quint.validate: Validation of a Qualitative Interaction Tree

In quint: Qualitative Interaction Trees

Validation of a Qualitative Interaction Tree

Description

A bootstrap-based validation procedure to estimate the optimism in the effect sizes of a QUINT tree which gives insight in the generalizability of the results.

Usage

quint.validate(object, B = 10, allresults = FALSE)

Arguments

object	a (pruned) QUINT tree object of class quint.
B	number of bootstrap samples. Default number is 10; for better accuracy B=1000 is recommended.
allresults	option to return an extended list of output. Default is set to FALSE. See Value section for details.

Details

In this procedure bootstrap trees are grown of the same leaf size as the (pruned) QUINT tree. The bootstrap samples are drawn from the data used to grow the original tree. For every bootstrap tree the largest and smallest (i.e., largest negative) treatment mean differences (or treatment effect sizes) of two leaves are saved. Treatment mean differences in the leaves are then predicted using the original data set as input for each bootstrapped tree. From these predictions, the largest and smallest treatment mean differences are saved. For each bootstrap tree, the largest predicted treatment effect is subtracted from the largest treatment effect in the bootstrap sample. The average of these values is the bias (i.e., the optimism) for the largest treatment effects. This is done likewise for the smallest treatment effects. Subsequently, the bias is computed as the difference between the bias for the largest effects minus the bias for the smallest effects.

The details of this validaton procedure are described in Appendix C of Dusseldorp & Van Mechelen (2014).

Value

Returns a list with the following components:

estopt	the estimated optimism for either the treatment effect size (biasd) or the raw treatment mean difference (biasdif).
li	a data frame with leaf information output similar to the leaf information output of the (pruned) QUINT tree object. An extra column is added for the bias-corrected differences in treatment outcomes (d or diff). The bias-corrected values are only computed for the leaves with the most extreme values, i.e. the largest and smallest treatment effects. Hence, the other leaves get the value NA in this column.
optd	a matrix with computed estimated optimism of the treatment effect size per bootstrapp tree. The first column contains the difference between the largest and smallest effect size of the bootstrapped tree. The second column contains the difference between the largest and smallest predicted effect size. Returned when allresults is set to TRUE and crit='es' is specified in the QUINT object.
optdif	a matrix with computed estimated optimism of the raw mean difference bootstrapped tree. The first column contains the difference between the largest and smallest raw mean difference of the bootstrapped tree. The second column contains the difference between the largest and smallest predicted raw mean difference. Returned when allresults is set to TRUE and crit='es' is specified in the QUINT object.
resultd	a vector with the estimated overall mean optimism, the mean bias for the smallest and for the largest effect size. Returned when allresults is set to TRUE and crit="es".
resultdif	a vector with the estimated overall mean optimism, the mean bias for the smallest and largest raw mean difference. Returned when allresults is set to TRUE and crit="dm".

References

Dusseldorp E. and Van Mechelen I. (2014). Qualitative interaction trees: a tool to identify qualitative treatment-subgroup interactions. Statistics in Medicine, 33(2), 219-237. DOI: 10.1002/sim.5933.

See Also

quint, prune.quint, quint.control, quint.bootstrapCI

Examples

## Not run: data(bcrp)
formula1<- I(cesdt1-cesdt3)~cond | nationality+marital+wcht1+age+
  trext+comorbid+disopt1+uncomt1+negsoct1

set.seed(10)
control1<-quint.control(maxl=5,B=2)
quint1<-quint(formula1, data= subset(bcrp,cond<3),control=control1) #Grow a QUINT tree

prquint1<-prune(quint1) #Prune tree to optimal size

set.seed(3)
valquint1<-quint.validate(prquint1, B = 5) #estimate the optimism by bootstrapping 5 times
valquint1
## End(Not run)

quint documentation built on July 2, 2022, 1:07 a.m.