study, but we also saw some limitations in their study design itself.
Lu: Right, they admit that their test focused on observational studies and didn't use multiple imputation, which is a very common way to fill in missing values when you're dealing with data.
Tom: It’s a strong theoretical extension of how we define causal questions when data is messy, so now we need to look at how robust these estimators actually are when you run them against messy production data.
Kirsten E. Landsiedel, Rachel Abbott, Atukunda Mucunguzi, Florence Mwangwa, Elijah Kakande, Edwin D. Charlebois, Carina Marquez, Moses R. Kamya†,∗, Laura B. Balzer
School of Public Health, University of California Berkeley · Division of HIV, Infectious Diseases and Global Medicine at University of California San Francisco Division of HIV Infectious Diseases and Global Medicine at University of California San Francisco Infectious Diseases Research Collaboration in Kampala Uganda Center for AIDS Prevention at University of California San Francisco Department of Medicine Makerere University in Kampala Uganda
stat.ME
Submitted: 2025-06-03
Updated: 2026-06-16
Comments: 17 pages of main text (double-spaced; including 4 figures) + 20 pages of supplementary material (double-spaced; 1 figure; 2 tables) + 83 references
Journal ref: American Journal of Epidemiology, October 2026
DOI: 10.1093/aje/kwag281
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 79/100
The gist: Missing data are ubiquitous in public health research When estimating causal effects, there are well-established methods to address bias to due missing outcomes The authors demonstrate how this…
Key concepts
- Counterfactual Strata Effects
- This is a type of causal question focusing on a specific group—the 'stratum'—that is either missing data or directly influenced by the exposure. It allows researchers to estimate the effect of an exposure on this particular sub-population, like people who are at risk for an outcome.
- TMLE with Super Learner
- This is a statistical method used to estimate causal effects. TMLE is a doubly robust procedure, meaning it provides reliable estimates even if some assumptions are slightly wrong. Combining it with Super Learner enhances its ability to handle complex data structures and improve accuracy in estimation.
- Missingness DAGs
- Directed Acyclic Graphs (DAGs) are diagrams used to map out the relationships between variables in a study, especially when data is missing. These graphs help researchers identify potential sources of bias and determine if a causal effect can be reliably identified, guiding how to handle missing information.
Terminology
Summary
Missing data are ubiquitous in public health research When estimating causal effects, there are well-established methods to address bias to due missing outcomes The authors demonstrate how this framework can be extended to missing exposures. p. 2 Causal Inference with Missing Exposures and Missing
Outcomes
Abstract
Missing data are ubiquitous in public health research. When estimating causal effects, there are
well-established methods to address bias to due missing outcomes. We
demonstrate how this framework can be extended to missing exposures. We further extend this
framework to incorporate missingness on the baseline outcome, which induces missingness on the
population of interest (e. We further extend this
framework to incorporate missingness on the baseline outcome, which induces missingness on the
population of interest (e. , persons at-risk). , persons at-risk). To do so, we highlight Counterfactual Strata Effects, a
general class of causal estimands where the focus population is subject to missingness and/or impacted
by the exposure.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Framework Extension
The standard Causal Roadmap framework can be extended to address multi-source missingness in real-world (i.e., observational) settings. p. 3 Here, we discuss how this framework can be extended to address multi-source missingness in real-world
(i. , observational) settings. , observational) settings. To do so, we present a series of causal models and identification results,
building from simple to complex. To do so, we present a series of causal models and identification results,
building from simple to complex. First, we discuss how the Roadmap approach for missing and censored
outcomes can be extended to address missingness on the exposure. First, we discuss how the Roadmap approach for missing and censored
outcomes can be extended to address missingness on the exposure. Second, we present how this approach
can be applied with missing exposures and outcomes. Second, we present how this approach
can be applied with missing exposures and outcomes. Finally, we address missingness on the exposure, the
