Configuring the connection to the server

We need to tell R how to connect to the server where the data are. CohortMethod uses the DatabaseConnector package, which provides the createConnectionDetails function. Type ?createConnectionDetails for the specific settings required for the various database management systems (DBMS). For example, one might connect to a PostgreSQL database using this code:

library(CohortMethod)
connectionDetails <- createConnectionDetails(dbms = "postgresql",
                                             server = "localhost/ohdsi",
                                             user = "joe",
                                             password = "supersecret")

cdmDatabaseSchema <- "my_cdm_data"
cohortDatabaseSchema <- "my_results"
cohortTable <- "my_cohorts"
options(sqlRenderTempEmulationSchema = NULL)

The last few lines define the cdmDatabaseSchema, cohortDatabaseSchema, and cohortTable variables. We’ll use these later to tell R where the data in CDM format live, and where we want to write intermediate tables. Note that for Microsoft SQL Server, databaseschemas need to specify both the database and the schema, so for example cdmDatabaseSchema <- "my_cdm_data.dbo". For database platforms that do not support temp tables, such as Oracle, it is also necessary to provide a schema where the user has write access that can be used to emulate temp tables. PostgreSQL supports temp tables, so we can set options(sqlRenderTempEmulationSchema = NULL) (or not set the sqlRenderTempEmulationSchema at all.)

Preparing the exposures and outcome(s)

We need to define the exposures and outcomes for our study. Here, we will define our exposures using the OHDSI Capr package. We define two exposure cohorts, one for celecoxib and one for diclofenac. It is often a good idea to restrict your analysis to a specific indication, to maximize the comparability of the two cohorts. In this case, we will restrict to osteoarthritis of the knee. We will create a cohort for this indication, starting a the first ever diagnosis, and ending at observation period end.

library(Capr)

celecoxibConceptId <- 1118084
diclofenacConceptId <- 1124300
osteoArthritisOfKneeConceptId <- 4079750

celecoxib <- cs(
  descendants(celecoxibConceptId),
  name = "Celecoxib"
)

celecoxibCohort <- cohort(
  entry = entry(
    drugExposure(celecoxib)
  ),
  exit = exit(endStrategy = drugExit(celecoxib,
                                     persistenceWindow = 30,
                                     surveillanceWindow = 0))
)

diclofenac  <- cs(
  descendants(diclofenacConceptId),
  name = "Diclofenac"
)

diclofenacCohort <- cohort(
  entry = entry(
    drugExposure(diclofenac)
  ),
  exit = exit(endStrategy = drugExit(diclofenac,
                                     persistenceWindow = 30,
                                     surveillanceWindow = 0))
)

osteoArthritisOfKnee <- cs(
  descendants(osteoArthritisOfKneeConceptId),
  name = "Osteoarthritis of knee"
)

osteoArthritisOfKneeCohort <- cohort(
  entry = entry(
    conditionOccurrence(osteoArthritisOfKnee, firstOccurrence())
  ),
  exit = exit(
    endStrategy = observationExit()
  )
)
# Note: this will automatically assign cohort IDs 1,2, and 3, respectively:
exposuresAndIndicationCohorts <- makeCohortSet(celecoxibCohort, 
                                               diclofenacCohort, 
                                               osteoArthritisOfKneeCohort)

We’ll pull the outcome definition from the OHDSI PhenotypeLibrary:

library(PhenotypeLibrary)
outcomeCohorts <- getPlCohortDefinitionSet(77) # GI bleed

We combine the exposure and outcome cohort definitions, and use CohortGenerator to generate the cohorts:

allCohorts <- bind_rows(outcomeCohorts,
                        exposuresAndIndicationCohorts)

library(CohortGenerator)
cohortTableNames <- getCohortTableNames(cohortTable = cohortTable)
createCohortTables(connectionDetails = connectionDetails,
                   cohortDatabaseSchema = cohortDatabaseSchema,
                   cohortTableNames = cohortTableNames)
generateCohortSet(connectionDetails = connectionDetails,
                  cdmDatabaseSchema = cdmDatabaseSchema,
                  cohortDatabaseSchema = cohortDatabaseSchema,
                  cohortTableNames = cohortTableNames,
                  cohortDefinitionSet = allCohorts)

