In OHDSI/Legend: Large-Scale Evidence Generation and Evaluation in a Network of Databases

library(DatabaseConnector)
library(knitr)
library(xtable)
library(ggplot2)
# library(kableExtra)
source("DataPulls.R")
source("PlotsAndTables.R")
options(knitr.kable.NA = '')

extraCriteria <- list(
  depression = "Further, we exclude patients with diagnoses of bipolar disorder or schizophrenia on or prior to their index date."
)

extraCovariates <- list(
  depression = "Prior number of depression treatments (1, 2, 3, 4, 5 or higher)"
)

negativeControls <- list(
  hypertension = 76
)

totalOutcomes <- list(
  hypertension = 57
)

totalExposures <- list(
  hypertension = "56 different anti-hypertensive medications, representing 15 different drug-class levels and 7 major drug class levels",
  stuff = "at least 2500 patients in both target and comparator cohors"
)

  abbr <- read.table(header = TRUE, sep = ",", stringsAsFactors = FALSE, text = "
name,shortName
ACE inhibitors,ACEIs
Aldosterone antagonist diuretics,AADs
Alpha-1 blockers,A1Bs
Angiotensin receptor blockers (ARBs),ARBs
Beta blockers - cardioselective,cBBs
Beta blockers - cardioselective and vasodilatory,cvBBs
Beta blockers - combined alpha and beta receptor,aBBs
Beta blockers - intrinsic sympathomimetic activity,sBBs
Beta blockers - noncardioselective,nBBs
Dihydropyridine calcium channel blockers (dCCB),dCCBs
Direct vasodilators,DVs
Loop diuretics,LDs
Nodihydropyridine calcium channel blockers (ndCCB),nCCBs
Potassium sparing diurectics,KSDs
Thiazide or thiazide-like diuretics,TZs")

  abbreviations <- abbr$shortName
  names(abbreviations) <- abbr$name

if (is.null(params$load)) {
  connectionDetails <- createConnectionDetails(dbms = "postgresql",
                                               server = paste(Sys.getenv("shinydbServer"),
                                                              Sys.getenv("shinydbDatabase"),
                                                              sep = "/"),
                                               port = Sys.getenv("shinydbPort"),
                                               user = Sys.getenv("shinydbUser"),
                                               password = Sys.getenv("shinydbPw"),
                                               schema = Sys.getenv("shinydbSchema"))
  connection <- connect(connectionDetails)

  targetName <- getExposureName(connection = connection, exposureId = params$targetId)
  comparatorName <- getExposureName(connection = connection, exposureId = params$comparatorId)
  outcomeName <- getOutcomeName(connection = connection, outcomeId = params$outcomeId)
  analyses <- getAnalyses(connection = connection)
  databaseDetails <- getDatabaseDetails(connection = connection,
                                        databaseId = params$databaseId)
  studyPeriod <- getStudyPeriod(connection = connection,
                                targetId = params$targetId,
                                comparatorId = params$comparatorId,
                                databaseId = params$databaseId)

  mainResults <- getMainResults(connection = connection,
                                targetIds = params$targetId,
                                comparatorIds = params$comparatorId,
                                outcomeIds = params$outcomeId,
                                databaseIds = params$databaseId,
                                analysisIds = c(1, 2, 3, 4))

  subgroupResults <- getSubgroupResults(connection = connection,
                                        targetIds = params$targetId,
                                        comparatorIds = params$comparatorId,
                                        outcomeIds = params$outcomeId,
                                        databaseIds = params$databaseId)

  controlResults <- getControlResults(connection = connection,
                                      targetId = params$targetId,
                                      comparatorId = params$comparatorId,
                                      analysisId = 1,
                                      databaseId = params$databaseId)

  attrition <- getAttrition(connection = connection,
                            targetId = params$targetId,
                            comparatorId = params$comparatorId,
                            outcomeId = params$outcomeId,
                            analysisId = 1,
                            databaseId = params$databaseId)

  followUpDist <- getCmFollowUpDist(connection = connection,
                                    targetId = params$targetId,
                                    comparatorId = params$comparatorId,
                                    outcomeId = params$outcomeId,
                                    databaseId = params$databaseId,
                                    analysisId = 1)

