RESEARCH DESIGN Data I study the degree of local poverty and economic activity at the village (phum) level in Kampong Speu province. Villages are the smallest unit; in Kampong Speu province, the median size of a village was 401 persons and 78 households in 1998. The compact size of villages provide a reasonable approximation of households distance to the border that separated the West and Southwest.Footnote 6 I compare the Southwest and the West within Kampong Speu for three reasons. First, the regions represent the ideological polarization within the regime; Mok was the staunchest ally of the party’s genocidal faction, and Sy was a typical moderate communist. Second, the border between zones did not overlap with other political boundaries unlike other DK zones, meaning a discrete change in the outcomes can be attributed to different zone leadership rather than differences between provinces. This mitigates the “compound treatment problem” commonly found in research designs that rely on geography. Third, the density of observations within a narrow bandwidth around the border means villages in the Southwest are being compared to an appropriate counterfactual, which is not true at other boundaries where villages are further from borders, or where the natural borders were excessively wide, like divisions created using the Mekong River. Poverty data are from the Cambodian National Poverty Identification System (IDPoor). The data are collected in 2011 through a 16 question household survey conducted by elected village representatives who use questions regarding assets, living standards, means of transport, employment, education, and health to score household poverty on a continuous scale which is subsequently used to construct poverty categories. I measure the percentage of poor households. Further detail on the data collection process of IDPoor is available in Figure A.1 in the Supplementary Material. Second, I use data on the nighttime luminosity of villages to estimate GDP within a $ 2\hskip0.3em \mathrm{km}\times 2 $ km grid cell surrounding the village center (Ghosh et al. Reference Ghosh, Powell, Elvidge, Baugh, Sutton and Anderson2010). Luminosity proxies both formal and informal economic activity, which is important given the role of informal commerce. I use the inverse hyperbolic sine (IHS) transformation to account for skewness and zeros.Footnote 7 Empirical Strategy The nature of assignment into the Southwest versus West suggests a regression discontinuity (RD) approach which compares modern outcomes between nearby villages on either side of the boundary. The RD estimator is (1) $$ {\displaystyle \begin{array}{l}{y}_v=\alpha +\gamma {\mathrm{SW}\ \mathrm{Zone}}_v+f\left({\mathrm{geographic}\ \mathrm{location}}_v\right)\\ {}\hskip1.7pc +\hskip2px \sum_{s=1}^nse{g}_v^s+{\varepsilon}_v\hskip5.55pt for\hskip5.55pt v\in bw,\end{array}} $$ where $ {y}_v $ is the outcome of interest, and $ {\mathrm{SW}\ \mathrm{Zone}}_v $ is a binary indicator scored 1 if a village was in the Southwest and 0 otherwise. $ f({\mathrm{geographic}\ \mathrm{location}}_v) $ is the forcing variable, which I define as the distance between the village v and the border that divided the Southwest and West.Footnote 8 I evaluate Equation 1 using distance in a linear and quadratic form. I follow Dell (Reference Dell2010) and include boundary segment fixed effects $ {\sum}_{s=1}^nse{g}_v^s $ , which are computed by splitting the border into n segments s and then scoring 1 if a village is closest to segment s. Segment fixed effects compare villages that lie in the same neighborhood around the border, avoiding imprecise comparisons that may occur if villages have similar absolute distances at extreme ends of the boundary.Footnote 9 The robust error term is $ {\varepsilon}_v $ . I adjust standard errors for spatial heteroskedasticity and autocorrelation (HAC), following the data-driven procedure developed by Kelly (Reference Kelly2020) for selecting a HAC spatial kernel based on the spatial structure of each respective outcome. I estimate Equation 1 within narrow MSE optimal bandwidths $ bw $ (Calonico, Cattaneo, and Titiunik Reference Calonico, Cattaneo and Titiunik2014), however, results are robust to alternative bandwidth choices. I include the distance to the provincial capital as an adjusting covariate in some regressions to account for proximity to the provincial center. The RD approach is advantageous relative to a selection on observables strategy. Since DK demographically targeted repression against the more well-off, repression levels are likely correlated with positive developmental trajectories, leading to an upward biased estimate of the effect on repression. Since administrative microdata before the regime was destroyed by DK cadres, one cannot adequately adjust for pre-existing development levels. Since my RD design compares villages which were arbitrarily split into