Survival analysis endures as an old, yet active research field with applications that spread across many domains. Continuing improvements in data acquisition techniques pose constant challenges in applying existing survival analysis methods to these emerging data sets. In this paper, we present tools for fitting regularized Cox survival analysis models on high-dimensional, massive sample-size (HDMSS) data using a variant of the cyclic coordinate descent optimization technique tailored for the sparsity that HDMSS data often present. Experiments on two real data examples demonstrate that efficient analyses of HDMSS data using these tools result in improved predictive performance and calibration.

Objective

The current observational research literature shows extensive publication bias and contradiction. The Observational Health Data Sciences and Informatics (OHDSI) initiative seeks to improve research reproducibility through open science.

Methods

OHDSI has created an international federated data source of electronic health records and administrative claims that covers nearly 10% of the world's population. Using a common data model with a practical schema and extensive vocabulary mappings, data from around the world follow the identical format. OHDSI's research methods emphasize reproducibility, with a large-scale approach to addressing confounding using propensity score adjustment with extensive diagnostics; negative and positive control hypotheses to test for residual systematic error; a variety of data sources to assess consistency and generalizability; a completely open approach including protocol, software, models, parameters, and raw results so that studies can be externally verified; and the study of many hypotheses in parallel so that the operating characteristics of the methods can be assessed.

Results

OHDSI has already produced findings in areas like hypertension treatment that are being incorporated into practice, and it has produced rigorous studies of COVID-19 that have aided government agencies in their treatment decisions, that have characterized the disease extensively, that have estimated the comparative effects of treatments, and that the predict likelihood of advancing to serious complications.

Conclusions

OHDSI practices open science and incorporates a series of methods to address reproducibility. It has produced important results in several areas, including hypertension therapy and COVID-19 research.

Clinical trials often lack power to identify rare adverse drug events (ADEs) and therefore cannot address the threat rare ADEs pose, motivating the need for new ADE detection techniques. Emerging national patient claims and electronic health record databases have inspired post-approval early detection methods like the Bayesian self-controlled case series (BSCCS) regression model. Existing BSCCS models do not account for multiple outcomes, where pathology may be shared across different ADEs. We integrate a pathology hierarchy into the BSCCS model by developing a novel informative hierarchical prior linking outcome-specific effects. Considering shared pathology drastically increases the dimensionality of the already massive models in this field. We develop an efficient method for coping with the dimensionality expansion by reducing the hierarchical model to a form amenable to existing tools. Through a synthetic study we demonstrate decreased bias in risk estimates for drugs when using conditions with different true risk and unequal prevalence. We also examine observational data from the MarketScan Lab Results dataset, exposing the bias that results from aggregating outcomes, as previously employed to estimate risk trends of warfarin and dabigatran for intracranial hemorrhage and gastrointestinal bleeding. We further investigate the limits of our approach by using extremely rare conditions. This research demonstrates that analyzing multiple outcomes simultaneously is feasible at scale and beneficial.

Often the literature makes assertions of medical product effects on the basis of ' p < 0.05'. The underlying premise is that at this threshold, there is only a 5% probability that the observed effect would be seen by chance when in reality there is no effect. In observational studies, much more than in randomized trials, bias and confounding may undermine this premise. To test this premise, we selected three exemplar drug safety studies from literature, representing a case-control, a cohort, and a self-controlled case series design. We attempted to replicate these studies as best we could for the drugs studied in the original articles. Next, we applied the same three designs to sets of negative controls: drugs that are not believed to cause the outcome of interest. We observed how often p < 0.05 when the null hypothesis is true, and we fitted distributions to the effect estimates. Using these distributions, we compute calibrated p-values that reflect the probability of observing the effect estimate under the null hypothesis, taking both random and systematic error into account. An automated analysis of scientific literature was performed to evaluate the potential impact of such a calibration. Our experiment provides evidence that the majority of observational studies would declare statistical significance when no effect is present. Empirical calibration was found to reduce spurious results to the desired 5% level. Applying these adjustments to literature suggests that at least 54% of findings with p < 0.05 are not actually statistically significant and should be reevaluated.