If all went well, we now have a table with the cohorts of interest. We can see how many entries per cohort:

connection <- DatabaseConnector::connect(connectionDetails)
sql <- "SELECT cohort_definition_id, COUNT(*) AS count
FROM @cohortDatabaseSchema.@cohortTable
GROUP BY cohort_definition_id"
cohortCounts <- DatabaseConnector::renderTranslateQuerySql(
  connection = connection,
  sql = sql,
  cohortDatabaseSchema = cohortDatabaseSchema,
  cohortTable = cohortTable
  )
DatabaseConnector::disconnect(connection)

##   cohort_concept_id   count
## 1                 1  917230
## 2                 2 1791695
## 3                 3  993116
## 4                77 1123643

Extracting the data from the server

Now we can tell CohortMethod to extract the cohorts, construct covariates, and extract all necessary data for our analysis.

Important: The target and comparator drug must not be included in the covariates, including any descendant concepts. You will need to manually add the drugs and descendants to the excludedCovariateConceptIds of the covariate settings. In this example code we exclude the concepts for celecoxib and diclofenac and specify addDescendantsToExclude = TRUE:

# Define which types of covariates must be constructed:
covSettings <- createDefaultCovariateSettings(
  excludedCovariateConceptIds = c(diclofenacConceptId, celecoxibConceptId),
  addDescendantsToExclude = TRUE
)

#Load data:
cohortMethodData <- getDbCohortMethodData(
  connectionDetails = connectionDetails,
  cdmDatabaseSchema = cdmDatabaseSchema,
  targetId = 1,
  comparatorId = 2,
  outcomeIds = 77,
  exposureDatabaseSchema = cohortDatabaseSchema,
  exposureTable = cohortTable,
  outcomeDatabaseSchema = cohortDatabaseSchema,
  outcomeTable = cohortTable,
  nestingCohortDatabaseSchema = cohortDatabaseSchema,
  nestingCohortTable = cohortTable,
  getDbCohortMethodDataArgs = createGetDbCohortMethodDataArgs(
    removeDuplicateSubjects = "keep first, truncate to second",
    firstExposureOnly = TRUE,
    washoutPeriod = 365,
    restrictToCommonPeriod = TRUE,
    nestingCohortId = 3,
    covariateSettings = covSettings
  )
)
cohortMethodData

There are many parameters, but they are all documented in the CohortMethod manual. The createDefaultCovariateSettings function is described in the FeatureExtraction package. In short, we are pointing the function to the table created earlier and indicating which concept IDs in that table identify the target, comparator, nesting cohort and outcome. We instruct that the default set of covariates should be constructed, including covariates for all conditions, drug exposures, and procedures that were found on or before the index date. To customize the set of covariates, please refer to the FeatureExtraction package vignette by typing vignette("UsingFeatureExtraction", package="FeatureExtraction").

We let the CohortMethod package construct our sets of new users: - For those patients who have exposure to both the target and the comparator, we keep whichever is their first, and truncate their exposure at the start of the second (if the first occurrences of the target and comparator start simultaneously the patient is removed). - We restrict to the first exposure overall (in this case redundant with the previous step). - We require a washout period of 365 days, meaning any patients with less than 365 days of prior observation are removed. - We restrict to the period in time when both drugs were observed in the database. This can be especially helpful when one of the exposures is new to the market. - We restrict to the nesting cohort, in this case our indication. Only exposures within the nesting cohort are kept.

We can see how many persons and exposures were left after each of these steps:

getAttritionTable(cohortMethodData)

The cohortMethodData() function extracts all data about the exposures, outcomes, and covariates from the server and stores them in thecohortMethodDataobject. This object uses theAndromedapackage to store information in a way that ensures R does not run out of memory, even when the data are large. We can use the genericsummary()` function to view some more information of the data we extracted:

summary(cohortMethodData)

saveCohortMethodData(cohortMethodData, "coxibVsNonselVsGiBleed.zip")

Fitting a propensity model

We can fit a propensity model using the covariates constructed by the getDbcohortMethodData() function:

ps <- createPs(cohortMethodData = cohortMethodData, population = studyPop)

The createPs() function uses the Cyclops package to fit a large-scale regularized logistic regression.

