Faculty Bibliography

IMPORTANCE:Communication and adoption of modern study design and analytical techniques is of high importance for the improvement of clinical research from observational data. OBJECTIVE:To compare a modern method for statistical inference, including a target trial emulation framework and doubly robust estimation, with approaches common in the clinical literature, such as Cox proportional hazards models. DESIGN, SETTING, AND PARTICIPANTS:This retrospective cohort study used longitudinal electronic health record data for outcomes at 28-days from time of hospitalization within a multicenter New York, New York, hospital system. Participants included adult patients hospitalized between March 1 and May 15, 2020, with COVID-19 and not receiving corticosteroids for chronic use. Data were analyzed from October 2021 to March 2022. EXPOSURES:Corticosteroid exposure was defined as more than 0.5 mg/kg methylprednisolone equivalent in a 24-hour period. For target trial emulation, exposures were corticosteroids for 6 days if and when a patient met criteria for severe hypoxia vs no corticosteroids. For approaches common in clinical literature, treatment definitions used for variables in Cox regression models varied by study design (no time frame, 1 day, and 5 days from time of severe hypoxia). MAIN OUTCOMES AND MEASURES:The main outcome was 28-day mortality from time of hospitalization. The association of corticosteroids with mortality for patients with moderate to severe COVID-19 was assessed using the World Health Organization (WHO) meta-analysis of corticosteroid randomized clinical trials as a benchmark. RESULTS:A total of 3298 patients (median [IQR] age, 65 [53-77] years; 1970 [60%] men) were assessed, including 423 patients who received corticosteroids at any point during hospitalization and 699 patients who died within 28 days of hospitalization. Target trial emulation analysis found corticosteroids were associated with a reduced 28-day mortality rate, from 32.2%; (95% CI, 30.9%-33.5%) to 25.7% (95% CI, 24.5%-26.9%). This estimate is qualitatively identical to the WHO meta-analysis odds ratio of 0.66 (95% CI, 0.53-0.82). Hazard ratios using methods comparable with current corticosteroid research range in size and direction, from 0.50 (95% CI, 0.41-0.62) to 1.08 (95% CI, 0.80-1.47). CONCLUSIONS AND RELEVANCE:These findings suggest that clinical research based on observational data can be used to estimate findings similar to those from randomized clinical trials; however, the correctness of these estimates requires designing the study and analyzing the data based on principles that are different from the current standard in clinical research.

BACKGROUND:Although there is consensus that having a "high-enough" dose of buprenorphine (BUP-NX) or methadone is important for reducing relapse to opioid use, there is debate about what this dose is and how it should be attained. We estimated the extent to which different dosing strategies would affect risk of relapse over 12 weeks of treatment, separately for BUP-NX and methadone. METHODS:This was a secondary analysis of three comparative effectiveness trials. We examined four dosing strategies: 1) increasing dose in response to participant-specific opioid use, 2) increasing dose weekly until some minimum dose (16 mg BUP, 100 mg methadone) was reached, 3) increasing dose weekly until some minimum and increasing dose in response to opioid use thereafter (referred to as the "hybrid strategy"), and 4) keeping dose constant after the first 2 weeks of treatment. We used a longitudinal sequentially doubly robust estimator to estimate contrasts between dosing strategies on risk of relapse. RESULTS:For BUP-NX, increasing dose following the hybrid strategy resulted in the lowest risk of relapse. For methadone, holding dose constant resulted in greatest risk of relapse; the other three strategies performed similarly. For example, the hybrid strategy reduced week 12 relapse risk by 13 % (RR: 0.87, 95 %CI: 0.83-0.95) and by 20 % (RR: 0.80, 95 %CI: 0.71-0.90) for BUP-NX and methadone respectively, as compared to holding dose constant. CONCLUSIONS:Doses should be targeted toward minimum thresholds and, in the case of BUP-NX, raised when patients continue to use opioids.

The rapid finding of effective therapeutics requires efficient use of available resources in clinical trials. Covariate adjustment can yield statistical estimates with improved precision, resulting in a reduction in the number of participants required to draw futility or efficacy conclusions. We focus on time-to-event and ordinal outcomes. When more than a few baseline covariates are available, a key question for covariate adjustment in randomised studies is how to fit a model relating the outcome and the baseline covariates to maximise precision. We present a novel theoretical result establishing conditions for asymptotic normality of a variety of covariate-adjusted estimators that rely on machine learning (e.g.,

Violations of the positivity assumption (also called the common support condition) challenge health policy research and can result in significant bias, large variance, and invalid inference. We define positivity in the single- and multiple-timepoint (i.e., longitudinal) health policy evaluation setting, and discuss real-world threats to positivity. We show empirical evidence of the practical positivity violations that can result when attempting to estimate the effects of health policies (in this case, Naloxone Access Laws). In such scenarios, an alternative is to estimate the effect of a shift in law enactment (e.g., the effect if enactment had been delayed by some number of years). Such an effect corresponds to what is called a modified treatment policy, and dramatically weakens the required positivity assumption, thereby offering a means to estimate policy effects even in scenarios with serious positivity problems. We apply the approach to define and estimate the longitudinal effects of Naloxone Access Laws on opioid overdose rates.

