Search

Obtaining causally interpretable meta-analysis results is challenging when there are differences in the distribution of effect modifiers between eligible trials. To overcome this, recent work on transportability methods has considered standardizing results of individual studies over the case-mix of a target population, prior to pooling them as in a classical random-effect meta-analysis. One practical challenge, however, is that case-mix standardization often requires individual participant data (IPD) on outcome, treatments and case-mix characteristics to be fully accessible in every eligible study, along with IPD case-mix characteristics for a random sample from the target population. In this paper, we aim to develop novel strategies to integrate aggregated-level data from eligible trials with non-accessible IPD into a causal meta-analysis, by extending moment-based methods frequently used for population-adjusted indirect comparison in health technology assessment. Since valid inference for these moment-based methods by M-estimation theory requires additional aggregated data that are often unavailable in practice, computational methods to address this concern are also developed. We assess the finite-sample performance of the proposed approaches by simulated data, and then apply these on real-world clinical data to investigate the effectiveness of risankizumab versus ustekinumab among patients with moderate to severe psoriasis.

When estimating heterogeneous treatment effects, missing outcome data can complicate treatment effect estimation, causing certain subgroups of the population to be poorly represented. In this work, we discuss this commonly overlooked problem and consider the impact that missing at random (MAR) outcome data has on causal machine learning estimators for the conditional average treatment effect (CATE). We then propose two de-biased machine learning estimators for the CATE, the mDR-learner and mEP-learner, which address the issue of under-representation by integrating inverse probability of censoring weights into the DR-learner and EP-learner respectively. We show that under reasonable conditions, these estimators are oracle efficient, and illustrate their favorable performance through simulated data settings, comparing them to existing CATE estimators, including comparison to estimators which use common missing data techniques. Guidance on the implementation of these estimators is provided and we present an example of their application using the ACTG175 trial, exploring treatment effect heterogeneity when comparing Zidovudine mono-therapy against alternative antiretroviral therapies among HIV-1-infected individuals., Comment: 34 pages, 6 figures, 4 tables

While synthetic data hold great promise for privacy protection, their statistical analysis poses significant challenges that necessitate innovative solutions. The use of deep generative models (DGMs) for synthetic data generation is known to induce considerable bias and imprecision into synthetic data analyses, compromising their inferential utility as opposed to original data analyses. This bias and uncertainty can be substantial enough to impede statistical convergence rates, even in seemingly straightforward analyses like mean calculation. The standard errors of such estimators then exhibit slower shrinkage with sample size than the typical 1 over root-$n$ rate. This complicates fundamental calculations like p-values and confidence intervals, with no straightforward remedy currently available. In response to these challenges, we propose a new strategy that targets synthetic data created by DGMs for specific data analyses. Drawing insights from debiased and targeted machine learning, our approach accounts for biases, enhances convergence rates, and facilitates the calculation of estimators with easily approximated large sample variances. We exemplify our proposal through a simulation study on toy data and two case studies on real-world data, highlighting the importance of tailoring DGMs for targeted data analysis. This debiasing strategy contributes to advancing the reliability and applicability of synthetic data in statistical inference., Comment: Accepted for the 38th Conference on Neural Information Processing Systems (NeurIPS 2024), joint first authors

The ICH E9 (R1) addendum on estimands, coupled with recent advancements in causal inference, has prompted a shift towards using model-free treatment effect estimands that are more closely aligned with the underlying scientific question. This represents a departure from traditional, model-dependent approaches where the statistical model often overshadows the inquiry itself. While this shift is a positive development, it has unintentionally led to the prioritization of an estimand's ability to perfectly answer the key scientific question over its practical learnability from data under plausible assumptions. We illustrate this by scrutinizing assumptions in the recent clinical trials literature on principal stratum estimands, demonstrating that some popular assumptions are not only implausible but often inevitably violated. We advocate for a more balanced approach to estimand formulation, one that carefully considers both the scientific relevance and the practical feasibility of estimation under realistic conditions.