To fit the propensity model, Cyclops needs to know the hyperparameter value which specifies the variance of the prior. By default Cyclops will use cross-validation to estimate the optimal hyperparameter. However, be aware that this can take a really long time. You can use the prior and control parameters of the createPs() to specify Cyclops behavior, including using multiple CPUs to speed-up the cross-validation.

Propensity score diagnostics

We can compute the area under the receiver-operator curve (AUC) for the propensity score model:

computePsAuc(ps)

We can also plot the propensity score distribution. By default, the plotPS() function shows the preference score, a transformation of the propensity score that adjusts for differences in sizes between the target and comparator:

plotPs(ps,
       targetLabel = "Celexocib",
       comparatorLabel = "Diclofenac",
       showCountsLabel = TRUE,
       showAucLabel = TRUE,
       showEquipoiseLabel = TRUE)

It is also possible to inspect the propensity model itself by showing the covariates that have non-zero coefficients:

getPsModel(ps, cohortMethodData)

One advantage of using the regularization when fitting the propensity model is that most coefficients will shrink to zero and fall out of the model. It is a good idea to inspect the remaining variables for anything that should not be there, for example variations of the drugs of interest that we forgot to exclude.

Finally, we can inspect the percent of the population in equipoise, meaning they have a preference score between 0.3 and 0.7:

CohortMethod::computeEquipoise(ps)

A low equipoise indicates there is little overlap between the target and comparator populations.

Using the propensity score

We can use the propensity scores to trim, stratify, match, or weigh our population. For example, one could trim to equipoise, meaning only subjects with a preference score between 0.3 and 0.7 are kept:

trimmedPop <- trimByPs(ps,
                       trimByPsArgs = createTrimByPsArgs(
                         equipoiseBounds = c(0.3, 0.7)
                       ))
# Note: we need to also provide the original PS object so the preference score
# is computed using the original relative sizes of the cohorts:
plotPs(trimmedPop, ps)

Instead (or additionally), we could stratify the population based on the propensity score:

stratifiedPop <- stratifyByPs(ps, 
                              stratifyByPsArgs = createStratifyByPsArgs(
                                numberOfStrata = 5
                              ))
plotPs(stratifiedPop)

We can also match subjects based on propensity scores. In this example, we’re using one-to-one matching. By default, createPs() will use a caliper of 0.2 on the standardized logit scale:

matchedPop <- matchOnPs(ps,
                        matchOnPsArgs = createMatchOnPsArgs(
                          maxRatio = 1
                        ))
plotPs(matchedPop, ps)

We can see the effect of trimming and/or matching on the population using the getAttritionTable function:

getAttritionTable(matchedPop)

Or, if we like, we can plot an attrition diagram:

drawAttritionDiagram(matchedPop)

Evaluating covariate balance

To evaluate whether our use of the propensity score is indeed making the two cohorts more comparable, we can compute the covariate balance before and after trimming, matching, and/or stratifying:

balance <- computeCovariateBalance(matchedPop, cohortMethodData)

plotCovariateBalanceScatterPlot(balance, 
                                showCovariateCountLabel = TRUE, 
                                showMaxLabel = TRUE)

plotCovariateBalanceOfTopVariables(balance)

The ‘before matching’ population is the population as extracted by the getDbCohortMethodData function, so before any further filtering steps. We typically consider the populations to be balanced if, after PS adjustment, all covariates have a standardized difference of means smaller than 0.1.

Inspecting select population characteristics

It is customary to include a table in your paper that lists some select population characteristics before and after matching/stratification/trimming. This is usually the first table, and so will be referred to as ‘table 1’. To generate this table, you can use the createCmTable1 function:

createCmTable1(balance)

Generalizability

The goal of any propensity score adjustments is typically to make the target and comparator cohorts comparably, to allow proper causal inference. However, in doing so, we often need to modify our population, for example dropping subjects that have no counterpart in the other exposure cohort. The population we end up estimating an effect for may end up being very different from the population we started with. An important question is: how different? And it what ways? If the populations before and after adjustment are very different, our estimated effect may not generalize to the original population (if effect modification is present). The getGeneralizabilityTable() function informs on these differences:

getGeneralizabilityTable(balance)

In this case, because we used PS matching, we are likely aiming to estimate the average treatment effect in the treated (ATT). For this reason, the getGeneralizabilityTable() function automatically selected the target cohort as the basis for evaluating generalizability: it shows, for each covariate, the mean value before and PS adjustment in the target cohort. Also shown is the standardized difference of mean, and the table is reverse sorted by the absolute standard difference of mean (ASDM).