  balance <- getCovariateBalance(connection = connection,
                                 targetId = params$targetId,
                                 comparatorId = params$comparatorId,
                                 databaseId = params$databaseId,
                                 analysisId = 2)

  popCharacteristics <- getCovariateBalance(connection = connection,
                                            targetId = params$targetId,
                                            comparatorId = params$comparatorId,
                                            databaseId = params$databaseId,
                                            analysisId = 1,
                                            outcomeId = params$outcomeId)

  ps <- getPs(connection = connection,
              targetIds = params$targetId,
              comparatorIds = params$comparatorId,
              databaseId = params$databaseId)

  kaplanMeier <- getKaplanMeier(connection = connection,
                                targetId = params$targetId,
                                comparatorId = params$comparatorId,
                                outcomeId = params$outcomeId,
                                databaseId = params$databaseId,
                                analysisId = 2)

  DatabaseConnector::disconnect(connection)

  originalTargetName <- targetName
  originalComparatorName <- comparatorName

  if (params$targetId < 5000) {
    names <- unlist(strsplit(targetName, " & "))
    newName <- paste(
    sapply(names, function(name) {
      if (name %in% names(abbreviations)) {
        name <- abbreviations[name]
      }
      name
    }), collapse = " & ")

    targetName <- newName
  }

  if (params$comparatorId < 5000) {
    names <- unlist(strsplit(comparatorName, " & "))
    newName <- paste(
    sapply(names, function(name) {
      if (name %in% names(abbreviations)) {
        name <- abbreviations[name]
      }
      name
    }), collapse = " & ")

    comparatorName <- newName
  }

  if (!is.null(params$save)) {
    save(targetName, comparatorName, outcomeName, analyses, databaseDetails,
         studyPeriod, mainResults, subgroupResults, controlResults,
         attrition, followUpDist, balance, popCharacteristics, ps, kaplanMeier,
         file = params$save)
  }

  if (!is.null(params$save)) {
    save(targetName, comparatorName, outcomeName, analyses, databaseDetails,
         studyPeriod, mainResults, subgroupResults, controlResults,
         attrition, followUpDist, balance, popCharacteristics, ps, kaplanMeier,
         originalTargetName, originalComparatorName,
         file = params$save)
  }
} else {
  load(params$load)
}

targetName <- uncapitalize(targetName)
comparatorName <- uncapitalize(comparatorName)
outcomeName <- uncapitalize(outcomeName)
indicationName <- uncapitalize(params$indicationId)

databaseName <- databaseDetails$databaseName

minYear <- substr(studyPeriod$minDate, 1, 4)
maxYear <- substr(studyPeriod$maxDate, 1, 4)

coverage <- getCoverage(controlResults)

\dropcap{T}he Large-scale Evidence Generation and Evaluation in a Network of Databases (LEGEND) project aims to generate reliable evidence on the effects of medical interventions using observational healthcare data to support clinical decision making. LEGEND follows ten guiding principles (see Supporting Information); chief among these stand that we generate evidence at large-scale to achieve completeness and faciliate analysis of the overall distribution of effect size estimates across treatments and outcomes. We also generate evidence consistently by applying a systematic approach across all research questions and disseminate evidence regardless on the estimates effects to avoid publication bias. These aims help overcome the questionable reliable of observational research \citep{schuemie2018improving}. This LEGEND document reports the risk of r outcomeName between new users of r targetName and r comparatorName treated for r params$indicationId.

