1 Setup
2 OECD growth, 2000–2024
3 OECD convergence
4 Global convergence (advantage of backwardness)
5 Conditional convergence: NIQ interaction
6 Robustness: same model across income measures
7 Spatial robustness
8 Outliers from the GDP interaction model
9 Session

1 Setup

Data sources:

GDP per capita (PPP, constant 2021 international $) — World Bank via OWID
GDP/GNI per capita in USD constant — World Bank
Household / total final consumption per capita — World Bank
National IQ — Jensen & Kirkegaard 2025 (d_NIQ_S.rds)
Nighttime lights (DMSP 1992-2013, VIIRS 2013-2024) — Yao country panels

All growth rates are CAGR computed on the earliest available value in 2000–2005 and the latest available value in 2020–2024.

m %>% summarise(
  countries = n(),
  has_GDP_PPP    = sum(!is.na(`GDP/cap PPP const`)),
  has_GNI_PPP    = sum(!is.na(`GNI/cap PPP const`)),
  has_HHcons_PPP = sum(!is.na(`HH cons/cap PPP`)),
  has_NIQ        = sum(!is.na(NIQ)),
  has_VIIRS      = sum(!is.na(viirs_growth)),
  oecd_members   = sum(is_oecd)
) %>% kable()

countries	has_GDP_PPP	has_GNI_PPP	has_HHcons_PPP	has_NIQ	has_VIIRS	oecd_members
193	193	118	131	176	142	38

2 OECD growth, 2000–2024

2.1 GDP per capita

oecd <- m %>% filter(is_oecd, !is.na(`GDP/cap PPP const`)) %>%
  mutate(country = entity,
         g = `GDP/cap PPP const`,
         color = ifelse(g < median(g), "below", "above"))
ggplot(oecd, aes(x = reorder(country, g), y = g, fill = color)) +
  geom_col() +
  geom_hline(yintercept = median(oecd$g), lty = "dashed", lwd = 0.4) +
  geom_text(aes(label = sprintf("%.2f", g)), hjust = -0.1, size = 3) +
  coord_flip() +
  scale_fill_manual(values = c(above = "#4c72b0", below = "#c44e52"), guide = "none") +
  labs(x = NULL,
       y = "Annualized GDP per capita growth, 2000–2024 (%, PPP)",
       title = sprintf("OECD GDP per capita growth, 2000–2024 (median = %.2f%%)",
                       median(oecd$g))) +
  theme_minimal(base_size = 11)

2.2 Household consumption per capita (avoids tax-haven distortion)

GDP-based ranking puts Ireland near the top because of multinational IP-onshoring and aircraft-leasing accounting that doesn’t reach households. Re-ranking on household consumption per capita (PPP) strips most of that distortion: the same spending data that consumers actually do, divided by population.

oecd_c <- m %>% filter(is_oecd, !is.na(`HH cons/cap PPP`)) %>%
  mutate(country = entity,
         g = `HH cons/cap PPP`,
         color = ifelse(g < median(g), "below", "above"))
ggplot(oecd_c, aes(x = reorder(country, g), y = g, fill = color)) +
  geom_col() +
  geom_hline(yintercept = median(oecd_c$g), lty = "dashed", lwd = 0.4) +
  geom_text(aes(label = sprintf("%.2f", g)), hjust = -0.1, size = 3) +
  coord_flip() +
  scale_fill_manual(values = c(above = "#4c72b0", below = "#c44e52"), guide = "none") +
  labs(x = NULL,
       y = "Annualized household consumption per capita growth, 2000–2024 (%, PPP)",
       title = sprintf("OECD HH consumption per capita growth, 2000–2024 (median = %.2f%%)",
                       median(oecd_c$g)),
       subtitle = "Consumption-based: largely immune to tax-haven inflation") +
  theme_minimal(base_size = 11)

