Empirical method validation using negative control outcomes

Implements OHDSI-style empirical calibration for MR estimates. For each negative control outcome (an outcome with no expected causal relationship to the exposure), the same allele score model is fitted to obtain a null MR estimate. The distribution of these null estimates is then used to assess systematic bias and, optionally, to calibrate the primary estimate's p-value and confidence interval via the EmpiricalCalibration package.

Usage

runNegativeControlAnalysis(
  cohortData,
  instrumentTable,
  covariateData = NULL,
  negativeControlColumns = NULL,
  primaryEstimate = NULL,
  modelBackend = "glm"
)

Arguments

cohortData: Data frame. Output of buildMRCohort, expected to contain negative control outcome columns (named nc_outcome_<id> or as specified in negativeControlColumns).
instrumentTable: Data frame. Output of getMRInstruments.
covariateData: Optional covariate data for adjustment. Same format as accepted by fitOutcomeModel.
negativeControlColumns: Optional character vector of column names in cohortData that contain negative control outcomes. If NULL, auto-detects columns matching nc_outcome_*.
primaryEstimate: Optional list. Output of computeMREstimate. If provided, the primary estimate is calibrated using the negative control distribution.
modelBackend: Character. Model fitting backend: "glm" or "cyclops". Default is "glm".

Value

A list with class "medusaNegativeControls" containing:

ncEstimates: Data frame with one row per negative control outcome: outcome_id, beta_ZY, se_ZY, beta_MR, se_MR, pval, log_rr, se_log_rr.
calibration: Output of EmpiricalCalibration::fitSystematicErrorModel() if the package is available, otherwise NULL.
calibratedPrimary: Named list with calibratedP, calibratedCiLower, calibratedCiUpper for the primary estimate, or NULL if primaryEstimate was not provided or calibration is unavailable.
biasDetected: Logical. TRUE if the systematic error model indicates the null distribution mean is significantly different from zero.

Details

Run Negative Control Outcome Analysis for Empirical Calibration

Negative control outcomes should be conditions that are biologically implausible as consequences of the exposure. A well-calibrated MR analysis should produce null estimates for these outcomes. Systematic deviation from the null suggests residual pleiotropy, population stratification, or other bias.

When EmpiricalCalibration is installed, the function fits a systematic error model to the negative control estimates and uses it to produce calibrated p-values and confidence intervals for the primary estimate. This approach is described in Schuemie et al. (2014, 2018) and is standard practice in OHDSI observational studies.

References

Schuemie, M. J., et al. (2014). Interpreting observational studies: why empirical calibration is needed to correct p-values. Statistics in Medicine, 33(2), 209-218.

Schuemie, M. J., et al. (2018). Empirical confidence interval calibration for population-level effect estimation studies in observational healthcare data. PNAS, 115(11), 2571-2577.

Examples

simData <- simulateMRData(n = 2000, nSnps = 5, seed = 123)
cohortData <- simulateNegativeControlOutcomes(simData$data, nControls = 10)
ncResults <- runNegativeControlAnalysis(
  cohortData = cohortData,
  instrumentTable = simData$instrumentTable
)
#> Running negative control analysis for 10 outcomes...
#>   10 of 10 negative control models fitted successfully.
#>   Package 'EmpiricalCalibration' not installed. Skipping systematic error model. Install with: remotes::install_github('ohdsi/EmpiricalCalibration')
#> Negative control analysis complete.
ncResults$ncEstimates
#>    outcome_id     beta_ZY     se_ZY     beta_MR     se_MR      pval      log_rr
#> 1           1 -0.15699438 0.2728961 -0.49566773 0.8624621 0.5654858 -0.49566773
#> 2           2  0.08931716 0.2343064  0.28199503 0.7400863 0.7031811  0.28199503
#> 3           3 -0.73834441 0.4406275 -2.33112483 1.4029781 0.0966021 -2.33112483
#> 4           4 -0.07535182 0.3249300 -0.23790321 1.0260469 0.8166437 -0.23790321
#> 5           5 -0.45467701 0.4086319 -1.43552094 1.2949883 0.2676371 -1.43552094
#> 6           6 -0.43102268 0.2793367 -1.36083873 0.8882859 0.1255267 -1.36083873
#> 7           7 -0.05332100 0.2544409 -0.16834679 0.8034363 0.8340318 -0.16834679
#> 8           8  0.08689478 0.3269849  0.27434702 1.0325888 0.7904791  0.27434702
#> 9           9 -0.22835265 0.2461892 -0.72096237 0.7793050 0.3548956 -0.72096237
#> 10         10  0.01321082 0.3272555  0.04170962 1.0332269 0.9677995  0.04170962
#>    se_log_rr
#> 1  0.8624621
#> 2  0.7400863
#> 3  1.4029781
#> 4  1.0260469
#> 5  1.2949883
#> 6  0.8882859
#> 7  0.8034363
#> 8  1.0325888
#> 9  0.7793050
#> 10 1.0332269