Data

# === SAMPLE FILTERS (adjust here) ===
MIN_CAPACITY_MCM <- 1       # Set to NA to skip capacity filter
REQUIRE_CAPACITY <- TRUE   # Set to TRUE to drop dams without capacity_mcm
analysis_data <- readRDS(file.path(strategic_dir,
                                    "strategic_placement_analysis_data.rds"))

# Apply sample filters
n_full <- nrow(analysis_data)
if (REQUIRE_CAPACITY) {
  analysis_data <- analysis_data %>% filter(!is.na(capacity_mcm))
}
if (!is.na(MIN_CAPACITY_MCM)) {
  analysis_data <- analysis_data %>%
    filter(is.na(capacity_mcm) | capacity_mcm >= MIN_CAPACITY_MCM)
}

# Ensure decade variable exists
analysis_data <- analysis_data %>%
  mutate(decade = floor(as.numeric(year_fin) / 10) * 10)

# Basin grouping for FE
analysis_data <- analysis_data %>%
  mutate(
    basin_id = as.character(main_riv),
    basin_group = ifelse(
      dam_id %in% (analysis_data %>%
                    group_by(main_riv) %>%
                    filter(n() >= 5) %>%
                    pull(dam_id)),
      as.character(main_riv),
      "Other basins"
    )
  )

# Check which version columns are available
has_v1 <- "suitability_ratio_v1" %in% names(analysis_data)
has_v3 <- "suitability_ratio_v3" %in% names(analysis_data)

# Check if constrained (same-country) columns are available
has_constrained <- "suitability_ratio_c" %in% names(analysis_data) &&
  sum(!is.na(analysis_data$suitability_ratio_c)) > 30
has_constrained_v1 <- has_constrained &&
  "suitability_ratio_c_v1" %in% names(analysis_data)
has_constrained_v3 <- has_constrained &&
  "suitability_ratio_c_v3" %in% names(analysis_data)

# Transboundary subset with along-river border distance
tb_data <- analysis_data %>%
  filter(transboundary == 1, !is.na(dam_dist_border))

# Hydro/irrigation-only subsample
hydro_irrig_data <- analysis_data %>%
  filter(hydropower == 1 | irrigation == 1)

# Version labels
versions <- c("v1", "v2", "v3")[c(has_v1, TRUE, has_v3)]

1. Sample Overview

This section describes the analysis sample after filtering from the Global Dam Tracker (GDAT) universe. Dams are retained if they can be matched to a HydroRIVERS segment, have a recorded completion year between 1950–2025, and have a known primary purpose. The suitability computation considers all candidate river segments within a +/-200 km buffer along the river network, scoring each on hydropower potential (flow, catchment area, hydraulic head) and population exposure at the time of construction.

tibble(
  Statistic = c("Total dams", "Transboundary", "Domestic",
                "TB with along-river border distance",
                "Hydropower dams", "Irrigation dams",
                "Hydro or Irrigation subsample",
                "Mean candidates per dam",
                "Median candidates per dam"),
  Value = c(
    nrow(analysis_data),
    sum(analysis_data$transboundary),
    sum(1 - analysis_data$transboundary),
    nrow(tb_data),
    sum(analysis_data$hydropower),
    sum(analysis_data$irrigation),
    nrow(hydro_irrig_data),
    sprintf("%.0f", mean(analysis_data$n_candidates, na.rm = TRUE)),
    sprintf("%.0f", median(analysis_data$n_candidates, na.rm = TRUE))
  )
) %>%
  kbl(caption = "Sample Summary") %>%
  kable_styling(bootstrap_options = c("striped", "hover"),
                full_width = FALSE)
Sample Summary
Statistic Value
Total dams 4901
Transboundary 2634
Domestic 2267
TB with along-river border distance 745
Hydropower dams 732
Irrigation dams 1054
Hydro or Irrigation subsample 1786
Mean candidates per dam 156
Median candidates per dam 59 
analysis_data %>%
  count(purpose_category, transboundary) %>%
  mutate(transboundary = ifelse(transboundary == 1,
                                 "Transboundary", "Domestic")) %>%
  pivot_wider(names_from = transboundary, values_from = n, values_fill = 0) %>%
  mutate(Total = Domestic + Transboundary) %>%
  arrange(desc(Total)) %>%
  kbl(caption = "Dams by Purpose and Transboundary Status") %>%
  kable_styling(bootstrap_options = c("striped", "hover"),
                full_width = FALSE)
Dams by Purpose and Transboundary Status
purpose_category Domestic Transboundary Total
Other 692 688 1380
Flood Control 489 656 1145
Irrigation 456 598 1054
Hydropower 336 396 732
Water Supply 294 296 590 
analysis_data %>%
  count(decade, transboundary) %>%
  mutate(transboundary = ifelse(transboundary == 1,
                                 "Transboundary", "Domestic")) %>%
  pivot_wider(names_from = transboundary, values_from = n, values_fill = 0) %>%
  mutate(Total = Domestic + Transboundary) %>%
  kbl(caption = "Dams by Decade") %>%
  kable_styling(bootstrap_options = c("striped", "hover"),
                full_width = FALSE)
Dams by Decade
decade Domestic Transboundary Total
1950 363 402 765
1960 671 705 1376
1970 538 651 1189
1980 383 436 819
1990 220 220 440
2000 67 131 198
2010 24 63 87
2020 1 26 27

2. Suitability Ratios Across Specifications

I compute three alternative suitability indices to ensure the results are not driven by a particular functional form:

  • V1 (Baseline): Raw product of discharge, upstream catchment area, and hydraulic head. Heavily right-skewed because a few candidate sites with large rivers dominate the score.
  • V2 (Gradient-weighted): Applies square-root transforms to flow and catchment area, giving less weight to extreme values. This is the primary specification.
  • V3 (Percentile Ranks): Converts each component to its within-river-system percentile rank before combining. This ensures each component contributes roughly equally and yields ratios on a more interpretable 0–1 scale.

The suitability ratio divides a dam’s actual site score by the best available score within +/-200 km along the river network. A ratio of 1 means the dam sits at the optimal site; lower values indicate the dam was placed at a less suitable location relative to nearby alternatives.

suit_summary <- analysis_data %>%
  mutate(tb_label = ifelse(transboundary == 1,
                            "Transboundary", "Domestic")) %>%
  group_by(tb_label) %>%
  summarise(
    n = n(),
    across(
      all_of(paste0("suitability_ratio_", versions)),
      list(mean = ~mean(.x, na.rm = TRUE),
           median = ~median(.x, na.rm = TRUE)),
      .names = "{.col}_{.fn}"
    ),
    .groups = "drop"
  )

suit_summary %>%
  kbl(digits = 3,
      caption = "Suitability Ratio by Specification and TB Status") %>%
  kable_styling(bootstrap_options = c("striped", "hover"),
                full_width = FALSE) %>%
  add_header_above(c(" " = 2,
                      "V1 (Baseline)" = 2,
                      "V2 (Gradient)" = 2,
                      "V3 (Percentile)" = 2))
Suitability Ratio by Specification and TB Status
V1 (Baseline)
V2 (Gradient)
V3 (Percentile)
tb_label n suitability_ratio_v1_mean suitability_ratio_v1_median suitability_ratio_v2_mean suitability_ratio_v2_median suitability_ratio_v3_mean suitability_ratio_v3_median
Domestic 2267 0.146 0.009 0.241 0.111 0.696 0.706
Transboundary 2634 0.080 0.002 0.140 0.057 0.647 0.655

Quick t-tests (no controls)

trb <- analysis_data %>% filter(transboundary == 1)
dom <- analysis_data %>% filter(transboundary == 0)

for (v in versions) {
  col <- paste0("suitability_ratio_", v)
  tt <- t.test(trb[[col]], dom[[col]])
  cat(sprintf(
    "**%s**: TRB mean = %.3f, DOM mean = %.3f, diff = %.3f (p = %.4f)  \n",
    toupper(v), tt$estimate[1], tt$estimate[2],
    tt$estimate[1] - tt$estimate[2], tt$p.value
  ))
}

V1: TRB mean = 0.080, DOM mean = 0.146, diff = -0.066 (p = 0.0000)
V2: TRB mean = 0.140, DOM mean = 0.241, diff = -0.101 (p = 0.0000)
V3: TRB mean = 0.647, DOM mean = 0.696, diff = -0.049 (p = 0.0000)

Across all three specifications, transboundary dams have lower mean suitability ratios than domestic dams. The gap is largest for V1 and V2, which are more sensitive to extreme suitability scores on large rivers. V3 compresses the distribution by construction (percentile ranks are bounded between 0 and 1), so the raw gap is smaller, but it remains significant. This pattern is consistent with the hypothesis that transboundary dams face political or strategic constraints that lead to placement at sites that are sub-optimal from a purely engineering perspective.