baseline outcome, and endline outcome.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Counterfactual Strata Effects Definition
Counterfactual Strata Effects are defined for a group which is subject to missingness and/or influenced by the exposure. The authors illustrate practical relevance with SEARCH-TB, which aimed to evaluate the effect of alcohol use on incident tuberculosis (TB) infection. p. 4 To do so, we use Counterfactual Strata Effects: a general class
of causal estimands defined for a group which is subject to missingness and/or influenced by the
exposure. [5, 25–34]
We illustrate practical relevance with SEARCH-TB, which aimed to evaluate the effect of alcohol use on
incident tuberculosis (TB) infection. [11, 12, 16–20, 38, 39]
To define causal effects when the outcome is subject to missingness, we consider counterfactuals indexed by
both the exposure and the outcome measurement indicator. Specifically, let Y
∗ = Y
A=a,∆Y =1 be the
counterfactual outcome for a given participant if, possibly contrary-to-fact, their exposure were at level
A = a and missingness on the outcome prevented. When estimating causal effects, there are
well-established methods to address bias to due missing outcomes. Commonly, causal estimands are
defined under hypothetical interventions to “set” the exposure and to prevent missingness.Where the paper says it
Where the paper says it
Where the paper says it
Statistical Estimation and Inference
To estimate the causal estimand, the authors highlight the use of TMLE with Super Learner for estimation and inference. TMLE is a doubly robust procedure and asymptotically efficient under certain conditions. p. 3 , not equal to single regression
coefficient). Sixth, we conduct estimation and inference with inverse weighting, standardization, or doubly
robust methods, such as targeted minimum loss-based estimation (TMLE). To do so, we highlight Counterfactual Strata Effects, a
general class of causal estimands where the focus population is subject to missingness and/or impacted
by the exposure. They are termed such because the estimand involves conditioning on a counterfactual
variable.Where the paper says it
Where the paper says it
Real-World Evidence and Performance
The results of SEARCH-TB’s investigation of the effect of alcohol use on incident TB infection among adults in rural Eastern Uganda have been previously published. The primary analysis using TMLE with clustering by household found that alcohol use was associated with a 49% increase in the risk of incident TB: risk ratio (RR)=1. p. 2 We demonstrate with a real-data example to investigate the effect of alcohol
consumption on the risk of incident tuberculosis (TB) infection in rural Uganda. We highlight the use
of TMLE with Super Learner for estimation and inference and discuss the practical consequences of our
approach. For simplicity, we focus
on defining and identifying causal estimands under a single level of the exposure, but our results naturally
generalize to causal effects defined in terms of contrasts of counterfactual outcome distributions under two
levels of the exposure (i. , the average treatment effect or causal risk ratio). We demonstrate with a real-data example to investigate the effect of alcohol
consumption on the risk of incident tuberculosis (TB) infection in rural Uganda. We highlight the use
of TMLE with Super Learner for estimation and inference and discuss the practical consequences of our
approach.Where the paper says it
Where the paper says it
Where the paper says it
Limitations and Future Directions
Limitations include focusing on cross-sectional or prospective studies rather than case-control studies. They did not consider multiple imputation, which is a common approach for missing data. p. 2 We highlight the use
of TMLE with Super Learner for estimation and inference and discuss the practical consequences of our
approach. Keywords: Causal Inference, Counterfactual Strata Effects, Missing Data, Real-world Evidence, Super
Learner, Targeted Minimum Loss-based Estimation, TMLE, Tuberculosis
3 If these
assumptions hold, we can rewrite E(Y
∗
) in terms of the longitudinal G-computation formula:
E{E[E(Y | ∆Y = 1, L1, A = a, ∆A = 1, L0) | A = a, ∆A = 1, L0)]}, shown in terms of iterated expectations
and with proof in Appendix S3. [9, 39, 40]
Counterfactual Strata Effects
We now introduce missingness on the outcome at baseline, which induces missingness on the focus
population. assumptions can be represented as Y
∗ ⊥ A | L and Y
∗ ⊥ ∆Y | A = a, L, respectively. These conditions are
often referred to as the “sequential randomization assumption” or “sequential exchangeability” and can be
