📊 statistics

A Unified Three-Stage Weighting Framework for Causal Inference and Mediation Analysis under Case-Control Sampling

This paper proposes a unified three-stage weighting framework that corrects for outcome-dependent sampling bias in case-control studies to enable accurate estimation of total and pathway-specific causal effects, including mediation effects, within a marginal structural modeling context.

Original authors: Tarikul Islam, Mahbub A. H. M. Latif

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tarikul Islam, Mahbub A. H. M. Latif

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand why people get a specific disease, like heart trouble. The gold standard for this research is to follow a huge group of healthy people for years, watching who gets sick and who stays healthy. This is like watching a whole forest grow from seedlings to see which trees get hit by lightning.

However, this "watching the whole forest" approach is expensive and takes forever, especially if the disease is rare (like finding a specific tree struck by lightning in a massive forest).

So, scientists often use a shortcut called a Case-Control Study. Instead of watching the whole forest, they go to the hospital and grab two groups:

The "Cases": People who already have the disease.
The "Controls": People who don't have the disease.

They then look back in time to see what these two groups did differently (like smoking, diet, or genetics).

The Problem: A Distorted Mirror

The paper argues that this shortcut creates a distorted mirror. Because the researchers specifically picked people based on whether they were sick or not, the data they collect doesn't look like the real world.

The Real World: Maybe only 5% of people have the disease.
The Study Data: Because they picked equal numbers of sick and healthy people, their data looks like 50% of people have the disease.

If you try to calculate the "real" risk or how a disease spreads through a chain of events (like: Smoking $\to$ High Blood Pressure $\to$ Heart Disease) using this distorted data, your results will be wrong. It's like trying to guess the average height of all humans by measuring only basketball players and jockeys; your average will be skewed.

Furthermore, most existing methods to fix this distortion require you to already know the exact percentage of people in the real world who have the disease. But often, scientists don't know this number. It's like trying to fix a broken map without knowing where the destination actually is.

The Solution: The "Three-Stage Weighting" Framework

The authors propose a new method called the Three-Stage Weighting Framework. Think of this as a three-step recipe to turn your distorted "hospital sample" back into a realistic "population model."

Stage 1: Guessing the Missing Piece (Prevalence Recovery)

Since we don't know the true disease rate in the real world, the authors use a clever trick. They take the data they have (the hospital patients) and mix it with a "synthetic" group of people generated from public records (like census data) that represents the general population's background (age, race, location).

They use a computer algorithm (like a smart sorting machine) to figure out: "How different is the background of my hospital patients compared to the real world?"
By analyzing these differences, the algorithm can mathematically guess the true disease rate in the real world, even without being told the answer upfront.

Stage 2: Rebalancing the Scale (Population Reconstruction)

Now that they have a good guess of the true disease rate, they go back to their hospital data and apply weights.

If the real world has very few sick people, but the study has many, they tell the computer: "Count each sick person in our study as only a tiny fraction of a person."
If the real world has many healthy people, but the study has few, they say: "Count each healthy person as a whole bunch of people."

This step effectively "un-distorts" the mirror, making the study data look statistically identical to the real population.

Stage 3: Tracing the Path (Causal & Mediation Analysis)

Now that the data looks like the real world, they can finally ask the big questions:

Total Effect: How much does smoking cause heart disease?
Mediation (The "How"): How much of that damage is direct, and how much happens through high blood pressure?

They use special "causal weights" to untangle these chains. It's like separating the threads of a knot to see exactly which string is pulling which part of the puzzle. This allows them to calculate not just if smoking causes heart disease, but how it does so (directly vs. through blood pressure).

The Results: Does it Work?

The authors tested this method in two ways:

Computer Simulations: They created fake data where they knew the "true" answer. When they used old methods, the answers were wrong. When they used their new three-stage method, the answers were spot-on, even when the disease was rare.
Real Data Test: They applied it to real health survey data (NHANES) regarding smoking and heart disease. They successfully reconstructed the real-world disease rate from a small, biased sample and calculated the causal effects.

The Bottom Line

This paper provides a new toolkit for scientists. It allows them to use the efficient, cheaper "Case-Control" study design (picking people from hospitals) without falling into the trap of distorted results.

Most importantly, it removes the need to know the "true disease rate" beforehand. It lets scientists figure out the rate themselves using available background data, fix the distortion, and then accurately measure how diseases happen and spread through different pathways. It turns a broken, biased snapshot into a clear, reliable picture of reality.

Technical Summary: A Unified Three-Stage Weighting Framework for Causal Inference and Mediation Analysis under Case–Control Sampling

1. Problem Statement

Case–control studies are a cornerstone of epidemiologic research, offering substantial efficiency gains for studying rare outcomes or when prospective follow-up is impractical. However, the retrospective, outcome-dependent sampling mechanism inherent to these designs distorts the marginal population outcome distribution. While conditional associations (e.g., odds ratios) remain identifiable, marginal causal estimands—such as the Average Treatment Effect (ATE) and pathway-specific mediation effects—cannot be directly recovered from case–control data without additional assumptions.

Existing methods for causal inference in case–control settings face three primary limitations:

Dependence on Known Prevalence: Most weighting-based approaches require prior knowledge of the population outcome prevalence, which is frequently unavailable or estimated with high uncertainty.
Restrictive Assumptions: Mediation analyses often rely on rare disease approximations or strong parametric assumptions that are difficult to verify empirically.
Lack of Unified Framework: Current literature largely focuses on total effects, lacking a unified strategy to simultaneously estimate total and pathway-specific mediation effects (direct, indirect, and interaction) under retrospective sampling without specialized likelihood formulations.