3. Suitability Ratio Density Plots

The density plots below visualize the full distribution of suitability ratios for domestic and transboundary dams. Key patterns to look for: (1) whether the transboundary distribution is shifted leftward (toward worse sites), and (2) whether the shapes differ—for instance, a heavier left tail for transboundary dams would suggest a subset of particularly poorly-sited dams.

plot_density <- function(data, col, vlabel) {
  ggplot(data, aes(x = .data[[col]],
                    fill = factor(transboundary))) +
    geom_density(alpha = 0.6) +
    geom_vline(xintercept = 1, linetype = "dashed", color = "black") +
    scale_fill_manual(
      values = c("0" = "steelblue", "1" = "coral"),
      labels = c("Domestic", "Transboundary"),
      name = ""
    ) +
    labs(
      title = paste("Distribution of Suitability Ratios -", vlabel),
      subtitle = "Ratio of 1 = dam at optimal location; lower = worse site quality",
      x = "Suitability Ratio (Actual / Optimal)",
      y = "Density"
    ) +
    theme_minimal(base_size = 13) +
    theme(legend.position = "bottom")
}

V3 (Percentile Ranks)

plot_density(analysis_data, "suitability_ratio_v3",
             "V3 (Percentile Ranks)")

V2 (Gradient-weighted)

plot_density(analysis_data, "suitability_ratio_v2",
             "V2 (Gradient-weighted)")

V1 (Baseline)

plot_density(analysis_data, "suitability_ratio_v1",
             "V1 (Baseline)")

4. Suitability by Dam Purpose

Different dam purposes may have different site selection criteria. Hydropower dams, for example, are strongly constrained by topography and river flow, so siting distortions may be more visible. Irrigation dams may have more flexibility in placement. The bar charts below show mean suitability ratios by purpose category, separately for domestic and transboundary dams, with 95% confidence intervals.

plot_purpose <- function(data, col, vlabel) {
  pdata <- data %>%
    group_by(transboundary, purpose_category) %>%
    summarise(
      mean_suit = mean(.data[[col]], na.rm = TRUE),
      se_suit = sd(.data[[col]], na.rm = TRUE) / sqrt(n()),
      n = n(),
      .groups = "drop"
    ) %>%
    filter(n >= 5) %>%
    mutate(tb_label = ifelse(transboundary == 1,
                              "Transboundary", "Domestic"))

  ggplot(pdata, aes(x = purpose_category, y = mean_suit,
                     fill = tb_label)) +
    geom_col(position = position_dodge(width = 0.8), width = 0.7) +
    geom_errorbar(
      aes(ymin = mean_suit - 1.96 * se_suit,
          ymax = mean_suit + 1.96 * se_suit),
      position = position_dodge(width = 0.8), width = 0.3
    ) +
    scale_fill_manual(values = c("Domestic" = "steelblue",
                                  "Transboundary" = "coral")) +
    labs(
      title = paste("Suitability by Purpose -", vlabel),
      subtitle = "Error bars show 95% confidence intervals",
      x = "Dam Purpose", y = "Mean Suitability Ratio", fill = ""
    ) +
    theme_minimal(base_size = 13) +
    theme(legend.position = "bottom",
          axis.text.x = element_text(angle = 45, hjust = 1))
}

V3 (Percentile Ranks)

plot_purpose(analysis_data, "suitability_ratio_v3",
             "V3 (Percentile Ranks)")

V2 (Gradient-weighted)

plot_purpose(analysis_data, "suitability_ratio_v2",
             "V2 (Gradient-weighted)")

V1 (Baseline)

plot_purpose(analysis_data, "suitability_ratio_v1",
             "V1 (Baseline)")

5. Main Regression Results (V2 Baseline)

I present a sequence of increasingly demanding specifications using V2 (gradient-weighted suitability) as the baseline outcome. The progression from simple OLS to country fixed effects allows me to assess how much of the transboundary gap is explained by observable controls, geographic sorting across continents, and country-level heterogeneity. Controls include log reservoir capacity, and indicators for hydropower and irrigation as primary purposes.

model1 <- feols(
  suitability_ratio ~ transboundary,
  data = analysis_data
)

model2 <- feols(
  suitability_ratio ~ transboundary + log_capacity + hydropower + irrigation,
  data = analysis_data
)

model3 <- feols(
  suitability_ratio ~ transboundary + log_capacity + hydropower + irrigation |
    continent + decade,
  data = analysis_data,
  cluster = ~continent
)

model4 <- feols(
  suitability_ratio ~ transboundary + log_capacity + hydropower + irrigation |
    admin0 + decade,
  data = analysis_data,
  cluster = ~admin0
)
etable(model1, model2, model3, model4,
       headers = c("Simple", "Controls", "Continent FE", "Country FE"),
       title = "Effect of Transboundary Status on Suitability Ratio (V2)")
##                              model1              model2             model3
##                              Simple            Controls       Continent FE
## Dependent Var.:   suitability_ratio   suitability_ratio  suitability_ratio
##                                                                           
## Constant         0.2408*** (0.0050)  0.1152*** (0.0075)                   
## transboundary   -0.1013*** (0.0068) -0.0910*** (0.0065) -0.0794** (0.0144)
## log_capacity                         0.0312*** (0.0017)  0.0318** (0.0055)
## hydropower                           0.0561*** (0.0101)   0.0574. (0.0244)
## irrigation                              0.0121 (0.0081)   -0.0127 (0.0332)
## Fixed-Effects:  ------------------- ------------------- ------------------
## continent                        No                  No                Yes
## decade                           No                  No                Yes
## admin0                           No                  No                 No
## _______________ ___________________ ___________________ __________________
## S.E. type                       IID                 IID      by: continent
## Observations                  4,901               4,901              4,901
## R2                          0.04288             0.14023            0.18403
## Within R2                        --                  --            0.12103
## 
##                              model4
##                          Country FE
## Dependent Var.:   suitability_ratio
##                                    
## Constant                           
## transboundary   -0.0655*** (0.0074)
## log_capacity     0.0307*** (0.0027)
## hydropower          0.0369 (0.0258)
## irrigation         -0.0285 (0.0217)
## Fixed-Effects:  -------------------
## continent                        No
## decade                          Yes
## admin0                          Yes
## _______________ ___________________
## S.E. type                by: admin0
## Observations                  4,901
## R2                          0.26802
## Within R2                   0.09185
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
b1 <- coef(model1)["transboundary"]
se1 <- se(model1)["transboundary"]
b2 <- coef(model2)["transboundary"]
b3 <- coef(model3)["transboundary"]
p3 <- pvalue(model3)["transboundary"]
b4 <- coef(model4)["transboundary"]
p4 <- pvalue(model4)["transboundary"]

b_cap3 <- coef(model3)["log_capacity"]
b_hydro3 <- coef(model3)["hydropower"]

cat(sprintf(
  "The transboundary coefficient is negative and highly significant across all specifications. The simple OLS estimate (%.3f, p < 0.001) shrinks modestly with controls (%.3f) and further with continent and decade fixed effects (%.3f, p = %.3f), but remains significant even with country fixed effects (%.3f, p = %.3f). This last specification compares transboundary and domestic dams *within the same country*, controlling for any country-level differences in engineering capacity, regulatory environment, or river quality. The fact that the penalty persists at the within-country level suggests it is not driven by transboundary dams being concentrated in countries with generally worse dam sites.\n\n",
  b1, b2, b3, p3, b4, p4
))

The transboundary coefficient is negative and highly significant across all specifications. The simple OLS estimate (-0.101, p < 0.001) shrinks modestly with controls (-0.091) and further with continent and decade fixed effects (-0.079, p = 0.005), but remains significant even with country fixed effects (-0.065, p = 0.000). This last specification compares transboundary and domestic dams within the same country, controlling for any country-level differences in engineering capacity, regulatory environment, or river quality. The fact that the penalty persists at the within-country level suggests it is not driven by transboundary dams being concentrated in countries with generally worse dam sites.

cat(sprintf(
  "The control variables behave as expected: larger dams (higher log capacity) are placed at better sites (%.3f, p < 0.01 across specifications), consistent with larger projects receiving more careful siting. Hydropower dams tend to be at somewhat better sites (%.3f in the continent FE model), consistent with stronger engineering constraints on hydropower siting. The irrigation indicator is small and insignificant.\n",
  b_cap3, b_hydro3
))

The control variables behave as expected: larger dams (higher log capacity) are placed at better sites (0.032, p < 0.01 across specifications), consistent with larger projects receiving more careful siting. Hydropower dams tend to be at somewhat better sites (0.057 in the continent FE model), consistent with stronger engineering constraints on hydropower siting. The irrigation indicator is small and insignificant.

6. Heterogeneity by Dam Purpose

I explore whether the transboundary penalty varies by dam purpose. Hydropower dams are most constrained by physical geography (they need specific combinations of flow and head), so any political distortion of siting should be most visible for these dams. Irrigation dams may have more locational flexibility, potentially diffusing the signal.