more or less extreme administrative zones, in expectation, confounding factors ought to be similar between villages within a narrow bandwidth, conditional on location. So long as the design assumptions hold, my strategy identifies a local average treatment effect (LATE) of exposure to an extreme DK administrator. I turn to discussing these assumptions now. Design Assumptions and Inferential Threats My design relies on two core assumptions: the smoothness of confounders at the discontinuity, and the absence of strategic line drawing to sort observations in a particular way. Balance Tests The assumption that confounders are smooth at the discontinuity is reasonable given the arbitrary placement of the road with respect to Kampong Speu. The historical record suggests one confounding feature: the Pol Pot faction wished to divide territory in a way to give Mok control over more productive areas and Sy with less well-endowed land, a goal that would be accomplished by splitting the relatively soil-poor Kampong Speu in two, leaving Mok with the entirety of the soil-rich Kampot province and leaving Sy with the more desolate Koh Kong and the more rugged areas in the far-north of Kampong Speu (Chandler, Kiernan, and Boua Reference Chandler, Kiernan and Boua1988; Vickery Reference Vickery1984). Geographic variables are “slow moving” in the sense that they vary little within small bandwidths; given the arbitrary placement of the road, there is no reason to believe differences in productivity are sharp at the discontinuity despite the fact Mok was given more rich endowments in the far South outside of Kampong Speu. I test balance on agro-economic, climatic, and topographic area when crossing from one side of the border to another on $ 1\hskip0.3em \mathrm{km}\times 1 $ km grid cells. Figure 4 shows the mean and variance of temperature and rainfall, forest and cropland cover, ruggedness, elevation, and soil fertility are similar on either side of the boundary. Critically, since the Cambodian economy was agriculture based in the lead-up to the Khmer Rouge, the similarity of factors of production on either side of the border suggest villages had similar access to sources of productive income. DK destroyed micro data from the 1962 Cambodian census, making balance tests on predetermined socioeconomic variables difficult. I use three data sources to probe for pre-existing economic differences from satellite data: estimated population, built-up areas, and road networks in 1975. As Figure 4 shows the equivalence confidence interval contains zero for these outcomes, however, the substantive size of the estimate could suggest some degree of initial imbalance. There is no reason to regard this imbalance as systematic evidence of manipulation rather than chance: first, it would make little sense for a border to be strategically drawn to give built-up areas—the locations DK was most concerned about repressing—to a moderate. Second, to the extent road density could be higher in the former Southwest, increased market access should boost development for former-Southwest villages, suggesting bias in the opposite direction of my main prediction. Balance tests are substantively similar with nonparametric estimator within an MSE optimal bandwidth (Figure A.1 in the Supplementary Material). Sorting Test Another possibility is that the road was chosen as the border because it would give Mok more authority. This concern is somewhat mitigated by the fact that the border was a natural landmark: had DK drawn a particular line, the potential for strategic line drawing would be more first order. If it was the case NR4 was chosen strategically with respect to localities, one would expect a discontinuity in the number of villages at the boundary, specifically with more villages under the authority of Mok rather than Sy. I test for strategic sorting of villages along the border, and I find the density of the running variable is continuous (Figure A.2 in the Supplementary Material).
BASELINE RESULTS Table 2 contains results from Equation 1, which shows a substantively large and statistically significant difference between villages in the Southwest and West. Columns 1–4 refer to the poverty outcome, whereas 5–8 refer to night lights. Columns 1, 2, 5, and 6 use a linear forcing variable and columns 3, 4, 7, and 8 use the squared term. Descriptive RD plots are in E.6 in the Dataverse Appendix.Footnote 10 The results imply a meaningful increase in poverty and reduction in luminosity at the discontinuity. Conservative estimates in columns 1 and 3 suggest poverty increased by 4%—0.41 standard deviations and 20% of the control average—in the Southwest. Meanwhile, nighttime lights decreased by 0.6 standard deviations in the Southwest, which is consistent with the increase in poverty, and is robust to alternative aggregation techniques (Figure B.6 in the Supplementary Material). The inclusion of segment