Micro Enterprise Exposure and Bank
Profitability During the COVID Period

library(tidyverse)
library(fixest)
library(MatchIt)

load("brazilian_banks_201403_onwards.rda")

# ============================================================
# Data preparation function
# ------------------------------------------------------------
# Purpose:
#   Create a clean panel dataset for the corporate lending analysis.
#
# Main steps:
#   1. Construct Micro exposure variables
#   2. Apply basic sample restrictions
#   3. Drop observations with missing values in key variables
#   4. Create transformed balance-sheet and profitability measures
#   5. Define treatment status using micro exposure in a pre-treatment quarter
#   6. Return the final estimation dataset
# ============================================================

make_data <- function(
  pre_q_char,                         # e.g. "2019-12-31"
  data_raw = brazilian_banks_201403_onwards,
  EMP_COL = NULL
) {

  # ----------------------------------------------------------
  # 0. Convert the reference quarter to Date format
  #    and set the last allowed quarter in the sample
  # ----------------------------------------------------------
  pre_q  <- as.Date(pre_q_char)
  last_q <- as.Date("2022-06-30")

  # ----------------------------------------------------------
  # 1. Construct lending variables and apply initial filtering
  # ----------------------------------------------------------
  data <- data_raw %>%
    dplyr::mutate(
      micro = Credit_portfolio_of_micro_sized_borrower,
      small = Credit_portfolio_of_small_sized_borrower,
      med   = Credit_portfolio_of_medium_sized_borrower,
      large = Credit_portfolio_of_large_sized_borrower
    ) %>%
    dplyr::filter(
      !is.na(Legal_Person_Loans_Total),
      Legal_Person_Loans_Total != 0,
      Quarter <= last_q
    ) %>%
    dplyr::mutate(
      num = dplyr::coalesce(micro, 0),
      denom = dplyr::coalesce(micro, 0) +
                  dplyr::coalesce(small, 0) +
                  dplyr::coalesce(med, 0) +
                  dplyr::coalesce(large, 0),
      micro_exposure = num / Legal_Person_Loans_Total
    )

  # ----------------------------------------------------------
  # 2. Keep only observations with non-missing values in the
  #    variables required for the empirical analysis
  # ----------------------------------------------------------
  needed_vars <- c(
    "FinInst",
    "Quarter",
    "Net_Income_qtr",
    "Total_Assets",
    "Equity",
    "Total_Deposits",
    "Legal_Person_Loans_Total",
    "Income_Statement__Other_Operating_Income_and_Expenses__Administrative_Expenses_qtr",
    "TD",
    "Headquarters___State",
    "Segment",
    "micro_exposure"
  )

  if (!is.null(EMP_COL)) {
    needed_vars <- c(needed_vars, EMP_COL)
  }

  needed_vars <- intersect(needed_vars, names(data))

  data_clean <- data %>%
    dplyr::filter(
      dplyr::if_all(dplyr::all_of(needed_vars), ~ !is.na(.))
    )

  # ----------------------------------------------------------
  # 3. Create transformed variables used in the regressions
  # ----------------------------------------------------------
  df_final <- data_clean %>%
    dplyr::mutate(
      log_assets      = log(Total_Assets),
      eq_assets       = Equity / Total_Assets,
      dep_assets      = Total_Deposits / Total_Assets,
      loans_assets    = Legal_Person_Loans_Total / Total_Assets,
      adminexp_assets = -Income_Statement__Other_Operating_Income_and_Expenses__Administrative_Expenses_qtr / Total_Assets,
      state           = as.factor(Headquarters___State),
      segment         = as.factor(Segment),
      TD              = as.factor(TD),
      roa             = 400 * Net_Income_qtr / Total_Assets
    ) %>%
    dplyr::arrange(Quarter, FinInst)