As longitudinal data becomes more available in many settings, policy makers are increasingly interested in the effect of time-varying treatments (e.g. sustained treatment strategies). In settings such as this, the preferred analysis techniques are the g-methods, however these require the untestable assumption of no unmeasured confounding. Instrumental variable analyses can minimise bias through unmeasured confounding. Of these methods, the Two Stage Least Squares technique is one of the most well used in Econometrics, but it has not been fully extended, and evaluated, in full time-varying settings. This paper proposes a robust two stage least squares method for the econometric evaluation of time-varying treatment. Using a simulation study we found that, unlike standard two stage least squares, it performs relatively well across a wide range of circumstances, including model misspecification. It compares well with recent time-varying instrument approaches via g-estimation. We illustrate the methods in an evaluation of treatment intensification for Type-2 Diabetes Mellitus, exploring the exogeneity in prescribing preferences to operationalise a time-varying instrument., Comment: 34 Pages, 4 Figures, and 7 Tables

The STEP 1 randomized trial evaluated the effect of taking semaglutide vs placebo on body weight over a 68 week duration. As with any study evaluating an intervention delivered over a sustained period, non-adherence was observed. This was addressed in the original trial analysis within the Estimand Framework by viewing non-adherence as an intercurrent event. The primary analysis applied a treatment policy strategy which viewed it as an aspect of the treatment regimen, and thus made no adjustment for its presence. A supplementary analysis used a hypothetical strategy, targeting an estimand that would have been realised had all participants adhered, under the assumption that no post-baseline variables confounded adherence and change in body weight. In this paper we propose an alternative Instrumental Variable method to adjust for non-adherence which does not rely on the same `unconfoundedness' assumption and is less vulnerable to positivity violations (e.g., it can give valid results even under conditions where non-adherence is guaranteed). Unlike many previous Instrumental Variable approaches, it makes full use of the repeatedly measured outcome data, and allows for a time-varying effect of treatment adherence on a participant's weight. We show that it provides a natural vehicle for defining two distinct hypothetical estimands: the treatment effect if all participants would have adhered to semaglutide, and the treatment effect if all participants would have adhered to both semaglutide and placebo. When applied to the STEP 1 study, they both suggest a sustained, slowly decaying weight loss effect of semaglutide treatment.

Observational epidemiological studies commonly seek to estimate the causal effect of an exposure on an outcome. Adjustment for potential confounding bias in modern studies is challenging due to the presence of high-dimensional confounding, which occurs when there are many confounders relative to sample size or complex relationships between continuous confounders and exposure and outcome. Despite recent advances, limited evaluation, and guidance are available on the implementation of doubly robust methods, Augmented Inverse Probability Weighting (AIPW) and Targeted Maximum Likelihood Estimation (TMLE), with data-adaptive approaches and cross-fitting in realistic settings where high-dimensional confounding is present. Motivated by an early-life cohort study, we conducted an extensive simulation study to compare the relative performance of AIPW and TMLE using data-adaptive approaches in estimating the average causal effect (ACE). We evaluated the benefits of using cross-fitting with a varying number of folds, as well as the impact of using a reduced versus full (larger, more diverse) library in the Super Learner ensemble learning approach used for implementation. We found that AIPW and TMLE performed similarly in most cases for estimating the ACE, but TMLE was more stable. Cross-fitting improved the performance of both methods, but was more important for estimation of standard error and coverage than for point estimates, with the number of folds a less important consideration. Using a full Super Learner library was important to reduce bias and variance in complex scenarios typical of modern health research studies.

In May 2023, the U.S. Food and Drug Administration (FDA) released guidance for industry on "Adjustment for Covariates in Randomized Clinical Trials for Drugs and Biological Products". Covariate adjustment is a statistical analysis method for improving precision and power in clinical trials by adjusting for pre-specified, prognostic baseline variables. Though recommended by the FDA and the European Medicines Agency (EMA), many trials do not exploit the available information in baseline variables or make use only of the baseline measurement of the outcome. This is likely (partly) due to the regulatory mandate to pre-specify baseline covariates for adjustment, leading to challenges in determining appropriate covariates and their functional forms. We will explore the potential of automated data-adaptive methods, such as machine learning algorithms, for covariate adjustment, addressing the challenge of pre-specification. Specifically, our approach allows the use of complex models or machine learning algorithms without compromising the interpretation or validity of the treatment effect estimate and its corresponding standard error, even in the presence of misspecified outcome working models. This contrasts the majority of competing works which assume correct model specification for the validity of standard errors. Our proposed estimators either necessitate ultra-sparsity in the outcome model (which can be relaxed by limiting the number of predictors in the model) or necessitate integration with sample splitting to enhance their performance. As such, we will arrive at simple estimators and standard errors for the marginal treatment effect in randomized clinical trials, which exploit data-adaptive outcome predictions based on prognostic baseline covariates, and have low (or no) bias in finite samples even when those predictions are themselves biased.