Worldwide, hypertension stands as a leading cause of mortality, with an increasing prevalence over the last two decades \citep{forouzanfar2017global}. The 2017 American College of Cardiology (ACC) / American Heart Association (AHA) clinical practice guidelines define hypertension based on averaged blood pressure (BP) measured in a healthcare setting; systolic BP between 130 - 139 mmHg or diastolic BP between 80 - 89 mmHg characterize stage 1 hypertension, and systolic BP $>$ 140 mmHg or diastolic BP $>$ 90 mmHg mark stage 2 hypertension \citep{whelton20182017}. Elevated BPs contribute to approximately half of all stroke and ischemic heart disease deaths and 13\% of all forms of deaths globally \citep{who2009global}. While antihypertensive therapies carry well-established benefits in reducing BP and the risk of major cardiovascular events, the health benefits and drug safety concerns of any one class of antihypertensive drugs relative to other classes as first-line therapy remains debatable. Part of this debate arises from a paucity of large randomized controlled trials and observational studies providing head-to-head comparisons between all pairs of individual drugs and drug classes. One notable counter-example is the Antihypertensive and Lipid-Lowering Treatment to Prevent Heart Attack Trial (ALLHAT) \citep{allhat2002major} that randomized 33,357 patients to receive either amlodipine (a dihydropyridine calcium channel blocker), chlorthalidone (a thiazide or thiazide-like diuertic) or lisinopril (an angiotensin converting enzyme inhibitor). However, ALLHAT employed only a single drug representative per class and did not include several other important antihypertensive classes, such as angiotensin receptor blockers and beta-blockers. Reboussin et al.~provide a systematic review of randomized controlled trials examining comparative benefits and harms of various antihypertensives as first-line therapy \citep{reboussin2017systematic}. Further observational study can help refine first-line therapy recommendations in terms of both treatment effecacy and relative drug safety, such as those from the 2017 ACC/AHA guidelines and the 2018 European Society of Cardiology (ESC) and European Society of Hypertension (ESH) Guidelines for the management of arterial hypertension\citep{williams20182018}.

Methods

Data source

We conduct a new-user cohort study comparing new users of r targetName with new users of r comparatorName in the r databaseName (r params$databaseId) database encoded in the Observational Medical Outcomes Partnership (OMOP) common data model (CDM) version 5 \citep{hripcsak2015observational,overhage2012validation,ryan2013empirical}. r sub(paste0(".*\\(",databaseDetails$databaseId,"\\)"), paste0("The ", databaseDetails$databaseId), databaseDetails$description) The study period spans from r studyPeriod$minDate to r studyPeriod$maxDate.

Study design

This study follows a retrospective, observational, comparative cohort design \citep{ryan2013empirical} . We include patients who are first time users of r targetName or r comparatorName, and who have a diagnosis of r indicationName on or prior to treatment initation. We require that patients have continuous observation in the database for at least 365 days prior to treatment initiation. We exclude patients with prior r outcomeName events and less than 1 day at risk. r if (indicationName %in% names(extraCriteria)) { extraCriteria[indicationName] } Links to full cohort details, include concept codes, are provided in the Supporting Information. The outcome of interest is r outcomeName. We begin the outcome risk window 1 day after treatment initation and consider two design choices to define the window end. First, we end the outcome time-at-risk window at first cessation of continuous drug exposure, analogous to an on-treatment design and, second, we end the outcome time-at-risk window when the patient is no longer observable in the database, analogous to an intent-to-treat design. Continuous drug exposures are constructed from the available longitudinal data by considering sequential prescriptions that have fewer than 30 days gap between prescriptions.

Statistical analysis

We conduct our cohort study using the open-source OHDSI CohortMethod R package \citep{schuemie2018cohortmethod} , with large-scale analytics achieved through the Cyclops R package \citep{suchard2013massive} . We use propensity scores (PSs) -- estimates of treatment exposure probability conditional on pre-treatment baseline features in the one year prior to treatment initiation -- to control for potential measured confoudning and improve balance between the target (r targetName) and comparator (r comparatorName) cohorts \citep{rosenbaum1983central} . We use an expansive PS model that includes all available patient demographics, drug, condition and procedure covariates generated through the FeatureExtraction R package \citep{schuemie2018featureextration} instead of a prespecified set of investigator-selected confounders. We perform PS stratification or variable-ratio matching and then estimate comparative r targetName-vs-r comparatorName hazard ratios (HRs) using a Cox proportional hazards model. Detailed covariate and methods informations are provided in the Supporting Information. We present PS and covariate balance metrics to assess successful confounding control, and provide HR estimates and Kaplan-Meier survival plots for the outcome of r outcomeName. We additionally estimate HRs for pre-specified subgroups to evaluate interactions with the treatment effect. For efficiency reasons, we fit subgroup Cox models using PS stratification only.

