surveynnet: Neural Net for Complex Survey Data

In surveynnet: Neural Network for Complex Survey Data

Neural Net for Complex Survey Data

Description

The surveynnet package extends the functionality of nnet (Venables and Ripley, 2002), which already supports survey weights, by enabling it to handle clustered and stratified data. It achieves this by incorporating design effects through the use of effective sample sizes in the calculations, performed by the package described in Valliant et al. (2023), by following the methods outlined by Chen and Rust (2017) and Valliant et al. (2018).

Usage

surveynnet(x, y, weight, strat, clust, comp_cases = FALSE, ...)

Arguments

x	Matrix or data frame of predictors. Must not contain any missing values.
y	Vector of targets / response values. Must not contain any missing values.
weight	The weights for each sample.
strat	The stratum for each sample.
clust	The cluster for each sample.
comp_cases	If TRUE, filter out missing values from x, y, weight, strat, and clust. Default FALSE. Note that in either case, the dimensions of all data mentioned above must agree.
...	Additional arguments to be passed into PracTools::deffCR or nnet::nnet. See documentation of those packages and functions for more details. Note that for the neural net (nnet), the default here is set to 3 layers ("size" parameter) and maximum iterations ("maxit" parameter) is set to 2000.

Value

A list containing two objects:

• A dataframe with the fitted values of the neural nets, using: no weights ("fitted"), the user-inputted weights ("fitted_weighted"), and the new method that adjusts the weights by using a design effect incorporating cluster and strata ("fitted_deff").

• The fitted neural network object (from nnet), using the novel design-effect based weights; this can be used to predict the outcomes for new observations.

References

• Chen, S., and K. F. Rust. 2017."An Extension of Kish’s Formula for Design Effects to Two- and Three-Stage Designs with Stratification.”, Journal of Survey Statistics and Methodology,5 (2): 111–30.

• Valliant, R., J. A. Dever, and F. Kreuter. 2018. Practical Tools for Designing and Weighting Survey Samples .2nd ed. New York: Springer-Verlag.

Examples

# short example with body fat dataset
y <- body_fat$pct_body_fat
x <- body_fat[,c("Weight_kg", "Height_cm", "Age")]
weight <- body_fat$survey_wt
strat <- body_fat$stratum
clust <- body_fat$cluster
y[strat==1] <- y[strat==1] + 30*0.00015*rnorm(sum(strat==1))
y[strat==2] <- y[strat==2] + 30*0.15*rnorm(sum(strat==2))

myout <- surveynnet(x,y,weight = weight, strat = strat, clust=clust)
myout


# NHANES example
# Predicting Diastolic BP from BMI, Systolic BP and Height
# PLEASE NOTE: for this example, pass "nest=TRUE" into the
# "..." parameters of the main function `surveynnet`

x <- nhanes.demo[,c("BMXBMI", "BPXSY1", "BMXHT")]
weight <- nhanes.demo$WTMEC2YR
strat <- nhanes.demo$SDMVSTRA
clust <- nhanes.demo$SDMVPSU
y <- nhanes.demo$BPXDI1
myout <- surveynnet(x,y,weight = weight, strat = strat, clust=clust, nest=TRUE)
head(myout$results, 15)

surveynnet documentation built on April 3, 2025, 7:38 p.m.