Bias in studies of prenatal exposures using real-world data due to pregnancy identification method
summary
The gist
The findings are complex, revolving around how missing outcome data (e.g., delivery vs.
In short
The study investigated how to correctly identify pregnancies with prenatal exposures when outcome data (like delivery) is missing. Researchers simulated 10 million pregnancies and tested three identification methods: observed deliveries, observed deliveries plus miscarriages, and all pregnancies. Findings showed that identification strategy matters; bias was only reduced when missingness was due to measured variables, indicating that knowing the cause of missing data is crucial for accurate prenatal exposure estimates.
Key concepts
- Outcome-Based Identification
- Identifying a pregnancy based on a specific result, such as delivery or miscarriage. This method misses pregnancies where care occurred but the final outcome wasn't recorded, potentially introducing bias into studies about prenatal exposures.
- Prenatal Encounter Identification
- Using the prenatal visit itself to identify a pregnancy, even if the final outcome is unknown. This requires statistical methods to handle missing data properly, as it can introduce selection bias if not handled correctly.
- Missingness Mechanism
- The underlying reason why some pregnancy outcomes are missing. The study tested three scenarios: unobserved miscarriage (no known outcome), measured covariates (known factors causing missingness), and a mix of both. This mechanism determines how identification methods perform.
- Direct Standardization
- A non-parametric method used to estimate treatment effects when comparing groups with different exposure levels. It was used here across the three analytic samples to calculate the total effect of initiating a treatment on preeclampsia risk.
Terminology used across episodes
This episode discusses
- Bias in studies of prenatal exposures using real-world data due to pregnancy identification method · Paper Radio
- Healthy Live Births Should be Considered as Competing Events when Estimating the Total Effect of Prenatal Medication Use on Pregnancy Outcomes
- Treatments for pregestational chronic conditions during pregnancy: emulating a target trial with a treatment decision design
The paper
Bias in studies of prenatal exposures using real-world data due to pregnancy identification method · Read on arXiv
Chase D. Latour, Jessie K. Edwards, Michele Jonsson Funk, Elizabeth A. Suarez, Kim Boggess, Mollie E. Wood
Department of Epidemiology, University of North Carolina at Chapel Hill · Carolina Population Center, University of North Carolina at Chapel Hill · Center for Pharmacoepidemiology and Treatment Science, Rutgers Institute for Health, Health Care Policy and Aging Research · Department of Biostatistics and Epidemiology, Rutgers School of Public Health · Department of Obstetrics and Gynecology, University of North Carolina Chapel Hill
Background: Researchers typically identify pregnancies in healthcare data based on observed outcomes (e.g., delivery). This outcome-based approach misses pregnancies that received prenatal care but whose outcomes were not recorded (e.g., at-home miscarriage), potentially inducing selection bias in effect estimates for prenatal exposures. Alternatively, prenatal encounters can be used to identify pregnancies, including those with unobserved outcomes. However, this prenatal approach requires methods to address missing data. Methods: We simulated 10,000,000 pregnancies and estimated the total effect of initiating treatment on the risk of preeclampsia. We generated data for 36 scenarios in which we varied the effect of treatment on miscarriage and/or preeclampsia; the percentage with missing outcomes (5% or 20%); and the cause of missingness: (1) measured covariates, (2) unobserved miscarriage, and (3) a mix of both. We then created three analytic samples to address missing pregnancy outcomes: observed deliveries, observed deliveries and miscarriages, and all pregnancies. Treatment effects were estimated using non-parametric direct standardization. Results: Risk differences (RDs) and risk ratios (RRs) from the three analytic samples were similarly biased when all missingness was due to unobserved miscarriage (log-transformed RR bias range: -0.12-0.33 among observed deliveries; -0.11-0.32 among observed deliveries and miscarriages; and-0.11-0.32 among all pregnancies). When predictors of missingness were measured, only the all pregnancies approach was unbiased (-0.27-0.33; -0.29-0.03; and-0.02-0.01, respectively). Conclusions: When all missingness was due to miscarriage, the analytic samples returned similar effect estimates. Only among all pregnancies did bias decrease as the proportion of missingness due to measured variables increased.