Residual study bias from unmeasured and systematic sources can exist in observational studies after controlling for measured confounding \citep{schuemie2014interpreting,schuemie2016robust} . To estimate such residual bias, we conduct negative control outcome experiments with r length(unique(controlResults$outcomeName)) negative control outcomes identified through a data-rich algorithm \citep{voss2017accuracy}. We fit the negative control estimates to an empirical null distribution that characterizes the study residual bias and is an important artifact from which to assess the study design \citep{schuemie2018improving} . Using the empirical null distribution and synthetic positive controls \citep{schuemie2018empirical} , we additionally calibrate all HR estimates, their 95\% confidence intervals (CIs) and the $p$-value to reject the null hypothesis of no differential effect (HR = 1). Empirical calibration serves as an important diagnostic tool to evaluate if residual systematic error is sufficient to cast doubt on the accuracy of the unknown effect estimate.

Results

Population characteristics

```r}")} drawAttritionDiagram(attrition, targetName, comparatorName)

Figure \ref{fig:attrition} diagrams the inclusion of study subjects from the `r params$databaseId` database under the on-treatment with stratification design.
We augment these counts with cohort sizes we identify for the remaining designs in Table \ref{tab:power}.
This table also reports total patient follow-up time, numbers of `r outcomeName` events these patients experience and unadjusted incidence rates.
Table \ref{tab:demographics} compares base-line characteristics between patient cohorts.
\begin{table*}
\caption{\textbf{Patient cohorts.}
Target (T) cohort is `r targetName` new-users.  Comparative (C) cohort is `r comparatorName` new-users.
We report total number of patients, follow-up time (in years), number of `r outcomeName` events, and event incidence rate (IR) per 1,000 patient years (PY) in patient cohorts, as well as the their minimum detectable relative risk (MDRR).
Note that the IR does not account for any stratification or matching.
}\label{tab:power}
\vspace*{-0.5em}
\centering{
\rowcolors{2}{gray!6}{white}
\begin{tabular}{lrrrrrrrrr}
\hiderowcolors
\toprule
 &
\multicolumn{2}{c}{Patients} &
\multicolumn{2}{c}{PYs} &
\multicolumn{2}{c}{Events} &
\multicolumn{2}{c}{IR} &
\\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9}
\multicolumn{1}{c}{Design} &
\multicolumn{1}{c}{T} & \multicolumn{1}{c}{C} &
\multicolumn{1}{c}{T} & \multicolumn{1}{c}{C} &
\multicolumn{1}{c}{T} & \multicolumn{1}{c}{C} &
\multicolumn{1}{c}{T} & \multicolumn{1}{c}{C} &
\multicolumn{1}{c}{MDRR} \\
\midrule
\showrowcolors
```r
table <- preparePowerTable(mainResults, analyses)

print(xtable(table, format = "latex"),
      include.rownames = FALSE,
      include.colnames = FALSE,
      hline.after = NULL,
      only.contents = TRUE,
      add.to.row = list(pos = list(nrow(table)), command = c("\\bottomrule")),
      sanitize.text.function = identity)

\end{tabular} } \end{table*}

\begin{table} \caption{\textbf{Patient demographics.} We report the standardized difference of population means (StdDiff) before and after stratification for selected base-line patient characteristics.}\label{tab:demographics} \vspace*{-0.5em} \centerline{ \resizebox{0.5\textwidth}{!}{ \rowcolors{2}{gray!6}{white} \begin{tabular}{lrrrrrr} \hiderowcolors \toprule & \multicolumn{3}{c}{Before stratification} & \multicolumn{3}{c}{After stratification} \ \cmidrule(lr){2-4} \cmidrule(lr){5-7} \multicolumn{1}{c}{Characteristic} & \multicolumn{1}{c}{T (\%)} & \multicolumn{1}{c}{C (\%)} & \multicolumn{1}{c}{StdDiff} & \multicolumn{1}{c}{T (\%)} & \multicolumn{1}{c}{C (\%)} & \multicolumn{1}{c}{StdDiff} \ \midrule \showrowcolors

table <- prepareTable1(balance, pathToCsv = file.path(params$root, "Table1Specs.csv"))
table <- table[3:nrow(table),]

print(xtable(table, format = "latex", align = c("l","l","r","r","r","r","r","r")),
      include.rownames = FALSE,
      include.colnames = FALSE,
      hline.after = NULL,
      only.contents = TRUE,
      add.to.row = list(pos = list(nrow(table)), command = c("\\bottomrule")),
      sanitize.text.function = identity)