fixed effects and alternative functional forms of the running variable do not meaningfully impact the estimates, nor does adjusting for distance to Chbar Mon, the provincial capital. Threats to Inference Confounding (Observable and Unobservable): Sensitivity Analysis Although regression discontinuity is a credible research design, my study is observational, and adjusting for pre-existing covariates may be important to account for precision or omitted variable bias. First, I show results are robust to covariate adjustment: including the density of roads and built-up area in 1975 does not meaningfully impact the estimates (C.1 in the Supplementary Material). Second, I do a sensitivity analysis to assess how severe unobservable selection would need to be in order to overturn the main findings (Cinelli and Hazlett Reference Cinelli and Hazlett2020). In general, I find an unobservable feature would need to explain at least 10% of the residual variation in exposure to Mok and modern development. Since such a confounder would need to be up to four times stronger than built up area in 1975, a strong predictor, it is implausible that a covariate so large is driving the finding (B.4 in the Supplementary Material). Road Effect: A Placebo Case Study Roads themselves may be engines of commerce. To the extent distance to a road matters for development, my main specification flexibly controls for this by including a polynomial in village proximity to the highway. Since my focus is on the local discontinuity in development driven by being on one side of the road versus the other, for the road to explain away the main result, it must be the case that being on one side determines a change in development. Roads may divide areas which follow different developmental trajectories due to separation and isolation from one another, in which case, a spatial discontinuity in development may emerge by virtue of the border being a road rather than differences in administration. I test this possibility using National Road 3 (NR3) in Kampot Province as a placebo case study. NR3 bisected Kampot in a similar way as NR4, but the entire province was in the Southwest zone during DK. Available qualitative evidence suggests the road separated different communities before the regime; “[i]n some places the line of demarcation between the two kinds of peasantry was apparently quite clear…north of the road running between Chhouk and Kampot the population was isolated, hostile to everything urban, and, incidentally, revolutionary from long before 1970, while south of that road the peasants interacted with the market, were familiar with urban ways, and considered themselves part of wider Cambodian society” (Vickery Reference Vickery1984, 4). I find no evidence of a discontinuity along NR3, suggesting that roads do not generically predict discrete shifts in development across space. The absence of an effect in the context of NR3 increases our confidence that the main finding is driven by the administrative boundary NR4 represented rather than the fact it is a road (B.5 in the Supplementary Material). Civil Conflict Legacy: Falsification Test The legacy could be driven by civil conflict legacies rather than state repression. This explanation is implausible, since both zones were one combat theater during the civil war (1970–75). Civil war violence outside of state repression must change underlying factors of production to have a persistent effect (Blattman and Miguel Reference Blattman and Miguel2010). In Cambodia, the most plausible way this could occur would be if explosive remnants of war (ERW, landmines or bombs) were differently allocated between zones, degrading land (Lin Reference Lin2022). I find no evidence of differential ERW (B.5 in the Supplementary Material). Robustness I probe the robustness of my findings in several ways. DHS Wealth Data To validate my measure of poverty at the village level with administrative data, I replicate my findings using survey data at the individual level collected by the Demographic Health Survey from 2000 to 2014. I show rural household wealth is lower in the former Southwest (B.1 in the Supplementary Material). Two-Dimensional Forcing Variable Nonlinear spatial trends could be mistaken for discrete change in income levels if the univariate forcing variable masks higher-order changes across latitude–longitude space. I include a two-dimensional forcing variable (B.1 in the Supplementary Material) and estimate the RD along border points to account for this (B.2 in the Supplementary Material). Power Analysis One may be concerned that the number of observations is small in a narrow neighborhood, reducing the statistical power of the tests. Since the effect size I uncover is substantial, not many observations are required. I show that I have sufficient power within an MSE optimal bandwidth to detect the main effects (B.2 in the Supplementary Material). Falsification Test Spatial autocorrelation could explain the finding if village development clusters geographically. I show the bias-corrected CCT standard errors are robust to spatial noise simulations which create synthetic outcome data following the same spatial structure of the true data (Kelly Reference Kelly2020) (Figure B.8 in the Supplementary Material). Donut RD I estimate a series of donut-hole RDs, dropping observations close to the border, and find similar results even after 10% of the data are dropped (Figure B.8 in the Supplementary Material). The result guards against the possibility that when approaching the road, villages experience a differential positive development shock unrelated to DK.