2.3 GDP vs. consumption: which countries change most?

shift <- m %>% filter(is_oecd, !is.na(`GDP/cap PPP const`),
                      !is.na(`HH cons/cap PPP`)) %>%
  mutate(diff = `GDP/cap PPP const` - `HH cons/cap PPP`)

ggplot(shift, aes(x = reorder(entity, diff), y = diff,
                  fill = diff > 0)) +
  geom_col() +
  coord_flip() +
  scale_fill_manual(values = c(`TRUE` = "#dd8452", `FALSE` = "#4c72b0"),
                    labels = c("Consumption grew faster", "GDP grew faster"),
                    name = NULL) +
  geom_hline(yintercept = 0, lwd = 0.4) +
  geom_text(aes(label = sprintf("%+.2f", diff)),
            hjust = ifelse(shift$diff > 0, -0.1, 1.1), size = 3) +
  labs(x = NULL,
       y = "GDP CAGR − consumption CAGR (pp/yr)",
       title = "Where GDP growth doesn't translate to consumption growth",
       subtitle = "Positive = GDP outpaced consumption (tax-haven / corporate-savings)\nNegative = consumption outpaced GDP (sovereign wealth distributions)") +
  theme_minimal(base_size = 11)

3 OECD convergence

oecd_ok <- m %>% filter(is_oecd, !is.na(`GDP/cap PPP const`), !is.na(gdp_start))
fit_oecd <- lm(`GDP/cap PPP const` ~ log_gdp_start, data = oecd_ok)
r2 <- summary(fit_oecd)$r.squared
slope <- coef(fit_oecd)[2]

ggplot(oecd_ok, aes(x = gdp_start, y = `GDP/cap PPP const`)) +
  geom_smooth(method = "lm", se = FALSE, color = "black", lty = "dashed") +
  geom_point(aes(color = is_oecd_postcomm), size = 3, alpha = 0.9) +
  geom_text(aes(label = entity), size = 3, vjust = -0.9, alpha = 0.85) +
  scale_x_log10(labels = dollar) +
  scale_color_manual(values = c(`FALSE` = "steelblue", `TRUE` = "#c44e52"),
                     labels = c("Other OECD", "Post-communist OECD"),
                     name = NULL) +
  labs(x = "GDP per capita in 2000 (PPP, log scale)",
       y = "Annualized growth 2000-2024 (%)",
       title = sprintf("OECD convergence (slope = %.2f, R² = %.2f)", slope, r2),
       subtitle = "Post-communist members do most of the work") +
  theme_minimal(base_size = 11)

fit_full <- lm(`GDP/cap PPP const` ~ log_gdp_start, data = oecd_ok)
fit_pc   <- lm(`GDP/cap PPP const` ~ log_gdp_start, data = oecd_ok %>% filter(is_oecd_postcomm))
fit_npc  <- lm(`GDP/cap PPP const` ~ log_gdp_start, data = oecd_ok %>% filter(!is_oecd_postcomm))
data.frame(
  group = c("Full OECD", "Post-communist", "Other OECD"),
  n     = c(nobs(fit_full), nobs(fit_pc), nobs(fit_npc)),
  slope = round(c(coef(fit_full)[2], coef(fit_pc)[2], coef(fit_npc)[2]), 2),
  R2    = round(c(summary(fit_full)$r.squared, summary(fit_pc)$r.squared,
                  summary(fit_npc)$r.squared), 3)
) %>% kable()

group	n	slope	R2
Full OECD	38	-1.58	0.520
Post-communist	8	-3.37	0.878
Other OECD	30	-1.02	0.337