\end{tabular} } } \end{table}

Patient characteristics balance

```r", "The preference score is a transformation of the propensity score that adjusts for prevalence differences between populations. A higher overlap indicates that subjects in the two populations are more similar in terms of their predicted probability of receiving one treatment over the other.\label{fig:ps}")} plotPs(ps, targetName, comparatorName)

Figure \ref{fig:ps} plots the preference score distributions, re-scalings of PS estimates to adjust for differential treatment prevalences, for patients treated with `r targetName` and `r comparatorName`.
We assess characteristics balance achieved through PS adjustment by comparing all characteristics' standardized difference (StdDiff) between treatment group means before and after PS trimming and stratification (Table \ref{tab:demographics}).
Figure \ref{fig:balance} plots StdDiff for all `r nrow(balance)` base-line patient features that serve as input for the PS model.
Before stratification, `r sum(na.omit(balance$beforeMatchingStdDiff) > 0.1)` features have a StdDiff $> 0.1$.  After stratification, the count is `r sum(na.omit(balance$afterMatchingStdDiff) > 0.1)`.
<!-- (TODO) Add judgement statement about measured confounding control using something like: `r judgePropensityScore(1.1,1.1)` -->

```r As a rule-of-thum, all values $<0.1$ is generall considered well-balance \\citep{austin2009balance}."}
plotCovariateBalanceScatterPlot(balance,
                                beforeLabel = "Before stratification",
                                afterLabel = "After stratification")

Outcome assessment

\begin{table*} \caption{Time-at-risk distributions as percentiles in the target and comparator cohorts after stratification.} \label{tab:fu} \centering{ \rowcolors{2}{gray!6}{white} \begin{tabular}{crrrrrrr} \hiderowcolors \toprule & \multicolumn{1}{c}{min} & \multicolumn{1}{c}{10\%} & \multicolumn{1}{c}{25\%} & \multicolumn{1}{c}{50\%} & \multicolumn{1}{c}{75\%} & \multicolumn{1}{c}{90\%} & \multicolumn{1}{c}{max} \ \midrule \showrowcolors

table <- prepareFollowUpDistTable(followUpDist)
table$Cohort <- c(targetName, comparatorName)

print(xtable(table, format = "latex"),
      include.rownames = FALSE,
      include.colnames = FALSE,
      hline.after = NULL,
      only.contents = TRUE,
      add.to.row = list(pos = list(nrow(table)), command = c("\\bottomrule")),
      sanitize.text.function = identity)

\end{tabular} } \end{table*}

Table \ref{tab:fu} details the time to first r outcomeName or censoring distributions for patients in the r targetName and r comparatorName cohorts. We report in Table \ref{tab:hr} estimated HRs comparing r targetName to r comparatorName for the on-treatment and intent-to-treat designs with stratification or matching. Figure \ref{fig:km} plots Kaplan-Meier survival curves for patients under the intent-to-treat design. To examine possible subgroup differences in treatment-effect, we include Table{tab:subgroups} that reports HR estimates separately for children (age $<$ 18), the elderly (age $\ge$ 65), female patients, pregnant women, patients with hepatic impairment and patients with renal impairment, using PS stratification.

```r This plot is adjusted for the propensity score stratification; the ", targetName, " curve shows the actual observed survival. The ", comparatorName, " curve applies reweighting to approximate the counterfactual of what ", targetName, " survival would look like had the ", targetName, " cohort been exposed to ", comparatorName, " instead. The shaded area denotes the 95\% CI.")} plotKm <- plotKaplanMeier(kaplanMeier, targetName, comparatorName)

\begin{table*}
\caption{
Hazard ratio (HR) estimates and their confidence intervals (CIs) and $p$-value to reject the null hypothesis of no difference (HR = 1) under various designs.
}
\label{tab:hr}
\centering{
\rowcolors{2}{gray!6}{white}
\begin{tabular}{lrrrr}
\hiderowcolors
\toprule
& \multicolumn{2}{c}{Uncalibrated} & \multicolumn{2}{c}{Calibrated} \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5}
\multicolumn{1}{c}{Design}
& \multicolumn{1}{c}{HR (95\% CI)} & \multicolumn{1}{c}{$p$}
& \multicolumn{1}{c}{HR (95\% CI)} & \multicolumn{1}{c}{$p$} \\
\midrule
\showrowcolors
```r
table <- mainResults
table$hr <- sprintf("%.2f (%.2f - %.2f)", mainResults$rr, mainResults$ci95Lb, mainResults$ci95Ub)
table$p <- sprintf("%.2f", table$p)
table$calHr <- sprintf("%.2f (%.2f - %.2f)", mainResults$calibratedRr, mainResults$calibratedCi95Lb, mainResults$calibratedCi95Ub)
table$calibratedP <- sprintf("%.2f", table$calibratedP)
table <- merge(table, analyses)
table <- table[, c("description", "hr", "p", "calHr", "calibratedP")]

