EHR (version 0.4-11)

analysisPheWAS: Statistical Analysis for PheWAS

Description

Implement three commonly used statistical methods to analyze data for Phenome Wide Association Study (PheWAS)

Usage

analysisPheWAS(
  method = c("firth", "glm", "lr"),
  adjust = c("PS", "demo", "PS.demo", "none"),
  Exposure,
  PS,
  demographics,
  phenotypes,
  data
)

Value

estimate

the estimate of log odds ratio.

stdError

the standard error.

statistic

the test statistic.

pvalue

the p-value.

Arguments

method

define the statistical analysis method from 'firth', 'glm', and 'lr'. 'firth': Firth's penalized-likelihood logistic regression; 'glm': logistic regression with Wald test, 'lr': logistic regression with likelihood ratio test.

adjust

define the adjustment method from 'PS','demo','PS.demo', and 'none'. 'PS': adjustment of PS only; 'demo': adjustment of demographics only; 'PS.demo': adjustment of PS and demographics; 'none': no adjustment.

Exposure

define the variable name of exposure variable.

PS

define the variable name of propensity score.

demographics

define the list of demographic variables.

phenotypes

define the list of phenotypes that need to be analyzed.

data

define the data.

Author

Leena Choi and Cole Beck

Details

Implements three commonly used statistical methods to analyze the associations between exposure (e.g., drug exposure, genotypes) and various phenotypes in PheWAS. Firth's penalized-likelihood logistic regression is the default method to avoid the problem of separation in logistic regression, which is often a problem when analyzing sparse binary outcomes and exposure. Logistic regression with likelihood ratio test and conventional logistic regression with Wald test can be also performed.

Examples

## use small datasets to run this example
data(dataPheWASsmall)
## make dd.base with subset of covariates from baseline data (dd.baseline.small)
## or select covariates with upper code as shown below
upper.code.list <- unique(sub("[.][^.]*(.).*", "", colnames(dd.baseline.small)) )
upper.code.list <- intersect(upper.code.list, colnames(dd.baseline.small))
dd.base <- dd.baseline.small[, upper.code.list]
## perform regularized logistic regression to obtain propensity score (PS) 
## to adjust for potential confounders at baseline
phenos <- setdiff(colnames(dd.base), c('id', 'exposure'))
data.x <- as.matrix(dd.base[, phenos])
glmnet.fit <- glmnet::cv.glmnet(x=data.x, y=dd.base[,'exposure'],
                                family="binomial", standardize=TRUE,
                                alpha=0.1)
dd.base$PS <- c(predict(glmnet.fit, data.x, s='lambda.min'))
data.ps <- dd.base[,c('id', 'PS')]
dd.all.ps <- merge(data.ps, dd.small, by='id')  
demographics <- c('age', 'race', 'gender')
phenotypeList <- setdiff(colnames(dd.small), c('id','exposure','age','race','gender'))
## run with a subset of phenotypeList to get quicker results
phenotypeList.sub <- sample(phenotypeList, 5)
results.sub <- analysisPheWAS(method='firth', adjust='PS', Exposure='exposure',
                              PS='PS', demographics=demographics, 
                              phenotypes=phenotypeList.sub, data=dd.all.ps)
## run with the full list of phenotype outcomes (i.e., phenotypeList)
# \donttest{
        results <- analysisPheWAS(method='firth', adjust='PS',Exposure='exposure',
                          PS='PS', demographics=demographics,
                          phenotypes=phenotypeList, data=dd.all.ps) 
         # }