Full Sample

hydro_data <- analysis_data %>% filter(hydropower == 1)
irrig_data <- analysis_data %>% filter(irrigation == 1)

model_hydro <- feols(
  suitability_ratio ~ transboundary + log_capacity | continent + decade,
  data = hydro_data, cluster = ~continent
)

model_irrig <- feols(
  suitability_ratio ~ transboundary + log_capacity | continent + decade,
  data = irrig_data, cluster = ~continent
)

model_purpose_interact <- feols(
  suitability_ratio ~ transboundary * hydropower +
    transboundary * irrigation + log_capacity | continent + decade,
  data = analysis_data, cluster = ~continent
)

Table 2: Full Sample

etable(model_hydro, model_irrig, model_purpose_interact,
       headers = c("Hydropower Only", "Irrigation Only", "Full Sample Interactions"),
       title = "Effects by Dam Purpose (Full Sample)")
##                                   model_hydro       model_irrig
##                               Hydropower Only   Irrigation Only
## Dependent Var.:             suitability_ratio suitability_ratio
##                                                                
## transboundary               -0.0485* (0.0145)  -0.0670 (0.0527)
## log_capacity               0.0219*** (0.0012)  0.0415. (0.0113)
## hydropower                                                     
## irrigation                                                     
## transboundary x hydropower                                     
## transboundary x irrigation                                     
## Fixed-Effects:             ------------------ -----------------
## continent                                 Yes               Yes
## decade                                    Yes               Yes
## __________________________ __________________ _________________
## S.E.: Clustered                 by: continent     by: continent
## Observations                              732             1,054
## R2                                    0.08431           0.19414
## Within R2                             0.03821           0.13432
## 
##                                   model_purpose_i..
##                            Full Sample Interactions
## Dependent Var.:                   suitability_ratio
##                                                    
## transboundary                     -0.0987* (0.0278)
## log_capacity                      0.0316** (0.0056)
## hydropower                          0.0316 (0.0295)
## irrigation                         -0.0417 (0.0618)
## transboundary x hydropower         0.0531. (0.0216)
## transboundary x irrigation          0.0541 (0.0660)
## Fixed-Effects:                    -----------------
## continent                                       Yes
## decade                                          Yes
## __________________________        _________________
## S.E.: Clustered                       by: continent
## Observations                                  4,901
## R2                                          0.18672
## Within R2                                   0.12392
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
b_hydro <- coef(model_hydro)["transboundary"]
p_hydro <- pvalue(model_hydro)["transboundary"]
se_hydro <- se(model_hydro)["transboundary"]
n_hydro <- nobs(model_hydro)
b_irrig <- coef(model_irrig)["transboundary"]
p_irrig <- pvalue(model_irrig)["transboundary"]
se_irrig <- se(model_irrig)["transboundary"]
n_irrig <- nobs(model_irrig)
b_interact_base <- coef(model_purpose_interact)["transboundary"]
p_interact_base <- pvalue(model_purpose_interact)["transboundary"]
b_interact_hydro <- coef(model_purpose_interact)["transboundary:hydropower"]
p_interact_hydro <- pvalue(model_purpose_interact)["transboundary:hydropower"]
b_interact_irrig <- coef(model_purpose_interact)["transboundary:irrigation"]
p_interact_irrig <- pvalue(model_purpose_interact)["transboundary:irrigation"]

cat(sprintf(
  "The split-sample results show that the transboundary penalty is present across dam types but with different precision. For hydropower dams alone (n = %d), the coefficient is %.3f (SE = %.3f, p = %.3f)---significant and negative. For irrigation dams (n = %d), the point estimate is actually larger in magnitude (%.3f) but not individually significant (SE = %.3f, p = %.3f), likely reflecting the small number of continent clusters rather than absence of an effect.\n\n",
  n_hydro, b_hydro, se_hydro, p_hydro,
  n_irrig, b_irrig, se_irrig, p_irrig
))

The split-sample results show that the transboundary penalty is present across dam types but with different precision. For hydropower dams alone (n = 732), the coefficient is -0.048 (SE = 0.015, p = 0.029)—significant and negative. For irrigation dams (n = 1054), the point estimate is actually larger in magnitude (-0.067) but not individually significant (SE = 0.053, p = 0.332), likely reflecting the small number of continent clusters rather than absence of an effect.

net_hydro <- b_interact_base + b_interact_hydro
net_irrig <- b_interact_base + b_interact_irrig

cat(sprintf(
  "The full-sample interaction model reveals a base transboundary penalty of %.3f (p = %.3f) for dams with purposes other than hydropower or irrigation. Both `transboundary x hydropower` (+%.3f, p = %.3f) and `transboundary x irrigation` (+%.3f, p = %.3f) are positive, partially offsetting the base penalty. The net transboundary effect for hydropower dams is approximately %.3f and for irrigation dams %.3f. The hydropower interaction is marginally significant, while the irrigation interaction is not---but both point in the same direction: the transboundary siting distortion may be somewhat smaller for dams whose purpose imposes stronger physical-geographic constraints on location.\n",
  b_interact_base, p_interact_base,
  b_interact_hydro, p_interact_hydro,
  b_interact_irrig, p_interact_irrig,
  net_hydro, net_irrig
))

The full-sample interaction model reveals a base transboundary penalty of -0.099 (p = 0.024) for dams with purposes other than hydropower or irrigation. Both transboundary x hydropower (+0.053, p = 0.070) and transboundary x irrigation (+0.054, p = 0.458) are positive, partially offsetting the base penalty. The net transboundary effect for hydropower dams is approximately -0.046 and for irrigation dams -0.045. The hydropower interaction is marginally significant, while the irrigation interaction is not—but both point in the same direction: the transboundary siting distortion may be somewhat smaller for dams whose purpose imposes stronger physical-geographic constraints on location.

Hydro/Irrigation Subsample

To obtain a cleaner comparison between dam types, I restrict to dams whose primary purpose is either hydropower or irrigation (n = 1786). In this subsample, the hydropower indicator distinguishes the two types directly.

model_hi_base <- feols(
  suitability_ratio ~ transboundary + log_capacity + hydropower |
    continent + decade,
  data = hydro_irrig_data,
  cluster = ~continent
)

model_hi_interact <- feols(
  suitability_ratio ~ transboundary * hydropower + log_capacity |
    continent + decade,
  data = hydro_irrig_data,
  cluster = ~continent
)

Table 2b: Hydro/Irrigation Only

etable(model_hydro, model_irrig, model_hi_base, model_hi_interact,
       headers = c("Hydropower Only", "Irrigation Only",
                    "H/I Combined", "H/I Interaction"),
       title = "Effects by Dam Purpose (Hydro/Irrigation Sample)")
##                                   model_hydro       model_irrig
##                               Hydropower Only   Irrigation Only
## Dependent Var.:             suitability_ratio suitability_ratio
##                                                                
## transboundary               -0.0485* (0.0145)  -0.0670 (0.0527)
## log_capacity               0.0219*** (0.0012)  0.0415. (0.0113)
## hydropower                                                     
## transboundary x hydropower                                     
## Fixed-Effects:             ------------------ -----------------
## continent                                 Yes               Yes
## decade                                    Yes               Yes
## __________________________ __________________ _________________
## S.E.: Clustered                 by: continent     by: continent
## Observations                              732             1,054
## R2                                    0.08431           0.19414
## Within R2                             0.03821           0.13432
## 
##                                model_hi_base model_hi_interact
##                                 H/I Combined   H/I Interaction
## Dependent Var.:            suitability_ratio suitability_ratio
##                                                               
## transboundary               -0.0579 (0.0326)  -0.0615 (0.0466)
## log_capacity               0.0323** (0.0063) 0.0322** (0.0062)
## hydropower                  0.0744. (0.0290)   0.0693 (0.0434)
## transboundary x hydropower                     0.0093 (0.0437)
## Fixed-Effects:             ----------------- -----------------
## continent                                Yes               Yes
## decade                                   Yes               Yes
## __________________________ _________________ _________________
## S.E.: Clustered                by: continent     by: continent
## Observations                           1,786             1,786
## R2                                   0.16371           0.16377
## Within R2                            0.13655           0.13662
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
b_hi_base <- coef(model_hi_base)["transboundary"]
p_hi_base <- pvalue(model_hi_base)["transboundary"]

has_hi_interact_coef <- "transboundary:hydropower" %in% names(coef(model_hi_interact))
if (has_hi_interact_coef) {
  b_hi_interact <- coef(model_hi_interact)["transboundary:hydropower"]
  p_hi_interact <- pvalue(model_hi_interact)["transboundary:hydropower"]
  cat(sprintf(
    "In the hydro/irrigation subsample, the base transboundary effect is %.3f (p = %.3f). The interaction `transboundary x hydropower` is %.3f (p = %.3f), indicating whether the transboundary penalty differs between hydropower and irrigation dams within this cleaner comparison group.\n",
    b_hi_base, p_hi_base, b_hi_interact, p_hi_interact
  ))
} else {
  cat(sprintf(
    "In the hydro/irrigation subsample, the transboundary effect is %.3f (p = %.3f).\n",
    b_hi_base, p_hi_base
  ))
}

In the hydro/irrigation subsample, the base transboundary effect is -0.058 (p = 0.150). The interaction transboundary x hydropower is 0.009 (p = 0.842), indicating whether the transboundary penalty differs between hydropower and irrigation dams within this cleaner comparison group.

7. Position Deviation from Optimal

Position deviation measures the distance (in km along the river) between where a dam was actually built and where the suitability-maximizing site would have been. Positive values mean the dam was placed upstream of the optimal site; negative values mean downstream. If transboundary dams are systematically displaced from optimal locations—particularly toward upstream positions where the river is still within the dam-building country’s territory—this would suggest strategic positioning motivations.

plot_posdev <- function(data, col, vlabel) {
  ggplot(data, aes(x = .data[[col]],
                    fill = factor(transboundary))) +
    geom_density(alpha = 0.6) +
    geom_vline(xintercept = 0, linetype = "dashed", color = "black") +
    scale_fill_manual(
      values = c("0" = "steelblue", "1" = "coral"),
      labels = c("Domestic", "Transboundary"),
      name = ""
    ) +
    coord_cartesian(xlim = c(-200, 200)) +
    labs(
      title = paste("Position Deviation from Optimal -", vlabel),
      subtitle = "Positive = upstream of optimal; negative = downstream",
      x = "Position Deviation (km)", y = "Density"
    ) +
    theme_minimal(base_size = 13) +
    theme(legend.position = "bottom")
}

V3 (Percentile Ranks)

plot_posdev(analysis_data, "position_deviation_km_v3",
            "V3 (Percentile Ranks)")

V2 (Gradient-weighted)

plot_posdev(analysis_data, "position_deviation_km_v2",
            "V2 (Gradient-weighted)")

V1 (Baseline)

plot_posdev(analysis_data, "position_deviation_km_v1",
            "V1 (Baseline)")

Table 3: Regressions (Full Sample)

I test whether transboundary status predicts the direction and magnitude of displacement from the optimal site. The first two models use position deviation (in km) as a continuous outcome; the last two model the binary indicator for whether the dam moved downstream. Columns 2 and 4 add interactions with dam purpose to test whether hydropower and irrigation dams exhibit different displacement patterns from other dam types.

model_posdev <- feols(
  position_deviation_km ~ transboundary + log_capacity + hydropower + irrigation |
    continent + decade,
  data = analysis_data, cluster = ~continent
)

model_posdev_interact <- feols(
  position_deviation_km ~ transboundary * hydropower +
    transboundary * irrigation + log_capacity | continent + decade,
  data = analysis_data, cluster = ~continent
)

logit_downstream <- feglm(
  moved_downstream ~ transboundary + log_capacity + hydropower + irrigation |
    continent + decade,
  data = analysis_data,
  family = binomial(link = "logit")
)

logit_downstream_interact <- feglm(
  moved_downstream ~ transboundary * hydropower +
    transboundary * irrigation + log_capacity | continent + decade,
  data = analysis_data,
  family = binomial(link = "logit")
)
etable(model_posdev, model_posdev_interact,
       logit_downstream, logit_downstream_interact,
       headers = c("Deviation (km)", "Deviation (Interact)",
                    "Pr(Downstream)", "Pr(Downstream, Interact)"),
       title = "Position Deviation from Optimal (V2)")
##                                     model_posdev model_posdev_interact
##                                   Deviation (km)  Deviation (Interact)
## Dependent Var.:            position_deviation_km position_deviation_km
##                                                                       
## transboundary                     19.50* (4.340)        21.75* (4.828)
## log_capacity                      -2.993 (1.969)        -2.970 (2.029)
## hydropower                       -16.31* (4.484)        -13.58 (6.979)
## irrigation                        37.80* (11.32)         41.40 (22.43)
## transboundary x hydropower                              -5.648 (6.656)
## transboundary x irrigation                              -6.684 (21.58)
## Fixed-Effects:             --------------------- ---------------------
## continent                                    Yes                   Yes
## decade                                       Yes                   Yes
## __________________________ _____________________ _____________________
## Family                                       OLS                   OLS
## S.E.: Clustered                    by: continent         by: continent
## Observations                               4,901                 4,901
## Squared Cor.                             0.09775               0.09816
## Pseudo R2                                0.00901               0.00905
## BIC                                     55,608.6              55,623.3
## 
##                               logit_downstream       logit_downstream..
##                                 Pr(Downstream) Pr(Downstream, Interact)
## Dependent Var.:               moved_downstream         moved_downstream
##                                                                        
## transboundary                 -0.1171 (0.1720)         -0.2796 (0.1868)
## log_capacity                   0.1579 (0.1049)          0.1552 (0.1083)
## hydropower                  0.6165*** (0.0675)       0.5677*** (0.1710)
## irrigation                 -0.5836*** (0.1203)          -1.042 (0.7818)
## transboundary x hydropower                              0.1383 (0.3708)
## transboundary x irrigation                               0.8162 (1.214)
## Fixed-Effects:             -------------------       ------------------
## continent                                  Yes                      Yes
## decade                                     Yes                      Yes
## __________________________ ___________________       __________________
## Family                                   Logit                    Logit
## S.E.: Clustered                  by: continent            by: continent
## Observations                             4,901                    4,901
## Squared Cor.                           0.02608                  0.02702
## Pseudo R2                              0.06709                  0.06909
## BIC                                    1,839.4                  1,852.8
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
b_pos <- coef(model_posdev)["transboundary"]
p_pos <- pvalue(model_posdev)["transboundary"]
b_pos_hydro <- coef(model_posdev)["hydropower"]
p_pos_hydro <- pvalue(model_posdev)["hydropower"]
b_pos_irrig <- coef(model_posdev)["irrigation"]
p_pos_irrig <- pvalue(model_posdev)["irrigation"]

b_pos_int_base <- coef(model_posdev_interact)["transboundary"]
b_pos_int_hydro <- coef(model_posdev_interact)["transboundary:hydropower"]
p_pos_int_hydro <- pvalue(model_posdev_interact)["transboundary:hydropower"]
b_pos_int_irrig <- coef(model_posdev_interact)["transboundary:irrigation"]
p_pos_int_irrig <- pvalue(model_posdev_interact)["transboundary:irrigation"]

b_logit <- coef(logit_downstream)["transboundary"]
p_logit <- pvalue(logit_downstream)["transboundary"]
b_logit_hydro <- coef(logit_downstream)["hydropower"]
b_logit_irrig <- coef(logit_downstream)["irrigation"]

cat("**Position deviation (km):**\n\n")

Position deviation (km):

cat(sprintf(
  "Transboundary dams are placed approximately %.1f km **%s** of their optimal site (p = %.3f). %s",
  abs(b_pos), ifelse(b_pos > 0, "upstream", "downstream"), p_pos,
  ifelse(b_pos > 0,
    "This is the key result---transboundary dams are systematically shifted upstream, consistent with capturing water before it crosses the border. ",
    "")
))

Transboundary dams are placed approximately 19.5 km upstream of their optimal site (p = 0.011). This is the key result—transboundary dams are systematically shifted upstream, consistent with capturing water before it crosses the border.

cat(sprintf(
  "The control variables show sensible patterns: hydropower dams sit %.1f km downstream of optimal (p = %.3f), consistent with hydropower benefiting from cumulative flow further downstream. Irrigation dams sit %.1f km upstream of optimal (p = %.3f), consistent with diverting water at higher elevations for gravity-fed distribution.\n\n",
  abs(b_pos_hydro), p_pos_hydro,
  abs(b_pos_irrig), p_pos_irrig
))

The control variables show sensible patterns: hydropower dams sit 16.3 km downstream of optimal (p = 0.022), consistent with hydropower benefiting from cumulative flow further downstream. Irrigation dams sit 37.8 km upstream of optimal (p = 0.029), consistent with diverting water at higher elevations for gravity-fed distribution.

cat(sprintf(
  "**Interactions with purpose:** The interaction terms (`transboundary x hydropower` = %.1f, p = %.3f; `transboundary x irrigation` = %.1f, p = %.3f) are small and insignificant. This means the transboundary upstream shift is roughly the **same magnitude regardless of dam purpose**---it does not differ meaningfully between hydropower, irrigation, and other dams.\n\n",
  b_pos_int_hydro, p_pos_int_hydro,
  b_pos_int_irrig, p_pos_int_irrig
))

Interactions with purpose: The interaction terms (transboundary x hydropower = -5.6, p = 0.444; transboundary x irrigation = -6.7, p = 0.772) are small and insignificant. This means the transboundary upstream shift is roughly the same magnitude regardless of dam purpose—it does not differ meaningfully between hydropower, irrigation, and other dams.

cat(sprintf(
  "**Logit (Pr(Downstream)):** The transboundary coefficient is %.3f but not significant (p = %.3f), so there is no strong evidence that transboundary status shifts the binary probability of being downstream of optimal. The purpose controls are highly significant: hydropower dams are more likely to move downstream (%.3f, p < 0.001) while irrigation dams are less likely (%.3f, p < 0.001). Again, the interactions add nothing.\n",
  b_logit, p_logit, b_logit_hydro, b_logit_irrig
))

Logit (Pr(Downstream)): The transboundary coefficient is -0.117 but not significant (p = 0.496), so there is no strong evidence that transboundary status shifts the binary probability of being downstream of optimal. The purpose controls are highly significant: hydropower dams are more likely to move downstream (0.616, p < 0.001) while irrigation dams are less likely (-0.584, p < 0.001). Again, the interactions add nothing.

Table 3b: Hydro/Irrigation Subsample

model_posdev_hi <- feols(
  position_deviation_km ~ transboundary * hydropower + log_capacity |
    continent + decade,
  data = hydro_irrig_data,
  cluster = ~continent
)
etable(model_posdev_hi,
       title = "Position Deviation: Hydro vs Irrigation (H/I Sample)",
       headers = c("H/I Interact"))
##                                  model_posdev_hi
##                                     H/I Interact
## Dependent Var.:            position_deviation_km
##                                                 
## transboundary                      20.78 (19.40)
## hydropower                       -59.03* (13.06)
## log_capacity                    -0.8538 (0.8109)
## transboundary x hydropower        -1.782 (16.05)
## Fixed-Effects:             ---------------------
## continent                                    Yes
## decade                                       Yes
## __________________________ _____________________
## S.E.: Clustered                    by: continent
## Observations                               1,786
## R2                                       0.15757
## Within R2                                0.13650
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
b_pos_hi <- coef(model_posdev_hi)["transboundary"]
p_pos_hi <- pvalue(model_posdev_hi)["transboundary"]

has_pos_hi_interact <- "transboundary:hydropower" %in% names(coef(model_posdev_hi))
if (has_pos_hi_interact) {
  b_pos_hi_int <- coef(model_posdev_hi)["transboundary:hydropower"]
  p_pos_hi_int <- pvalue(model_posdev_hi)["transboundary:hydropower"]
  cat(sprintf(
    "Within the hydro/irrigation subsample, the base transboundary effect on position deviation is %.2f km (p = %.3f). The `transboundary x hydropower` interaction is %.2f km (p = %.3f), indicating whether hydropower dams are displaced differently from irrigation dams.\n",
    b_pos_hi, p_pos_hi, b_pos_hi_int, p_pos_hi_int
  ))
} else {
  cat(sprintf(
    "Within the hydro/irrigation subsample, the transboundary effect on position deviation is %.2f km (p = %.3f).\n",
    b_pos_hi, p_pos_hi
  ))
}

Within the hydro/irrigation subsample, the base transboundary effect on position deviation is 20.78 km (p = 0.344). The transboundary x hydropower interaction is -1.78 km (p = 0.917), indicating whether hydropower dams are displaced differently from irrigation dams.

