How Durable Are Democrats' Gains in Georgia?

2020 Presidential Vote

Alright, to start with we’ll pull demographic data by county from the latest American Community Survey using the amazing tidycensus package. I’ll download the sf geometries for some mapping as well.

## acs
acs <- get_acs(geography = "county",
               variables = c(tot_pop = "B01003_001", 
                             age_male = "B01002_002",
                             age_female = "B01002_003",
                             ba = "B15003_022",
                             ma = "B15003_023",
                             pd = "B15003_024",
                             phd = "B15003_025", 
                             poverty = "B17001_001",
                             medicaid = "C27007_001",
                             tot_white = "B02001_002",
                             tot_black = "B02001_003",
                             tot_ai = "B02001_004",
                             tot_asian = "B02001_005",
                             born_other_state = "B06001_025",
                             income = "B19127_001"),
              year = 2019,
              geometry = TRUE)

acs2 <- acs %>%
    st_drop_geometry() %>% 
    clean_names() %>%
    pivot_wider(names_from = variable, values_from = estimate, id_cols = name) %>%
    left_join(st_drop_geometry(acs) %>% 
                  clean_names() %>%
                  separate(name, into = c("county", "state"), 
                           sep = ", ", remove = FALSE) %>%
                  select(name, county, state)) %>%
    relocate(state, county) %>%
    distinct() %>%
    left_join(fips_codes %>% 
                  select(-state) %>% 
                  unite(fips_code, c(state_code, county_code), sep = ""), 
              by= c("state" = "state_name", "county")) %>%
    mutate(medicaid_perc = medicaid / tot_pop,
           pov_perc = poverty / tot_pop,
           college = ba + ma + pd + phd, 
           college_perc = college / tot_pop,
           white_perc = tot_white / tot_pop,
           black_perc = tot_black / tot_pop,
           ai_perc = tot_ai / tot_pop,
           asian_perc = tot_asian / tot_pop,
           other_state_perc = born_other_state / tot_pop,
           income_per_cap = income / tot_pop) 

Then I’ll pull in the 2020 presidential voting data by county from a few helpful GitHub sites. Note that Alaska presents some difficulties because its election returns are variously reported by either state house districts (where there are 40) and by boroughs (where there are 19, plus 11 census areas within an unorganized district). While I standardized the results for the 2016 and 2020 elections by district for Alaska, the ACP data is by borough (and 2020 election data hasn’t been updated by borough yet), so Alaska data isn’t included in the ACP comparison charts below.

states <- tibble(state.abb, state.name, state.region) %>%
    rename(state = state.name)

pres20 <- read_csv("https://github.com/tonmcg/US_County_Level_Election_Results_08-20/raw/master/2020_US_County_Level_Presidential_Results.csv") %>%
    mutate(election = 2020) %>%
    left_join(states, by = c("state_name" = "state")) %>%
    rename(fips_code = county_fips, county = county_name, 
           state_abbr = state.abb, state = state_name) %>%
    mutate(dem_margin = per_dem - per_gop)


alaska <- read_csv("https://github.com/MEDSL/county-returns/blob/master/countypres_2000-2016.csv?raw=true") %>%
    clean_names() %>%
    filter(state == "Alaska",
           year == 2016) %>%
    select(state, county, party, candidate_votes = candidatevotes, 
           total_votes = totalvotes) %>%
    mutate(party_perc = candidate_votes / total_votes) %>%
    pivot_wider(names_from = party, values_from = party_perc, id_cols = county) %>%
    rename(per_dem = democrat, per_gop = republican) %>%
    mutate(dem_margin = per_dem - per_gop,
           state = "Alaska",
           election = 2016) %>%
    select(-"NA") %>%
    filter(county != "District 99") %>%
    bind_cols(pres20 %>% filter(state == "Alaska") %>% select(fips_code))
  

pres16 <- read_csv("https://github.com/tonmcg/US_County_Level_Election_Results_08-20/raw/master/2016_US_County_Level_Presidential_Results.csv") %>%
    left_join(states, by = c("state_abbr" = "state.abb")) %>%
    rename(county = county_name) %>%
    mutate(per_point_diff = parse_number(per_point_diff),
           per_point_diff = per_point_diff / 100,
           election = 2016,
           fips_code = if_else(str_length(combined_fips) == 4, 
                                   paste0(0, combined_fips),
                                   as.character(combined_fips)),
           dem_margin = per_dem - per_gop,
           county = if_else(fips_code == "46113", "Oglala Lakota County", county),
           fips_code = if_else(fips_code == "46113", "46102", fips_code)) %>%
    select(-c(X1, combined_fips)) %>%
    distinct() %>%
    filter(state != "Alaska") %>%
    bind_rows(alaska)

## Warning: Missing column names filled in: 'X1' [1]

pres <- pres20 %>%
    bind_rows(pres16) %>%
    mutate(election = paste0("dem_perc", election),
           state_county = paste(state, county, sep = "_")) %>%
    pivot_wider(names_from = election, values_from = c(per_dem, total_votes, dem_margin),
                id_cols = fips_code) %>%
    left_join(pres20 %>% select(state, county, fips_code)) %>%
    rename(total_votes20 = 4, total_votes16 = 5,
           dem_perc20 = 2, dem_perc16 = 3,
           dem_margin20 = 6, dem_margin16 = 7) %>%
    mutate(dem_perc_dif = dem_perc20 - dem_perc16, 
           dem_margin_dif = (dem_margin20 - dem_margin16),
           total_votes_dif = total_votes20 - total_votes16) %>%
    relocate(state, county) 

One of the articles that made me want to dig into this data myself was from Bloomberg (as seen in New York Magazine), which used county classifications from the American Communities Project to see how types of counties shifted their votes from 2016 to 2020.