4 Global convergence (advantage of backwardness)

w <- m %>% filter(!is.na(`GDP/cap PPP const`), !is.na(gdp_start))
fit_w <- lm(`GDP/cap PPP const` ~ log_gdp_start, data = w)

region_colors <- c(
  "Africa"        = "#e64b35",
  "Asia"          = "#4dbbd5",
  "Europe"        = "#00a087",
  "North America" = "#3c5488",
  "South America" = "#f39b7f",
  "Oceania"       = "#8491b4"
)

label_set <- c("United States","China","India","Germany","Japan","United Kingdom",
               "France","Italy","Brazil","Russia","Canada","Australia","Mexico",
               "Spain","South Korea","Indonesia","Saudi Arabia","Turkey","Argentina",
               "Ireland","Vietnam","Bangladesh","Ethiopia","Rwanda","Cambodia",
               "Lithuania","Poland","Latvia","Estonia",
               "Equatorial Guinea","Venezuela","Zimbabwe","Yemen","Libya","Sudan",
               "Central African Republic","Burundi","Singapore","Norway","Luxembourg",
               "Qatar","United Arab Emirates","Iceland","Switzerland","Guyana",
               "Niger","Sierra Leone")

ggplot(w, aes(x = gdp_start, y = `GDP/cap PPP const`, color = owid_region)) +
  geom_smooth(method = "lm", se = FALSE, color = "black", lty = "dashed", lwd = 0.9) +
  geom_point(alpha = 0.85, size = 3) +
  geom_text(aes(label = ifelse(entity %in% label_set, entity, "")),
            size = 2.8, vjust = -0.9, hjust = 0.5, alpha = 0.85, show.legend = FALSE) +
  scale_x_log10(labels = dollar) +
  scale_color_manual(values = region_colors, name = NULL, na.value = "#999999") +
  labs(x = "GDP per capita in 2000 (PPP, log scale)",
       y = "Annualized growth 2000-2024 (%)",
       title = sprintf("Global convergence / advantage of backwardness (slope = %.2f, R² = %.2f, n = %d)",
                       coef(fit_w)[2], summary(fit_w)$r.squared, nobs(fit_w)),
       subtitle = "Unconditional convergence is weak globally — Asian and post-Soviet countries dominate the catch-up corner; Africa stays poor despite the gap.") +
  theme_minimal(base_size = 11) +
  theme(legend.position = "bottom")

5 Conditional convergence: NIQ interaction

5.1 Main interaction model

mw <- m %>% filter(!is.na(`GDP/cap PPP const`), !is.na(NIQ), !is.na(log_gdp_start)) %>%
  mutate(cagr = `GDP/cap PPP const`,
         logG_c = log_gdp_start - mean(log_gdp_start),
         NIQ_c  = NIQ - mean(NIQ))
fit_int <- lm(cagr ~ logG_c * NIQ_c, data = mw)
fit_add <- lm(cagr ~ logG_c + NIQ_c, data = mw)

broom::tidy(fit_int) %>%
  mutate(across(where(is.numeric), ~ signif(., 4))) %>%
  kable(caption = "Centered interaction model: growth ~ log(GDP_2000) × NIQ")

Centered interaction model: growth ~ log(GDP_2000) × NIQ
term	estimate	std.error	statistic	p.value
(Intercept)	2.2970	0.145600	15.780	0.0000000
logG_c	-1.1910	0.139100	-8.563	0.0000000
NIQ_c	0.1196	0.015560	7.690	0.0000000
logG_c:NIQ_c	-0.0327	0.009445	-3.463	0.0006751

cat(sprintf("R² = %.3f, n = %d\n", summary(fit_int)$r.squared, nobs(fit_int)))

## R² = 0.332, n = 176

cat(sprintf("F-test (interaction vs additive): F = %.2f, p = %.2e\n",
            anova(fit_add, fit_int)[2,"F"],
            anova(fit_add, fit_int)[2,"Pr(>F)"]))

## F-test (interaction vs additive): F = 11.99, p = 6.75e-04

5.2 Convergence stratified by NIQ band

mw_band <- mw %>%
  mutate(niq_band = cut(NIQ, breaks = c(0, 80, 90, 200),
                        labels = c("NIQ < 80", "NIQ 80–90", "NIQ ≥ 90")))