8. Transboundary Dams: Suitability vs Border Distance

For the subset of transboundary dams where I can measure along-river distance to the nearest international border crossing, I examine whether proximity to the border is associated with worse site quality. If the transboundary suitability penalty is driven by geographic constraints near borders (e.g., narrow valleys, unfavorable terrain at boundary locations), I would expect dams closer to the border to have systematically lower suitability. If instead the penalty reflects broader political-strategic considerations, the relationship between border distance and suitability may be weak or absent.

plot_border <- function(data, col, vlabel) {
  ggplot(data, aes(x = dam_dist_border, y = .data[[col]])) +
    geom_point(alpha = 0.5, color = "coral") +
    geom_smooth(method = "lm", se = TRUE, color = "darkred") +
    labs(
      title = paste("TB Dams: Suitability vs Border Distance -", vlabel),
      subtitle = "Along-river distance to nearest border crossing",
      x = "Distance to Border (km, along river)",
      y = "Suitability Ratio"
    ) +
    theme_minimal(base_size = 13)
}

V3 (Percentile Ranks)

plot_border(tb_data, "suitability_ratio_v3", "V3 (Percentile Ranks)")

V2 (Gradient-weighted)

plot_border(tb_data, "suitability_ratio_v2", "V2 (Gradient-weighted)")

V1 (Baseline)

plot_border(tb_data, "suitability_ratio_v1", "V1 (Baseline)")

Regressions

tb_data <- tb_data %>%
  mutate(
    log_dam_dist_border = log(dam_dist_border + 1),
    border_dist_bin = cut(dam_dist_border,
                          breaks = c(0, 50, 100, 200, 500, Inf),
                          labels = c("<50km", "50-100km", "100-200km",
                                     "200-500km", ">500km"))
  )

model_tb_lin <- feols(
  suitability_ratio ~ dam_dist_border + log_capacity +
    hydropower + irrigation | continent + decade,
  data = tb_data, cluster = ~continent
)

model_tb_log <- feols(
  suitability_ratio ~ log_dam_dist_border + log_capacity +
    hydropower + irrigation | continent + decade,
  data = tb_data, cluster = ~continent
)

model_tb_bins <- feols(
  suitability_ratio ~ border_dist_bin + log_capacity +
    hydropower + irrigation | continent + decade,
  data = tb_data, cluster = ~continent
)
etable(model_tb_lin, model_tb_log, model_tb_bins,
       headers = c("Linear", "Log Distance", "Distance Bins"),
       title = "Transboundary Dams: Suitability vs Along-River Border Distance")
##                                model_tb_lin      model_tb_log     model_tb_bins
##                                      Linear      Log Distance     Distance Bins
## Dependent Var.:           suitability_ratio suitability_ratio suitability_ratio
##                                                                                
## dam_dist_border          -8.34e-6 (4.26e-5)                                    
## log_capacity               0.0367* (0.0073)  0.0366* (0.0073)  0.0381* (0.0078)
## hydropower                0.0672** (0.0095) 0.0658** (0.0100)  0.0694* (0.0212)
## irrigation                 -0.0195 (0.0259)  -0.0201 (0.0235)  -0.0131 (0.0264)
## log_dam_dist_border                           0.0008 (0.0007)                  
## border_dist_bin50-100km                                        -0.0161 (0.0301)
## border_dist_bin100-200km                                       -0.0197 (0.0127)
## border_dist_bin200-500km                                       -0.0252 (0.0163)
## border_dist_bin>500km                                          -0.0213 (0.0415)
## Fixed-Effects:           ------------------ ----------------- -----------------
## continent                               Yes               Yes               Yes
## decade                                  Yes               Yes               Yes
## ________________________ __________________ _________________ _________________
## S.E.: Clustered               by: continent     by: continent     by: continent
## Observations                            745               745               712
## R2                                  0.19433           0.19431           0.19271
## Within R2                           0.17247           0.17245           0.17301
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
b_lin <- coef(model_tb_lin)["dam_dist_border"]
p_lin <- pvalue(model_tb_lin)["dam_dist_border"]
b_log <- coef(model_tb_log)["log_dam_dist_border"]
p_log <- pvalue(model_tb_log)["log_dam_dist_border"]

cat(sprintf(
  "The linear distance coefficient is %.4f (p = %.3f), and the log specification yields %.4f (p = %.3f). ",
  b_lin, p_lin, b_log, p_log
))

The linear distance coefficient is -0.0000 (p = 0.857), and the log specification yields 0.0008 (p = 0.334).

cat("None of the border distance specifications produce significant coefficients, and the binned specification shows no monotonic pattern across distance categories. This null result is informative: the transboundary suitability penalty is not concentrated among dams near the border. Instead, it appears to be a broad feature of transboundary dam siting---consistent with political-strategic considerations that affect project selection and approval processes rather than purely geographic constraints at border locations.\n")

None of the border distance specifications produce significant coefficients, and the binned specification shows no monotonic pattern across distance categories. This null result is informative: the transboundary suitability penalty is not concentrated among dams near the border. Instead, it appears to be a broad feature of transboundary dam siting—consistent with political-strategic considerations that affect project selection and approval processes rather than purely geographic constraints at border locations.

9. Sensitivity: Transboundary Coefficient Across Specifications

A key concern is whether the transboundary suitability gap is an artifact of the suitability index construction. To address this, I run the same regression model (continent and decade fixed effects, standard errors clustered by continent) using each of the three suitability specifications as the dependent variable. If the transboundary coefficient is consistently negative across specifications with very different functional forms, this increases confidence that the result reflects a real phenomenon rather than a measurement artifact.

Note that the V3 (percentile) coefficient is expected to be smaller in magnitude since percentile ranks mechanically compress the scale. The relevant comparison is the sign, significance, and consistency of the coefficient across specifications.

sens_models <- list()
sens_coefs <- list()

for (v in versions) {
  col <- paste0("suitability_ratio_", v)
  fml <- as.formula(paste(col,
    "~ transboundary + log_capacity + hydropower + irrigation",
    "| continent + decade"))
  m <- feols(fml, data = analysis_data, cluster = ~continent)
  sens_models[[v]] <- m
  sens_coefs[[v]] <- tibble(
    spec = v,
    coef = coef(m)["transboundary"],
    se = se(m)["transboundary"],
    pval = pvalue(m)["transboundary"]
  )
}

sens_df <- bind_rows(sens_coefs) %>%
  mutate(
    spec_label = case_when(
      spec == "v1" ~ "V1 (Baseline)",
      spec == "v2" ~ "V2 (Gradient)",
      spec == "v3" ~ "V3 (Percentile)"
    ),
    sig = case_when(
      pval < 0.01 ~ "***",
      pval < 0.05 ~ "**",
      pval < 0.10 ~ "*",
      TRUE ~ ""
    )
  )
sens_df %>%
  select(Specification = spec_label,
         Coefficient = coef, SE = se,
         `p-value` = pval, Sig = sig) %>%
  kbl(digits = 4,
      caption = "Transboundary Coefficient Across Suitability Specifications") %>%
  kable_styling(bootstrap_options = c("striped", "hover"),
                full_width = FALSE)
Transboundary Coefficient Across Suitability Specifications
Specification Coefficient SE p-value Sig
V1 (Baseline) -0.0527 0.0141 0.0202 **
V2 (Gradient) -0.0794 0.0144 0.0053 ***
V3 (Percentile) -0.0275 0.0107 0.0616 
ggplot(sens_df, aes(x = spec_label, y = coef)) +
  geom_point(size = 4, color = "coral") +
  geom_errorbar(aes(ymin = coef - 1.96 * se,
                     ymax = coef + 1.96 * se),
                width = 0.2, color = "coral") +
  geom_hline(yintercept = 0, linetype = "dashed", color = "gray50") +
  labs(
    title = "Transboundary Effect Across Suitability Specifications",
    subtitle = "95% CI; negative = lower suitability for transboundary dams",
    x = "Specification",
    y = "Coefficient on Transboundary"
  ) +
  theme_minimal(base_size = 13)

all_neg <- all(sens_df$coef < 0)
cat(sprintf(
  "The transboundary coefficient is %s across all three specifications. ",
  ifelse(all_neg, "negative", "not consistently negative")
))

The transboundary coefficient is negative across all three specifications.

v2_row <- sens_df %>% filter(spec == "v2")
v3_row <- sens_df %>% filter(spec == "v3")
cat(sprintf(
  "The V2 gradient-weighted specification shows an effect of %.3f (p = %.3f), while V3 percentile ranks show %.3f (p = %.3f). The consistency across specifications with very different distributional properties supports the interpretation that transboundary dams are genuinely placed at less suitable sites.\n",
  v2_row$coef, v2_row$pval, v3_row$coef, v3_row$pval
))

The V2 gradient-weighted specification shows an effect of -0.079 (p = 0.005), while V3 percentile ranks show -0.028 (p = 0.062). The consistency across specifications with very different distributional properties supports the interpretation that transboundary dams are genuinely placed at less suitable sites.

10. Constrained (Same-Country) Suitability Analysis

A natural concern is that the transboundary suitability penalty is mechanical: perhaps the best engineering site within +/-200 km happens to be across the border, and the country simply cannot build there. To address this, I construct a constrained suitability ratio (\(SR_c\)) that compares each dam’s actual site only to the best available site within the dam’s own country. By construction, \(SR_c \geq SR_u\) since the constrained optimal is weakly worse than the unconstrained optimal.

If country borders explain the penalty, \(SR_c\) should be much closer to 1 than \(SR_u\), and the transboundary coefficient on \(SR_c\) should attenuate substantially. If instead the penalty reflects within-country choices, the attenuation should be small.

# Constrained columns available -- proceed with analysis
constrained_versions <- c("v1", "v2", "v3")[c(has_constrained_v1, TRUE, has_constrained_v3)]

Descriptive Comparison

constrained_summary <- analysis_data %>%
  filter(!is.na(suitability_ratio_c), !is.na(optimal_outside_country)) %>%
  group_by(transboundary) %>%
  summarise(
    n = n(),
    mean_sr_u = mean(suitability_ratio, na.rm = TRUE),
    mean_sr_c = mean(suitability_ratio_c, na.rm = TRUE),
    median_sr_u = median(suitability_ratio, na.rm = TRUE),
    median_sr_c = median(suitability_ratio_c, na.rm = TRUE),
    pct_optimal_outside = mean(optimal_outside_country, na.rm = TRUE) * 100,
    .groups = "drop"
  ) %>%
  mutate(
    tb_label = ifelse(transboundary == 1, "Transboundary", "Domestic"),
    sr_gap = mean_sr_c - mean_sr_u
  )

constrained_summary %>%
  select(Status = tb_label, n,
         `Mean SR_u` = mean_sr_u, `Mean SR_c` = mean_sr_c,
         `Median SR_u` = median_sr_u, `Median SR_c` = median_sr_c,
         `Gap (SR_c - SR_u)` = sr_gap,
         `% Optimal Outside` = pct_optimal_outside) %>%
  kbl(digits = 3,
      caption = "Unconstrained vs Constrained Suitability Ratios (V2)") %>%
  kable_styling(bootstrap_options = c("striped", "hover"),
                full_width = FALSE)
Unconstrained vs Constrained Suitability Ratios (V2)
Status n Mean SR_u Mean SR_c Median SR_u Median SR_c Gap (SR_c - SR_u) % Optimal Outside
Domestic 2266 0.241 0.244 0.110 0.111 0.003 0.000
Transboundary 2634 0.140 0.152 0.057 0.059 0.013 9.377 
forced_domestic <- analysis_data %>%
  filter(optimal_outside_country == TRUE, !is.na(suitability_ratio_c))
if (nrow(forced_domestic) > 0) {
  cat(sprintf(
    "Among dams whose unconstrained optimal is outside the country (n = %d): mean $SR_u$ = %.3f, mean $SR_c$ = %.3f, mean gap = %.3f. These are dams 'forced' to use a domestic site that is worse than the cross-border alternative.\n",
    nrow(forced_domestic),
    mean(forced_domestic$suitability_ratio, na.rm = TRUE),
    mean(forced_domestic$suitability_ratio_c, na.rm = TRUE),
    mean(forced_domestic$suitability_ratio_c - forced_domestic$suitability_ratio, na.rm = TRUE)
  ))
}

Among dams whose unconstrained optimal is outside the country (n = 247): mean \(SR_u\) = 0.175, mean \(SR_c\) = 0.308, mean gap = 0.133. These are dams ‘forced’ to use a domestic site that is worse than the cross-border alternative.

% Optimal Outside Country

p_outside_data <- analysis_data %>%
  filter(!is.na(suitability_ratio_c), !is.na(optimal_outside_country)) %>%
  group_by(transboundary) %>%
  summarise(
    n = n(),
    pct_outside = mean(optimal_outside_country, na.rm = TRUE) * 100,
    .groups = "drop"
  ) %>%
  mutate(label = ifelse(transboundary == 1, "Transboundary", "Domestic"))

ggplot(p_outside_data, aes(x = label, y = pct_outside, fill = label)) +
  geom_col(width = 0.6) +
  geom_text(aes(label = sprintf("%.1f%%", pct_outside)),
            vjust = -0.5, size = 4) +
  scale_fill_manual(values = c("Domestic" = "steelblue",
                                "Transboundary" = "coral")) +
  labs(
    title = "Percentage of Dams with Optimal Location Outside Country",
    subtitle = "i.e., country-constrained optimal differs from unconstrained optimal",
    x = "", y = "% of Dams"
  ) +
  theme_minimal(base_size = 13) +
  theme(legend.position = "none") +
  ylim(0, max(p_outside_data$pct_outside) * 1.3)

For domestic dams, the unconstrained and constrained optimals are virtually always the same (candidates rarely cross borders). For transboundary dams, only a modest share have their best site outside the country. This immediately suggests that country borders are not the primary driver of the suitability penalty.

Constrained Density Plots

plot_constrained_density <- function(data, col_c, col_u, vlabel) {
  plot_data <- data %>%
    filter(!is.na(.data[[col_c]])) %>%
    select(transboundary, SR_u = !!sym(col_u), SR_c = !!sym(col_c)) %>%
    pivot_longer(cols = c(SR_u, SR_c),
                 names_to = "measure", values_to = "ratio")

  ggplot(plot_data, aes(x = ratio,
                         fill = factor(transboundary),
                         linetype = measure)) +
    geom_density(alpha = 0.3) +
    geom_vline(xintercept = 1, linetype = "dashed", color = "black") +
    scale_fill_manual(
      values = c("0" = "steelblue", "1" = "coral"),
      labels = c("Domestic", "Transboundary"),
      name = ""
    ) +
    scale_linetype_manual(
      values = c("SR_u" = "solid", "SR_c" = "dashed"),
      labels = c(expression(SR[u]~"(unconstrained)"),
                 expression(SR[c]~"(same country)")),
      name = ""
    ) +
    labs(
      title = paste("Unconstrained vs Constrained Suitability -", vlabel),
      subtitle = "Solid = unconstrained; Dashed = constrained to same country",
      x = "Suitability Ratio", y = "Density"
    ) +
    theme_minimal(base_size = 13) +
    theme(legend.position = "bottom")
}

V3 (Percentile Ranks)

plot_constrained_density(analysis_data,
                          "suitability_ratio_c_v3", "suitability_ratio_v3",
                          "V3 (Percentile Ranks)")

V2 (Gradient-weighted)

plot_constrained_density(analysis_data,
                          "suitability_ratio_c", "suitability_ratio",
                          "V2 (Gradient-weighted)")

V1 (Baseline)

plot_constrained_density(analysis_data,
                          "suitability_ratio_c_v1", "suitability_ratio_v1",
                          "V1 (Baseline)")

Constrained SR Density (Standalone)

p6_data <- analysis_data %>%
  filter(!is.na(suitability_ratio_c), !is.na(optimal_outside_country))

ggplot(p6_data, aes(x = suitability_ratio_c,
                     fill = factor(transboundary))) +
  geom_density(alpha = 0.6) +
  geom_vline(xintercept = 1, linetype = "dashed", color = "black") +
  scale_fill_manual(
    values = c("0" = "steelblue", "1" = "coral"),
    labels = c("Domestic", "Transboundary"),
    name = ""
  ) +
  labs(
    title = "Constrained Suitability Ratio (Same-Country Optimal)",
    subtitle = "Ratio of 1 = dam at best location within its own country",
    x = expression(SR[c] ~ "(Actual / Optimal in same country)"),
    y = "Density"
  ) +
  theme_minimal(base_size = 13) +
  theme(legend.position = "bottom")

SR_u vs SR_c Scatter

ggplot(p6_data, aes(x = suitability_ratio, y = suitability_ratio_c,
                     color = factor(transboundary))) +
  geom_point(alpha = 0.3, size = 1.5) +
  geom_abline(intercept = 0, slope = 1, linetype = "dashed", color = "gray40") +
  scale_color_manual(
    values = c("0" = "steelblue", "1" = "coral"),
    labels = c("Domestic", "Transboundary"),
    name = ""
  ) +
  labs(
    title = "Unconstrained vs Constrained Suitability Ratio",
    subtitle = "Points above diagonal = constrained optimal worse than unconstrained (optimal outside country)",
    x = expression(SR[u] ~ "(actual / optimal anywhere)"),
    y = expression(SR[c] ~ "(actual / optimal in same country)")
  ) +
  coord_equal(xlim = c(0, 1.05), ylim = c(0, 1.05)) +
  theme_minimal(base_size = 13) +
  theme(legend.position = "bottom")

Most points lie on the 45-degree line, meaning the unconstrained and constrained optimals are the same segment. Points above the diagonal are dams where the best site is outside the country; these are predominantly transboundary dams (orange).

Constrained Regressions (V2)

model_c1 <- feols(
  suitability_ratio_c ~ transboundary,
  data = analysis_data
)

model_c2 <- feols(
  suitability_ratio_c ~ transboundary + log_capacity + hydropower + irrigation,
  data = analysis_data
)

model_c3 <- feols(
  suitability_ratio_c ~ transboundary + log_capacity + hydropower + irrigation |
    continent + decade,
  data = analysis_data,
  cluster = ~continent
)

model_c4 <- feols(
  suitability_ratio_c ~ transboundary + log_capacity + hydropower + irrigation |
    admin0 + decade,
  data = analysis_data,
  cluster = ~admin0
)

etable(model_c1, model_c2, model_c3, model_c4,
       headers = c("Simple", "Controls", "Continent FE", "Country FE"),
       title = "Effect of Transboundary Status on Constrained Suitability Ratio (SR_c)")
##                            model_c1            model_c2            model_c3
##                              Simple            Controls        Continent FE
## Dependent Var.: suitability_ratio_c suitability_ratio_c suitability_ratio_c
##                                                                            
## Constant         0.2437*** (0.0052)  0.1201*** (0.0079)                    
## transboundary   -0.0914*** (0.0071) -0.0816*** (0.0068)  -0.0670** (0.0138)
## log_capacity                         0.0301*** (0.0017)   0.0302** (0.0064)
## hydropower                           0.0673*** (0.0106)    0.0669. (0.0250)
## irrigation                             0.0161. (0.0085)    -0.0137 (0.0330)
## Fixed-Effects:  ------------------- ------------------- -------------------
## continent                        No                  No                 Yes
## decade                           No                  No                 Yes
## admin0                           No                  No                  No
## _______________ ___________________ ___________________ ___________________
## S.E. type                       IID                 IID       by: continent
## Observations                  4,901               4,901               4,901
## R2                          0.03240             0.12264             0.16998
## Within R2                        --                  --             0.10300
## 
##                            model_c4
##                          Country FE
## Dependent Var.: suitability_ratio_c
##                                    
## Constant                           
## transboundary   -0.0534*** (0.0089)
## log_capacity     0.0293*** (0.0031)
## hydropower          0.0421 (0.0260)
## irrigation         -0.0294 (0.0215)
## Fixed-Effects:  -------------------
## continent                        No
## decade                          Yes
## admin0                          Yes
## _______________ ___________________
## S.E. type                by: admin0
## Observations                  4,901
## R2                          0.25465
## Within R2                   0.07633
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
cat("**Side-by-side comparison: Unconstrained vs Constrained**\n\n")
## **Side-by-side comparison: Unconstrained vs Constrained**
etable(model3, model_c3, model4, model_c4,
       headers = c("SR_u (Cont. FE)", "SR_c (Cont. FE)",
                    "SR_u (Country FE)", "SR_c (Country FE)"),
       title = "Unconstrained vs Constrained Suitability Ratio")
##                             model3            model_c3              model4
##                    SR_u (Cont. FE)     SR_c (Cont. FE)   SR_u (Country FE)
## Dependent Var.:  suitability_ratio suitability_ratio_c   suitability_ratio
##                                                                           
## transboundary   -0.0794** (0.0144)  -0.0670** (0.0138) -0.0655*** (0.0074)
## log_capacity     0.0318** (0.0055)   0.0302** (0.0064)  0.0307*** (0.0027)
## hydropower        0.0574. (0.0244)    0.0669. (0.0250)     0.0369 (0.0258)
## irrigation        -0.0127 (0.0332)    -0.0137 (0.0330)    -0.0285 (0.0217)
## Fixed-Effects:  ------------------ ------------------- -------------------
## continent                      Yes                 Yes                  No
## decade                         Yes                 Yes                 Yes
## admin0                          No                  No                 Yes
## _______________ __________________ ___________________ ___________________
## S.E.: Clustered      by: continent       by: continent          by: admin0
## Observations                 4,901               4,901               4,901
## R2                         0.18403             0.16998             0.26802
## Within R2                  0.12103             0.10300             0.09185
## 
##                            model_c4
##                   SR_c (Country FE)
## Dependent Var.: suitability_ratio_c
##                                    
## transboundary   -0.0534*** (0.0089)
## log_capacity     0.0293*** (0.0031)
## hydropower          0.0421 (0.0260)
## irrigation         -0.0294 (0.0215)
## Fixed-Effects:  -------------------
## continent                        No
## decade                          Yes
## admin0                          Yes
## _______________ ___________________
## S.E.: Clustered          by: admin0
## Observations                  4,901
## R2                          0.25465
## Within R2                   0.07633
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
coef_u <- coef(model3)["transboundary"]
coef_c <- coef(model_c3)["transboundary"]
p_c3 <- pvalue(model_c3)["transboundary"]
attenuation <- (1 - coef_c / coef_u) * 100

coef_u4 <- coef(model4)["transboundary"]
coef_c4 <- coef(model_c4)["transboundary"]
p_c4 <- pvalue(model_c4)["transboundary"]
attenuation4 <- (1 - coef_c4 / coef_u4) * 100

cat(sprintf(
  "The constrained regressions mirror the unconstrained results closely. The transboundary coefficient on $SR_c$ is negative and significant across all four specifications, following the same attenuation pattern as $SR_u$ when moving from simple OLS to country fixed effects.\n\n"
))

The constrained regressions mirror the unconstrained results closely. The transboundary coefficient on \(SR_c\) is negative and significant across all four specifications, following the same attenuation pattern as \(SR_u\) when moving from simple OLS to country fixed effects.

cat(sprintf(
  "**With continent + decade FE:** The transboundary coefficient attenuates from %.4f ($SR_u$) to %.4f ($SR_c$, p = %.3f), a **%.1f%% reduction**. This means approximately **%.0f%% of the siting penalty occurs within the dam's own country**, not due to border constraints.\n\n",
  coef_u, coef_c, p_c3, attenuation, 100 - attenuation
))

With continent + decade FE: The transboundary coefficient attenuates from -0.0794 (\(SR_u\)) to -0.0670 (\(SR_c\), p = 0.008), a 15.6% reduction. This means approximately 84% of the siting penalty occurs within the dam’s own country, not due to border constraints.

cat(sprintf(
  "**With country FE:** The coefficient moves from %.4f ($SR_u$) to %.4f ($SR_c$, p = %.3f), an attenuation of **%.1f%%**. Even in the most demanding specification---comparing transboundary and domestic dams within the same country, using only same-country alternatives as the benchmark---the penalty remains large and highly significant. This rules out the mechanical explanation that transboundary dams look worse simply because the best site is across the border.\n",
  coef_u4, coef_c4, p_c4, attenuation4
))

With country FE: The coefficient moves from -0.0655 (\(SR_u\)) to -0.0535 (\(SR_c\), p = 0.000), an attenuation of 18.3%. Even in the most demanding specification—comparing transboundary and domestic dams within the same country, using only same-country alternatives as the benchmark—the penalty remains large and highly significant. This rules out the mechanical explanation that transboundary dams look worse simply because the best site is across the border.

Optimal Outside Country (LPM)

model_outside <- feols(
  optimal_outside_country ~ transboundary + log_capacity + hydropower + irrigation |
    continent + decade,
  data = analysis_data %>% filter(!is.na(optimal_outside_country)),
  cluster = ~continent
)

etable(model_outside,
       headers = "Pr(Optimal Outside Country)",
       title = "Does Transboundary Status Predict Optimal Being Outside Country?")
##                               model_outside
##                 Pr(Optimal Outside Country)
## Dependent Var.:     optimal_outside_country
##                                            
## transboundary               0.1053 (0.0639)
## log_capacity               -0.0096 (0.0053)
## hydropower                 0.0455* (0.0153)
## irrigation                  0.0126 (0.0141)
## Fixed-Effects:      -----------------------
## continent                               Yes
## decade                                  Yes
## _______________     _______________________
## S.E.: Clustered               by: continent
## Observations                          4,900
## R2                                  0.10226
## Within R2                           0.06362
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Sensitivity: Constrained Coefficient Across Specifications

c_sens_models <- list()
c_sens_coefs <- list()

for (v in constrained_versions) {
  if (v == "v2") {
    col_c <- "suitability_ratio_c"
    col_u <- "suitability_ratio"
  } else {
    col_c <- paste0("suitability_ratio_c_", v)
    col_u <- paste0("suitability_ratio_", v)
  }
  if (!col_c %in% names(analysis_data)) next

  fml_u <- as.formula(paste(col_u,
    "~ transboundary + log_capacity + hydropower + irrigation | continent + decade"))
  fml_c <- as.formula(paste(col_c,
    "~ transboundary + log_capacity + hydropower + irrigation | continent + decade"))

  m_u <- feols(fml_u, data = analysis_data, cluster = ~continent)
  m_c <- feols(fml_c, data = analysis_data, cluster = ~continent)

  c_sens_coefs[[v]] <- tibble(
    spec = v,
    coef_u = coef(m_u)["transboundary"],
    se_u = se(m_u)["transboundary"],
    coef_c = coef(m_c)["transboundary"],
    se_c = se(m_c)["transboundary"],
    attenuation_pct = (1 - coef(m_c)["transboundary"] /
                            coef(m_u)["transboundary"]) * 100
  )
}