evaluated graphically through the sequential backdoor criterion. , respectively. If these assumptions hold,
we have equivalence between our causal estimand E(Y
∗
) and the statistical estimand given by the
G-computation formula: E[E(Y | ∆Y = 1, A = a, L)].Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Supplementary Material
The supplementary materials include Appendix S1: The classic point-treatment problem, Appendix S2: Missing confounders, and Appendix S9: Missingness DAGs. These appendices provide proofs for the identification results and discuss methods to handle missing confounders like multiple imputation. p. 3 To do so, we present a series of causal models and identification results,
building from simple to complex. First, we discuss how the Roadmap approach for missing and censored
outcomes can be extended to address missingness on the exposure. First, we specify our causal question: how would
expected outcomes differ if all did versus did not have the exposure of interest and censoring were
prevented. Then, we specify a causal model, such as a directed acyclic graph (DAG) or non-parametric
structural equation model (NPSEM), to represent the data generating process for the baseline and
time-varying confounders, exposures, and outcomes. We demonstrate with a real-data example to investigate the effect of alcohol
consumption on the risk of incident tuberculosis (TB) infection in rural Uganda. We highlight the use
of TMLE with Super Learner for estimation and inference and discuss the practical consequences of our
approach. A) B) C)
Figure 1: Causal models with missingness on the outcome (Panels A-B) or missingness on the exposure (Panel
C): L=baseline covariates, ∆A=indicator of exposure measurement, A=observed exposure, Υ=underlying
outcome, ∆Y =indicator of outcome measurement, and Y = observed outcome. For ease of presentation, the
models are shown without dependence between the unmeasured variables, which are omitted on the graph.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
The paper cites numerous foundational works related to causal inference, missing data methods, and targeted estimation techniques The references cover topics such as the Causal Roadmap, Targeted Minimum Loss-based Estimation (TMLE), and Principal Strata Effects. The references also include works on missing data strategies like multiple imputation and missing confounders. p. 5 Appendix S2 provides a brief review of approaches to handle missingness on the baseline confounders L. Missing Outcomes
Consider a study of the effect of alcohol use on incident TB infection. Causal Inference with Missing Exposures and Missing
Outcomes
Kirsten E. Landsiedel1
, Rachel Abbott2
, Atukunda Mucunguzi3
, Florence Mwangwa3
, Elijah Kakande3
,
Edwin D.Where the paper says it
Where the paper says it
Appendix S7 Simulation Results
Table 1 shows the performance of various estimators across different components of the ratio-based statistical estimands. TMLE had negligible bias for all components; its bias-to-SD ratio remained below 0. p. 2 Causal Inference with Missing Exposures and Missing
Outcomes
Abstract
Missing data are ubiquitous in public health research. When estimating causal effects, there are
well-established methods to address bias to due missing outcomes. Causal Inference with Missing Exposures and Missing
Outcomes
Abstract
Missing data are ubiquitous in public health research. When estimating causal effects, there are
well-established methods to address bias to due missing outcomes. Causal Inference with Missing Exposures and Missing
Outcomes
Abstract
Missing data are ubiquitous in public health research. When estimating causal effects, there are
well-established methods to address bias to due missing outcomes. For simplicity, we focus
on defining and identifying causal estimands under a single level of the exposure, but our results naturally
generalize to causal effects defined in terms of contrasts of counterfactual outcome distributions under two
levels of the exposure (i. , the average treatment effect or causal risk ratio). For simplicity, we focus
on defining and identifying causal estimands under a single level of the exposure, but our results naturally
generalize to causal effects defined in terms of contrasts of counterfactual outcome distributions under two
levels of the exposure (i. , the average treatment effect or causal risk ratio).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Table 2 Performance Details
Table 2 provides further details on estimator performance with TMLE, G-comp, and IPW for the exposure-specific components of the ratio-based statistical estimands. The analysis shows that differences in performance between estimators are due to differences in statistical estimation.