MECHANISMS Repression and Poverty Traps: Conceptual Framework Violence targeted at higher educated segments of society creates a skill gap between generations, leaving younger people without mentors who can transfer skills and knowledge. This type of violence can be found across autocracies and coerced labor regimes: autocrats prefer low-skilled loyalists to competent persons to extend their survival (Egorov and Sonin Reference Egorov and Sonin2011) and under coerced labor, principals are more violent toward productive and skilled persons, who have a larger outside option (Acemoglu and Wolitzky Reference Acemoglu and Wolitzky2011). The mechanism I study is therefore plausible in other contexts. In the next sections, I show how poverty becomes self-perpetuating and persistent due to repression: a poverty trap. Educational attainment is lower in the former Southwest, with a particularly strong drop among the age cohort whose secondary schooling years were interrupted by the regime. This created an intergenerational poverty trap in two ways. Individuals in the Southwest were pushed into informal employment, which earns less income. The evidence is consistent with a model of poverty wherein individuals at time t remain poor in $ t+1 $ because their low income forces them into making decisions that reinforce their poverty, such as working low-earning jobs (Banerjee and Duflo Reference Banerjee and Duflo2011). Further, children born after the regime in the Southwest have worse health outcomes, which strongly predicts future income. This evidence highlights the intergenerational nature of the poverty trap: although children born in the former Southwest were never repressed by DK, they face deprivation because their parents were driven into poverty, perpetuating the cycle. Human Capital Declined after DK First, I evaluate whether trends in schooling by age changed in the Southwest zone. If human capital differentially declined in the Southwest due to the Khmer Rouge, one should observe relatively similar levels of schooling among age cohorts who finished schooling before the regime along with a sharp decline in schooling among villagers whose school-age years overlapped with the regime. I test this argument using a difference-in-differences design leveraging microdata on individual schooling and age from Cambodian Labor Force Surveys in 2000/2001. The common trends assumption is plausible in this setting, since all villagers educated before DK were in the same province, meaning the institutional differences between regions only emerged after the regime. I placebo test this assumption, regressing schooling on a series of separate cohort dummies among persons past schooling age in 1975, finding no evidence of large or significant breaks in educational trends (Table D.6 in the Supplementary Material). I estimate the following model: (2) $$ \begin{array}{rl}{y}_{i,(v),(c)}={\displaystyle \sum_{c
e 20-24}^C}{\beta}_c\left(\right.{\mathrm{Cohort}}_{i,(c)}\times 1\{S{W}_{i,(v)}\}\left)\right.+{\mu}_v\hskip4.5em +\hskip0.35em {\lambda}_{iy}+{\displaystyle \sum_{k=1}^K}{\alpha}_k{x}_i^k+{\epsilon}_{i,(v),(c).}& \end{array} $$ The outcome is years of schooling, $ {y}_{i,(v),(c)} $ , measured for individual i in village v among age cohort c. The coefficients of interest are $ {\beta}_c $ , which capture the differential effect of an individual living in the Southwest zone in an age cohort who would have had primary or secondary schooling after DK, relative to the age cohort who would have completed primary or secondary schooling before the regime (aged 20–24 in 1975). Note $ {\beta}_c $ captures both pre-trends in educational differences among older cohorts and the dynamic effect of having overlapped with Mok’s rule during schooling age. I include birth-decade-by-commune fixed effects $ {\lambda}_{iy} $ which absorb decadal educational trends over space, village fixed-effects ( $ {\mu}_v $ ) to adjust for village-invariant factors, and K individual controls $ {\sum}_{k=1}^K{\alpha}_k{x}_i^k $ including age and its square and respondent gender. Standard errors are clustered at the village. Figure 5 graphically displays the identification approach and results. Figure 