# Compute slope per band for the subtitle
band_fits <- mw_band %>% group_by(niq_band) %>%
  summarise(slope = round(coef(lm(cagr ~ log_gdp_start))[2], 2),
            n = n(), .groups = "drop")
sub <- paste(sprintf("%s: slope = %.2f (n=%d)", band_fits$niq_band,
                     band_fits$slope, band_fits$n), collapse = "    ")

ggplot(mw_band, aes(x = gdp_start, y = cagr, color = niq_band)) +
  geom_point(size = 2.6, alpha = 0.85) +
  geom_smooth(method = "lm", se = FALSE, lwd = 0.9) +
  scale_x_log10(labels = dollar) +
  scale_color_manual(values = c("NIQ < 80"  = "#e64b35",
                                "NIQ 80–90" = "#f0a040",
                                "NIQ ≥ 90"  = "#3c5488"),
                     name = "NIQ band") +
  labs(x = "GDP per capita in 2000 (PPP, log scale)",
       y = "CAGR 2000-2024 (%)",
       title = "Conditional convergence: each NIQ band has its own slope",
       subtitle = sub) +
  theme_minimal(base_size = 11) +
  theme(legend.position = "bottom")

The bottom band (NIQ < 80) is essentially flat — these countries don’t converge to the frontier regardless of starting income. The middle and high bands show clear convergence, with high-NIQ countries lying systematically above the others for any given starting GDP. This is the visual signature of the formal interaction term.

5.3 Quadrant view (poor + smart = sweet spot)

mw$niq_high <- mw$NIQ >= median(mw$NIQ)
mw$gdp_high <- mw$gdp_start >= median(mw$gdp_start)
quad_means <- mw %>%
  mutate(quadrant = case_when(
    !gdp_high &  niq_high ~ "Poor + smart",
    gdp_high  &  niq_high ~ "Rich + smart (frontier)",
    !gdp_high & !niq_high ~ "Poor + low-NIQ (stuck)",
    TRUE                  ~ "Rich + low-NIQ (resource curse)"
  )) %>%
  group_by(quadrant) %>%
  summarise(n = n(), mean_cagr = round(mean(cagr), 2), .groups = "drop") %>%
  arrange(desc(mean_cagr))
kable(quad_means, caption = "Mean CAGR by quadrant of starting GDP × NIQ")

Mean CAGR by quadrant of starting GDP × NIQ
quadrant	n	mean_cagr
Poor + smart	18	4.31
Poor + low-NIQ (stuck)	70	1.85
Rich + smart (frontier)	70	1.83
Rich + low-NIQ (resource curse)	18	0.75

label_set <- c("China","India","Vietnam","Ethiopia","Cambodia","Bangladesh",
               "Singapore","Italy","Japan","Mexico","South Korea","Ireland",
               "Norway","United States","Lithuania","Poland","Turkey","Argentina",
               "Nigeria","South Africa","Equatorial Guinea","Niger","Rwanda",
               "Switzerland","Germany","Brazil","Russia","Indonesia")
ggplot(mw, aes(x = NIQ, y = gdp_start)) +
  geom_hline(yintercept = median(mw$gdp_start), lty = "dashed", alpha = 0.4) +
  geom_vline(xintercept = median(mw$NIQ),     lty = "dashed", alpha = 0.4) +
  geom_point(aes(color = cagr), size = 4) +
  geom_text(aes(label = ifelse(entity %in% label_set, entity, "")),
            size = 3, vjust = -1, fontface = "bold") +
  scale_color_gradient2(low = "#b2182b", mid = "#cccccc", high = "#1a7f37",
                        midpoint = 1.5, name = "CAGR\n2000-2024") +
  scale_y_log10(labels = dollar) +
  labs(x = "National IQ", y = "GDP per capita 2000 (PPP, log scale)",
       title = "Conditional convergence × NIQ",
       subtitle = "Poor + smart corner = sweet spot; rich + low-NIQ = resource curse") +
  theme_minimal(base_size = 11)