The ACP data is really interesting – from the methodology page is sounds like they did some k-means clustering of 36 different county-level variables to find 15 different county types – things like “Exurbs” and “Military Hubs”, but also “Evangelical Hubs” and “Aging Farmlands”. These aren’t mutually-exclusive categories (i.e. there might be suburban counties that are tagged as Military Posts, for example), but they’re still useful categories for how we actually think about communities across the country.

For example, Chattahoochee County is home to Ft. Benning, a U.S. Army base near Columbus, GA. It’s a rural part of the state, but near medium-sized Columbus. With Ft. Benning in its borders, Chattahoochee County (and neighboring Muscogee County) has a distinct feel compared with other medium-sized cities in rural counties, and “Military Hub” captures its feel more than simply a classification based on just population density.

communities <- readxl::read_xlsx("C:\\Users\\ChadPeltier\\Downloads\\County-Type-Share.xlsx") %>%
    clean_names() 

communities_key <- communities %>%
    slice_head(n = 15) %>%
    select(key, new_names)
    
communities2 <- communities %>%
    left_join(communities_key, by = c("type_number_2" = "key")) %>%
    rename(acp_county_type = new_names.y, fips_code = fips) %>%
    select(3,8) %>%
    mutate(fips_code = if_else(str_length(fips_code) == 4, paste0(0, fips_code),
                              as.character(fips_code)),
           fips_code = as.character(fips_code))

However, it’s also useful to have an alternative county classification system with mutually-exclusive boundaries to use as well. Here’s the NCHS county codes data from the CDC:

nchs <- readxl::read_xlsx("C:\\Users\\ChadPeltier\\Downloads\\NCHSURCodes2013.xlsx") %>%
    clean_names() %>%
    rename(nchs_code = 7) %>%
    mutate(fips_code = if_else(str_length(fips_code) == 4, paste0(0, fips_code),
                              as.character(fips_code)),
           fips_code = as.character(fips_code),
           nchs_code = case_when(nchs_code == 1 ~ "Large central metro", 
                                 nchs_code == 2 ~ "Large fringe metro",
                                 nchs_code == 3 ~ "Medium metro",
                                 nchs_code == 4 ~ "Small metro",
                                 nchs_code == 5 ~ "Micropolitan",
                                 nchs_code == 6 ~ "Noncore")) %>%
    select(fips_code, nchs_code)

Finally we can combine all of these together into a single dataframe.

combined <- pres %>%
    left_join(acs2 %>% select(-c(state, county)), by = "fips_code") %>%
    left_join(communities2) %>%
    left_join(nchs) %>%
    mutate(turnout20 = total_votes20 / tot_pop,
           turnout16 = total_votes16 / tot_pop,
           turnout_dif = turnout20 - turnout16) %>%
    relocate(name, county, nchs_code, acp_county_type) 

ga <- combined %>%
    filter(state == "Georgia") %>%
    mutate(county = str_remove(county, " County"))

## Dem margin shift 2020-2016
p1 <- ga %>%
    group_by(acp_county_type) %>%
    summarize(dem_margin_dif = mean(dem_margin_dif, na.rm = TRUE)) %>%
    drop_na() %>%
    ggplot(aes(dem_margin_dif, reorder(acp_county_type, dem_margin_dif), 
               color = cut(dem_margin_dif, c(-Inf, 0, 0.03, Inf)))) +
    geom_point(size = 2) +
    geom_segment(aes(x = 0, xend = dem_margin_dif, y = acp_county_type, 
                     yend = acp_county_type), size = 1) +
    scale_x_continuous(labels = scales::percent_format()) +
    labs(y = "", x = "", title = "GA Democratic Shifts in 2020 by County",
         subtitle = "Difference in vote margin between 2016 and 2020") +
    scale_color_manual(values = c("(-Inf,0]" = "red",
                                  "(0,0.03]" = "purple",
                                  "(0.03, Inf]" = "royalblue")) + 
    theme(legend.position = "none") 