Quantile regression is a powerful tool for detecting exposure-outcome associations given covariates across different parts of the outcome's distribution, but has two major limitations when the aim is to infer the effect of an exposure. Firstly, the exposure coefficient estimator may not converge to a meaningful quantity when the model is misspecified, and secondly, variable selection methods may induce bias and excess uncertainty, rendering inferences biased and overly optimistic. In this paper, we address these issues via partially linear quantile regression models which parametrize the conditional association of interest, but do not restrict the association with other covariates in the model. We propose consistent estimators for the unknown model parameter by mapping it onto a nonparametric main effect estimand that captures the (conditional) association of interest even when the quantile model is misspecified. This estimand is estimated using the efficient influence function under the nonparametric model, allowing for the incorporation of data-adaptive procedures such as variable selection and machine learning. Our approach provides a flexible and reliable method for detecting associations that is robust to model misspecification and excess uncertainty induced by variable selection methods. The proposal is illustrated using simulation studies and data on annual health care costs associated with excess body weight.

Recent advances in generative models facilitate the creation of synthetic data to be made available for research in privacy-sensitive contexts. However, the analysis of synthetic data raises a unique set of methodological challenges. In this work, we highlight the importance of inferential utility and provide empirical evidence against naive inference from synthetic data, whereby synthetic data are treated as if they were actually observed. Before publishing synthetic data, it is essential to develop statistical inference tools for such data. By means of a simulation study, we show that the rate of false-positive findings (type 1 error) will be unacceptably high, even when the estimates are unbiased. Despite the use of a previously proposed correction factor, this problem persists for deep generative models, in part due to slower convergence of estimators and resulting underestimation of the true standard error. We further demonstrate our findings through a case study., Comment: Accepted for the 40th Conference on Uncertainty in Artificial Intelligence (UAI 2024), *joint first authors

Orthogonal meta-learners, such as DR-learner, R-learner and IF-learner, are increasingly used to estimate conditional average treatment effects. They improve convergence rates relative to na\"{\i}ve meta-learners (e.g., T-, S- and X-learner) through de-biasing procedures that involve applying standard learners to specifically transformed outcome data. This leads them to disregard the possibly constrained outcome space, which can be particularly problematic for dichotomous outcomes: these typically get transformed to values that are no longer constrained to the unit interval, making it difficult for standard learners to guarantee predictions within the unit interval. To address this, we construct orthogonal meta-learners for the prediction of counterfactual outcomes which respect the outcome space. As such, the obtained i-learner or imputation-learner is more generally expected to outperform existing learners, even when the outcome is unconstrained, as we confirm empirically in simulation studies and an analysis of critical care data. Our development also sheds broader light onto the construction of orthogonal learners for other estimands.

Recently, it has become common for applied works to combine commonly used survival analysis modeling methods, such as the multivariable Cox model and propensity score weighting, with the intention of forming a doubly robust estimator of an exposure effect hazard ratio that is unbiased in large samples when either the Cox model or the propensity score model is correctly specified. This combination does not, in general, produce a doubly robust estimator, even after regression standardization, when there is truly a causal effect. We demonstrate via simulation this lack of double robustness for the semiparametric Cox model, the Weibull proportional hazards model, and a simple proportional hazards flexible parametric model, with both the latter models fit via maximum likelihood. We provide a novel proof that the combination of propensity score weighting and a proportional hazards survival model, fit either via full or partial likelihood, is consistent under the null of no causal effect of the exposure on the outcome under particular censoring mechanisms if either the propensity score or the outcome model is correctly specified and contains all confounders. Given our results suggesting that double robustness only exists under the null, we outline two simple alternative estimators that are doubly robust for the survival difference at a given time point (in the above sense), provided the censoring mechanism can be correctly modeled, and one doubly robust method of estimation for the full survival curve. We provide R code to use these estimators for estimation and inference in the supporting information.

There are now many options for doubly robust estimation; however, there is a concerning trend in the applied literature to believe that the combination of a propensity score and an adjusted outcome model automatically results in a doubly robust estimator and/or to misuse more complex established doubly robust estimators. A simple alternative, canonical link generalized linear models (GLM) fit via inverse probability of treatment (propensity score) weighted maximum likelihood estimation followed by standardization (the g-formula) for the average causal effect, is a doubly robust estimation method. Our aim is for the reader not just to be able to use this method, which we refer to as IPTW GLM, for doubly robust estimation, but to fully understand why it has the doubly robust property. For this reason, we define clearly, and in multiple ways, all concepts needed to understand the method and why it is doubly robust. In addition, we want to make very clear that the mere combination of propensity score weighting and an adjusted outcome model does not generally result in a doubly robust estimator. Finally, we hope to dispel the misconception that one can adjust for residual confounding remaining after propensity score weighting by adjusting in the outcome model for what remains `unbalanced' even when using doubly robust estimators. We provide R code for our simulations and real open-source data examples that can be followed step-by-step to use and hopefully understand the IPTW GLM method. We also compare to a much better-known but still simple doubly robust estimator.