c_sens_df <- bind_rows(c_sens_coefs) %>%
  mutate(spec_label = case_when(
    spec == "v1" ~ "V1 (Baseline)",
    spec == "v2" ~ "V2 (Gradient)",
    spec == "v3" ~ "V3 (Percentile)"
  ))
c_sens_df %>%
  select(Specification = spec_label,
         `Coef (SR_u)` = coef_u, `SE (SR_u)` = se_u,
         `Coef (SR_c)` = coef_c, `SE (SR_c)` = se_c,
         `Attenuation %` = attenuation_pct) %>%
  kbl(digits = 4,
      caption = "Transboundary Coefficient: Unconstrained vs Constrained Across Specs") %>%
  kable_styling(bootstrap_options = c("striped", "hover"),
                full_width = FALSE)
Transboundary Coefficient: Unconstrained vs Constrained Across Specs
Specification Coef (SR_u) SE (SR_u) Coef (SR_c) SE (SR_c) Attenuation %
V1 (Baseline) -0.0527 0.0141 -0.0411 0.0136 21.9403
V2 (Gradient) -0.0794 0.0144 -0.0670 0.0138 15.5689
V3 (Percentile) -0.0275 0.0107 -0.0174 0.0147 36.9104 
c_plot_df <- c_sens_df %>%
  select(spec_label, coef_u, se_u, coef_c, se_c) %>%
  pivot_longer(
    cols = c(coef_u, coef_c),
    names_to = "type", values_to = "coef"
  ) %>%
  mutate(
    se = ifelse(type == "coef_u",
                c_sens_df$se_u[match(spec_label, c_sens_df$spec_label)],
                c_sens_df$se_c[match(spec_label, c_sens_df$spec_label)]),
    type_label = ifelse(type == "coef_u",
                         "Unconstrained (SR_u)", "Constrained (SR_c)")
  )

ggplot(c_plot_df, aes(x = spec_label, y = coef,
                       color = type_label, shape = type_label)) +
  geom_point(size = 4, position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = coef - 1.96 * se,
                     ymax = coef + 1.96 * se),
                width = 0.2,
                position = position_dodge(width = 0.4)) +
  geom_hline(yintercept = 0, linetype = "dashed", color = "gray50") +
  scale_color_manual(values = c("Unconstrained (SR_u)" = "coral",
                                  "Constrained (SR_c)" = "steelblue")) +
  labs(
    title = "Transboundary Coefficient: Unconstrained vs Constrained",
    subtitle = "95% CI; attenuation from SR_u to SR_c measures role of country borders",
    x = "Suitability Specification",
    y = "Coefficient on Transboundary",
    color = "", shape = ""
  ) +
  theme_minimal(base_size = 13) +
  theme(legend.position = "bottom")

Across all specifications, the constrained coefficient (\(SR_c\)) is only modestly smaller than the unconstrained coefficient (\(SR_u\)). The consistency of the small attenuation across V1, V2, and V3 reinforces the conclusion that country borders explain only a minor fraction of the transboundary siting penalty.

11. Robustness Checks

I conduct three robustness exercises: (1) using the suitability percentile (within-river-system rank) as the dependent variable instead of the ratio; (2) restricting to dams with at least 10 candidate segments (ensuring meaningful comparison sets); and (3) restricting to post-1970 dams (where data quality is higher and the international water governance regime was more established).

model_pctl <- feols(
  suitability_percentile ~ transboundary + log_capacity +
    hydropower + irrigation | continent + decade,
  data = analysis_data, cluster = ~continent
)

model_high_qual <- feols(
  suitability_ratio ~ transboundary + log_capacity +
    hydropower + irrigation | continent + decade,
  data = analysis_data %>% filter(n_candidates >= 10),
  cluster = ~continent
)

model_post1970 <- feols(
  suitability_ratio ~ transboundary + log_capacity +
    hydropower + irrigation | continent + decade,
  data = analysis_data %>% filter(decade >= 1970),
  cluster = ~continent
)
etable(model_pctl, model_high_qual, model_post1970,
       headers = c("Suitability Pctl", ">=10 Candidates", "Post-1970"),
       title = "Robustness Checks (V2)")
##                             model_pctl    model_high_qual     model_post1970
##                       Suitability Pctl    >=10 Candidates          Post-1970
## Dependent Var.: suitability_percentile  suitability_ratio  suitability_ratio
##                                                                             
## transboundary       -0.0698** (0.0145)  -0.0522* (0.0119) -0.0987** (0.0181)
## log_capacity         0.0638** (0.0099) 0.0368*** (0.0034)   0.0282* (0.0072)
## hydropower           0.1147** (0.0222)   0.0626* (0.0219)    0.0430 (0.0301)
## irrigation           -0.1817* (0.0484)   -0.0012 (0.0223)   -0.0121 (0.0392)
## Fixed-Effects:  ---------------------- ------------------ ------------------
## continent                          Yes                Yes                Yes
## decade                             Yes                Yes                Yes
## _______________ ______________________ __________________ __________________
## S.E.: Clustered          by: continent      by: continent      by: continent
## Observations                     4,901              4,743              2,760
## R2                             0.31140            0.21150            0.19817
## Within R2                      0.27009            0.15566            0.11289
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
b_pctl <- coef(model_pctl)["transboundary"]
p_pctl <- pvalue(model_pctl)["transboundary"]
b_hq <- coef(model_high_qual)["transboundary"]
p_hq <- pvalue(model_high_qual)["transboundary"]
b_70 <- coef(model_post1970)["transboundary"]
p_70 <- pvalue(model_post1970)["transboundary"]

cat(sprintf(
  "All three robustness checks confirm the main finding. The suitability percentile specification yields a coefficient of %.3f (p = %.3f), indicating that transboundary dams sit about %.0f percentile points lower in the within-river-system suitability distribution. Restricting to dams with at least 10 candidate sites produces a coefficient of %.3f (p = %.3f), ruling out the concern that the result is driven by dams with very few alternatives. The post-1970 subsample shows a coefficient of %.3f (p = %.3f).\n",
  b_pctl, p_pctl, abs(b_pctl) * 100,
  b_hq, p_hq,
  b_70, p_70
))

All three robustness checks confirm the main finding. The suitability percentile specification yields a coefficient of -0.070 (p = 0.009), indicating that transboundary dams sit about 7 percentile points lower in the within-river-system suitability distribution. Restricting to dams with at least 10 candidate sites produces a coefficient of -0.052 (p = 0.012), ruling out the concern that the result is driven by dams with very few alternatives. The post-1970 subsample shows a coefficient of -0.099 (p = 0.005).

Summary of Key Findings

# Main effect
main_effect <- coef(model3)["transboundary"]
main_pval <- pvalue(model3)["transboundary"]
country_effect <- coef(model4)["transboundary"]

cat(sprintf(
  "1. **Transboundary dams are systematically placed at less suitable sites** compared to domestic dams, with the gap robust across three different suitability index constructions (V1 baseline, V2 gradient-weighted, V3 percentile ranks).\n\n"
))
  1. Transboundary dams are systematically placed at less suitable sites compared to domestic dams, with the gap robust across three different suitability index constructions (V1 baseline, V2 gradient-weighted, V3 percentile ranks).
cat(sprintf(
  "2. **The penalty survives demanding specifications**: even comparing transboundary and domestic dams within the same country (country fixed effects), transboundary dams have suitability ratios approximately %.1f percentage points lower.\n\n",
  abs(country_effect) * 100
))
  1. The penalty survives demanding specifications: even comparing transboundary and domestic dams within the same country (country fixed effects), transboundary dams have suitability ratios approximately 6.5 percentage points lower.
cat(sprintf(
  "3. **The penalty is not about country borders**: constraining the optimal to the dam's own country ($SR_c$) attenuates the transboundary coefficient by only ~%.0f%%. The vast majority of the siting penalty reflects within-country choices, not an inability to access better sites across the border.\n\n",
  attenuation
))
  1. The penalty is not about country borders: constraining the optimal to the dam’s own country (\(SR_c\)) attenuates the transboundary coefficient by only ~16%. The vast majority of the siting penalty reflects within-country choices, not an inability to access better sites across the border.
cat("4. **The effect is present across dam types**, though the magnitude varies between hydropower and irrigation dams as shown in the interaction models.\n\n")
  1. The effect is present across dam types, though the magnitude varies between hydropower and irrigation dams as shown in the interaction models.
cat(sprintf(
  "5. **The post-1970 subsample** shows a transboundary coefficient of %.3f, suggesting the siting distortion has %s over time.\n\n",
  b_70,
  ifelse(abs(b_70) > abs(main_effect), "not diminished (and may have grown)",
         "modestly attenuated but persists")
))
  1. The post-1970 subsample shows a transboundary coefficient of -0.099, suggesting the siting distortion has not diminished (and may have grown) over time.
cat("6. **Border proximity does not explain the gap**: among transboundary dams, along-river distance to the border has no significant relationship with suitability, suggesting the penalty reflects broad political-strategic dynamics rather than geographic constraints at border locations.\n")
  1. Border proximity does not explain the gap: among transboundary dams, along-river distance to the border has no significant relationship with suitability, suggesting the penalty reflects broad political-strategic dynamics rather than geographic constraints at border locations.

Report generated with strategic_placement_report.Rmd