## `summarise()` ungrouping output (override with `.groups` argument)

p2 <- combined %>%
    group_by(acp_county_type) %>%
    summarize(dem_margin_dif = mean(dem_margin_dif, na.rm = TRUE)) %>%
    drop_na() %>%
    ggplot(aes(dem_margin_dif, reorder(acp_county_type, dem_margin_dif), 
               color = cut(dem_margin_dif, c(-Inf, 0, 0.03, Inf)))) +
    geom_point(size = 2) +
    geom_segment(aes(x = 0, xend = dem_margin_dif, y = acp_county_type, yend = acp_county_type), size = 1) +
    scale_x_continuous(labels = scales::percent_format()) +
    labs(y = "", x = "", title = "US Democratic Shifts in 2020 by County",
         subtitle = "Difference in vote margin between 2016 and 2020") +
    scale_color_manual(values = c("(-Inf,0]" = "red",
                                  "(0,0.03]" = "purple",
                                  "(0.03, Inf]" = "royalblue")) +
    theme(legend.position = "none") 

## `summarise()` ungrouping output (override with `.groups` argument)

## Dem Margin in 2020
p3 <- ga %>%
    group_by(acp_county_type) %>%
    summarize(dem_margin20 = mean(dem_margin20, na.rm = TRUE)) %>%
    drop_na() %>%
    ggplot(aes(dem_margin20, reorder(acp_county_type, dem_margin20), 
               color = cut(dem_margin20, c(-Inf, -0.025, 0.025, Inf)))) +
    geom_point(size = 2) +
    geom_segment(aes(x = 0, xend = dem_margin20, y = acp_county_type, yend = acp_county_type), size = 1) +
    scale_x_continuous(labels = scales::percent_format()) +
    labs(y = "", x = "", title = "GA 2020 Democrat Margin by County") +
    scale_color_manual(values = c("(-Inf,-0.025]" = "red",
                                  "(-0.025,0.025]" = "purple",
                                  "(0.025, Inf]" = "royalblue")) +
    theme(legend.position = "none") 

## `summarise()` ungrouping output (override with `.groups` argument)

p4 <- combined %>%
    group_by(acp_county_type) %>%
    summarize(dem_margin20 = mean(dem_margin20, na.rm = TRUE)) %>%
    drop_na() %>%
    ggplot(aes(dem_margin20, reorder(acp_county_type, dem_margin20), 
               color = cut(dem_margin20, c(-Inf, -0.025, 0.025, Inf)))) +
    geom_point(size = 2) +
    geom_segment(aes(x = 0, xend = dem_margin20, y = acp_county_type, yend = acp_county_type), size = 1) +
    scale_x_continuous(labels = scales::percent_format()) +
    labs(y = "", x = "", title = "US 2020 Democrat Margin by County") +
    scale_color_manual(values = c("(-Inf,-0.025]" = "red",
                                  "(-0.025,0.025]" = "purple",
                                  "(0.025, Inf]" = "royalblue")) +
    theme(legend.position = "none") 

## `summarise()` ungrouping output (override with `.groups` argument)

p1 + p2 + p3 + p4

#ggsave("test.png", width = 16/1.2, height = 9/1.2)

ga %>%
    ggplot(aes(dem_margin20, reorder(acp_county_type, dem_margin20), 
               color = cut(dem_margin20, c(-Inf, -0.025, 0.025, Inf)))) +
    geom_point(aes(alpha = 0.3, size = total_votes20)) + 
    #geom_boxplot() + 
    ggrepel::geom_text_repel(data = ga %>% filter(total_votes20 > 200000), aes(label = county)) + 
    scale_x_continuous(labels = scales::percent_format()) +
    labs(y = "", x = "", title = "GA 2020 Democrat Margin by County") +
    geom_vline(xintercept = 0, linetype = "dashed") + 
    scale_color_manual(values = c("(-Inf,-0.025]" = "red",
                                  "(-0.025,0.025]" = "purple",
                                  "(0.025, Inf]" = "royalblue")) +
    theme(legend.position = "none") 

## Turnout diff
ga %>%
    group_by(acp_county_type) %>%
    summarize(turnout_dif = mean(turnout_dif, na.rm = TRUE)) %>%
    drop_na() %>%
    ggplot(aes(turnout_dif, reorder(acp_county_type, turnout_dif), 
               color = turnout_dif)) +
    geom_point(size = 2) +
    geom_segment(aes(x = 0, xend = turnout_dif, y = acp_county_type, yend = acp_county_type), size = 1) +
    labs(y = "", x = "", title = "GA Turnout Shifts in 2020 by County",
         subtitle = "Difference in total turnout between 2016 and 2020") +
    theme(legend.position = "none")  +
    scale_x_continuous(labels = scales::percent_format())

