Eric J. Tchetgen Tchetgen’s research while affiliated with University of Pennsylvania and other places

Our objective is to construct well-calibrated prediction sets for a time-to-event outcome subject to right-censoring with guaranteed coverage. Our approach is inspired by modern conformal inference literature, in that, unlike classical frameworks, we obviate the need for a well-specified parametric or semi-parametric survival model to accomplish our goal. In contrast to existing conformal prediction methods for survival data, which restrict censoring to be of Type I, whereby potential censoring times are assumed to be fully observed on all units in both training and validation samples, we consider the more common right-censoring setting in which either only the censoring time or only the event time of primary interest is directly observed, whichever comes first. Under a standard conditional independence assumption between the potential survival and censoring times given covariates, we propose and analyze two methods to construct valid and efficient lower predictive bounds for the survival time of a future observation. The proposed methods build upon modern semiparametric efficiency theory for censored data, in that the first approach incorporates inverse-probability-of-censoring weighting (IPCW), while the second approach is based on augmented-inverse-probability-of-censoring weighting (AIPCW). For both methods, we formally establish asymptotic coverage guarantees, and demonstrate both via theory and empirical experiments that AIPCW substantially improves efficiency over IPCW in the sense that its coverage error bound is of second-order mixed bias type, that is \emph{doubly robust}, and therefore guaranteed to be asymptotically negligible relative to the coverage error of IPCW.

Background The impact of pre-infection vaccination on the risk of long COVID remains unclear in the pediatric population. We aim to assess the effectiveness of BNT162b2 on long COVID risks with various strains of the SARS-CoV-2 virus in children and adolescents, using comparative effectiveness methods. We further explore if such pre-infection vaccination can mitigate the risk of long COVID beyond its established protective benefits against SARS-CoV-2 infection using causal mediation analysis. Methods We conducted real-world vaccine effectiveness study and mediation analysis using data from twenty health systems in the RECOVER PCORnet electronic health record (EHR) Program. Three independent cohorts were constructed including adolescents (12–20 years) during the Delta phase (July 1–November 30, 2021), children (5–11 years) and adolescents (12–20 years) during the Omicron phase (January 1–November 30, 2022). The intervention is first dose of the BNT162b2 vaccine in comparison with no receipt of COVID-19 vaccine. The outcomes of interest include conclusive or probable diagnosis of long COVID following a documented SARS-CoV-2 infection, and body-system-specific condition clusters of post-acute sequelae of SARS-CoV-2 infection (PASC), such as cardiac, gastrointestinal, musculoskeletal, respiratory, and syndromic categories. The effectiveness was reported as (1-relative risk)∗100 and mediating effects were reported as relative risks. Findings 112,590 adolescents (88,811 vaccinated) were included in the cohort for the analysis against Delta variant, and 188,894 children (101,277 vaccinated), and 84,735 adolescents (37,724 vaccinated) were included for the analysis against Omicron variant. During the Delta period, the estimated effectiveness of the BNT162b2 vaccine against long COVID among adolescents was 95.4% (95% CI: 90.9%–97.7%). During the Omicron phase, the estimated effectiveness against long COVID among children was 60.2% (95% CI: 40.3%–73.5%) and 75.1% (95% CI: 50.4%–87.5%) among adolescents. The direct effect of vaccination, defined as the effect beyond their impact on SARS-CoV-2 infections, was found to be statistically non-significant in all three study cohorts, with estimated relative risk of 1.08 (95% CI: 0.75–1.55) in the Delta study among adolescents, 1.24 (95% CI: 0.92–1.66) among children and 0.91 (95% CI: 0.69–1.19) among adolescents in the Omicron studies. Meanwhile, the estimated indirect effects, which are effects through protecting SARS-CoV-2 infections, were estimated as 0.04 (95% CI: 0.03–0.05) among adolescents during Delta phase, 0.31 (95% CI: 0.23–0.42) among children and 0.21 (95% CI: 0.16–0.27) among adolescents during the Omicron period. Interpretation Our study suggests that BNT162b2 was effective in reducing risk of long COVID outcomes in children and adolescents during the Delta and Omicron periods. The mediation analysis indicates the vaccine’s effectiveness is primarily derived from its role in reducing the risk of SARS-CoV-2 infection. Funding 10.13039/100000002National Institutes of Health.

Many proposals for the identification of causal effects require an instrumental variable that satisfies strong, untestable unconfoundedness and exclusion restriction assumptions. In this paper, we show how one can potentially identify causal effects under violations of these assumptions by harnessing a negative control population or outcome. This strategy allows one to leverage sup-populations for whom the exposure is degenerate, and requires that the instrument-outcome association satisfies a certain parallel trend condition. We develop the semiparametric efficiency theory for a general instrumental variable model, and obtain a multiply robust, locally efficient estimator of the average treatment effect in the treated. The utility of the estimators is demonstrated in simulation studies and an analysis of the Life Span Study.