DOI: 10.1111/ppe.70043
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: "Bias in studies of prenatal exposures using real-world data due to pregnancy identification method".
Jane: Here's the gist of the paper. The findings are complex, revolving around how missing outcome data (e.g., delivery vs.
Tom: all pregnancies) affects estimates of treatment effects when the missingness mechanism is unobserved miscarriage versus measured covariates.
Jane: Here is a detailed synthesis combining the provided text and quotes: Detailed Summary of Findings: The study addresses a critical methodological challenge in analyzing prenatal exposures: how to identify pregnancies when outcomes (like delivery or miscarriage) are not fully recorded, which can introduce significant selection bias.
Tom: Researchers typically identify pregnancies based on observed outcomes (e.g., delivery), but this approach risks missing pregnancies that received prenatal care but whose final outcome was never documented, potentially biasing effect estimates for prenatal exposures.
Paper summary: Jane: The paper is testing how missing outcome data messes up our results when we try to study prenatal exposures, and they simulate a lot of scenarios to show this happens. They generate thirty-six different scenarios by changing the treatment effect on miscarriage or preeclampsia risk, and also varying the percentage of outcomes that are missing, whether it’s five percent or twenty percent.
Lu: They also vary the cause of that missingness, which is what makes it really interesting because they test three types: one where it’s measured covariates they knew about, another where the outcome was unobserved miscarriage, and a mix of both.
Tom: The main thing they conclude is that when the missingness comes from unobserved miscarriage, the risk differences and risk ratios from all three analytic samples are similar in their bias across observed deliveries, observed deliveries and miscarriages, and all pregnancies.
Meng: That suggests that if you can't measure why something is missing—like an unobserved miscarriage—the way you identify the pregnancy might not matter much for your final estimate of the treatment effect.
Jane: But then they show a different pattern when the missingness is due to measured variables, meaning it’s Missing At Random. In that case, only including all pregnancies seems to fix the bias across those scenarios.
Tom: That’s a key distinction here: if you know why data is missing, like if a known risk factor caused it, then including all pregnancies helps in the "all pregnancies" approach to get an unbiased estimate across all those conditions.
Lu: They also found that when everything was due to unobserved miscarriage, the bias in estimates for observed deliveries was similar across those three samples. Specifically, the log-transformed RR bias range stayed between negative zero point one two and positive zero point three three among observed deliveries.
Jane: So even when it’s just unobserved miscarriage, the identification method doesn't change much—the bias stays in that specific range no matter which sample you use.
Conclusion: Tom: So to wrap up this paper "Bias in studies of prenatal exposures using real-world data due to pregnancy identification method", they found that when the missingness is just due to unobserved miscarriage, the results are consistently biased regardless of whether you look at observed deliveries or all pregnancies.
Jane: But they found a point where including all pregnancies does matter, which is when the cause of missingness involves measured variables. In that specific case, only using all pregnancies gave them unbiased estimates across all thirty-six scenarios tested.
Lu: The paper points out that even when it’s just unobserved miscarriage, the bias level is consistent—it stays in that log-transformed RR bias range of negative zero point one two to positive zero point three three for observed deliveries, negative zero point one one to positive zero point thirty-two for observed deliveries and miscarriages, and negative zero point one one to positive zero point thirty-two for all pregnancies.
Meng: It suggests that if you can’t measure the missingness cause, like unobserved miscarriage, then your choice of identification strategy might not matter much in terms of bias. But if you *can* measure the reason for missingness, then including all pregnancies is necessary to keep things unbiased.
Jane: So what does this mean for people using these studies? It means that when we study prenatal exposures and we don't know exactly why data is missing, like an unobserved miscarriage, our identification method might not be the right one to use.
Tom: And if you do have measured variables causing the missingness, then sticking with the all pregnancies approach seems to be the most robust way to get a result that isn't skewed by that missing data.
Lu: The authors flag that their method stops working when they can’t measure the cause of missingness and include those unobserved outcomes in their analysis.
Jane: That’s a big limitation, so we have to be careful about how we frame our prenatal exposure studies with real-world data if we want reliable numbers.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates