In donboyd5/microweight: Tools for Weighting Microdata Files

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

options(tibble.print_min = 4L, tibble.print_max = 4L)

library(microweight)
library(tidyverse)
library(knitr)

Economists, policy analysts, and others often need to take a sample for one geographic area, such as for the nation, and construct weights that will make the sample consistent with known data for a different geographic area, such as a state.

The function geoweight creates these new weights.

Setting up our data

In this vignette we'll use small microdata files extracted from the American Community Survey in the United States: acs (10,000 individuals) and acsbig (100,000 individuals).

To keep things simple, we create a minimal subset with only these columns:

Column | Description ------------- | ------------- stabbr | state postal abbreviation pwgtp | person weight agep | person's age mar | marital status (1=married, 2=single, ...) pincp | personal income, total wagp | wage income ssp | Social Security income

The first few records have a lot of zeros so we'll look at the tail, which is more representative of typical records:

acs2 <- acs %>%
  select(stabbr, pwgtp, agep, mar, pincp, wagp, ssp)
tail(acs2)

Look at targets for a few variables:

totals <- acs2 %>%
  group_by(stabbr) %>%
  summarise(across(c(pincp, wagp, ssp), ~sum(. * pwgtp)), .groups = "drop")
# totals

set.seed(1234)
targets_df <- totals %>%
  mutate(across(c(pincp, wagp, ssp), ~ . * (1 + rnorm(n=1, 0, .02))))
targets_df

targets_df <- totals

A simple geographic weighting problem

We want weights that will hit these targets. To do that we must define the following inputs for geoweight. We will use default values for other arguments:

Item | Description ------------- | ------------- wh | vector of household total weights, length h (see h, s, k definitions below) xmat | h x k matrix of data for households targets | s x k matrix of desired target values

Where h, s, and k are defined as follows:

Index | Description ------------- | ------------- h | number of households (or individuals, records, tax returns, etc.) s | number of states (or other geographies or subgroups) k | number of characteristics each household has

Set up the problem

Let's define the first four, as they are easy to understand:

wh <- acs2$pwgtp

target_names <- c("pincp", "wagp", "ssp")
xmat <- acs2 %>%
  select(all_of(target_names)) %>%
  as.matrix

targets <- targets_df %>%
  select(all_of(target_names)) %>%
  as.matrix


#' p <- make_problem(h=10, s=3, k=2)
#' dw <- get_dweights(p$targets)
#'
#' res1 <- geoweight(wh = p$wh, xmat = p$xmat, targets = p$targets,
#'   dweights = dw, quiet=TRUE)
#'

Solve the problem

Now we are ready to reweight:

result <- geoweight(wh, xmat, targets)
# result <- geoweight(wh, xmat, targets, method='Newton')

Examine results

Look at important summary measures:

names(result)
result$solver_message
result$etime
result$sse_unweighted
result$sse_weighted

How do the targeted values compare to the original values? geoweight returns a matrix with percent differences:

result$targets_pctdiff %>% round(2)

Surprisingly (perhaps), the percentage differences are not all zero.

donboyd5/microweight documentation built on Aug. 17, 2020, 4:48 p.m.