print(xtable(table),
      include.rownames = FALSE,
      include.colnames = FALSE,
      hline.after = NULL,
      only.contents = TRUE,
      add.to.row = list(pos = list(nrow(table)), command = c("\\bottomrule")),
      sanitize.text.function = identity)

\end{tabular} } \end{table*}

\begin{table*} \caption{ Subgroup analyses. We report HR estimates, their 95\% CIs and uncalibrated and calibrated (cal) $p$-values to reject the null hypothesis of no difference in five pre-specified patient subgroups. } \label{tab:subgroups} \centering{ \rowcolors{2}{gray!6}{white} \begin{tabular}{lrrrrrrrr} \hiderowcolors \toprule & \multicolumn{2}{c}{Subjects} & \multicolumn{3}{c}{On-treatment} & \multicolumn{3}{c}{Intent-to-treat} \ \cmidrule(lr){2-3} \cmidrule(lr){4-6} \cmidrule(lr){7-9} \multicolumn{1}{c}{Subgroup} & \multicolumn{1}{c}{T} & \multicolumn{1}{c}{C} & \multicolumn{1}{c}{HR (95\% CI)} & \multicolumn{1}{c}{$p$} & \multicolumn{1}{c}{cal-$p$} & \multicolumn{1}{c}{HR (95\% CI)} & \multicolumn{1}{c}{$p$} & \multicolumn{1}{c}{cal-$p$} \ \midrule \showrowcolors

if (nrow(subgroupResults) > 0) {
  table <- prepareSubgroupTable(subgroupResults)
  print(xtable(table),
        include.rownames = FALSE,
        include.colnames = FALSE,
        hline.after = NULL,
        only.contents = TRUE,
        add.to.row = list(pos = list(nrow(table)), command = c("\\bottomrule")),
        sanitize.text.function = identity)
}

\end{tabular} } \end{table*}

Residual systematic error

In the absense of bias, we expect 95\% of negative and positive control estimate 95\% confidence intervals to include their presumed HR. In the case of negative controls, the presumed HR = 1. Figure \ref{fig:negatives} describes the negative and positive control estimates under the on-treatment with PS stratification design. Before calibration, negative and positive controls demonstrate r judgeCoverage(coverage[coverage$group == "Uncalibrated", "coverage"]) coverage. After calibration, controls demonstrate r judgeCoverage(coverage[coverage$group == "Calibrated", "coverage"]) coverage.

plot <- plotScatter(controlResults)
suppressMessages(ggsave(paste0(opts_chunk$get("fig.path"), "error.pdf"), plot,
                        width = 14, height = 4, units = "in"))

\begin{figure} \centerline{ \includegraphics[width=1.0\textwidth]{r opts_chunk$get("fig.path")error} } \caption{ \textbf{Evaluation of effect estimation between r targetName and r comparatorName new-users}. The top plots HRs and their corresponding standard errors before calibration for each negative and synthetic positive control. The bottom plots the same estimates after calibration. } \label{fig:negatives} \end{figure}

Conclusions

We find that r targetName has a r judgeHazardRatio(mainResults[params$primary,"calibratedCi95Lb"], mainResults[params$primary,"calibratedCi95Ub"]) risk of r outcomeName as compared to r comparatorName within the population that the r params$databaseId represents.

Supporting Information {#suppinfo}

Here we enumerate the guiding principles of LEGEND and provide linking details on study cohorts and design.

