Left Parties and Women’s Political Representation: A Specification-Sensitivity Analysis with sensiverse

remotes::install_github("mganslmeier/sensiverse")
library(readxl)
library(dplyr)
library(ggplot2)
library(stringr)
library(tidyr)
library(scales)
library(reticulate)
library(sensiverse)

theme_set(theme_minimal(base_size = 12))

1 Research Question & Design

We examine whether stronger left‐party representation in parliament is associated with greater women’s political representation and empowerment, and how sensitive that relationship is to different modelling choices.

We treat:

• Focus IV: left_strength — combined seat share of social democratic and left‐socialist parties (from CPDS data integrated in QoG)
  
• Outcomes (dependent variables):
  1. ipu_l_sw – Women’s share in the lower or single chamber of parliament
     
  2. wdi_wip – World Bank indicator of women in national parliaments
     
  3. vdem_gender – V-Dem index of women’s political empowerment

We vary:

• Samples: Nordic / Anglo / Continental democracies, plus two periods (1990–2008, post-2009)
  
• Fixed effects: none / country / year / both
  
• Standard error types: simple / robust / clustered
  
• Controls:
  • Human capital (economic modernization, pwt_hci)
    
  • Internet penetration (digital diffusion, wdi_internet)
    
  • Unemployment rate (labor-market structure, wdi_unempilo)
    
  • Liberal democracy index (vdem_libdem, democratization and governance)
    
  • Electoral system type (proportional vs. majoritarian, iaep_es)

2 Load QoG Dataset and Map Variable Names

# ---- Load QoG Basic (Jan25) ------------------------------------
path_qog <- "data/qog_bas_ts_jan25.dta"
qog <- haven::read_dta(path_qog)

# Helper: pick first present name among candidates (keeps script robust across QoG formats)
find_col <- function(df, cands, label) {
  hit <- intersect(cands, names(df))
  if (!length(hit)) stop(sprintf("Missing variable: %s (looked for: %s)", label, paste(cands, collapse=", ")))
  hit[1]
}

# Core IDs (QoG typically uses cname for country names)
country_col <- find_col(qog, c("cname","country","CNAME"), "country")
year_col    <- find_col(qog, c("year","Year"), "year")

# Focus IV: Left strength = social democratic + left-socialist seat shares
cpds_ls_col  <- find_col(qog, c("cpds_ls"),  "CPDS social democratic seats")
cpds_lls_col <- find_col(qog, c("cpds_lls"), "CPDS left-socialist seats")

# Outcomes (DVs)
ipu_l_col     <- find_col(qog, c("ipu_l_sw"),     "IPU women share lower/single house")
wdi_wip_col   <- find_col(qog, c("wdi_wip"),      "WDI women in parliament")
vdem_gender_col <- find_col(qog, c("vdem_gender"), "V-Dem women political empowerment")

# Controls
pwt_hci_col     <- find_col(qog, c("pwt_hci"),        "Human capital index (PWT)")
wdi_internet_col<- find_col(qog, c("wdi_internet"),   "Internet users (% pop)")
wdi_unemp_col   <- find_col(qog, c("wdi_unempilo"),   "Unemployment (ILO modeled)")
vdem_libdem_col <- find_col(qog, c("vdem_libdem"),    "Liberal democracy index")
iaep_es_col     <- find_col(qog, c("iaep_es"),        "Electoral system (IAEP)")

3 Construct Analysis Matrix and Samples

nordics     <- c("Denmark","Finland","Iceland","Norway","Sweden")
anglo       <- c("United Kingdom","United States","Canada","Australia","New Zealand","Ireland")
continental <- c("Germany","France","Netherlands","Belgium","Austria","Switzerland","Luxembourg")

dat <- qog %>%
  transmute(
    country = .data[[country_col]],
    year    = .data[[year_col]],

    # DVs
    DEP_ipu_l_sw     = .data[[ipu_l_col]],
    DEP_wdi_wip      = .data[[wdi_wip_col]],
    DEP_vdem_gender  = .data[[vdem_gender_col]],