Objective To identify predictors of HIV acquisition in Botswana. Design We applied machine learning approaches to identify HIV risk predictors using existing data from a large, well-characterized HIV incidence cohort. Methods We applied machine learning (randomForestSRC) to analyze data from a large population-based HIV incidence cohort enrolled in a cluster-randomized HIV prevention trial in 30 communities across Botswana. We sought to identify the most important risk factors for HIV acquisition, starting with 110 potential predictors. Results During a median 29-month follow-up of 8,551 HIV-negative adults, 147 (1.7%) acquired HIV. Our machine learning analysis found that for females, the most important variables for predicting HIV acquisition were the use of injectable hormonal contraception, frequency of sex in the prior 3 months with the most recent partner and residing in a community with HIV prevalence of 29% or higher. For the small proportion (0.3%) of females who had all three risk factors, their estimated probability of acquiring HIV during 29 months of follow-up was 34% (approximate annual incidence of 14%). For males, non-long-term relationships with the most recent partner and community HIV prevalence of 34% or higher were the most important HIV risk predictors. The 6% of males who had both risk factors had a 5.1% probability of acquiring HIV during the follow-up period (approximate annual incidence of 2.1%). Conclusions Machine learning approaches allowed us to analyze a large number of variables to efficiently identify key factors strongly predictive of HIV risk. These factors could help target HIV prevention interventions in Botswana. Clinical Trials Registration NCT01965470

The instrumental variable method is a prominent approach to recover under certain conditions, valid inference about a treatment causal effect even when unmeasured confounding might be present. In a groundbreaking paper, Imbens and Angrist (1994) established that a valid instrument nonparametrically identifies the average causal effect among compliers, also known as the local average treatment effect under a certain monotonicity assumption which rules out the existence of so-called defiers. An often-cited attractive property of monotonicity is that it facilitates a causal interpretation of the instrumental variable estimand without restricting the degree of heterogeneity of the treatment causal effect. In this paper, we introduce an alternative equally straightforward and interpretable condition for identification, which accommodates both the presence of defiers and heterogenous treatment effects. Mainly, we show that under our new conditions, the instrumental variable estimand recovers the average causal effect for the subgroup of units for whom the treatment is manipulable by the instrument, a subgroup which may consist of both defiers and compliers, therefore recovering an effect estimand we aptly call the Nudge Average Treatment Effect.

Negative controls are increasingly used to evaluate the presence of potential unmeasured confounding in observational studies. Beyond the use of negative controls to detect the presence of residual confounding, proximal causal inference (PCI) was recently proposed to de-bias confounded causal effect estimates, by leveraging a pair of treatment and outcome negative control or confounding proxy variables. While formal methods for statistical inference have been developed for PCI, these methods can be challenging to implement as they involve solving complex integral equations that are typically ill-posed. We develop a regression-based PCI approach, employing two-stage generalized linear regression models (GLMs) to implement PCI, which obviates the need to solve difficult integral equations. The proposed approach has merit in that (i) it is applicable to continuous, count, and binary outcomes cases, making it relevant to a wide range of real-world applications, and (ii) it is easy to implement using off-the-shelf software for GLMs. We establish the statistical properties of regression-based PCI and illustrate their performance in both synthetic and real-world empirical applications.

Constructing prediction sets with coverage guarantees for unobserved outcomes is a core problem in modern statistics. Methods for predictive inference have been developed for a wide range of settings, but usually only consider test data points one at a time. Here we study the problem of distribution-free predictive inference for a batch of multiple test points, aiming to construct prediction sets for functions -- such as the mean or median -- of any number of unobserved test datapoints. This setting includes constructing simultaneous prediction sets with a high probability of coverage, and selecting datapoints satisfying a specified condition while controlling the number of false claims. For the general task of predictive inference on a function of a batch of test points, we introduce a methodology called batch predictive inference (batch PI), and provide a distribution-free coverage guarantee under exchangeability of the calibration and test data. Batch PI requires the quantiles of a rank ordering function defined on certain subsets of ranks. While computing these quantiles is NP-hard in general, we show that it can be done efficiently in many cases of interest, most notably for batch score functions with a compositional structure -- which includes examples of interest such as the mean -- via a dynamic programming algorithm that we develop. Batch PI has advantages over naive approaches (such as partitioning the calibration data or directly extending conformal prediction) in many settings, as it can deliver informative prediction sets even using small calibration sample sizes. We illustrate that our procedures provide informative inference across the use cases mentioned above, through experiments on both simulated data and a drug-target interaction dataset.

Unmeasured confounding is one of the major concerns in causal inference from observational data. Proximal causal inference (PCI) is an emerging methodological framework to detect and potentially account for confounding bias by carefully leveraging a pair of negative control exposure (NCE) and outcome (NCO) variables, also known as treatment and outcome confounding proxies. Although regression-based PCI is well developed for binary and continuous outcomes, analogous PCI regression methods for right-censored time-to-event outcomes are currently lacking. In this paper, we propose a novel two-stage regression PCI approach for right-censored survival data under an additive hazard structural model. We provide theoretical justification for the proposed approach tailored to different types of NCOs, including continuous, count, and right-censored time-to-event variables. We illustrate the approach with an evaluation of the effectiveness of right heart catheterization among critically ill patients using data from the SUPPORT study. Our method is implemented in the open-access R package 'pci2s'.

To develop effective HIV prevention strategies that can guide public health policy it is important to identify the main sources of infection in HIV prevention studies. Accordingly, we devised a statistical approach that leverages deep– (or next generation) sequenced pathogen data to estimate the relative contribution of different sources of infection in community–randomized trials of infectious disease prevention. We applied this approach to the Botswana Combination Prevention Project (BCPP) and estimated that 90% [95% Confidence Interval (CI): 81 – 93] of new infections that occurred in individuals in communities that received combination prevention (including universal HIV test–and–treat) originated from individuals residing in communities outside of the trial area. We estimate that the relative impact of the intervention was greater in rural geographically isolated communities with limited opportunity for imported infections compared to communities neighboring major urban centers. Treating people with HIV limits the spread of infection to uninfected individuals; accordingly, counterfactual modeling scenarios estimated that a nationwide application of the intervention could have reduced transmissions to recipients in trial communities by 59% [3 – 87], much higher than the observed 30% reduction. Our results suggest that the impact of the BCPP trial intervention was substantially limited by sources of transmission outside the trial area, and that the impact of the intervention could be considerably larger if applied nationally. We recommend that studies of infectious disease prevention consider the impact of sources of transmission beyond the reach of the intervention when designing and evaluating interventions to inform public health programs.