LEGEND principles

\begin{enumerate}[noitemsep] \item Evidence will be generated at large-scale. \item Dissemination of the evidence will not depend on the estimated effects. \item Evidence will be generated by consistently applying a systematic approach across all research questions. \item Evidence will be generated using a pre-specified analysis design. \item Evidence will be generated using open source software that is freely available to all. \item Evidence generation process will be empirically evaluated by including control research questions where the true effect size is known. \item Evidence will be generated using best-practices. \item LEGEND will not be used to evaluate methods. \item Evidence will be updated on a regular basis. \item No patient-level data will be shared between sites in the network, only aggregated data. \end{enumerate}

Study cohorts

Please see the LEGEND r indicationName Study protocol (https://github.com/OHDSI/Legend/tree/master/Documents) for complete specification of the r targetName, r comparatorName and r outcomeName cohorts using ATLAS (http://www.ohdsi.org/web/atlas).

Negative controls

We selected negative controls using a process similar to that outlined by \cite{voss2017accuracy}. We first construct a list of all conditions that satisfy the following criteria with respect to all drug exposures in the LEGEND r indicationName study: \begin{itemize}[noitemsep] \item No Medline abstract where the MeSH terms suggests a drug-condition association \citep{winnenburg2015leveraging}, \item No mention of the drug-condition pair on a US product label in the Adverse Drug Reactions'' orPostmarketing'' section \citep{duke2013consistency}, \item No US spontaneous reports suggesting that the pair is in an adverse event relationship \citep{evans2001use,banda2016curated}, \item OMOP vocabulary does not suggest that the drug is indicated for the condition, \item Vocabulary conditional concepts are usable (i.e., not too broad, not suggestive of an adverse event relationship, no pregnancy related), and \item Exact condition concept itself is used in patient level data. \end{itemize}

negatives <- controlResults[controlResults$effectSize == 1.0 & !is.na(controlResults$rr), ]

We optimize remaining condition concepts, such that parent concepts remove children as defined by the OMOP vocabulary and perform manual review to exclude any pairs that may still be in a causal relationship or too similar to the study outcome. For r indicationName, this process led to a candidate list of r negativeControls[indicationName] negative controls for which table can be found in study protocol (https://github.com/OHDSI/Legend/tree/master/Documents). In the comparison of r targetName and r comparatorName in the r params$databaseId database, r nrow(negatives) negative controls had sufficient outcomes to return estimable HRs. We list these conditions in Table \ref{tab:negatives}

\begin{table} \caption{Negative controls employed in the comparison of r targetName and r comparatorName in the r params$databaseId database.} \label{tab:negatives} \centering{ \rowcolors{2}{gray!6}{white} \begin{tabular}{l} \hiderowcolors \toprule \multicolumn{1}{c}{Condition} \ \midrule \showrowcolors

table <- data.frame(outcomeName = sort(negatives[,"outcomeName"]))
print(xtable(table),
      include.rownames = FALSE,
      include.colnames = FALSE,
      hline.after = NULL,
      only.contents = TRUE,
      add.to.row = list(pos = list(nrow(table)), command = c("\\bottomrule")),
      sanitize.text.function = identity)

\end{tabular} } \end{table}

Covariate sets

\begin{itemize}[noitemsep] \item Demographics (age in 5-year bands, gender, index year, index month) \item Conditions (condition occurrence in lookback window) \begin{itemize}[noitemsep] \item in 365 days prior to index date \item in 30 days prior to index date \end{itemize} \item Condition aggregation \begin{itemize}[noitemsep] \item SMOMED \end{itemize} \item Drugs (drug occurrence in lookback window) \begin{itemize}[noitemsep] \item in 365 days prior to index date \item in 30 days prior to index date \end{itemize} \item Drug aggregation \begin{itemize}[noitemsep] \item Ingredient \item ATC class \end{itemize} \item Risk Scores (Charlson comorbidity index) r if (indicationName %in% names(extraCovariates)) { paste("\\item", extraCovariates[indicationName]) } \end{itemize}

We exclude all covariates that occur in fewer than 0.1\% of patients within the target and comparator cohorts prior to model fitting for computational efficiency.

OHDSI/Legend documentation built on Dec. 29, 2020, 3:52 a.m.