Observational healthcare data, such as electronic health records and administrative claims, offer potential to estimate effects of medical products at scale. Observational studies have often been found to be nonreproducible, however, generating conflicting results even when using the same database to answer the same question. One source of discrepancies is error, both random caused by sampling variability and systematic (for example, because of confounding, selection bias, and measurement error). Only random error is typically quantified but converges to zero as databases become larger, whereas systematic error persists independent from sample size and therefore, increases in relative importance. Negative controls are exposure-outcome pairs, where one believes no causal effect exists; they can be used to detect multiple sources of systematic error, but interpreting their results is not always straightforward. Previously, we have shown that an empirical null distribution can be derived from a sample of negative controls and used to calibrate P values, accounting for both random and systematic error. Here, we extend this work to calibration of confidence intervals (CIs). CIs require positive controls, which we synthesize by modifying negative controls. We show that our CI calibration restores nominal characteristics, such as 95% coverage of the true effect size by the 95% CI. We furthermore show that CI calibration reduces disagreement in replications of two pairs of conflicting observational studies: one related to dabigatran, warfarin, and gastrointestinal bleeding and one related to selective serotonin reuptake inhibitors and upper gastrointestinal bleeding. We recommend CI calibration to improve reproducibility of observational studies.

Concerns over reproducibility in science extend to research using existing healthcare data; many observational studies investigating the same topic produce conflicting results, even when using the same data. To address this problem, we propose a paradigm shift. The current paradigm centres on generating one estimate at a time using a unique study design with unknown reliability and publishing (or not) one estimate at a time. The new paradigm advocates for high-throughput observational studies using consistent and standardized methods, allowing evaluation, calibration and unbiased dissemination to generate a more reliable and complete evidence base. We demonstrate this new paradigm by comparing all depression treatments for a set of outcomes, producing 17 718 hazard ratios, each using methodology on par with current best practice. We furthermore include control hypotheses to evaluate and calibrate our evidence generation process. Results show good transitivity and consistency between databases, and agree with four out of the five findings from clinical trials. The distribution of effect size estimates reported in the literature reveals an absence of small or null effects, with a sharp cut-off at p = 0.05. No such phenomena were observed in our results, suggesting more complete and more reliable evidence.This article is part of a discussion meeting issue 'The growing ubiquity of algorithms in society: implications, impacts and innovations'.

BACKGROUND:Uncertainty remains about the optimal monotherapy for hypertension, with current guidelines recommending any primary agent among the first-line drug classes thiazide or thiazide-like diuretics, angiotensin-converting enzyme inhibitors, angiotensin receptor blockers, dihydropyridine calcium channel blockers, and non-dihydropyridine calcium channel blockers, in the absence of comorbid indications. Randomised trials have not further refined this choice. METHODS:We developed a comprehensive framework for real-world evidence that enables comparative effectiveness and safety evaluation across many drugs and outcomes from observational data encompassing millions of patients, while minimising inherent bias. Using this framework, we did a systematic, large-scale study under a new-user cohort design to estimate the relative risks of three primary (acute myocardial infarction, hospitalisation for heart failure, and stroke) and six secondary effectiveness and 46 safety outcomes comparing all first-line classes across a global network of six administrative claims and three electronic health record databases. The framework addressed residual confounding, publication bias, and p-hacking using large-scale propensity adjustment, a large set of control outcomes, and full disclosure of hypotheses tested. FINDINGS:Using 4·9 million patients, we generated 22 000 calibrated, propensity-score-adjusted hazard ratios (HRs) comparing all classes and outcomes across databases. Most estimates revealed no effectiveness differences between classes; however, thiazide or thiazide-like diuretics showed better primary effectiveness than angiotensin-converting enzyme inhibitors: acute myocardial infarction (HR 0·84, 95% CI 0·75-0·95), hospitalisation for heart failure (0·83, 0·74-0·95), and stroke (0·83, 0·74-0·95) risk while on initial treatment. Safety profiles also favoured thiazide or thiazide-like diuretics over angiotensin-converting enzyme inhibitors. The non-dihydropyridine calcium channel blockers were significantly inferior to the other four classes. INTERPRETATION:This comprehensive framework introduces a new way of doing observational health-care science at scale. The approach supports equivalence between drug classes for initiating monotherapy for hypertension-in keeping with current guidelines, with the exception of thiazide or thiazide-like diuretics superiority to angiotensin-converting enzyme inhibitors and the inferiority of non-dihydropyridine calcium channel blockers. FUNDING:US National Science Foundation, US National Institutes of Health, Janssen Research & Development, IQVIA, South Korean Ministry of Health & Welfare, Australian National Health and Medical Research Council.