2. Methodology: The Three-Stage Weighting (3S-Weighting) Framework

The authors propose a modular, three-stage framework to reconstruct the target population distribution from retrospective case–control data, enabling valid causal and mediation inference using standard marginal structural modeling techniques. The method relies on externally available marginal covariate information (e.g., from census data) but does not require prior knowledge of the population outcome prevalence.

Stage 1: Population Outcome Prevalence Estimation
The first stage addresses the unknown population prevalence ( $\pi$ ) using density-ratio learning and label-shift correction.

Synthetic Population Generation: A synthetic population sample is generated using externally available marginal covariate distributions ( $P(X)$ ) to approximate the least informative joint distribution satisfying these constraints.
Density Ratio Estimation: The ratio between the target population covariate distribution and the observed case–control distribution, $r(X) = P_{Pop}(X)/P_{CC}(X)$ , is estimated by training a probabilistic classifier to distinguish between the synthetic sample and the observed case–control sample.
Label-Shift Correction: An initial outcome regression model is fitted on the case–control data. A label-shift correction factor, $\alpha(\pi)$ , is applied to adjust the intercept, recovering the population risk function $P_{Pop}(Y=1|X)$ .
Fixed-Point Iteration: The prevalence $\pi$ is estimated by solving a fixed-point equation where the expected population risk (weighted by the density ratio) equals the prevalence estimate itself.

Stage 2: Target Population Reconstruction
Once $\hat{\pi}$ is estimated, design-based weights are constructed to transform the retrospective sample into a pseudo-population that mimics the target population's marginal outcome distribution.

Design Weights: Cases ( $Y=1$ ) are weighted by $\hat{\pi}/\hat{p}_{CC}$ and controls ( $Y=0$ ) by $(1-\hat{\pi})/(1-\hat{p}_{CC})$ , where $\hat{p}_{CC}$ is the observed case proportion. This step restores the population outcome prevalence to $\hat{\pi}$ .

Stage 3: Causal and Mediation Weighting
Within the reconstructed pseudo-population, stabilized causal weights are applied to adjust for confounding and mediator imbalance.

Total Effect: Stabilized inverse probability of treatment weights (IPTW) are applied to estimate the Total Effect (TE).
Mediation Decomposition: Following the three-way decomposition framework (VanderWeele, 2014), specific weights are constructed to estimate:
- Pure Direct Effect (PDE): Weights adjust for treatment assignment and the mediator distribution under the control condition.
- Total Direct Effect (TDE): Weights adjust for treatment assignment and the mediator distribution under the treatment condition.
- Indirect and Interaction Effects: Derived multiplicatively from the TE, PDE, and TDE.
Estimation: Weighted marginal structural models (MSM) with log links are fitted to estimate relative risks (RR) for all causal components.

The framework is also extended to longitudinal dynamic settings involving time-varying exposures and mediators, where the three stages are applied sequentially across risk sets to reconstruct cumulative incidence functions.

3. Key Results

Simulation Studies:
Monte Carlo simulations were conducted under a counterfactual mediation framework with binary exposure, mediator, and outcome.

Bias Reduction: Conventional analyses ignoring retrospective sampling produced substantial bias in both total and mediation effect estimates. In contrast, the proposed 3S-weighting framework consistently recovered target population causal parameters across various sampling scenarios (sample sizes ranging from 500 to 8,000).
Consistency: As sample size increased, the empirical bias, standard deviation (SD), and root mean squared error (RMSE) for all estimands (TE, PDE, PIE, InE) decreased monotonically, demonstrating the consistency of the estimators.
Precision: Significant efficiency gains were observed, with SD reductions of over 70% at larger sample sizes compared to the smallest scenario.

Empirical Application (NHANES):
The framework was applied to 2017–2018 National Health and Nutrition Examination Survey (NHANES) data to analyze the causal mediation of smoking on cardiovascular disease (CVD) via hypertension.

Prevalence Recovery: The method successfully reconstructed the population CVD prevalence ( $\hat{\pi} = 0.114$ ) from a balanced case–control sample, closely matching the true population prevalence ($0.111$).
Effect Estimates: The estimated total effect of smoking on CVD was $RR_{TE} = 1.16$ . The analysis suggested that the association operated primarily through direct pathways rather than mediation via hypertension ( $RR_{PIE} = 1.03$ ).
Comparison: The proposed method yielded a smaller total effect estimate on the odds ratio scale (1.18) compared to a standard adjusted logistic regression (1.74). The authors attribute this attenuation to the proposed method's correction for outcome-dependent sampling, which targets a marginal population estimand rather than a conditional association.

4. Significance and Claims

The paper claims to provide a practical and theoretically grounded approach for causal inference in case–control studies that overcomes the "prevalence gap." Its primary contributions are:

Elimination of Prevalence Assumption: Unlike prior weighting methods, this framework does not require the population outcome prevalence to be known a priori; it estimates it internally using external covariate information.
Unified Mediation Analysis: It offers a unified strategy to estimate total, direct, indirect, and interaction effects simultaneously under retrospective sampling, avoiding rare disease approximations.
Modularity: The approach separates prevalence recovery, population reconstruction, and causal estimation, allowing it to be combined with standard weighting-based causal estimators without requiring specialized outcome-dependent likelihood formulations.
Broad Applicability: By integrating prevalence recovery and population reconstruction, the method expands the scope of causal questions addressable with case–control data, making it suitable for settings where full cohort follow-up is infeasible but auxiliary population data exists.

The authors conclude that while the method relies on standard causal assumptions (consistency, exchangeability, positivity) and assumes transportability of the covariate-outcome relationship, it represents a significant step toward bridging the gap between traditional case–control methodology and modern causal inference.