Anes 2024 is out! How to analyze the data with R | R-Bloggers

Sample

The 2024 monster was designed to represent American citizens aged 18 or older. Respondents were recruited from both personal and web monsters, with careful notes about suitability and exclusions.

Cod book

The code book decodes variable names in something human readable.

Here are two variables that we will use:

• V240002b: Mode or interview: Post-election: This variable catches the mode in which the respondent completed the majority of the survey after the elections.
  • -7. Insufficient parital, interview deleted
  • -6. No post -interview
  • 1. Face-to-face (FTF)
  • 2. Internet (web)
  • 3. Paper (papi)
  • 4. Telephone
  • 5. Video
• V241229: Plus: how often the government in Washington trust to do what is good [REVISED]:
  • -9. Refused
  • -8. Don’t know
  • -1. Not applicable
  • 1. Always
  • 2. Generally
  • 3. About half the time
  • 4. Some of the time
  • 5. Never

Let’s code again V241229 For easier analysis:

anes_2024_code <-
  anes_2024 |> 
  mutate(
    TrustGovernment = factor(
      case_when(
        V241229 == 1 ~ "Always",
        V241229 == 2 ~ "Most of the time",
        V241229 == 3 ~ "About half the time",
        V241229 == 4 ~ "Some of the time",
        V241229 == 5 ~ "Never",
        .default = NA
      ),
      levels = c("Always", "Most of the time", "About half the time", "Some of the time", "Never")
    )
  )

See more about codebooks.

Find the right weight

The Anes -documentation provides an overview of which weight variables are in which example:

Because we look at the fresh personal monster after the elections, we use weight variable V240101b.

An important remark: Anes weights count up to the sample size, not to the population size. Let’s confirm:

anes_2024_code |> filter(V240101b > 0) |> summarize(sum = sum(V240101b))

# A tibble: 1 × 1
    sum
  
1   925

This corresponds to the 925 Respondents shown in the example table of the documentation.

Adjustment for the population

If we want to draw conclusions all over the American population, we have to adjust the weights. The target population for this sample is around 232.5 million American citizens aged 18 or older, but that is just a rough number. According to the documentation, we can get the population numbers for a little more accurate from the estimates of the American Community Survey of 2023 American Community.

Thank you, Stephanie, for writing the code for the estimates of the population!

library(tidycensus)

varlist_2023 <- load_variables(2023, "acs1")

citizen_pop_18_plus <-
  get_acs(
    geography = "state",
    variable = "S2901_C01_001",
    year = 2023,
    survey = "acs1"
  )

pums_vars_2023 <- pums_variables |>
  as_tibble() |>
  filter(year == 2023, survey == "acs1")

pums_vars_2023 |>
  filter(var_code %in% c("AGEP", "TYPEHUGQ", "CIT")) |>
  select(var_code, var_label, val_min, val_max, val_label) |>
  print(n = 50)

# Find those who are 18+ and in group quarters
get_citizen_pop_18_plus_gq <- function(state) {
  get_pums(
    variables = c("AGEP", "TYPEHUGQ"),
    state = state,
    year = 2023,
    survey = "acs1",
    variables_filter = list(
      TYPEHUGQ = 2:3 ,
      CIT = 1:4,
      AGEP = (18:200)
    )
  ) |>
    summarize(estimate = sum(PWGTP), .by = STATE)
}

citizen_pop_18_plus_gq_l <- c(state.abb, "DC") |>
  map(get_citizen_pop_18_plus_gq)

citizen_pop_18_plus_gq_df <-
  citizen_pop_18_plus_gq_l |>
  list_rbind() |>
  rename(estimate_gq = estimate)

state_pops <- citizen_pop_18_plus |>
  select(GEOID, NAME, estimate_tot = estimate) |>
  full_join(citizen_pop_18_plus_gq_df, by = c("GEOID" = "STATE")) |>
  filter(NAME != "Puerto Rico") |>
  mutate(estimate_scope = estimate_tot - estimate_gq)

# Total without AK and HI
targetpop <-
  state_pops |>
  filter(!GEOID %in% c("02", "15")) |>
  summarize(TargetPop = sum(estimate_scope))

# Total with AK and HI
state_pops |>
  summarize(TargetPop = sum(estimate_scope))

The estimated target population (targetpop) Before 2024 is 232,449,541. Now we can save the Anes weights so that they get the population size:

anes_adjwgt <- anes_2024_code |>
  mutate(Weight = V240101b / sum(V240101b, na.rm = TRUE) * targetpop)

Layers and clusters

The documentation also tells us which variables of the layers and PSU (cluster) match each weight.

For V240101B these are:

• Layers: V240101d
• PSU: V240101c

More about Test design can be found in this chapter.

Put all the way together

Finally, we can build the design object. We will also filter to respondents who have actually completed the personal interview after the elections (V240002b == 1):

options(survey.lonely.psu = "adjust")

anes_des <- anes_adjwgt |>
  filter(V240002b == 1) |>
  as_survey_design(
    weights = Weight,
    strata = V240101d,
    ids = V240101c,
    nest = TRUE
  )

anes_des

And there we go! anes_des is our fully specified survey; From now on we use it every estimate that we calculate reflects the design of the survey.