5a shows the average years of schooling by cohort per zone, illustrating a (fitted) parallel trend in the pre period. Education levels then sharply decline in the former Southwest zone among 12- to 17-year olds (in 1975 years). Figure 5b corroborates the descriptive trend, showing the absence of a difference among older age cohorts between zones and a transitory decline in educational attainment for persons who were schooling age when the regime began. The evidence suggests schooling fell among school-aged Cambodians in the Southwest. If human capital decline persists overtime, one should observe lower human capital levels at the village level between zones. Table 3 shows the share of persons who have never attended school increases, whereas the literacy rate (of those over 15 years of age) declines. The estimates are largely commensurate with one another; whereas the percentage of persons who never attended school increases by 7% in the baseline, the share of literate persons over 15 declines by 8%. The results are substantively large, near 0.5 $ \sigma $ in the baseline estimates. In Table D.1 in the Supplementary Material, I show years of education decline by a year on average, and school attendance declines by 6% in 1998. The result is robust to adjusting for distance to schools (Table D.2 in the Supplementary Material). The education gap persists into 2008 through tertiary and secondary education (Table D.4 in the Supplementary Material). The impacts of repression on schooling could vary by gender: parents may invest less in schooling daughters in the aftermath of repression, pushing female family members into care-taking roles instead of education. The pattern of lower educational attainment for females is consistent across Cambodia. Notably, this mechanism would not be a rival account to a poverty trap, rather, it would be a dimension by which the poverty trap is perpetuated. I find no consistent evidence of this pattern (Table D.5 in the Supplementary Material). Labor Market Outcomes Qualitative evidence suggests the absence of educated and skilled persons drove individuals to work in low-paying jobs (Jeong Reference Jeong2014). In the Cambodian context, less schooling strongly predicts individuals being “own account workers”—self employed, typically working agricultural jobs on small-scale family farms or in otherwise informal roles. This line of work is highly labor intensive, has a low level of productivity, and involves low levels of skill and technology (Arnold Reference Arnold2008). If the decline in education caused by DK repression reshaped local labor markets by pushing individuals into low-earning informal agricultural work, one may expect an increase in the probability an individual is an own-account worker and a commensurate decline in earnings and productivity. Table 4 shows findings consistent with this pattern. Rural persons in the Southwest are far more likely to be own account workers—self-employed informal laborers (column 1). Consistent with broader patterns of employment and earnings, column 2 shows lower income from work as well. Finally, column 3 shows earnings per hour are also lower, meaning productivity for workers also diminishes. The evidence is consistent with qualitative accounts of how state repression shaped labor markets and workers in the wake of the human capital shock from the regime. In Table D.8 in the Supplementary Material, I show own-account workers are less educated and earn less. Intergenerational Consequences Economic impacts can reverberate across generations by reducing child health. Childhood health determines later-life income levels and is partially determined by maternal economic well-being (Bleakley Reference Bleakley2010). I evaluate how the DK shock impacted subsequent generations by exploring the health of children between zones with four rounds (2000, 2005, 2010, and 2014) of Demographic Health Survey (DHS) data. Parents who lost schooling and were therefore poor as a result of Mok’s rule may have less healthy children as a result of their low income. Finding worse health outcomes for youth may highlight how repression creates persistent, negative human capital consequences beyond schooling and across generations that were never exposed to violence. DHS randomly selects a subset of respondents and measures three critical dimensions of child health for persons aged 3–5: height for age (a measure of stunting), weight for age (a measure of wasting), and weight for height (a measure of being underweight). I create an index of health scores using the first principal component of these measures and evaluate each measure individually, and then estimate a version of Equation 1 which includes wave year fixed effects and maternal controls among rural households to maximize comparability. Table 5 shows the health index declines by nearly 0.8 $ \sigma $ . The effect is driven by underweight and