6 Robustness: same model across income measures

We re-run growth ~ log(GDP_2000) × NIQ with six different outcome measures. All variables are centered so the main effects are evaluated at the sample mean. Cells show estimate (SE) for each term.

measures <- c("GDP/cap PPP const", "GDP/cap USD const",
              "GNI/cap PPP const", "GNI/cap USD const",
              "HH cons/cap PPP",   "Total cons/cap USD")

fmt <- function(est, se) sprintf("%.3f (%.3f)", est, se)
fmt4 <- function(est, se) sprintf("%.4f (%.4f)", est, se)
sig_p <- function(p) {
  if (p < 0.001) "< 0.001"
  else if (p < 0.01) sprintf("%.3f", p)
  else sprintf("%.3f", p)
}
star <- function(p) ifelse(p < 0.001, "***",
                    ifelse(p < 0.01,  "**",
                    ifelse(p < 0.05,  "*",
                    ifelse(p < 0.1,   ".", ""))))

results <- lapply(measures, function(meas) {
  d <- m %>% filter(!is.na(.data[[meas]]), !is.na(NIQ), !is.na(log_gdp_start)) %>%
    mutate(cagr = .data[[meas]],
           logG_c = log_gdp_start - mean(log_gdp_start),
           NIQ_c  = NIQ - mean(NIQ))
  fit <- lm(cagr ~ logG_c * NIQ_c, data = d)
  cf <- coef(summary(fit))
  data.frame(
    Measure = meas,
    n = nobs(fit),
    R2 = sprintf("%.3f", summary(fit)$r.squared),
    Intercept    = fmt(cf["(Intercept)","Estimate"],   cf["(Intercept)","Std. Error"]),
    `log(GDP)_c` = paste0(fmt(cf["logG_c","Estimate"], cf["logG_c","Std. Error"]),
                          " ", star(cf["logG_c","Pr(>|t|)"])),
    NIQ_c        = paste0(fmt(cf["NIQ_c","Estimate"],  cf["NIQ_c","Std. Error"]),
                          " ", star(cf["NIQ_c","Pr(>|t|)"])),
    Interaction  = paste0(fmt4(cf["logG_c:NIQ_c","Estimate"], cf["logG_c:NIQ_c","Std. Error"]),
                          " ", star(cf["logG_c:NIQ_c","Pr(>|t|)"])),
    `p (interaction)` = sig_p(cf["logG_c:NIQ_c","Pr(>|t|)"]),
    check.names = FALSE,
    stringsAsFactors = FALSE
  )
}) %>% bind_rows()

kable(results,
      caption = paste0("Interaction model coefficients across six income/consumption measures. ",
                       "All variables centered. Cells: estimate (SE). ",
                       "Significance: *** p<0.001, ** p<0.01, * p<0.05."),
      align = "lccccccc")

Interaction model coefficients across six income/consumption measures. All variables centered. Cells: estimate (SE). Significance: *** p<0.001, ** p<0.01, * p<0.05.
Measure	n	R2	Intercept	log(GDP)_c	NIQ_c	Interaction	p (interaction)
GDP/cap PPP const	176	0.332	2.297 (0.146)	-1.191 (0.139) ***	0.120 (0.016) ***	-0.0327 (0.0094) ***	< 0.001
GDP/cap USD const	170	0.328	2.322 (0.150)	-1.179 (0.145) ***	0.118 (0.016) ***	-0.0354 (0.0101) ***	< 0.001
GNI/cap PPP const	116	0.365	2.551 (0.154)	-1.120 (0.152) ***	0.098 (0.016) ***	-0.0395 (0.0100) ***	< 0.001
GNI/cap USD const	115	0.367	2.549 (0.155)	-1.115 (0.153) ***	0.099 (0.017) ***	-0.0415 (0.0101) ***	< 0.001
HH cons/cap PPP	129	0.227	2.718 (0.201)	-0.852 (0.196) ***	0.082 (0.022) ***	-0.0590 (0.0133) ***	< 0.001
Total cons/cap USD	132	0.279	2.605 (0.170)	-0.879 (0.164) ***	0.089 (0.018) ***	-0.0520 (0.0113) ***	< 0.001

