In ttrodrigz/iterake: Tools for iterative raking

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

iterake

Overview

iterake's main utility is creating row weights using a process called iterative raking. Iterative raking (also known as rim weighting), is one of several methods used to correct the deviation between the marginal proportions in a sample and a known population, or, universe as it was first referred to (Deming & Stephan 1940) for a given set of variables.

iterake is designed with speed and simplicity in mind. The weighting algorithm is powered by data.table and takes advantage of its fast grouping and joining.

Workflow

The weighting process with iterake is fairly straightforward, we suggest:

1. Use the universe() function to build your population.
   1. The univerise is constructed with one or more categories where the marginal probabilites are known. These categories are built with the category() function.
   2. If you want to use the natural marginal proportions from an existing dataset as your targets, you can use inherit_category(). Just make sure the name given to the category matches the existing data and the data you intend to weight.
2. Compare the marginal proportions in your sample with the population with compare_margins() function.
3. If needed, create weights for your data using iterake().
4. Use compare_margins() again to verify that the weighted proportions in your sample now match the population.
5. Check the performance of the weighting model with weight_stats().

Installation

``` {r, eval = FALSE}

Install the development version from GitHub

install.packages("remotes") remotes::install_github("ttrodrigz/iterake")

## Motivating Example
Say you have conducted a study by randomly sampling 400 individuals from a population. You were dilligent in monitoring the responses to make sure the makeup of the sample adequately reflected the population. But, due to chance, slightly too many males and individuals under 50 years of age entered the sample.

You know from experts in your field that 60% of the population from which you sampled are female, and 20% of the population are less than 50 years old. Let's build a data set to use as an example:


```r
library(tibble)

N <- 400

set.seed(101)

df <- tibble(
  id = 1:N,
  Sex = sample(
    x = c("Male", "Female"),
    size = N,
    replace = TRUE,
    prob = c(0.42, 0.58)
  ),
  Under50 = sample(
    x = c(T, F),
    size = N,
    replace = TRUE,
    prob = c(0.22, 0.78)
  )

)

df

Step 1: Build the universe

Simply supply the data you intend on weighting, and build weighting categories by using the category() function.

library(iterake)

uni <- universe(

  data = df,

  category(
    name = "Sex",
    buckets = c("Male", "Female"),
    targets = c(0.4, 0.6)
  ),

  category(
    name = "Under50",
    buckets = c(TRUE, FALSE),
    targets = c(0.2, 0.8)
  )

)

Step 2: Compare marginal proportions prior to weighting

This is the time to inspect the differences in proportions between the sample and the population. A large discrepancy will require extreme weights, and in some cases the algorithm may not even converge. Before you decide to weight, keep in mind that weighting the data decreases accuracy. In some cases it is best to deal with the fact your sample doesn't perfectly match the population.

compare_margins(universe = uni)

Step 3: Weight the data

If weighting is necessary, pass the universe object to iterake().

df_wgt <- iterake(universe = uni)

df_wgt

Step 4: Compare marginal proportions after weighting

````r compare_margins( universe = uni, data = df_wgt, weight = weight, plot = TRUE )

#### **Step 5:** Inspect weights
Again, weights much higher or lower than 1 are undesirable, check the output with
`weight_stats()` to inspect the quality of the weights. Details about what each of the statistics mean can be found in the documentation.

```r
weight_stats(df_wgt[["weight"]])

ttrodrigz/iterake documentation built on March 27, 2024, 12:48 a.m.