Research
Research
Research on the political economy of development has generally relied on two primary sources of data on economic development: national accounts data collected by governments, and subnational wealth estimates based on nighttime luminosity (NTL) captured by high-orbit satellites. Recently, advances in machine learning architectures have made it possible to predict microspatial wealth with daytime satellite imagery, yielding performance far superior to nightlights. Here, we use a novel, high-resolution (1km) panel dataset (2003-2020) generated by a convolutional neural net (CNN) trained on daytime imagery to revisit key descriptive and model-based results on the causes and consequences of economic growth in Africa. We make three key contributions.
First, we show that concerns about non-random bias in national-level GDP are well-founded: estimates in resource-rich countries, as well as those with low statistical capacity are, significantly larger than those derived directly from our CNN-generated measure. We then revisit estimates of the association between development and several key explanatory variables, including regime type, strength of the rule of law, perceptions of corruption, and ethnic fractionalization; in some, but not all cases, we find economically and statistically meaningful differences in estimated effects.
Second, we are able to quantify the amount of wealth that NTL-based measures fail to capture due to strong left-censoring induced by the combination of a high luminosity threshold and selective use of electricity in poor communities. We find that NTL-based measures miss as much as 75 percent of countries’ total wealth, and that this figure is only partially explained by a country’s level of development as measured by GDP. Finally, we revisit several NTL-based results from the literature on the effects of ethnic inequality, coethnic favoritism, and national institutions on economic growth. We find substantively and statistically meaningful differences in estimated effects when using our CNN-derived measure compared to those that rely on nighttime lights. Our results suggest significant revisions to the prevailing narrative of the causes and consequences of Africa’s relatively poor economic performance.
How does climate volatility alter citizen demands, change voting behavior, and affect the long-term reputation of elected (and unelected) officials? Does this effect come primarily through the economic damages caused by climate volatility, or through alternative channels? Are they persistent or transitory? As climate volatility becomes more extreme, so too will its destabilizing impact on politics. Yet we know relatively little about effects on voting behavior, particularly in the developing world, and even less about downstream effects on the reputation of candidates and political institutions. Exploring the mechanisms behind these effects is also difficult due to a lack of data with the spatial and temporal resolution necessary for credible subnational analysis.
Here, we provide some of the first large-scale evidence on climate volatility’s effect on several measures of political accountability by combining several sources of survey data with high-resolution meteorological and climatic data. We also utilize a novel source of subnational economic data generated by combining remote sensing data with a convolutional neural network to generate annual, high-resolution estimates of growth at the 1x1km level for all of Africa. This ML-generated measure is a considerable improvement over nightlights-based alternatives, and permits credible mediation analysis linking negative political outcomes to climate volatility through reductions in economic growth. We supplement our focus on Africa with companion estimates from Latin America, exploiting variation in national-level institutions to examine whether they can explain the substantial effect heterogeneity we observe in our reduced-form results.
Survey harmonization — the process by which survey questions are made comparable across time and space — is a common task in applied social science research. However, most survey harmonization projects require extensive manual coding, making them prohibitively time-consuming and resource-intensive. In this paper, we develop an automated approach to survey harmonization. Specifically, we use sentence embeddings and a simple greedy algorithm to match questions and response levels across surveys. We apply our method to more than two decades of data from the Afrobarometer project. On the Afrobarometer data, our approach achieves near-human performance, and is able to combine thousands of survey questions in seconds. This work stands to significantly advance the practice of survey research. It also holds promise for other quantitative social science fields, including economic history, conflict studies, and public health research.
What do voters in Africa learn from poor performance? Existing observational work suggests that, when clarity of responsibility is sufficiently high, voters will reward politicians for increases in the quality of local public goods. At the same time, several large-scale information-accountability interventions have found that the provision of performance information has a null or, at best, weakly positive effect on voting behavior. I suggest that these seemingly contradictory findings are due to two complementary features of state and party strength. First, when the state lacks the technical expertise or resources to deliver large-scale public goods, poor performance is a noisy and potentially uninformative signal of candidate quality. Second, voter priors about candidate quality are both low in absolute terms and strongly determined by candidate behavior during pre-election campaigns. As a result, performance information often confirms voter priors about (low) candidate quality, yielding no change in voting behavior. To test this argument, I field a modified conjoint design that exploits the sequential revelation of information during the pre- and post-election periods. For policymakers considering information interventions, the design allows for a more nuanced understanding of when and under what conditions performance information will be valuable to voters, and when the reputational effects of that information will yield changes in voting behavior.
