logistf (version 1.26.0)

CLIP.confint: Confidence Intervals after Multiple Imputation: Combination of Likelihood Profiles

Description

This function implements the new combination of likelihood profiles (CLIP) method described in Heinze, Ploner and Beyea (2013). This method is useful for computing confidence intervals for parameters after multiple imputation of data sets, if the normality assumption on parameter estimates and consequently the validity of applying Rubin's rules (pooling of variances) is in doubt. It consists of combining the profile likelihoods into a posterior. The function CLIP.confint searches for those values of a regression coefficient, at which the cumulative distribution function of the posterior is equal to the values specified in the argument ci.level (usually 0.025 and 0.975). The search is performed using R's optimize function.

Usage

CLIP.confint(
  obj = NULL,
  variable = NULL,
  data,
  firth = TRUE,
  weightvar = NULL,
  control = logistf.control(),
  ci.level = c(0.025, 0.975),
  pvalue = TRUE,
  offset = NULL,
  bound.lo = NULL,
  bound.up = NULL,
  legacy = FALSE
)

Value

An object of class CLIP.confint, with items:

variable

The variable(s) which were analyzed

estimate

The pooled estimate (average over imputations)

ci

The confidence interval(s)

pvalue

The p-value(s)

imputations

The number of imputed datasets

ci.level

The confidence level (input)

bound.lo

The bounds used for finding the lower confidence limit; usually not of interest. May be useful for error-tracing.

bound.up

The bounds used for finding the upper confidence limit

iter

The number of iterations (for each variable and each tail)

call

The call object

Arguments

obj

Either a list of logistf fits (on multiple imputed data sets), or the result of analysis of a mice (multiply imputed) object using with.mids

variable

The variable of interest, for which confidence intervals should be computed. If missing, confidence intervals for all variables will be computed.

data

A list of data set corresponding to the model fits. Can be left blank if obj was obtained with the dataout=TRUE option or if obj was obtained by mice

firth

If TRUE, applies the Firth correction. Should correspond to the entry in obj.

weightvar

An optional weighting variable for each observation.

control

Control parameters for logistf, usually obtained by logistf.control()

ci.level

The two confidence levels for each tail of the posterior distribution.

pvalue

If TRUE, will also compute a P-value from the posterior.

offset

An optional offset variable

bound.lo

Bounds (vector of length 2) for the lower limit. Can be left blank. Use only if problems are encountered.

bound.up

Bounds (vector of length 2) for the upper limit. Can be left blank. Use only if problems are encountered.

legacy

If TRUE, will use pure R code for all model fitting. Can be slow. Not recommended.

Author

Georg Heinze and Meinhard Ploner

Details

For each confidence limit, this function performs a binary search to evaluate the combined posterior, which is obtained by first transforming the imputed-data likelihood profiles into cumulative distribution functions (CDFs), and then averaging the CDFs to obtain the CDF of the posterior. Usually, the binary search manages to find the confidence intervals very quickly. The number of iterations (mean and maximum) will be supplied in the output object. Further details on the method can be found in Heinze, Ploner and Beyea (2013).

References

Heinze G, Ploner M, Beyea J (2013). Confidence intervals after multiple imputation: combining profile likelihood information from logistic regressions. Statistics in Medicine, to appear.

See Also

logistf() for Firth's bias-Reduced penalized-likelihood logistic regression.

Examples

#generate data set with NAs 
freq=c(5,2,2,7,5,4)
y<-c(rep(1,freq[1]+freq[2]), rep(0,freq[3]+freq[4]), rep(1,freq[5]), rep(0,freq[6]))
x<-c(rep(1,freq[1]), rep(0,freq[2]), rep(1,freq[3]), rep(0,freq[4]), 
rep(NA,freq[5]),rep(NA,freq[6]))
toy<-data.frame(x=x,y=y)

# impute data set 5 times
set.seed(169)
toymi<-list(0)
for(i in 1:5){
  toymi[[i]]<-toy
  y1<-toymi[[i]]$y==1 & is.na(toymi[[i]]$x)
  y0<-toymi[[i]]$y==0 & is.na(toymi[[i]]$x) 
  xnew1<-rbinom(sum(y1),1,freq[1]/(freq[1]+freq[2]))
  xnew0<-rbinom(sum(y0),1,freq[3]/(freq[3]+freq[4]))
  toymi[[i]]$x[y1==TRUE]<-xnew1
  toymi[[i]]$x[y0==TRUE]<-xnew0
  }
  
 # logistf analyses of each imputed data set
 fit.list<-lapply(1:5, function(X) logistf(data=toymi[[X]], y~x, pl=TRUE))
  
 # CLIP confidence limits
 CLIP.confint(obj=fit.list, data = toymi)