## `summarise()` ungrouping output (override with `.groups` argument)

ga %>% 
    summarize(maj_black = mean(dem_margin_dif[black_perc > 0.5]),
              rural = mean(dem_margin_dif[nchs_code == "Noncore"]),
              suburban = mean(dem_margin_dif[acp_county_type == "Urban Suburbs"]),
              urban = mean(dem_margin_dif[acp_county_type == "Big Cities"]),
              white_working = mean(dem_margin_dif[white_perc >= .8 & college_perc <= .25]),
              maj_college = mean(dem_margin_dif[college >= 0.5])) %>%
    pivot_longer(everything()) %>%
    ggplot(aes(value, reorder(name, value), 
               color = cut(value, c(-Inf, -0.012, 0.012, Inf)))) +
    geom_point(size = 2) +
    geom_segment(aes(x = 0, xend = value, y = name, yend = name), size = 1) +
    scale_x_continuous(labels = scales::percent_format()) +
    labs(y = "", x = "", title = "GA Democrat Margin 2020 vs. 2016 by Group") +
    scale_color_manual(values = c("(-Inf,-0.012]" = "red",
                                  "(-0.012,0.012]" = "purple",
                                  "(0.012, Inf]" = "royalblue")) +
    theme(legend.position = "none") 

acs_sf <- acs %>%
    clean_names() %>%
    separate(name, into = c("county", "state"), 
                           sep = ", ", remove = FALSE) %>%
                  select(name, county, state) %>%
    filter(state == "Georgia") %>% 
    left_join(fips_codes %>% 
                  select(-state) %>% 
                  rename(state = state_name) %>% 
                  unite(fips_code, state_code, county_code, sep = "")) %>%
    select(fips_code, geometry) %>% 
    distinct(fips_code)

ga_sf <- acs_sf %>%
    left_join(ga)

ga_sf %>%
    filter(acp_county_type %in% c("Urban Suburbs", "Exurbs", "Big Cities")) %>% 
    ggplot() +
    geom_sf() +
    geom_sf(aes(fill = acp_county_type)) +
    theme(panel.grid.major = element_blank(),
          panel.grid.minor = element_blank(),
          panel.border = element_blank(),
          panel.background = element_blank(),
          axis.ticks = element_blank(),
          axis.text = element_blank()) +
    labs(x = "", y = "") + 
    #scale_fill_viridis_d(direction = -1) +
    geom_text(aes(label = county, geometry = geometry),
              #color = "white",
              size = 3,
              stat = "sf_coordinates") +
     coord_sf(xlim = c(-85.484619,-83.111573), 
             ylim = c(33.275436,34.524662)) 

## Warning in st_point_on_surface.sfc(sf::st_zm(x)): st_point_on_surface may not
## give correct results for longitude/latitude data

p1 <- ga_sf %>%
    ggplot() +
    geom_sf() +
    geom_sf(aes(fill = dem_margin20)) +
    theme_void() +
    scale_fill_gradient2(low = "#ca0020", high = "#0571b0",
                         labels = scales::percent_format()) +
    labs(title = "Democratic Margin 2020") +
    theme(legend.title = element_blank())

p2 <- ga_sf %>%
    ggplot() +
    geom_sf() +
    geom_sf(aes(fill = dem_margin_dif)) +
    theme_void() +
    scale_fill_gradient2(low = "#ca0020", high = "#0571b0",
                         labels = scales::percent_format()) +
    labs(title = "Democratic Margin Difference 2020-2016") +
    theme(legend.title = element_blank())

p3 <- ga_sf %>%
    ggplot() +
    geom_sf() +
    geom_sf(aes(fill = turnout20)) +
    theme_void() +
    scale_fill_fermenter(palette = "Blues", direction = 1,
                         labels = scales::percent_format()) + 
    labs(title = "Turnout 2020") +
    theme(legend.title = element_blank())


p1 + p2

#ggsave("ga_maps2.png", width = 16/1.2, height = 9/1.2)

p3 

#ggsave("ga_maps.png", width = 16/1.2, height = 9/1.2)

ga %>%
    select(county, acp_county_type, dem_margin20, dem_margin_dif) %>%
    arrange(desc(dem_margin_dif)) %>%
    # slice_head(n = 10) %>% 
    gt(rowname_col = vars(county)) %>%
    tab_header(title = "Georgia 2020 Presidential by County") %>%
    cols_label(acp_county_type = "County Type", county = "County", 
               dem_margin20 = "Democrat Margin 2020", 
               dem_margin_dif = "Democrat Margin 2020 - 2016") %>%
    fmt_percent(columns = vars(dem_margin20, dem_margin_dif)) %>%
    data_color(columns = vars(dem_margin20),
               colors = scales::col_bin(palette = c("red2", "blue2"), 
                                        bins = 2,
                                        domain = c(-40,40)))

ga %>%
  select("dem_margin_dif", "income_per_cap", "black_perc", "college_perc",
         "white_perc", "other_state_perc", "pov_perc") %>%
  ggpairs(progress = FALSE) 

ACS stats over time

One of my main questions is whether demographic trends, particularly in the Atlanta suburbs, might give some indication for whether these swing voters might be more solidly democratic, or if we can expect some combination of ticket-splitting and/or reversion to the (Republican) mean. So we can pull ACS data from 2014 in addition to 2019 to get a sense for how things are changing.

acs_time <- map_dfr(c(2014, 2019), ~ get_acs(geography = "county",
               variables = c(tot_pop = "B01003_001", 
                             age_male = "B01002_002",
                             age_female = "B01002_003",
                             ba = "B15003_022",
                             ma = "B15003_023",
                             pd = "B15003_024",
                             phd = "B15003_025", 
                             poverty = "B17001_001",
                             medicaid = "C27007_001",
                             tot_white = "B02001_002",
                             tot_black = "B02001_003",
                             tot_ai = "B02001_004",
                             tot_asian = "B02001_005",
                             born_other_state = "B06001_025",
                             income = "B19127_001"),
              year = .x) %>%
                  mutate(year = .x))

acs_time2 <- acs_time %>%
    clean_names() %>%
    unite(name_year, c(name, year), na.rm = TRUE, sep = ", ") %>%
    pivot_wider(names_from = variable, values_from = estimate, id_cols = name_year) %>%
    separate(name_year, into = c("county", "state", "year"), sep = ", ") %>%
    distinct() %>%
    left_join(fips_codes %>% select(-state) %>% unite(fips_code, c(state_code, county_code), sep = ""), by= c("state" = "state_name", "county")) %>%
    mutate(medicaid_perc = medicaid / tot_pop,
           pov_perc = poverty / tot_pop,
           college = ba + ma + pd + phd, 
           college_perc = college / tot_pop,
           white_perc = tot_white / tot_pop,
           black_perc = tot_black / tot_pop,
           ai_perc = tot_ai / tot_pop,
           asian_perc = tot_asian / tot_pop,
           other_state_perc = born_other_state / tot_pop,
           income_per_cap = income / tot_pop,
           ga = if_else(state == "Georgia", "Georgia", "All other states")) %>%
    left_join(communities2) 

The first plot below compares Georgia with other states according to six demographic variables, while the second chart looks only at Georgia’s Big Cities, Suburbs, and Exurbs:

## change over time plots
for_plot <- acs_time2 %>%
    select(county, state, year, ga, acp_county_type, college_perc, white_perc, black_perc,
           born_other_state, other_state_perc, income_per_cap) %>%
    mutate(ga = factor(ga, levels = c("All other states", "Georgia"))) %>% 
    group_by(state, year) %>%
    summarize(college_perc = mean(college_perc, na.rm = TRUE),
              white_perc = mean(white_perc, na.rm = TRUE),
              black_perc = mean(black_perc, na.rm = TRUE),
              born_other_state = sum(born_other_state), 
              other_state_perc = mean(other_state_perc, na.rm = TRUE),
              income_per_cap = mean(income_per_cap, na.rm = TRUE)) %>% 
    pivot_longer(college_perc:income_per_cap) 

## `summarise()` regrouping output by 'state' (override with `.groups` argument)

for_plot %>%
    ggplot(aes(year, value, group = state)) +
    geom_point(aes(color = "#F8766D")) + 
    geom_point(data = for_plot %>% filter(state == "Georgia"), aes(color = "#00BFC4"),
               size = 2.5) +
    geom_line(aes(color = "#F8766D")) + 
    geom_line(data = for_plot %>% filter(state == "Georgia"), aes(color = "#00BFC4"),
              size = 1.5) +
    facet_wrap(~ name, scales = "free_y") +
    theme(legend.position = "none") +
    labs(y = "", x = "", title = "Georgia Demographic Change 2014-2019 vs. the US")

## Warning: Removed 4 rows containing missing values (geom_point).

## GA change by county type
acs_time2 %>%
    select(county, state, year, ga, acp_county_type, college_perc, white_perc, black_perc,
           other_state_perc, income_per_cap) %>%
    filter(state == "Georgia",
           acp_county_type %in% c("Big Cities", "Exurbs", "Urban Suburbs")) %>% 
    group_by(acp_county_type, year) %>%
    summarize(college_perc = mean(college_perc, na.rm = TRUE),
              white_perc = mean(white_perc, na.rm = TRUE),
              black_perc = mean(black_perc, na.rm = TRUE),
              other_state_perc = mean(other_state_perc, na.rm = TRUE),
              income_per_cap = mean(income_per_cap, na.rm = TRUE)) %>% 
    pivot_longer(college_perc:other_state_perc) %>%
    drop_na(acp_county_type) %>%
    ggplot(aes(year, value, group = acp_county_type, color = acp_county_type)) +
    geom_point() + 
    geom_line() + 
    facet_wrap(~ name, scales = "free_y") +
    scale_y_continuous(labels = scales::percent_format(1L)) +
    labs(y = "", x = "", title = "Georgia Demographic 2014-2019 Change by County Type")

## `summarise()` regrouping output by 'acp_county_type' (override with `.groups` argument)

The first chart shows that Georgia counties have a substantial percentage of Black residents and residents born in another state, although these populations haven’t grown significantly in over the last five years. It has seen a marked increase in income per capita as well as in the percentage of college-educated adults (trends in line with the rest of the country). The second chart shows that many college-educated Georgians live in Big Cities, Exurbs, and Urban Suburbs, with increases in all three over the five year window.