  # ----------------------------------------------------------
  # 4. Define treatment status based on micro exposure in the
  #    chosen pre-treatment quarter
  #
  #    Control group: micro exposure < 5%
  #    Treated group: micro exposure >= 50%
  #    Middle group: dropped
  # ----------------------------------------------------------
  df_pre <- df_final %>%
    dplyr::filter(Quarter == pre_q) %>%
    dplyr::mutate(
      treatment = dplyr::case_when(
        micro_exposure < 0.05 ~ 0L,
        micro_exposure >= 0.50 ~ 1L,
        TRUE ~ NA_integer_
      )
    )

  selected_banks <- df_pre %>%
    dplyr::filter(!is.na(treatment)) %>%
    dplyr::distinct(FinInst, treatment)

  # ----------------------------------------------------------
  # 5. Merge treatment labels back to the full panel and keep
  #    only the variables needed for the final dataset
  # ----------------------------------------------------------
  df_out <- df_final %>%
    dplyr::inner_join(selected_banks, by = "FinInst") %>%
    dplyr::select(
      FinInst,
      Quarter,
      treatment,
      log_assets,
      eq_assets,
      dep_assets,
      loans_assets,
      adminexp_assets,
      roa,
      state,
      segment,
      TD,
      micro_exposure,
      dplyr::any_of(EMP_COL)
    ) %>%
    dplyr::arrange(Quarter, FinInst)

  return(df_out)
}

# Create the final dataset using 2019 Q4
df_out <- make_data(pre_q_char = "2019-12-31")

df_out %>%
  group_by(Quarter, treatment) %>%
  summarise(n = n(), .groups = "drop") %>%
  tidyr::pivot_wider(
    names_from = treatment,
    values_from = n,
    names_prefix = "treat_"
  )

plot_means <- function(
  data,
  varname,
  shock_date = as.Date("2019-12-31"),
  pretty_name = NULL
) {
  
  if (!("Quarter" %in% names(data))) stop("The dataset must contain a 'Quarter' column.")
  if (!("treatment" %in% names(data))) stop("The dataset must contain a 'treatment' column.")
  if (!(varname %in% names(data))) stop(paste0("Variable '", varname, "' not found."))
  
  if (is.null(pretty_name)) pretty_name <- varname
  
  df_plot <- data %>%
    dplyr::group_by(Quarter, treatment) %>%
    dplyr::summarise(
      mean_value = mean(.data[[varname]], na.rm = TRUE),
      .groups = "drop"
    ) %>%
    dplyr::mutate(
      treatment = factor(treatment, levels = c(0, 1), labels = c("Control", "Treated"))
    )
  
  p <- ggplot2::ggplot(
    df_plot,
    ggplot2::aes(
      x = Quarter,
      y = mean_value,
      color = treatment,
      group = treatment
    )
  ) +
    ggplot2::geom_line(linewidth = 0.9) +
    ggplot2::geom_point(size = 1.8) +
    ggplot2::geom_vline(
      xintercept = shock_date,
      linetype = "dashed",
      color = "black",
      linewidth = 0.6
    ) +
    ggplot2::scale_x_date(
      date_breaks = "1 year",
      date_labels = "%Y"
    ) +
    ggplot2::labs(
      title = pretty_name,
      x = NULL,
      y = NULL,
      color = NULL
    ) +
    ggplot2::theme_minimal(base_size = 12) +
    ggplot2::theme(
      plot.title = ggplot2::element_text(face = "bold", hjust = 0.5),
      axis.text.x = ggplot2::element_text(angle = 45, hjust = 1),
      legend.position = "bottom",
      panel.grid.minor = ggplot2::element_blank()
    )
  
  return(p)
}

library(patchwork)

vars <- c(
  "roa",
  "log_assets",
  "eq_assets",
  "dep_assets",
  "loans_assets",
  "adminexp_assets"
)

plots <- lapply(vars, function(v) {
  plot_means(df_out, v, pretty_name = v)
})

combined_plot <- wrap_plots(plots, ncol = 2) +
  plot_layout(guides = "collect") &
  theme(legend.position = "bottom")

combined_plot

ggplot2::ggsave(
  "bank_variables_over_time.png",
  plot = combined_plot,
  width = 14,
  height = 12,
  dpi = 300
)