Background/UNASSIGNED:Observational research provides a unique opportunity to learn causal effects when randomized trials are unavailable, but obtaining the correct estimates hinges on a multitude of design and analysis choices. We illustrate the advantages of modern causal inference methods and compare to standard research practice to estimate the effect of corticosteroids on mortality in hospitalized COVID-19 patients in an observational dataset. We use several large RCTs to benchmark our results. Methods/UNASSIGNED:Our retrospective data consists of 3,298 COVID-19 patients hospitalized at New York-Presbyterian March 1-May 15, 2020. We design our study using the target trial framework. We estimate the effect of an intervention consisting of six days of corticosteroids administered at the time of severe hypoxia and contrast with an intervention consisting of no corticosteroids. The dataset includes dozens of time-varying confounders. We estimate the causal effects using a doubly robust estimator where the probabilities of treatment, outcome, and censoring are estimated using flexible regressions via super learning. We compare these analyses to standard practice in clinical research, consisting of two approaches: (i)Cox models for an exposure of corticosteroids receipt within various time windows of hypoxia, and (ii)Cox time-varying model where the exposure is daily administration of corticosteroids beginning at hospitalization. Results/UNASSIGNED:Our target trial emulation estimates corticosteroids to reduce 28-day mortality from 32% (95% confidence interval: 31-34) to 23% (21-24). This is qualitatively identical to the WHO's RCT meta-analysis result. Hazard ratios from the Cox models range in size and direction from 0.50 (0.41-0.62) to 1.08 (0.80-1.47) and all study designs suffer from various forms of bias. Conclusion/UNASSIGNED:We demonstrate that clinical research based on observational data can unveil true causal relations. However, the correctness of these effect estimates requires designing and analyzing the data based on principles which are different from the current standard in clinical research. Widespread communication and adoption of these design and analytical techniques is of high importance for the improvement of clinical research.

OBJECTIVES/OBJECTIVE:To derive models that identify patients with COVID-19 at high risk for stroke. MATERIALS AND METHODS/METHODS:We used data from the AHA's Get With The Guidelines® COVID-19 Cardiovascular Disease Registry to generate models for predicting stroke risk among adults hospitalized with COVID-19 at 122 centers from March 2020-March 2021. To build our models, we used data on demographics, comorbidities, medications, and vital sign and laboratory values at admission. The outcome was a cerebrovascular event (stroke, TIA, or cerebral vein thrombosis). First, we used Cox regression with cross validation techniques to identify factors associated with the outcome in both univariable and multivariable analyses. Then, we assigned points for each variable based on corresponding coefficients to create a prediction score. Second, we used machine learning techniques to create risk estimators using all available covariates. RESULTS:Among 21,420 patients hospitalized with COVID-19, 312 (1.5%) had a cerebrovascular event. Using traditional Cox regression, we created/validated a COVID-19 stroke risk score with a C-statistic of 0.66 (95% CI, 0.60-0.72). The CANDLE score assigns 1 point each for prior cerebrovascular disease, afebrile temperature, no prior pulmonary disease, history of hypertension, leukocytosis, and elevated systolic blood pressure. CANDLE stratified risk of an acute cerebrovascular event according to low- (0-1: 0.2% risk), medium- (2-3: 1.1% risk), and high-risk (4-6: 2.1-3.0% risk) groups. Machine learning estimators had similar discriminatory performance as CANDLE: C-statistics, 0.63-0.69. CONCLUSIONS:We developed a practical clinical score, with similar performance to machine learning estimators, to help stratify stroke risk among patients hospitalized with COVID-19.

The same intervention can produce different effects in different sites. Existing transport mediation estimators can estimate the extent to which such differences can be explained by differences in compositional factors and the mechanisms by which mediating or intermediate variables are produced; however, they are limited to consider a single, binary mediator. We propose novel nonparametric estimators of transported interventional (in)direct effects that consider multiple, high-dimensional mediators and a single, binary intermediate variable. They are multiply robust, efficient, asymptotically normal, and can incorporate data-adaptive estimation of nuisance parameters. They can be applied to understand differences in treatment effects across sites and/or to predict treatment effects in a target site based on outcome data in source sites.