Senate results vs pres

Next we can compare Ossoff and Warnock’s special election margins over Biden in Georgia. While late polls in Georgia indicated that Ossoff and Warnock might be favored by a few points (1.8-2.1 points), there was still a lot of uncertainty and apprehension among Georgia Democrats. Even while the Georgia polls were among the most accurate in Georgia of any state for the November election, many top-tier pollsters sat out the January Senate elections, and Democrats’ leads were still within the margin of error. Further, given that Biden ran ahead of both Warnock and Ossoff in November and that there was every reason to expect GOP voters to be highly motivated given Trump’s recent loss (both overall and in the state!), Democrats’ expectations were mostly held in check.

ossoff <- read_csv("C:\\Users\\ChadPeltier\\Downloads\\detailxls\\ossoff.csv") %>%
    row_to_names(row_number = 2) %>%
    clean_names() %>%
    rename_with(.cols = 2:7, .fn = ~paste0("perdue_", .x)) %>%
    rename_with(.cols = 8:12, .fn = ~paste0("ossoff_", .x)) %>%
    select(county, contains("total")) %>%
    mutate(across(2:4, as.numeric),
           ossoff_percent = ossoff_total_votes_2 / total,
           perdue_percent = perdue_total_votes / total,
           ossoff_margin = ossoff_percent - perdue_percent) %>%
    select(county, ossoff_percent, ossoff_margin)

## Warning: Missing column names filled in: 'X2' [2], 'X3' [3], 'X4' [4], 'X5' [5],
## 'X6' [6], 'X7' [7], 'X8' [8], 'X9' [9], 'X10' [10], 'X11' [11], 'X12' [12],
## 'X13' [13]

## Warning in row_to_names(., row_number = 2): Row 2 does not provide unique names.
## Consider running clean_names() after row_to_names().

warnock <- read_csv("C:\\Users\\ChadPeltier\\Downloads\\detailxls\\warnock.csv") %>%
    row_to_names(row_number = 2) %>%
    clean_names() %>%
    rename_with(.cols = 2:7, .fn = ~paste0("loeffler_", .x))%>%
    rename_with(.cols = 8:12, .fn = ~paste0("warnock_", .x)) %>%
    select(county, contains("total")) %>%
    mutate(across(2:4, as.numeric),
           warnock_percent = warnock_total_votes_2 / total,
           loeffler_percent = loeffler_total_votes / total,
           warnock_margin = warnock_percent - loeffler_percent) %>%
    select(county, warnock_percent, warnock_margin)

## Warning: Missing column names filled in: 'X2' [2], 'X3' [3], 'X4' [4], 'X5' [5],
## 'X6' [6], 'X7' [7], 'X8' [8], 'X9' [9], 'X10' [10], 'X11' [11], 'X12' [12],
## 'X13' [13]

## Warning: Row 2 does not provide unique names. Consider running clean_names()
## after row_to_names().

senate <- ossoff %>%
    left_join(warnock)

ga_senate <- ga %>%
    left_join(senate) %>%
    mutate(ossoff_biden_margin = ossoff_percent - dem_perc20,
           warnock_biden_margin = warnock_percent - dem_perc20)

Nevertheless, the polls went 2/2 in Georgia for the 2020 election cycle.

## counties where ossoff / warnock outperformed Biden 
ga_senate_sf <- acs_sf %>%
    left_join(ga_senate)

## Joining, by = "fips_code"

p5 <- ga_senate_sf %>%
    ggplot() +
    geom_sf() +
    geom_sf(aes(fill = ossoff_biden_margin)) +
    theme_void() +
    scale_fill_gradient2(low = "#ca0020", high = "#0571b0",
                         labels = scales::percent_format()) +
    labs(title = "Ossoff Margin Over Biden (2021)") +
    theme(legend.title = element_blank())


p6 <- ga_senate_sf %>%
    ggplot() +
    geom_sf() +
    geom_sf(aes(fill = warnock_biden_margin)) +
    theme_void() +
    scale_fill_gradient2(low = "#ca0020", high = "#0571b0",
                         labels = scales::percent_format()) +
    labs(title = "Warnock Margin Over Biden (2021)") +
    theme(legend.title = element_blank())

p1 + p2 

p5 + p6

#ggsave("senate.png", width = 16/1.2, height = 9/1.2)

Here we can really see the impact that south Atlanta counties has for Democrats in 2020. Clayton, Douglas, Newton, and Henry counties all voted for Warnock in January with 3% margins over their margins for Biden in November.

While there wasn’t a ton of ticket splitting in November, Warnock did outperform Ossoff in most counties across Georgia. There are various reasons for this that will take more detailed data to disentangle, but it is generally thought that Perdue, as a non-appointed incumbent, would outperform Loeffler, who also shifted heavily to the right in her wide-open primary where she had to defeat GOP representative Doug Collins (I think we have to acknowledge the likely role that sexism plays in Perdue’s margin, too). Also, some percentage of swing voters might like both Democratic candidates, but would rather the Democrats not take unified control of government.