Weighted average derivative effects (WADEs) are nonparametric estimands with uses in economics and causal inference. Debiased WADE estimators typically require learning the conditional mean outcome as well as a Riesz representer (RR) that characterises the requisite debiasing corrections. RR estimators for WADEs often rely on kernel estimators, introducing complicated bandwidth-dependant biases. In our work we propose a new class of RRs that are isomorphic to the class of WADEs and we derive the WADE weight that is optimal, in the sense of having minimum nonparametric efficiency bound. Our optimal WADE estimators require estimating conditional expectations only (e.g. using machine learning), thus overcoming the limitations of kernel estimators. Moreover, we connect our optimal WADE to projection parameters in partially linear models. We ascribe a causal interpretation to WADE and projection parameters in terms of so-called incremental effects. We propose efficient estimators for two WADE estimands in our class, which we evaluate in a numerical experiment and use to determine the effect of Warfarin dose on blood clotting function.

Motivated by applications in personalized medicine and individualized policymaking, there is a growing interest in techniques for quantifying treatment effect heterogeneity in terms of the conditional average treatment effect (CATE). Some of the most prominent methods for CATE estimation developed in recent years are T-Learner, DR-Learner and R-Learner. The latter two were designed to improve on the former by being Neyman-orthogonal. However, the relations between them remain unclear, and likewise the literature remains vague on whether these learners converge to a useful quantity or (functional) estimand when the underlying optimization procedure is restricted to a class of functions that does not include the CATE. In this article, we provide insight into these questions by discussing DR-Learner and R-Learner as special cases of a general class of weighted Neyman-orthogonal learners for the CATE, for which we moreover derive oracle bounds. Our results shed light on how one may construct Neyman-orthogonal learners with desirable properties, on when DR-Learner may be preferred over R-Learner (and vice versa), and on novel learners that may sometimes be preferable to either of these. Theoretical findings are confirmed using results from simulation studies on synthetic data, as well as an application in critical care medicine., Comment: 70 pages, 8 figures

When choosing estimands and estimators in randomized clinical trials, caution is warranted as intercurrent events, such as - due to patients who switch treatment after disease progression, are often extreme. Statistical analyses may then easily lure one into making large implicit extrapolations, which often go unnoticed. We will illustrate this problem of implicit extrapolations using a real oncology case study, with a right-censored time-to-event endpoint, in which patients can cross over from the control to the experimental treatment after disease progression, for ethical reasons. We resolve this by developing an estimator for the survival risk ratio contrasting the survival probabilities at each time t if all patients would take experimental treatment with the survival probabilities at those times t if all patients would take control treatment up to time t, using randomization as an instrumental variable to avoid reliance on no unmeasured confounders assumptions. This doubly robust estimator can handle time-varying treatment switches and right-censored survival times. Insight into the rationale behind the estimator is provided and the approach is demonstrated by re-analyzing the oncology trial.

We respond to comments on our paper, titled "Instrumental variable estimation of the causal hazard ratio.", Comment: To appear in Biometrics

In the standard difference-in-differences research design, the parallel trends assumption may be violated when the relationship between the exposure trend and the outcome trend is confounded by unmeasured confounders. Progress can be made if there is an exogenous variable that (i) does not directly influence the change in outcome means (i.e. the outcome trend) except through influencing the change in exposure means (i.e. the exposure trend), and (ii) is not related to the unmeasured exposure - outcome confounders on the trend scale. Such exogenous variable is called an instrument for difference-in-differences. For continuous outcomes that lend themselves to linear modelling, so-called instrumented difference-in-differences methods have been proposed. In this paper, we will suggest novel multiplicative structural mean models for instrumented difference-in-differences, which allow one to identify and estimate the average treatment effect on count and rare binary outcomes, in the whole population or among the treated, when a valid instrument for difference-in-differences is available. We discuss the identifiability of these models, then develop efficient semi-parametric estimation approaches that allow the use of flexible, data-adaptive or machine learning methods to estimate the nuisance parameters. We apply our proposal on health care data to investigate the risk of moderate to severe weight gain under sulfonylurea treatment compared to metformin treatment, among new users of antihyperglycemic drugs.