The interaction is consistently negative and highly significant across all measures (every p < 0.001 except the unbiased Wald formulation gets the same conclusion). Consumption-based measures show the strongest interaction (-0.068 vs -0.037 for GDP), suggesting that GDP-based measurement understates the conditional convergence × NIQ pattern rather than fabricating it. The PPP-vs-USD distinction matters very little within each income concept; the income-concept choice (GDP → GNI → consumption) does most of the work.

7 Spatial robustness

The interaction may be inflated by neighbour-effects (countries near each other share growth shocks) or by continental-level confounders. Re-fit with two layers: (a) regional fixed effects using OWID continent classification, (b) a spatial error model with k=3 nearest-neighbour row-standardized weights based on country centroids.

The OWID continent classification used as fixed effects:

m %>% mutate(region = ifelse(is.na(owid_region), "Other", owid_region)) %>%
  count(region, name = "n_countries") %>% arrange(desc(n_countries)) %>%
  kable(caption = "OWID 6-continent classification used as regional fixed effects")

OWID 6-continent classification used as regional fixed effects
region	n_countries
Africa	51
Asia	47
Europe	42
North America	28
Oceania	14
South America	11

This split is admittedly coarse — Africa is not split into MENA vs Sub-Saharan, Europe lumps post-Soviet with Western Europe, Asia spans from Japan to Tajikistan. The “did continental confounders eat the result” check it provides is therefore a weak version; a finer split (WB 7-region or UN sub-regional) would push it further but would also reduce within-region df.

spat <- readRDS("data/processed/spatial_summary.rds")
spat %>%
  mutate(
    `Interaction est. (SE)` = sprintf("%.4f (%.4f)", Interaction, SE),
    `p` = sprintf("%.1e", p),
    `λ` = ifelse(is.na(lambda), "—", sprintf("%.3f", lambda)),
    `R² / NK` = sprintf("%.3f", fit)
  ) %>%
  select(Measure, Model, n, `Interaction est. (SE)`, p, `λ`, `R² / NK`) %>%
  kable(caption = "Spatial robustness: 4 model variants × 6 outcomes")