## where warnock beat ossoff
ga_senate_sf %>%
    mutate(warnock_ossoff_margin = warnock_biden_margin - ossoff_biden_margin) %>%
    ggplot() +
    geom_sf() +
    geom_sf(aes(fill = warnock_ossoff_margin)) +
    theme_void() +
    scale_fill_gradient2(low = "#ca0020", high = "#0571b0",
                         labels = scales::percent_format()) +
    labs(title = "Warnock Margin Over Ossoff (2021)") +
    theme(legend.title = element_blank())

p7 <- ga_senate_sf %>%
    group_by(acp_county_type) %>%
    summarize(ossoff_biden_margin = mean(ossoff_biden_margin, na.rm = TRUE)) %>%
    drop_na() %>%
    ggplot(aes(ossoff_biden_margin, reorder(acp_county_type, ossoff_biden_margin), 
               color = cut(ossoff_biden_margin, c(-Inf, -0.0025, 0.0025, Inf)))) +
    geom_point(size = 2) +
    geom_segment(aes(x = 0, xend = ossoff_biden_margin, y = acp_county_type, yend = acp_county_type), size = 1) +
    scale_x_continuous(labels = scales::percent_format()) +
    labs(y = "", x = "", title = "Ossoff 2021 Runoff Margin Over Biden") +
    scale_color_manual(values = c("(-Inf,-0.0025]" = "red",
                                  "(-0.0025,0.0025]" = "purple",
                                  "(0.0025, Inf]" = "royalblue")) +
    theme(legend.position = "none") 

## `summarise()` ungrouping output (override with `.groups` argument)

p8 <- ga_senate_sf %>%
    group_by(acp_county_type) %>%
    summarize(warnock_biden_margin = mean(warnock_biden_margin, na.rm = TRUE)) %>%
    drop_na() %>%
    ggplot(aes(warnock_biden_margin, reorder(acp_county_type, warnock_biden_margin), 
               color = cut(warnock_biden_margin, c(-Inf, -0.0025, 0.0025, Inf)))) +
    geom_point(size = 2) +
    geom_segment(aes(x = 0, xend = warnock_biden_margin, y = acp_county_type, yend = acp_county_type), size = 1) +
    scale_x_continuous(labels = scales::percent_format()) +
    labs(y = "", x = "", title = "Warnock 2021 Runoff Margin Over Biden") +
    scale_color_manual(values = c("(-Inf,-0.0025]" = "red",
                                  "(-0.0025,0.0025]" = "purple",
                                  "(0.0025, Inf]" = "royalblue")) +
    theme(legend.position = "none") 

## `summarise()` ungrouping output (override with `.groups` argument)

p7 + p8

Interestingly, Warnock’s advantage relative to Ossoff was probably most critical in Atlanta itself, where Ossoff actually performed worse than Biden! Further, and importantly for Democrats moving forward, Ossoff and Warnock both outperformed Biden in the suburbs and exurbs.

Survey

Finally, while this data is interesting, it’s also at the county level and doesn’t give us a clear picture of how individual voters within those counties, with their unique combinations of demographic factors and political opinions, might actually feel about Democratic candidates and Democratic policies.

While Ossoff and Warnock’s margins over Biden in January and demographic shifts might suggest Democrats have a solid shot of holding on to Georgia in 2022 and 2024, actual polling data would be critical to understanding how much Democrats’ gains were limited to anti-Trump voters and not actual conversions.

The UCLA / Democracy Fund Nationscape data can give us at least a little more information here. I’ll read in data from the last month of surveys in June 2020, bind them, and then add some Bloomberg CityLab Congressional District data to classify districts on an urban-rural scale, similar to the process above for counties. Then I’ll make chart for three survey questions, with results aggregated by district type.

Note that thanks to the excellent srvyr package, I was able to incorporate the survey’s weights, which hopefully gives a more accurate picture of the actual US population – although it should be noted that 2020 polls still were affected by signficant non-respondent bias that mostly went Trump’s way, even after weighting for factors like education.

path <- "C:\\Users\\ChadPeltier\\Downloads\\Nationscape_phase2\\Nationscape-DataRelease_WeeklyMaterials_DTA\\phase_2_v20200814"

files <- tibble(files = list.files(path))

files <- files %>%
    mutate(files2 = paste0(path, "\\", files))

files2 <- map(files$files2, ~ list.files(.x, full.names = TRUE)) %>%
    enframe() %>%
    unnest(value) %>%
    mutate(value = str_replace(value, "\\/", "\\\\"),
           date = str_extract(value, "(?<=ns)\\d+"),
           date = ymd(date)) %>%
    filter(str_detect(value, "\\.dta$"),
           date > ymd("2020-06-01"))

nationscape <- map_dfr(files2$value, haven::read_dta)

## add citylab cdi data
cdi <- read_csv("https://raw.githubusercontent.com/theatlantic/citylab-data/master/citylab-congress/citylab_cdi.csv") %>%
    clean_names() %>%
    mutate(cd = str_remove(cd, "-"),
           cd = str_replace(cd, "AL", "00"))

## final cleaning, replacing 888s, dropping NAs for ID and weight, changing col types 
nationscape2 <- nationscape %>%
    left_join(cdi %>% select(cd, cluster), by = c("congress_district" = "cd")) %>%
    mutate(across(everything(), as.character),
           across(everything(), ~ str_replace_all(., "888", NA_character_)),
           weight = as.numeric(weight),
           cluster = factor(cluster, levels = c("Pure rural", "Rural-suburban   mix",
                                                "Sparse suburban", "Dense suburban",
                                                "Urban-suburban mix", "Pure urban"))) %>%
    drop_na(response_id, weight) 

First, the survey asked “Would you consider voting for Trump?” This question gives us a baseline for how different types of districts would vote.