Appendix S8 Accounting for Outcome Dependence
This section outlines an approach to account for the dependence of participant outcomes within groups or clusters. It suggests specifying causal models at the household-level if dependence is assumed within households. p. 6 A) B) C)
Figure 1: Causal models with missingness on the outcome (Panels A-B) or missingness on the exposure (Panel
C): L=baseline covariates, ∆A=indicator of exposure measurement, A=observed exposure, Υ=underlying
outcome, ∆Y =indicator of outcome measurement, and Y = observed outcome. For ease of presentation, the
models are shown without dependence between the unmeasured variables, which are omitted on the graph. We demonstrate with a real-data example to investigate the effect of alcohol
consumption on the risk of incident tuberculosis (TB) infection in rural Uganda. We highlight the use
of TMLE with Super Learner for estimation and inference and discuss the practical consequences of our
approach.Where the paper says it
Where the paper says it
Appendix S9 Missingness DAGs
This appendix discusses the use of missingness graphs to represent studies with missingness on multiple variables and to assess whether causal effects can be identified. It demonstrates how the Causal Roadmap can be used to define, identify, and estimate causal estimands including Counterfactual Strata Effects. p. 2 , persons at-risk). To do so, we highlight Counterfactual Strata Effects, a
general class of causal estimands where the focus population is subject to missingness and/or impacted
by the exposure. [1–8] There is rich history of methods research to address the potential for bias when
participants with measured outcomes differ meaningfully from those with missing or censored
outcomes. [9–20] The Causal Roadmap provides one such approach.Where the paper says it
Where the paper says it
Appendix S1 Classic Point-Treatment Problem
Figure 5 shows the causal models for a classic point-treatment problem with complete measurement of the baseline covariates L, the exposure A, and the outcome Y. Identification requires no unmeasured confounding, corresponding to the assumption that baseline covariates L capture all joint causes of the exposure A and outcome Y. p. 5 To identify
this causal estimand and express it as a function of the observed data distribution, we would need the
baseline covariates L to be sufficient to control for confounding and differential outcome missingness. Specifically, we need L to capture all the common causes of the exposure and outcome, and, among those
with the exposure of interest, all the common causes of the outcome and its measurement.Where the paper says it
Appendix S4 Alternative Time-Ordering
This section presents an alternative timeline where the exposure was measured after the baseline outcome, and both are still subject to missingness. The target estimand is again given by P(Y∗1 = 1 Y∗0 = 0). p. 5 Specifically, let Y
∗ = Y
A=a,∆Y =1 be the
counterfactual outcome for a given participant if, possibly contrary-to-fact, their exposure were at level
A = a and missingness on the outcome prevented. Then our causal estimand E(Y
∗
) is the expected
counterfactual outcome if all participants had exposure A = a and their outcomes measured. , respectively. If these
assumptions hold, we can rewrite E(Y
∗
) in terms of the longitudinal G-computation formula:
E{E[E(Y | ∆Y = 1, L1, A = a, ∆A = 1, L0) | A = a, ∆A = 1, L0)]}, shown in terms of iterated expectations
and with proof in Appendix S3.Where the paper says it
Where the paper says it
Appendix S5 Principal Strata Effects
Counterfactual Strata Effects are distinct from Principal Strata Effects. They enable defining the effect of alcohol on incident TB among the counterfactual population of persons at-risk at baseline under alcohol use A = a. p. 5 Missing Outcomes
Consider a study of the effect of alcohol use on incident TB infection. Suppose that we have a
representative cohort of persons without TB at baseline, but some participants did not test for TB at the
end of follow-up.Where the paper says it
Appendix S7 Simulation Study
The simulation study calculated the true values of the causal estimands to be Ψ∗(P; 1) = 0.510, the expected outcome under no exposure Ψ∗(P; 0) = 0. p. 5 Specifically, let Y
∗ = Y
A=a,∆Y =1 be the
counterfactual outcome for a given participant if, possibly contrary-to-fact, their exposure were at level
A = a and missingness on the outcome prevented. Then our causal estimand E(Y
∗
) is the expected
counterfactual outcome if all participants had exposure A = a and their outcomes measured. For simplicity, we focus
on defining and identifying causal estimands under a single level of the exposure, but our results naturally