Evidence derived from existing health-care data, such as administrative claims and electronic health records, can fill evidence gaps in medicine. However, many claim such data cannot be used to estimate causal treatment effects because of the potential for observational study bias; for example, due to residual confounding. Other concerns include P hacking and publication bias. In response, the Observational Health Data Sciences and Informatics international collaborative launched the Large-scale Evidence Generation and Evaluation across a Network of Databases (LEGEND) research initiative. Its mission is to generate evidence on the effects of medical interventions using observational health-care databases while addressing the aforementioned concerns by following a recently proposed paradigm. We define 10 principles of LEGEND that enshrine this new paradigm, prescribing the generation and dissemination of evidence on many research questions at once; for example, comparing all treatments for a disease for many outcomes, thus preventing publication bias. These questions are answered using a prespecified and systematic approach, avoiding P hacking. Best-practice statistical methods address measured confounding, and control questions (research questions where the answer is known) quantify potential residual bias. Finally, the evidence is generated in a network of databases to assess consistency by sharing open-source analytics code to enhance transparency and reproducibility, but without sharing patient-level information. Here we detail the LEGEND principles and provide a generic overview of a LEGEND study. Our companion paper highlights an example study on the effects of hypertension treatments, and evaluates the internal and external validity of the evidence we generate.

Healthcare professionals increasingly rely on observational healthcare data, such as administrative claims and electronic health records, to estimate the causal effects of interventions. However, limited prior studies raise concerns about the real-world performance of the statistical and epidemiological methods that are used. We present the "OHDSI Methods Benchmark" that aims to evaluate the performance of effect estimation methods on real data. The benchmark comprises a gold standard, a set of metrics, and a set of open source software tools. The gold standard is a collection of real negative controls (drug-outcome pairs where no causal effect appears to exist) and synthetic positive controls (drug-outcome pairs that augment negative controls with simulated causal effects). We apply the benchmark using four large healthcare databases to evaluate methods commonly used in practice: the new-user cohort, self-controlled cohort, case-control, case-crossover, and self-controlled case series designs. The results confirm the concerns about these methods, showing that for most methods the operating characteristics deviate considerably from nominal levels. For example, in most contexts, only half of the 95% confidence intervals we calculated contain the corresponding true effect size. We previously developed an "empirical calibration" procedure to restore these characteristics and we also evaluate this procedure. While no one method dominates, self-controlled methods such as the empirically calibrated self-controlled case series perform well across a wide range of scenarios.

Objectives

To demonstrate the application of the Large-scale Evidence Generation and Evaluation across a Network of Databases (LEGEND) principles described in our companion article to hypertension treatments and assess internal and external validity of the generated evidence.

Materials and methods

LEGEND defines a process for high-quality observational research based on 10 guiding principles. We demonstrate how this process, here implemented through large-scale propensity score modeling, negative and positive control questions, empirical calibration, and full transparency, can be applied to compare antihypertensive drug therapies. We assess internal validity through covariate balance, confidence-interval coverage, between-database heterogeneity, and transitivity of results. We assess external validity through comparison to direct meta-analyses of randomized controlled trials (RCTs).

Results

From 21.6 million unique antihypertensive new users, we generate 6 076 775 effect size estimates for 699 872 research questions on 12 946 treatment comparisons. Through propensity score matching, we achieve balance on all baseline patient characteristics for 75% of estimates, observe 95.7% coverage in our effect-estimate 95% confidence intervals, find high between-database consistency, and achieve transitivity in 84.8% of triplet hypotheses. Compared with meta-analyses of RCTs, our results are consistent with 28 of 30 comparisons while providing narrower confidence intervals.

Conclusion

We find that these LEGEND results show high internal validity and are congruent with meta-analyses of RCTs. For these reasons we believe that evidence generated by LEGEND is of high quality and can inform medical decision-making where evidence is currently lacking. Subsequent publications will explore the clinical interpretations of this evidence.