## not_trump
nationscape2 %>%
    mutate(ga = if_else(state == "GA", 1, 0),
           consider_trump = case_when(consider_trump == 1 ~ "Yes",
                                      consider_trump == 2 ~ "No",
                                      consider_trump == 999 ~ "Don't know")) %>%
    filter(
           #cluster %in% c("Dense suburban", "Urban-suburban mix", "Sparse suburban"),
           !is.na(consider_trump)) %>%
    as_survey_design(weights = weight) %>%
    group_by(cluster, consider_trump) %>%
    summarize(mean = survey_mean()) %>%
    mutate(upper = mean + mean_se,
           lower = mean - mean_se,
           consider_trump = factor(consider_trump, 
                                   levels = c("No", "Don't know", "Yes"))) %>%
    drop_na(cluster) %>%
    ggplot(aes(mean, cluster, color = consider_trump)) +
    geom_point() +
    geom_errorbar(aes(xmin = lower, xmax = upper), width = 0.2) +
    labs(y = "", x = "% of U.S.",
         title = "Would you consider voting for Trump?") +
    scale_x_continuous(labels = scales::percent) +
    theme(legend.title = element_blank())

This is one of the clearest charts for demonstrating the geographic partisan divide in the US right now. As districts get more rural, the percentage of respondents who would consider voting for Trump increases.

nationscape2 %>%
    mutate(ga = if_else(state == "GA", 1, 0),
           group_favorability_democrats = case_when(group_favorability_democrats == 1 ~ "Very favorable",
                                      group_favorability_democrats == 2 ~ "Somewhat favorable",
                                      group_favorability_democrats == 999 ~ "Haven't heard enough",
                                      group_favorability_democrats == 3 ~ "Somewhat unfavorable",
                                      group_favorability_democrats == 4 ~ "Very unfavorable"),
           group_favorability_democrats = factor(
              group_favorability_democrats, 
              levels = c("Very unfavorable", "Somewhat unfavorable", "Haven't heard enough",
                         "Somewhat favorable", "Very favorable"))) %>%
    filter(consider_trump %in% c(2, 999),
           !is.na(group_favorability_democrats)) %>%
    as_survey_design(weights = weight) %>%
    group_by(cluster, group_favorability_democrats) %>%
    summarize(mean = survey_mean()) %>%
    mutate(upper = mean + mean_se,
           lower = mean - mean_se) %>%
    drop_na(cluster) %>%
    ggplot(aes(mean, cluster, color = group_favorability_democrats)) +
    geom_point() +
    geom_errorbar(aes(xmin = lower, xmax = upper), width = 0.2) +
    labs(y = "", x = "% of U.S.", 
         title = "How favorable is your impression of Democrats?",
         subtitle = "For individuals who don't know or would not consider voting for Trump") +
    scale_x_continuous(labels = scales::percent) +
    theme(legend.title = element_blank())

Next, the chart above looks at whether respondents who don’t know or definitely would not vote for Trump actually like Democrats. I was hoping this view might give some indication whether anti-Trump and swing voters are actually fans of Democrats themselves or just anti-Trump. This suggests that individuals who are anti-Trump are mostly still for Democrats, too. Across all district types, majorities (60-70%) of anti-Trump and undecided respondents have somewhat or very favorable opinions of Democrats.

nationscape2 %>%
    mutate(ga = if_else(state == "GA", 1, 0),
           public_option = case_when(public_option == 1 ~ "Agree",
                                     public_option == 999 ~ "Not sure",
                                     public_option == 2 ~ "Disagree"),
           public_option = factor(public_option, levels = c("Disagree", "Not sure", "Agree"))) %>%
    filter(consider_trump %in% c(2, 999),
           !is.na(public_option)) %>%
    as_survey_design(weights = weight) %>%
    group_by(cluster, public_option) %>%
    summarize(mean = survey_mean()) %>%
    mutate(upper = mean + mean_se,
           lower = mean - mean_se) %>%
    drop_na(cluster) %>%
    ggplot(aes(mean, cluster, color = public_option)) +
    geom_point() +
    geom_errorbar(aes(xmin = lower, xmax = upper), width = 0.2) +
    labs(y = "", x = "% of U.S.", 
         title = "The government should provide a public option",
         subtitle = "For individuals who don't know or would not consider voting for Trump") +
    scale_x_continuous(labels = scales::percent) +
    theme(legend.title = element_blank())

Finally, how do those same groups of voters feel about Democratic policies? Again, anti-Trump and undecided voters overwhelmingly like at least one major Democratic policy priority – a public option for health insurance. Majorities of these voters from every district type are in favor of a public option, with another 20% who aren’t sure.

Overall, while Democratic strategists in Georgia and around the country will work on this question for the next four years, there are decent reasons to expect Georgia to still be a blue-leaning state in 2022 and 2024.