reweight (version 1.2.1)

reweight: Adjust survey respondent weights

Description

Adjusts the weights of survey repondents so that the marginal distributions of certain variables fit more closely to those from a more precise source (e.g. Census Bureau's data).

Usage

reweight(ori, mar, raw=NA, wgt=NA, unique=T, bound=c(0, 100),
trace=F, tolerance=0.1, penalty=0, ...)
"print"(x, ...)                                       
"summary"(object, ...)  
"plot"(x, ...)

Arguments

ori

a matrix containing the factor levels. The levels should start from 1 and count upwards as in $1,2,\ldots$.

mar

a vector giving the marginal distributions for each of the factors that are listed in the same order as in ori.

raw

a vector of the raw counts of survey respondents corresponding to each line in ori. Default is all one's.

wgt

a vector of the original weights corresponding to each line in ori. Default is the value of raw.

unique

whether the factor level combination in ori matrix is already unique. If FALSE, it will be made so by aggregating the non-unique lines, along with raw and wgt. Default is TRUE.

bound

a vector of two elements giving the lower and upper bounds for the final weight ratios. The extreme weight ratio is reset to either the lower or upper bound. Default is c(0,100). In application a more practical bound might be c(0.5,2).

trace

if TRUE, show the path of the golden selection search of best regularization parameter $r$. Default is FALSE.

tolerance

the percentage of the smallest eigenvalue that is to be used as the lower start point of the golden selection searhch. Default is 0.1.

penalty

measures the strength of a penalty term (it puts penalty if the number of zero weighting ratios is large) in the GCV function. It is a multiplicative factor of the form $(1+q)^p$, where $p$ is the penalty and $q$ is the percentage of zero weighting ratios. Default is 0 (no penalty).

x

an object returned by calling reweight.

object

an object returned by calling reweight.

...

parameters to be passed to the generic fucntion.

Value

None.

Details

ori, raw, wgt typically come from survey data with categorical responses. The intent is to adjust the wgt so that the survey sample is more representative of the universe from where it comes. It is accomplished by fitting the marginal distributions of the sample to those of the universe, or those from a more precise source (e.g. census data). The method is based on the Tikhonov regularization.

The print method prints out the weight ratios, along with their corresponding factor level combinations. This data can then be matched back to the original sample data to adjust the original weights (by multipling each original weight with the weight ratio).

The summary method prints out various running statistics.

The plot method makes a panel of four diagnostic plots.

References

Feiming Chen (2006) A Heuristic Method for Weighting Survey Respondents. JSM 2006 Proceedings.

See Also

pumswgt

Examples

data(pumswgt)
r1 <- reweight(pumswgt$ori,pumswgt$mar,pumswgt$wgt)
plot(r1)
summary(r1)
r2 <- reweight(pumswgt$ori,pumswgt$mar,pumswgt$raw)
plot(r2)
summary(r2)
w <- print(r2)