Package 'svyVGAM'

Data from NHANES: number of sex partners

Description

These data are from the NHANES 2003-2004 survey in the US. They provide an example of overdispersed count data that motivates a two-component zero-inflation model

Usage

data("nhanes_sxq")

Format

A data frame with 2992 observations on the following 7 variables.

SDMVPSU

Primary Sampling Unit

SDMVSTRA

stratum

WTINT2YR

weights

malepartners

lifetime number of male sexual partners

RIDAGEYR

age in years

DMDEDUC

level of education: 1=less than high school, 2=high school, 3-more than high school, 7=refused

RIDRETH1

Race/ethnicity: 1=Mexican American, 2=Other Hispanic, 4=non-Hispanic White, 5=non-Hispanic Black, 5=Other

Source

NHANES files demo_c.xpt and sxq_c.xpt

See Also

Construction of the data set is described by https://notstatschat.rbind.io/2015/05/26/zero-inflated-poisson-from-complex-samples/

Examples

data(nhanes_sxq)
nhdes = svydesign(id=~SDMVPSU,strat=~SDMVSTRA,weights=~WTINT2YR, nest=TRUE, data=nhanes_sxq)
svy_vglm(malepartners~RIDAGEYR+factor(RIDRETH1)+DMDEDUC, zipoisson(), design=nhdes, crit = "coef")

---

Design-based inference for vector generalised linear models

Description

This function provides design-based (survey) inference for Thomas Yee's vector generalised linear models. It works by calling vglm with sampling weights, and then either using resampling (replicate weights) or extracting the influence functions and using a Horvitz-Thompson-type sandwich estimator.

Usage

svy_vglm(formula, family, design, ...)

Arguments

formula	Model formula, as for vglm
family	Model family, as for vglm
design	Survey design object
...	Other arguments to pass to vglm

Value

An S3 object of class svy_glm with print, coef and vcov methods, containing the design in the design component and a fitted vglm object in the fit component.

See Also

nhanes_sxq

vglm

svydesign svrepdesign

Examples

data(api)
dclus2<-svydesign(id=~dnum+snum, fpc=~fpc1+fpc2, data=apiclus2)

## Ordinary Gaussian regression
m1<-svyglm(api00~api99+mobility+ell, design=dclus2,family=gaussian)
## same model, but with the variance as a second parameter
m2<-svy_vglm(api00~api99+mobility+ell, design=dclus2,family=uninormal())
m1
m2
SE(m1)
SE(m2)

summary(m1)
summary(m2)

## Proportional odds model
dclus2<-update(dclus2, mealcat=as.ordered(cut(meals,c(0,25,50,75,100))))
a<-svyolr(mealcat~avg.ed+mobility+stype, design=dclus2)
b<-svy_vglm(mealcat~avg.ed+mobility+stype, design=dclus2, family=propodds())
a
b
SE(a)
SE(b) #not identical, because svyolr() uses approximate Hessian


## Zero-inflated Poisson
data(nhanes_sxq)
nhdes = svydesign(id=~SDMVPSU,strat=~SDMVSTRA,weights=~WTINT2YR,
       nest=TRUE, data=nhanes_sxq)

sv1<-svy_vglm(malepartners~RIDAGEYR+factor(RIDRETH1)+DMDEDUC,
      zipoisson(), design=nhdes, crit = "coef")
sv1
summary(sv1)

## Multinomial
## Reference group (non-Hispanic White) average older and more educated
## so coefficients are negative
mult_eth<- svy_vglm(RIDRETH1~RIDAGEYR+DMDEDUC,
     family=multinomial(refLevel=3), design=nhdes)

## separate logistic regressions are close but not identical
two_eth<-svyglm(I(RIDRETH1==1)~RIDAGEYR+DMDEDUC, family=quasibinomial,
    design=subset(nhdes, RIDRETH1 %in% c(1,3)))

summary(mult_eth)
summary(two_eth)

---