The recognition that personalised treatment decisions lead to better clinical outcomes has sparked recent research activity in the following two domains. Policy learning focuses on finding optimal treatment rules (OTRs), which express whether an individual would be better off with or without treatment, given their measured characteristics. OTRs optimize a pre-set population criterion, but do not provide insight into the extent to which treatment benefits or harms individual subjects. Estimates of conditional average treatment effects (CATEs) do offer such insights, but valid inference is currently difficult to obtain when data-adaptive methods are used. Moreover, clinicians are (rightly) hesitant to blindly adopt OTR or CATE estimates, not least since both may represent complicated functions of patient characteristics that provide little insight into the key drivers of heterogeneity. To address these limitations, we introduce novel nonparametric treatment effect variable importance measures (TE-VIMs). TE-VIMs extend recent regression-VIMs, viewed as nonparametric analogues to ANOVA statistics. By not being tied to a particular model, they are amenable to data-adaptive (machine learning) estimation of the CATE, itself an active area of research. Estimators for the proposed statistics are derived from their efficient influence curves and these are illustrated through a simulation study and an applied example.

The problem of how to best select variables for confounding adjustment forms one of the key challenges in the evaluation of exposure effects in observational studies, and has been the subject of vigorous recent activity in causal inference. A major drawback of routine procedures is that there is no finite sample size at which they are guaranteed to deliver exposure effect estimators and associated confidence intervals with adequate performance. In this work, we will consider this problem when inferring conditional causal hazard ratios from observational studies under the assumption of no unmeasured confounding. The major complication that we face with survival data is that the key confounding variables may not be those that explain the censoring mechanism. In this paper, we overcome this problem using a novel and simple procedure that can be implemented using off-the-shelf software for penalized Cox regression. In particular, we will propose tests of the null hypothesis that the exposure has no effect on the considered survival endpoint, which are uniformly valid under standard sparsity conditions. Simulation results show that the proposed methods yield valid inferences even when covariates are high-dimensional.

Longitudinal observational patient data can be used to investigate the causal effects of time-varying treatments on time-to-event outcomes. Several methods have been developed for controlling for the time-dependent confounding that typically occurs. The most commonly used is inverse probability weighted estimation of marginal structural models (MSM-IPTW). An alternative, the sequential trials approach, is increasingly popular, in particular in combination with the target trial emulation framework. This approach involves creating a sequence of `trials' from new time origins, restricting to individuals as yet untreated and meeting other eligibility criteria, and comparing treatment initiators and non-initiators. Individuals are censored when they deviate from their treatment status at the start of each `trial' (initiator/non-initiator) and this is addressed using inverse probability of censoring weights. The analysis is based on data combined across trials. We show that the sequential trials approach can estimate the parameter of a particular MSM, and compare it to a MSM-IPTW with respect to the estimands being identified, the assumptions needed and how data are used differently. We show how both approaches can estimate the same marginal risk differences. The two approaches are compared using a simulation study. The sequential trials approach, which tends to involve less extreme weights than MSM-IPTW, results in greater efficiency for estimating the marginal risk difference at most follow-up times, but this can, in certain scenarios, be reversed at late time points. We apply the methods to longitudinal observational data from the UK Cystic Fibrosis Registry to estimate the effect of dornase alfa on survival.

The (weighted) average treatment effect is commonly used to quantify the main effect of a binary exposure on an outcome. Extensions to continuous exposures, however, either quantify the effects of interventions that are rarely relevant (e.g., applying the same exposure level uniformly in the population), or consider shift interventions that are rarely intended, raising the question how large a shift to consider. Average derivative effects (ADEs) instead express the effect of an infinitesimal shift in each subject's exposure level, making inference less prone to extrapolation. ADEs, however, are rarely considered in practice because their estimation usually requires estimation of (a) the conditional density of exposure given covariates, and (b) the derivative of (a) w.r.t. exposure. Here, we introduce a class of estimands which can be inferred without requiring estimates of (a) and (b), but which reduce to ADEs when the exposure obeys a specific distribution determined by the choice of estimand in the class. We moreover show that when the exposure does not obey this distribution, our estimand represents an ADE w.r.t. an `intervention' exposure distribution. We identify the `optimal' estimand in our class and propose debiased machine learning estimators, by deriving influence functions under the nonparametric model.