Fitting a simple outcome model

In theory we could fit an outcome model without using the propensity scores. In this example we are fitting an outcome model using a Cox regression:

outcomeModel <- fitOutcomeModel(
  population = studyPop,
  fitOutcomeModelArgs = createFitOutcomeModelArgs(
    modelType = "cox"
  )
)
outcomeModel

But of course we want to make use of the matching done on the propensity score:

outcomeModel <- fitOutcomeModel(
  population = matchedPop,
  fitOutcomeModelArgs = createFitOutcomeModelArgs(
    modelType = "cox"
  )
)
outcomeModel

Note that we define the sub-population to be only those in the matchedPop object, which we created earlier by matching on the propensity score.

Instead of matching or stratifying we can also perform Inverse Probability of Treatment Weighting (IPTW):

outcomeModel <- fitOutcomeModel(
  population = ps,
  fitOutcomeModelArgs = createFitOutcomeModelArgs(
    modelType = "cox",
    inversePtWeighting = TRUE,
    bootstrapCi = TRUE
  )
)
outcomeModel

Note that in this case we may want to use a bootstrap to compute the confidence interval.

Adding interaction terms

We may be interested whether the effect is different across different groups in the population. To explore this, we may include interaction terms in the model. In this example we include three interaction terms:

interactionCovariateIds <- c(8532001, 201826210, 21600960413) 
# 8532001 = Female
# 201826210 = Type 2 Diabetes
# 21600960413 = Concurent use of antithrombotic agents
outcomeModel <- fitOutcomeModel(
  population = matchedPop,
  cohortMethodData = cohortMethodData,
  fitOutcomeModelArgs = createFitOutcomeModelArgs(
    modelType = "cox",
    interactionCovariateIds = interactionCovariateIds
  )
)
outcomeModel

It is prudent to verify that covariate balance has also been achieved in the subgroups of interest. For example, we can check the covariate balance in the subpopulation of females:

balanceFemale <- computeCovariateBalance(
  population,
  cohortMethodData,
  computeCovariateBalanceArgs = createComputeCovariateBalanceArgs(
    subgroupCovariateId = 8532001
  )
)
plotCovariateBalanceScatterPlot(balanceFemale)

Adding covariates to the outcome model

One final refinement would be to use the same covariates we used to fit the propensity model to also fit the outcome model. This way we are more robust against misspecification of the model, and more likely to remove bias. For this we use the regularized Cox regression in the Cyclops package. (Note that the treatment variable is automatically excluded from regularization.)

outcomeModel <- fitOutcomeModel(
  population = matchedPop,
  cohortMethodData = cohortMethodData,
  fitOutcomeModelArgs = createFitOutcomeModelArgs(
    modelType = "cox",
    useCovariates = TRUE,
  )
)
outcomeModel

Inspecting the outcome model

We can inspect more details of the outcome model:

exp(coef(outcomeModel))

exp(confint(outcomeModel))

We can also see the covariates that ended up in the outcome model:

getOutcomeModel(outcomeModel, cohortMethodData)

Kaplan-Meier plot

We can create the Kaplan-Meier plot:

plotKaplanMeier(matchedPop, includeZero = FALSE)

Note that the Kaplan-Meier plot will automatically adjust for any stratification, matching, or trimming that may have been applied.

Time-to-event plot

We can also plot time-to-event, showing both events before and after the index date, and events during and outside the defined time-at-risk window. This plot can provide insight into the temporal pattern of the outcome relative to the exposures:

plotTimeToEvent(
  cohortMethodData = cohortMethodData,
  outcomeId = 77,
  minDaysAtRisk = 1,
  riskWindowStart = 0,
  startAnchor = "cohort start",
  riskWindowEnd = 30,
  endAnchor = "cohort end"
)

Note that this plot does not show any adjustment for the propensity score.