    # Focus IV: Left strength = cpds_ls + cpds_lls (sum available parts), then scale
    left_strength_raw = dplyr::coalesce(.data[[cpds_ls_col]], 0) +
                        dplyr::coalesce(.data[[cpds_lls_col]], 0),
    CONT_left   = as.numeric(scale(left_strength_raw)),

    # Controls (scaled for comparability)
    CONT_hci       = as.numeric(scale(.data[[pwt_hci_col]])),
    CONT_internet  = as.numeric(scale(.data[[wdi_internet_col]])),
    CONT_unemp     = as.numeric(scale(.data[[wdi_unemp_col]])),
    CONT_libdem    = as.numeric(scale(.data[[vdem_libdem_col]])),
    # PR vs non-PR: IAEP electoral system = 3 (PR) or 4 (Mixed) -> treat as PR=1 (common practice in comp. pol.)
    pr_flag        = dplyr::case_when(.data[[iaep_es_col]] %in% c(3,4) ~ 1L, 
                                      .data[[iaep_es_col]] %in% c(1,2) ~ 0L, 
                                      TRUE ~ NA_integer_),
    CONT_pr        = as.numeric(scale(pr_flag)),
    FEgroup = .data[[country_col]],
    FEtime  = .data[[year_col]]
  ) %>%
  mutate(
    SAMPLE_all         = 1L,
    SAMPLE_1990_2008   = as.integer(year >= 1990 & year <= 2008),
    SAMPLE_post2009    = as.integer(year >= 2009),
    SAMPLE_nordics     = as.integer(country %in% nordics),
    SAMPLE_anglo       = as.integer(country %in% anglo),
    SAMPLE_continental = as.integer(country %in% continental)
  ) %>%
  # Require focus IV + at least one DV observed
  filter(!is.na(CONT_left) & (!is.na(DEP_ipu_l_sw) & !is.na(DEP_wdi_wip) & !is.na(DEP_vdem_gender)))

4 Declare the Specification Universe

specs <- list(
  dep  = c("DEP_ipu_l_sw","DEP_wdi_wip","DEP_vdem_gender"),
  cont = c("CONT_left","CONT_hci","CONT_internet","CONT_unemp","CONT_libdem","CONT_pr"),
  fes  = c("none","group","time","time+group"),
  set  = c("simple","robust","clustered"),
  sample = c("SAMPLE_all", "SAMPLE_1990_2008","SAMPLE_post2009",
             "SAMPLE_nordics","SAMPLE_anglo","SAMPLE_continental")
)

5 Estimate the OLS Universe

res <- estimate_model_space(
  df        = dat,
  focus_var = "CONT_left",
  specs     = specs,
  space_n   = 5000
)

6 Filter Best-Fit Models by AIC

res_filt <- filter_model_space(res, top_aic_pct = 0.25)

7 Inspect Sign Shares

shares_all <- calculate_sign_shares(res, dimension = NULL)
shares_all

for (d in c("dep","fes","set","sample")) {
  s <- calculate_sign_shares(res, dimension = d)
  print(plot_sign_share(s, dimension = d))
}

8 Uncertainty Sources — Nearest Neighbour (One-Step Flips)

nn_out <- find_uncertainty_source_neigh(
  specs   = specs,
  df      = dat,
  focus   = "CONT_left",
  n_draws = 400,
  seed    = 2025
)
plot_importance_neigh(nn_out)

9 Uncertainty Sources — Multinomial Logit

mlogit_out <- find_uncertainty_source_mlogit(res, focus="CONT_left")
plot_importance_mlogit(mlogit_out, aggregate=TRUE)

plot_importance_mlogit(mlogit_out, aggregate=FALSE)

10 Neural Net + SHAP Explanation

# install.packages(c("keras","vip"))
# keras::install_keras()
# use_python("~/miniforge3/bin/python") # make sure you use the correct python and tensorflow version
# use_condaenv("tf_env")

nn_out <- find_uncertainty_source_neuronet(
  res    = res,
  focus  = "CONT_left",
  epochs = 15,
  batch_size = 256,
  verbose = 0
)

p1 <- plot_importance_neuronet(nn_out, aggregate = TRUE)
p2 <- plot_importance_neuronet(nn_out, aggregate = FALSE)

plot(p1)
plot(p2)