# ============================================================
# Construct a pre–post dataset
# ------------------------------------------------------------
# The function extracts bank characteristics in a pre-period
# and combines them with profitability in a later post-period.
# It then computes the change in ROA between the two periods.
# ============================================================

make_pre_post_df <- function(df, pre_q_char, post_q_char) {

  pre_q  <- as.Date(pre_q_char)
  post_q <- as.Date(post_q_char)

  # --- 1. Data from the pre-treatment quarter ---
  df_pre <- df %>%
    dplyr::filter(Quarter == pre_q)

  # --- 2. ROA from the post-treatment quarter ---
  df_post <- df %>%
    dplyr::filter(Quarter == post_q) %>%
    dplyr::select(FinInst, roa_post = roa)

  # --- 3. Merge the two datasets by bank ---
  df_out <- df_pre %>%
    dplyr::inner_join(df_post, by = "FinInst") %>%
    dplyr::mutate(
      roa_diff = roa_post - roa   # change in ROA between periods
    )

  return(df_out)
}

# Create the pre–post dataset
pre_post_df <- make_pre_post_df(
  df_out,
  pre_q_char  = "2019-12-31",
  post_q_char = "2021-09-30"
)

# ============================================================
# Collapse small categorical groups
# ------------------------------------------------------------
# Categories with very few observations are grouped into
# an "Other" category to avoid extremely small cells.
# ============================================================

collapse_categories <- function(x, min_n = 5) {
  tab <- table(x)
  valid_levels <- names(tab)[tab >= min_n]

  x_new <- ifelse(x %in% valid_levels, as.character(x), "Other")
  factor(x_new)
}

# Apply the collapsing to categorical variables
pre_post_df <- pre_post_df %>%
  dplyr::mutate(
    state   = collapse_categories(state,   min_n = 5),
    segment = collapse_categories(segment, min_n = 5)
  )

did_simple <- lm(roa_diff  ~ treatment, data = pre_post_df)
summary(did_simple)

did_simple <- lm(roa_diff  ~ treatment + segment + log_assets + eq_assets + TD + loans_assets + dep_assets + state, data = pre_post_df)
summary(did_simple)

library(cobalt)

psm_seg <- matchit(
  treatment ~ log_assets + eq_assets + dep_assets +
    loans_assets + adminexp_assets,
  data   = pre_post_df,
  method = "nearest",
  ratio  = 5,
  caliper = 0.05,
  distance = "logit"
)

summary(psm_seg2)$sum.matched

love.plot(psm_seg, binary = "std")

matched_df <- match.data(psm_seg)

model_ate <- lm(roa_diff ~ treatment, data = matched_df, weights = weights)

summary(model_ate)

model_ate <- lm(roa_diff ~ treatment + log_assets + eq_assets + loans_assets + dep_assets, data = matched_df, weights = weights)

summary(model_ate)

plot_matched_quarterly_means <- function(
  df_out,
  matched_df,
  varname,
  shock_date = as.Date("2019-12-31"),
  pretty_name = NULL
) {
  # ----------------------------------------------------------
  # 0. Basic checks
  # ----------------------------------------------------------
  if (!(varname %in% names(df_out))) {
    stop(paste("Variable", varname, "not found in df_out."))
  }
  if (!("weights" %in% names(matched_df))) {
    stop("matched_df must contain a 'weights' column.")
  }
  if (!("FinInst" %in% names(matched_df))) {
    stop("matched_df must contain a 'FinInst' column.")
  }
  if (!("treatment" %in% names(matched_df))) {
    stop("matched_df must contain a 'treatment' column.")
  }
  if (!("Quarter" %in% names(df_out))) {
    stop("df_out must contain a 'Quarter' column.")
  }
  if (!("treatment" %in% names(df_out))) {
    stop("df_out must contain a 'treatment' column.")
  }

  if (is.null(pretty_name)) pretty_name <- varname

  # ----------------------------------------------------------
  # 1. Keep only matched banks
  # ----------------------------------------------------------
  matched_banks <- unique(matched_df$FinInst)