wasting children, rather than stunting, which suggests childhood food poverty drives health differences. In Table D.7 in the Supplementary Material, I show the effect is orders of magnitude larger for mothers without formal schooling, and that the difference between zones attenuates when mothers have some education, suggesting the education shock from the regime plays a crucial role in explaining health differences. Alternative Mechanisms Social Capital Cultural persistence via social trust has been shown to be an important persistence channel linking coerced labor to modern development (Lichter, Löffler, and Siegloch Reference Lichter, Löffler and Siegloch2021; Lowes and Montero Reference Lowes and Montero2021; Nunn and Wantchekon Reference Nunn and Wantchekon2011). I use data from Cambodia’s Violence Against Women survey, which asks respondent’s three questions about communal trust and social cohesion including: (1) “Do neighbors in COMMUNITY NAME generally tend to know each other well?,” (2) “If there were a street fight in COMMUNITY NAME would people generally do something to stop it?,” and (3) “If someone in your family suddenly fell ill or had an accident, would your neighbors offer to help?” I code 1 if a respondent answers “yes” and 0 otherwise. There is little variation across villages in different zones along these dimensions, with respondent’s reporting affirmative answers at high and nearly identical rates between zones (Table 6). This suggests social capital cannot explain the developmental divergence. Property Rights Weak property rights institutions are another channel commonly cited in the literature (Acemoglu, Johnson, and Robinson Reference Acemoglu, Johnson and Robinson2001; Dell Reference Dell2010). Since neither the DK ban on private property nor collectivization has persisted, it is unlikely formal institutional persistence explains contemporary maldevelopment. However, the destruction of records of land ownership could have increased the contestability of land, creating tenure insecurity and poverty. Social conflict as a result of extractive institutions has been shown to be a crucial persistence mechanism (Guardado Reference Guardado2018). To measure respect for property rights, I collect data on the count of village land disputes from Commune Database Online. This measure captures a highly salient aspect of respect for property in the Cambodian context, where the highly agrarian economy has seen increasing poverty due to land grabs and unclear titling (Kerbo Reference Kerbo2014). I find no difference in land disputes at the discontinuity (Table 7). Political Legacy A third class of explanations relates to the political dominance of former regime elites; historical coercive institutions may create politically uncompetitive environments, which outlast initial conditions. As explained, this mechanism is unlikely, as Mok and his cadres were driven from the region during Vietnam’s invasion. I measure the competitiveness of commune council elections in 2012 and 2017 (Herfindahl index, higher values represent less competitiveness) and the vote share for the ruling CPP in 2012/2017. Data are only reported at the commune level, of which there are 87 in Kampong Speu. Since the units are much larger than villages, I include all communes in the RD, but also include district fixed effects to absorb spatial heterogeneity. Results in Table 8 show local elections are largely similar between zones. To the extent competitiveness of elections is different, elections appeared more competitive in the Southwest in 2017, where vote shares were less concentrated by 3.6 points. As such, political competition and partisan support are unlikely explanations. Another observable implication of elite capture and corruption—a key persistence mechanism in the study of extractive institutional persistence—is lower levels of public goods. In Table C.2 in the Supplementary Material, I show villages on either side of the border have similar access to hospitals, schools, and commune centers. Migration Migration could explain my findings if a substantial proportion of wealthy individuals fled the (former) Southwest zone after the regime for the West. While this would not invalidate the design, it would mean the primary persistence channel was selective migration of more well-off civilians rather than an intergenerational human capital shock. I treat migration as a post-treatment variable, and back out how large selective migration would need to be in order to explain the result under weak assumptions and a conservative trimming exercise. Selective migrants would need to occupy the entire top 25% of the wealth distribution in the rural West to explain the result—an implausibly high proportion considering low rates of rural—rural migration (See Table C.4 and Figure C.1 in the Supplementary Material for results and Section G in the Dataverse Appendix for explanation).