deff {PracTools}

Design effects of various types

Description

Compute the Kish, Henry, Spencer, or Chen-Rust design effects.

Usage

deff(w, x=NULL, y=NULL, p=NULL, strvar=NULL, clvar=NULL, Wh=NULL, nest=FALSE, type)

Arguments

w

vector of weights for a sample

x

matrix of covariates used to construct a GREG estimator of the total of y. This matrix does not include the intercept. Used only for Henry deff.

y

vector of the sample values of an analysis variable

p

vector of 1-draw selection probabilities, i.e., the probability that each unit would be selected in a sample of size 1. Used only for Spencer deff.

strvar

vector of stratum identifiers; equal in length to that of w. Used only for Chen-Rust deff.

clvar

vector of cluster identifiers; equal in length to that of w. Used only for Chen-Rust deff.

Wh

vector of the proportions of elements that are in each stratum; length is number of strata. Used only for Chen-Rust deff.

nest

Are cluster IDs numbered within strata (TRUE or FALSE)? If TRUE, cluster IDs can be restarted within strata, e.g., 1,2,3,1,2,3,...

type

type of allocation; must be one of "kish", "henry", "spencer", "cr"

Details

deff calls one of deffK, deffH, deffS, or deffCR depending on the value of the type parameter. The Kish design effect is the ratio of the variance of an estimated mean in stratified simple random sampling without replacement (stsrswor) to the variance of the estimated mean in srswor, assuming that all stratum unit variances are equal. In that case, proportional allocation with equal weighting is optimal. deffK equals 1 + relvar(w) where relvar is relvariance of the vector of survey weights. This measure is not appropriate in samples where unequal weighting is more efficient than equal weighting.

The Henry design effect is the ratio of the variance of the general regression (GREG) estimator of a total of y to the variance of the estimated total in srswr. Calculations for the Henry deff are done as if the sample is selected in a single-stage and with replacement. Varying selection probabilities can be used. The model for the GREG is assumed to be y = \alpha + \beta x + \epsilon, i.e., the model has an intercept.

The Spencer design effect is the ratio of the variance of the pwr-estimator of the total of y, assuming that a single-stage sample is selected with replacement, to the variance of the total estimated in srswr. Varying selection probabilities can be used.

The Chen-Rust deff accounts for stratification, clustering, and unequal weights, but does not account for the use of any auxiliary data in the estimator of a mean. The Chen-Rust deff returned here is appropriate for stratified, two-stage sampling.

Value

Numeric design effect for types kish, henry, spencer. For type cr a list with components:

strata components

Matrix with deff's due to weighting, clustering, and stratification for each stratum

overall deff

Design effect for full sample accounting for weighting, clustering, and stratification

Author(s)

Richard Valliant, Jill A. Dever, Frauke Kreuter

References

Chen, S. and Rust, K. (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), 111-130.

Henry, K.A., and Valliant, R. (2015). A Design Effect Measure for Calibration Weighting in Single-stage Samples. Survey Methodology, 41, 315-331.

Kish, L. (1965). Survey Sampling. New York: John Wiley & Sons.

Kish, L. (1992). Weighting for unequal Pi. Journal of Official Statistics, 8, 183-200.

Park, I., and Lee, H. (2004). Design Effects for the Weighted Mean and Total Estimators under Complex Survey Sampling. Survey Methodology, 30, 183-193.

Spencer, B. D. (2000). An Approximate Design Effect for Unequal Weighting When Measurements May Correlate With Selection Probabilities. Survey Methodology, 26, 137-138.

Valliant, R., Dever, J., Kreuter, F. (2018, chap. 14). Practical Tools for Designing and Weighting Survey Samples, 2nd edition. New York: Springer.

See Also

deffK, deffH, deffS, deffCR

Examples

require(reshape)      # has function that allows renaming variables
require(sampling)

set.seed(-500398777)
    # generate population using HMT function
pop.dat <- as.data.frame(HMT())
mos <- pop.dat$x
pop.dat$prbs.1d <- mos / sum(mos)
    # select pps sample
n <- 80
pk <- pop.dat$prbs.1d
sam <- UPrandomsystematic(pk)
sam <- sam==1

sam.dat <- pop.dat[sam, ]
dsgn.wts <- 1/pk[sam]
deff(w=dsgn.wts, type="kish")
deff(w=dsgn.wts, y=sam.dat$y, p=sam.dat$prbs.1d, type="spencer")
deff(w=dsgn.wts, x=sam.dat$x, y=sam.dat$y, type="henry")


data(MDarea.popA)
Ni <- table(MDarea.popA$TRACT)
m <- 10
probi <- m*Ni / sum(Ni)
    # select sample of clusters
set.seed(-780087528)
sam <- cluster(data=MDarea.popA, clustername="TRACT", size=m, method="systematic",
                pik=probi, description=TRUE)
    # extract data for the sample clusters
samclus <- getdata(MDarea.popA, sam)
samclus <- rename(samclus, c("Prob" = "pi1"))
    # treat sample clusters as strata and select srswor from each
nbar <- 4
s <- strata(data = as.data.frame(samclus), stratanames = "TRACT",
            size = rep(nbar,m), method="srswor")
    # extracts the observed data
samdat <- getdata(samclus,s)
samdat <- rename(samdat, c("Prob" = "pi2"))
    # add a fake stratum ID
H <- 2
nh <- m * nbar / H
stratum <- NULL
for (h in 1:H){
    stratum <- c(stratum, rep(h,nh))
}
wt <- 1/(samdat$pi1*samdat$pi2) * runif(m*nbar)
samdat <- cbind(subset(samdat, select = -c(Stratum)), stratum, wt)
deff(w = samdat$wt, y=samdat$y2, strvar = samdat$stratum, clvar = samdat$TRACT, Wh=NULL, type="cr")