  # ----------------------------------------------------------
  # 2. Extract treatment and matching weights
  # ----------------------------------------------------------
  weights_info <- matched_df %>%
    dplyr::select(FinInst, treatment, weights)

  # ----------------------------------------------------------
  # 3. Rebuild the full panel for matched banks only
  # ----------------------------------------------------------
  df_panel <- df_out %>%
    dplyr::filter(FinInst %in% matched_banks) %>%
    dplyr::left_join(weights_info, by = c("FinInst", "treatment"))

  # ----------------------------------------------------------
  # 4. Compute weighted quarterly means
  # ----------------------------------------------------------
  df_plot <- df_panel %>%
    dplyr::group_by(Quarter, treatment) %>%
    dplyr::summarise(
      weighted_mean = weighted.mean(.data[[varname]], weights, na.rm = TRUE),
      .groups = "drop"
    ) %>%
    dplyr::mutate(
      treatment = factor(treatment, levels = c(0, 1), labels = c("Control", "Treated"))
    )

  # ----------------------------------------------------------
  # 5. Plot
  # ----------------------------------------------------------
  p <- ggplot2::ggplot(
    df_plot,
    ggplot2::aes(
      x = Quarter,
      y = weighted_mean,
      color = treatment,
      group = treatment
    )
  ) +
    ggplot2::geom_line(linewidth = 0.9) +
    ggplot2::geom_point(size = 1.8) +
    ggplot2::geom_vline(
      xintercept = shock_date,
      linetype = "dashed",
      color = "black",
      linewidth = 0.6
    ) +
    ggplot2::scale_x_date(
      date_breaks = "1 year",
      date_labels = "%Y"
    ) +
    ggplot2::labs(
      title = pretty_name,
      x = NULL,
      y = NULL,
      color = NULL
    ) +
    ggplot2::theme_minimal(base_size = 12) +
    ggplot2::theme(
      plot.title = ggplot2::element_text(face = "bold", hjust = 0.5),
      axis.text.x = ggplot2::element_text(angle = 45, hjust = 1),
      legend.position = "bottom",
      panel.grid.minor = ggplot2::element_blank()
    )

  return(p)
}

library(patchwork)

pretty_names <- c(
  roa = "ROA",
  log_assets = "Log assets",
  eq_assets = "Equity / Assets",
  dep_assets = "Deposits / Assets",
  loans_assets = "Loans / Assets",
  adminexp_assets = "Administrative expenses / Assets"
)

vars <- names(pretty_names)

plots <- lapply(vars, function(v) {
  plot_matched_quarterly_means(
    df_out = df_out,
    matched_df = matched_df,
    varname = v,
    pretty_name = pretty_names[[v]]
  )
})

combined_matched_plot <- wrap_plots(plots, ncol = 2) +
  patchwork::plot_layout(guides = "collect") &
  ggplot2::theme(legend.position = "bottom")

combined_matched_plot

df_out <- df_out %>%
  dplyr::mutate(
    post = ifelse(Quarter > as.Date("2019-12-31"), 1L, 0L)
  )

fe_model1 <- feols(
  roa ~ treatment * post + log_assets + eq_assets  + loans_assets + dep_assets + state + TD + segment  | FinInst + Quarter,
  data = df_out
)
summary(fe_model1)

fe_model2 <- feols(
  roa ~ treatment * post | FinInst + Quarter,
  data = df_out
)
summary(fe_model2)

fe_model3 <- feols(
  roa ~ treatment * post | FinInst + Quarter,
  data = df_panel
)
summary(fe_model3)

df_panel <- df_out %>%
  dplyr::filter(FinInst %in% matched_banks) %>%
  dplyr::left_join(weights_info, by = c("FinInst", "treatment"))

fe_model4 <- feols(
  roa ~ treatment * post + log_assets + eq_assets  + loans_assets + dep_assets + TD + state  | FinInst + Quarter,
  data = df_panel
)
summary(fe_model4)