The optimal moment to start renal replacement therapy in a patient with acute kidney injury (AKI) remains a challenging problem in intensive care nephrology. Multiple randomised controlled trials have tried to answer this question, but these can, by definition, only analyse a limited number of treatment initiation strategies. In view of this, we use routinely collected observational data from the Ghent University Hospital intensive care units (ICUs) to investigate different pre-specified timing strategies for renal replacement therapy initiation based on time-updated levels of serum potassium, pH and fluid balance in critically ill patients with AKI with the aim to minimize 30-day ICU mortality. For this purpose, we apply statistical techniques for evaluating the impact of specific dynamic treatment regimes in the presence of ICU discharge as a competing event. We discuss two approaches, a non-parametric one - using an inverse probability weighted Aalen-Johansen estimator - and a semiparametric one - using dynamic-regime marginal structural models. Furthermore, we suggest an easy to implement cross-validation technique that can be used for the out-of-sample performance assessment of the optimal dynamic treatment regime. Our work illustrates the potential of data-driven medical decision support based on routinely collected observational data., Comment: 49 pages, 4 figures

Due to concerns about parametric model misspecification, there is interest in using machine learning to adjust for confounding when evaluating the causal effect of an exposure on an outcome. Unfortunately, exposure effect estimators that rely on machine learning predictions are generally subject to so-called plug-in bias, which can render naive $p$-values and confidence intervals invalid. Progress has been made via proposals like targeted minimum loss estimation and more recently double machine learning, which rely on learning the conditional mean of both the outcome and exposure. Valid inference can then be obtained so long as both predictions converge (sufficiently fast) to the truth. Focusing on partially linear regression models, we show that a specific implementation of the machine learning techniques can yield exposure effect estimators that have small bias even when one of the first-stage predictions does not converge to the truth. The resulting tests and confidence intervals are doubly robust. We also show that the proposed estimators may fail to be regular when only one nuisance parameter is consistently estimated; nevertheless, we observe in simulation studies that our proposal can lead to reduced bias and improved confidence interval coverage in moderate-to-large samples., Comment: 46 pages, 9 figures

Evaluation of treatment effects and more general estimands is typically achieved via parametric modelling, which is unsatisfactory since model misspecification is likely. Data-adaptive model building (e.g. statistical/machine learning) is commonly employed to reduce the risk of misspecification. Naive use of such methods, however, delivers estimators whose bias may shrink too slowly with sample size for inferential methods to perform well, including those based on the bootstrap. Bias arises because standard data-adaptive methods are tuned towards minimal prediction error as opposed to e.g. minimal MSE in the estimator. This may cause excess variability that is difficult to acknowledge, due to the complexity of such strategies. Building on results from non-parametric statistics, targeted learning and debiased machine learning overcome these problems by constructing estimators using the estimand's efficient influence function under the non-parametric model. These increasingly popular methodologies typically assume that the efficient influence function is given, or that the reader is familiar with its derivation. In this paper, we focus on derivation of the efficient influence function and explain how it may be used to construct statistical/machine-learning-based estimators. We discuss the requisite conditions for these estimators to perform well and use diverse examples to convey the broad applicability of the theory.

Systematic reviews and meta-analyses of mediation studies are increasingly being implemented in practice. Nonetheless, the methodology for conducting such review and analysis is still in a development phase, with much room for improvement. In this paper, we highlight and discuss challenges that investigators face in mediation systematic reviews and meta-analyses, then propose ways of accommodating these in practice., Comment: Currently under peer-review at the American Journal of Epidemiology

Inferring the causal effect of a non-randomly assigned exposure on an outcome requires adjusting for common causes of the exposure and outcome to avoid biased conclusions. Notwithstanding the efforts investigators routinely make to measure and adjust for such common causes (or confounders), some confounders typically remain unmeasured, raising the prospect of biased inference in observational studies. Therefore, it is crucial that investigators can practically assess their substantive conclusions' relative (in)sensitivity to potential unmeasured confounding. In this article, we propose a sensitivity analysis strategy that is informed by the stability of the exposure effect over different, well-chosen subsets of the measured confounders. The proposal entails first approximating the process for recording confounders to learn about how the effect is potentially affected by varying amounts of unmeasured confounding, then extrapolating to the effect had hypothetical unmeasured confounders been additionally adjusted for. A large set of measured confounders can thus be exploited to provide insight into the likely presence of unmeasured confounding bias, albeit under an assumption about how data on the confounders are recorded. The proposal's ability to reveal the true effect and ensure valid inference after extrapolation is empirically compared with existing methods using simulation studies. We demonstrate the procedure using two different publicly available datasets commonly used for causal inference.