generalize to causal effects defined in terms of contrasts of counterfactual outcome distributions under two
levels of the exposure (i. , the average treatment effect or causal risk ratio). Causal Inference with Missing Exposures and Missing
Outcomes
Abstract
Missing data are ubiquitous in public health research. When estimating causal effects, there are
well-established methods to address bias to due missing outcomes.Where the paper says it
Where the paper says it
Where the paper says it
Table 1 Simulation Results
Table 1 shows the performance of various estimators across different components of the ratio-based statistical estimands. TMLE had negligible bias for all components; its bias-to-SD ratio remained below 0.2. G-comp exhibited meaningful bias, as indicated by the bias-to-SD ratio, for all components except for the numerator under A = 1. p. 2 Causal Inference with Missing Exposures and Missing
Outcomes
Abstract
Missing data are ubiquitous in public health research. When estimating causal effects, there are
well-established methods to address bias to due missing outcomes.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Counterfactual Strata Effects can be used to precisely define causal estimands when
the focus population is subject to missingness and/or impacted by the exposure.
This allows for defining causal effects likethe effect of alcohol use on incident tuberculosis (TB) infection
among persons who are at-risk of TB at baseline. -
The framework can be extended to incorporate missingness on the baseline outcome, which
induces missingness on the population of interest (e.g., persons at-risk).
This enables researchers to explicitly define, identify, and estimate parameters in scenarios wherethe exposure influences the baseline risk of the outcome.
-
The paper demonstrates how
TMLE with Super Learner for estimation and inference
can be used to robustly and efficiently estimate statistical estimands. This machine learning-assisted approach avoids introducing new modeling assumptions during estimation while supporting valid statistical inference under reasonable conditions. -
The system can handle multi-source missingness by using causal models to
reflect the data generating process, generate counterfactuals, and aid in the identification of the corresponding causal estimand.
This focuses efforts only onthe portion of the observed data distribution relevant for the statistical estimand,
improving efficiency relative to approaches excluding participants with missing data. -
The system can perform sensitivity analyses by comparing results from different estimation approaches like TMLE, G-computation, and Inverse Probability Weighting (IPW). This comparison reveals how
theoretical properties (e.g., double robustness) translate into practical impact
on the final estimates across various scenarios.
Abstract
Missing data are ubiquitous in public health research. When estimating causal effects, there are well-established methods to address bias to due missing outcomes. Commonly, causal estimands are defined under hypothetical interventions to "set" the exposure and to prevent missingness. We demonstrate how this framework can be extended to missing exposures. We further extend this framework to incorporate missingness on the baseline outcome, which induces missingness on the population of interest (e.g., persons at-risk). To do so, we highlight Counterfactual Strata Effects, a general class of causal estimands where the focus population is subject to missingness and/or impacted by the exposure. They are termed such because the estimand involves conditioning on a counterfactual variable.For each setting, we present the causal model, relevant counterfactuals, causal estimand, and identification result. We demonstrate with a real-data example to investigate the effect of alcohol consumption on the risk of incident tuberculosis (TB) infection in rural Uganda. We highlight the use of TMLE with Super Learner for estimation and inference and discuss the practical consequences of our approach.
Sources
- Statistical Analysis Plan for SEARCH Phase I: Health Outcomes among Adults
- When exposure affects subgroup membership: Framing relevant causal questions in perinatal epidemiology and beyond
- A causal framework for evaluating the total effect of strategies aiming to expand screening and to improve outcomes
- Efficient Targeted Maximum Likelihood Estimators for Two-Phase Design Problems
- The Missing Covariate Indicator Method is Nearly Valid Almost Always
- Assessing treatment effects in observational data with missing confounders: A comparative study of practical doubly-robust and traditional missing data methods
- Causal Inference in Randomized Trials with Partial Clustering
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes