Routines for Performing Empirical Calibration of Observational Study Estimates

Routines for performing empirical calibration of observational study estimates. By using a set of negative control hypotheses we can estimate the empirical null distribution of a particular observational study setup. This empirical null distribution can be used to compute a calibrated p-value, which reflects the probability of observing an estimated effect size when the null hypothesis is true taking both random and systematic error into account. A similar approach can be used to calibrate confidence intervals, using both negative and positive controls. For more details, see Schuemie et al. (2013) and Schuemie et al. (2018) .

EmpiricalCalibration

EmpiricalCalibration is part of the OHDSI Methods Library.

Introduction

This R package contains routines for performing empirical calibration of observational study estimates. By using a set of negative control hypotheses we can estimate the empirical null distribution of a particular observational study setup. This empirical null distribution can be used to compute a calibrated p-value, which reflects the probability of observing an estimated effect size when the null hypothesis is true taking both random and systematic error into account, as described in the paper Interpreting observational studies: why empirical calibration is needed to correct p-values.

Also supported is empirical calibration of confidence intervals, based on the results for a set of negative and positive controls, as described in the paper Empirical confidence interval calibration for population-level effect estimation studies in observational healthcare data.

Features

• Estimate the empirical null distribution given the effect estimates of a set of negative controls.
• Estimate the calibrated p-value of a given hypothesis given the estimated empirical null distribution.
• Estimate a systematic error distribution given the effect estimates for a set of negative and positive controls.
• Estimate the calibrated confidence interval for a given estimate given the systematic error distribution.
• Produce various plots for evaluating the empirical calibration.
• Contains the data sets from the papers for illustration.

Screenshots and examples

data(sccs) #Load one of the included data sets
negatives <- sccs[sccs$groundTruth == 0,] #Select the negative controls
null <- fitNull(logRr = negatives$logRr, seLogRr = negatives$seLogRr) #Fit the null distribution
positive <- sccs[sccs$groundTruth == 1,]  #Select the positive control
plotCalibrationEffect(logRrNegatives = negatives$logRr,
                      seLogRrNegatives = negatives$seLogRr,
                      logRrPositives = positive$logRr,
                      seLogRrPositives = positive$seLogRr,
                      null = null)
 
#Compute the calibrated p-value:
calibrateP(null = null, logRr = positive$logRr, seLogRr = positive$seLogRr) #Compute calibrated p-value
[1] 0.8390598

Technology

This is a pure R package.

System requirements

Requires R (version 3.1.0 or newer).

Installation

In R, use the following commands to install the latest stable version from CRAN:

install.packages("EmpiricalCalibration")

To install the latest development version directly from GitHub, use:

install.packages("devtools")
library(devtools)
install_github("ohdsi/EmpiricalCalibration")

User Documentation

• Vignette: Empirical calibration of p-values
• Vignette: Empirical calibration of confidence intervals
• Package manual: EmpiricalCalibration.pdf

Support

• Developer questions/comments/feedback: OHDSI Forum
• We use the GitHub issue tracker for all bugs/issues/enhancements

License

EmpiricalCalibration is licensed under Apache License 2.0

Development

This package has been developed in RStudio.

Development status

This package is ready for use.

Acknowledgements

Martijn Schuemie is the author of this package.