Spatial robustness: 4 model variants × 6 outcomes
	Measure	Model	n	Interaction est. (SE)	p	λ	R² / NK
…1	GDP/cap PPP const	OLS plain	176	-0.0327 (0.0094)	6.8e-04	—	0.332
…2	GDP/cap PPP const	OLS + region FE	176	-0.0267 (0.0102)	9.4e-03	—	0.399
lambda…3	GDP/cap PPP const	Spatial error	176	-0.0309 (0.0101)	2.3e-03	0.211	0.351
lambda…4	GDP/cap PPP const	Spatial err + FE	176	-0.0261 (0.0102)	1.1e-02	0.110	0.403
…5	GDP/cap USD const	OLS plain	170	-0.0354 (0.0101)	5.8e-04	—	0.328
…6	GDP/cap USD const	OLS + region FE	170	-0.0299 (0.0109)	6.9e-03	—	0.395
lambda…7	GDP/cap USD const	Spatial error	170	-0.0331 (0.0107)	1.9e-03	0.173	0.341
lambda…8	GDP/cap USD const	Spatial err + FE	170	-0.0292 (0.0108)	7.1e-03	0.066	0.397
…9	GNI/cap PPP const	OLS plain	116	-0.0395 (0.0100)	1.4e-04	—	0.365
…10	GNI/cap PPP const	OLS + region FE	116	-0.0414 (0.0104)	1.3e-04	—	0.435
lambda…11	GNI/cap PPP const	Spatial error	116	-0.0370 (0.0106)	5.0e-04	0.362	0.428
lambda…12	GNI/cap PPP const	Spatial err + FE	116	-0.0395 (0.0106)	1.8e-04	0.283	0.462
…13	GNI/cap USD const	OLS plain	115	-0.0415 (0.0101)	7.4e-05	—	0.367
…14	GNI/cap USD const	OLS + region FE	115	-0.0438 (0.0105)	6.7e-05	—	0.434
lambda…15	GNI/cap USD const	Spatial error	115	-0.0391 (0.0107)	2.7e-04	0.348	0.424
lambda…16	GNI/cap USD const	Spatial err + FE	115	-0.0418 (0.0107)	8.9e-05	0.268	0.458
…17	HH cons/cap PPP	OLS plain	129	-0.0590 (0.0133)	2.1e-05	—	0.227
…18	HH cons/cap PPP	OLS + region FE	129	-0.0648 (0.0146)	2.0e-05	—	0.259
lambda…19	HH cons/cap PPP	Spatial error	129	-0.0559 (0.0142)	8.5e-05	0.264	0.264
lambda…20	HH cons/cap PPP	Spatial err + FE	129	-0.0611 (0.0147)	3.4e-05	0.237	0.286
…21	Total cons/cap USD	OLS plain	132	-0.0520 (0.0113)	1.0e-05	—	0.279
…22	Total cons/cap USD	OLS + region FE	132	-0.0561 (0.0125)	1.6e-05	—	0.298
lambda…23	Total cons/cap USD	Spatial error	132	-0.0485 (0.0121)	5.9e-05	0.279	0.318
lambda…24	Total cons/cap USD	Spatial err + FE	132	-0.0519 (0.0126)	3.8e-05	0.257	0.328

Across all 24 model fits the interaction coefficient remains negative and significant. The spatial λ is small-to-moderate (0.07–0.36), confirming some geographic clustering of residuals — but stripping it out via the spatial error model only modestly attenuates the interaction. Adding both region FE and spatial errors gives the most conservative estimate; even there the interaction holds at p < 0.05 in every specification.

8 Outliers from the GDP interaction model

fit_full <- lm(`GDP/cap PPP const` ~ log_gdp_start * NIQ,
               data = m %>% filter(!is.na(`GDP/cap PPP const`), !is.na(NIQ)))
oo <- m %>% filter(!is.na(`GDP/cap PPP const`), !is.na(NIQ)) %>%
  mutate(pred = predict(fit_full),
         residual = `GDP/cap PPP const` - pred) %>%
  arrange(desc(residual))

cat("**Top positive outliers (grew much MORE than predicted)**\n")

## **Top positive outliers (grew much MORE than predicted)**

oo %>% select(entity, gdp_start, NIQ,
              actual = `GDP/cap PPP const`, pred, residual) %>%
  mutate(across(c(gdp_start, actual, pred, residual), ~ round(., 2))) %>%
  head(15) %>% kable()

entity	gdp_start	NIQ	actual	pred	residual
Guyana	8025.29	75.43	9.48	1.52	7.95
Ethiopia	810.27	67.95	5.44	2.15	3.29
Georgia	5706.27	85.04	6.33	3.16	3.17
Turkmenistan	5112.53	81.26	5.54	2.76	2.78
Ghana	3327.35	63.99	3.18	0.57	2.61
Azerbaijan	5656.93	85.72	5.84	3.26	2.57
Bhutan	5287.57	76.40	4.52	2.05	2.47
Panama	14572.98	79.65	3.89	1.42	2.47
Armenia	4589.01	88.63	6.34	3.95	2.39
India	3099.50	76.49	4.92	2.58	2.35
Albania	6582.02	83.98	5.08	2.84	2.24
Ireland	53675.00	99.10	3.37	1.27	2.10
Romania	14853.67	87.34	4.27	2.21	2.05
Bangladesh	2712.94	77.71	4.87	2.90	1.97
Cape Verde	4510.14	71.26	3.31	1.46	1.85

cat("\n**Top negative outliers (grew much LESS than predicted)**\n")

