stepFWD: Customized stepwise regression with p-value and trend check

In andrija-djurovic/PDtoolkit: Collection of Tools for PD Rating Model Development and Validation

Customized stepwise regression with p-value and trend check

Description

stepFWD customized stepwise regression with p-value and trend check. Trend check is performed comparing observed trend between target and analyzed risk factor and trend of the estimated coefficients within the logistic regression. Note that procedure checks the column names of supplied db data frame therefore some renaming (replacement of special characters) is possible to happen. For details check help example.

Usage

stepFWD(
  start.model,
  p.value = 0.05,
  coding = "WoE",
  coding.start.model = TRUE,
  check.start.model = TRUE,
  db,
  offset.vals = NULL
)

Arguments

start.model	Formula class that represents starting model. It can include some risk factors, but it can be defined only with intercept (y ~ 1 where y is target variable).
p.value	Significance level of p-value of the estimated coefficients. For WoE coding this value is is directly compared to the p-value of the estimated coefficients, while for dummy coding multiple Wald test is employed and its p-value is used for comparison with selected threshold (p.value).
coding	Type of risk factor coding within the model. Available options are: "WoE" (default) and "dummy". If "WoE" is selected, then modalities of the risk factors are replaced by WoE values, while for "dummy" option dummies (0/1) will be created for n-1 modalities where n is total number of modalities of analyzed risk factor.
coding.start.model	Logical (TRUE or FALSE), if the risk factors from the starting model should be WoE coded. It will have an impact only for WoE coding option. Default is TRUE.
check.start.model	Logical (TRUE or FALSE), if risk factors from the starting model should be checked for p-value and trend in stepwise process. Default is TRUE. If FALSE is selected, then coding.start.model is forced to TRUE.
db	Modeling data with risk factors and target variable. All risk factors (apart from the risk factors from the starting model) should be categorized and as of character type.
offset.vals	This can be used to specify an a priori known component to be included in the linear predictor during fitting. This should be NULL or a numeric vector of length equal to the number of cases. Default is NULL.

Value

The command stepFWD returns a list of four objects.
The first object (model), is the final model, an object of class inheriting from "glm".
The second object (steps), is the data frame with risk factors selected at each iteration.
The third object (warnings), is the data frame with warnings if any observed. The warnings refer to the following checks: if risk factor has more than 10 modalities, if any of the bins (groups) has less than 5% of observations and if there are problems with WoE calculations.
The final, fourth, object dev.db returns the model development database.

Examples

suppressMessages(library(PDtoolkit))
data(loans)
#identify numeric risk factors
num.rf <- sapply(loans, is.numeric)
num.rf <- names(num.rf)[!names(num.rf)%in%"Creditability" & num.rf]
#discretized numeric risk factors using ndr.bin from monobin package
loans[, num.rf] <- sapply(num.rf, function(x) 
ndr.bin(x = loans[, x], y = loans[, "Creditability"])[[2]])
str(loans)
res <- stepFWD(start.model = Creditability ~ 1, 
	   p.value = 0.05, 
	   coding = "dummy",
	   db = loans)
summary(res$model)$coefficients
rf.check <- tapply(res$dev.db$Creditability, 
		 res$dev.db$Instalment_per_cent, 
		 mean)
rf.check
diff(rf.check)
res$steps
head(res$dev.db)

andrija-djurovic/PDtoolkit documentation built on Jan. 29, 2024, 9:35 a.m.