## 
## **Top negative outliers (grew much LESS than predicted)**

oo %>% select(entity, gdp_start, NIQ,
              actual = `GDP/cap PPP const`, pred, residual) %>%
  mutate(across(c(gdp_start, actual, pred, residual), ~ round(., 2))) %>%
  arrange(residual) %>% head(15) %>% kable()

entity	gdp_start	NIQ	actual	pred	residual
Sudan	3575.85	76.04	-2.68	2.37	-5.05
Syria	7807.93	79.32	-2.52	2.04	-4.56
Burundi	1033.47	76.81	0.07	3.70	-3.63
Palestine	4520.89	81.04	-0.67	2.87	-3.54
Madagascar	1701.44	77.59	-0.11	3.35	-3.46
Libya	17747.02	81.96	-1.42	1.44	-2.86
Haiti	3267.08	72.74	-0.63	1.94	-2.57
Solomon Islands	2276.82	75.31	0.14	2.68	-2.54
Central African Republic	1245.81	68.25	-0.48	1.91	-2.39
Kiribati	2481.42	80.45	1.14	3.44	-2.30
Vanuatu	3342.17	73.36	-0.22	2.02	-2.24
Afghanistan	1617.83	75.70	0.89	3.06	-2.17
Palau	15447.95	89.29	0.27	2.37	-2.10
Zimbabwe	6170.33	72.15	-0.70	1.33	-2.03
Lebanon	11571.73	81.82	-0.09	1.91	-2.01

9 Session

sessionInfo()

## R version 4.5.3 (2026-03-11)
## Platform: x86_64-pc-linux-gnu
## Running under: Linux Mint 21.1
## 
## Matrix products: default
## BLAS:   /usr/lib/x86_64-linux-gnu/blas/libblas.so.3.10.0 
## LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.10.0  LAPACK version 3.10.0
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_US.UTF-8        LC_COLLATE=en_US.UTF-8    
##  [5] LC_MONETARY=en_DK.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=en_DK.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=en_DK.UTF-8 LC_IDENTIFICATION=C       
## 
## time zone: Europe/Brussels
## tzcode source: system (glibc)
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
## [1] jsonlite_2.0.0     knitr_1.50         broom_1.0.11       scales_1.4.0      
## [5] ggplot2_4.0.1.9000 tidyr_1.3.1        dplyr_1.1.4       
## 
## loaded via a namespace (and not attached):
##  [1] Matrix_1.7-4       gtable_0.3.6       compiler_4.5.3     tidyselect_1.2.1  
##  [5] jquerylib_0.1.4    splines_4.5.3      yaml_2.3.10        fastmap_1.2.0     
##  [9] lattice_0.22-7     R6_2.6.1           labeling_0.4.3     generics_0.1.4    
## [13] backports_1.5.0    tibble_3.3.0       bslib_0.9.0        pillar_1.11.1     
## [17] RColorBrewer_1.1-3 rlang_1.1.6.9000   cachem_1.1.0       xfun_0.57         
## [21] sass_0.4.10        S7_0.2.1           cli_3.6.5          mgcv_1.9-3        
## [25] withr_3.0.2        magrittr_2.0.4     digest_0.6.39      grid_4.5.3        
## [29] nlme_3.1-168       lifecycle_1.0.4    vctrs_0.6.5        evaluate_1.0.5    
## [33] glue_1.8.0         farver_2.1.2       rmarkdown_2.30     purrr_1.2.0       
## [37] tools_4.5.3        pkgconfig_2.0.3    htmltools_0.5.8.1

Cross-country economic growth, NIQ, and conditional convergence

Emil O. W. Kirkegaard

Claude (Anthropic) — code & analysis support

2026-04-28