Steady improvements in ambient air quality in the US over the past several decades have led to large public health benefits, and the policies that helped drive these improvements are considered landmarks in successful environmental policymaking. However, recent trends in PM2.5 concentrations, a key pollutant, have stagnated or begun to reverse throughout much of the US. We quantify the contribution of wildfire smoke to these trends and find that since 2016, wildfire smoke has significantly slowed or reversed previous improvements in average annual PM2.5 concentrations in two-thirds of US states, eroding 23% of previous gains on average in those states (equivalent to 3.6 years of air quality progress) and over 50% in multiple western states. Smoke influence on trends in extreme PM2.5 concentrations is detectable by 2010 and is concentrated in a dozen western states. Wildfire-driven increases in ambient PM2.5 concentrations are unregulated under current air pollution law, and, absent additional intervention, wildfire's contribution to regional and national air quality trends is likely to grow as the climate continues to warm.
Paper available here.
Household electrification is thought to be an important part of a carbon neutral future, and could also have additional benefits to adopting households such as improved air quality. However, the effectiveness of specific electrification policies in reducing total emissions and boosting household livelihoods remains a crucial open question in both developed and developing countries. We investigated a transition of more than 750,000 households from gas to electric cookstoves - one of the most popular residential electrification strategies - in Ecuador following a program that promoted induction stoves, and assessed its impacts on electricity consumption, greenhouse gas emissions, and health. We estimate that the program resulted in a 5% increase in total residential electricity consumption between 2015 and 2021. By offsetting a commensurate amount of cooking gas combustion, we find that the program likely modestly reduced national greenhouse gas emissions, thanks in part to the country's electricity grid being 89% hydropower in later parts of the time period. Increased induction stove uptake was also associated with declines in all-cause and respiratory-related hospitalizations nationwide. These findings suggest that when the electricity grid is largely powered by renewables, gas-to-induction cooking transitions represent a promising way of amplifying the health and climate co-benefits of net-carbon-zero policies.
Paper available here.
Governments that rely on taxes, rather than aid and oil, have better governance on a range of measures, in part because direct taxation appears to increase citizens’ accountability demands. However, while existing evidence focuses on direct taxes, indirect taxes now comprise the majority of tax revenue worldwide. We argue here that taxation generates citizen accountability demands primarily when it is visible. Thus, relatively hidden indirect taxes are much less likely to mobilize popular pressure. Cross-national data analysis demonstrates that taxation’s effect on accountability applies mainly to direct taxes. Next, lab-in-the-field experiments in Uganda show that visibility critically determines whether taxation increases citizens’ ownership over budgets and willingness to punish low transfers from leaders. Finally, a survey experiment in Uganda indicates that even very common indirect taxes have low visibility to citizens. The findings suggest that growing budget reliance on indirect taxes may limit taxation's accountability dividends and thus may short circuit representation.
Paper available here.
In many regions of the world, sparse data on key economic outcomes inhibits the development, targeting, and evaluation of public policy. We demonstrate how advancements in satellite imagery and machine learning can help ameliorate these data and inference challenges. In the context of an expansion of the electrical grid across Uganda, we show how a combination of satellite imagery and computer vision can be used to develop local-level livelihood measurements appropriate for inferring the causal impact of electricity access on livelihoods. We then show how ML-based inference techniques deliver more reliable estimates of the causal impact of electrification than traditional alternatives when applied to these data. We estimate that grid access improves village-level asset wealth in rural Uganda by 0.17 standard deviations, more than doubling the growth rate over our study period relative to untreated areas. Our results provide country-scale evidence on the impact of a key infrastructure investment, and provide a low-cost, generalizable approach to future policy evaluation in data sparse environments.
Paper available here.
Conjoint analysis, which is a type of factorial design, has become popular among social scientists as a tool for measuring multidimensional preferences across several attributes. Because such experiments are based on multiple factors, each of which has several levels, researchers often focus on the average causal effect of a single attribute while marginalizing over the other attributes. What has been overlooked, however, is the fact that this so-called average marginal component effect (AMCE) critically relies upon the distribution of other attributes. This is problematic because most researchers use the uniform distribution when randomizing factors, and yet the population distribution of attributes in the real world is often far from uniform. Using an existing conjoint experiment and a simulation study, we demonstrate that the standard estimate of the AMCE can suffer from a substantial bias when the population distribution of attributes differs from the randomization distribution used in experiments. We address this problem by proposing a new experimental design and estimation method. The proposed methodologies are implemented through an open-source software package.
Paper available here.
Foreign aid may act much like oil money in reducing voters' willingness to to demand accountability from their government, enabling corruption, clientelism, and repression. This is an important causal mechanism connecting public budgets to quality of governance. Yet other scholarship counters that aid is more beneficial than oil, either indirectly because of donor oversight or directly because aid is more likely to produce citizen pressures on governments. Empirical work on the topic employs observational data at the national, macro level, and has left the question unresolved. At the micro level, in some countries citizens have experience with aid revenues and oil funds, thus possessing information about the political implications of these different revenue sources. This article provides the first experimental tests of the direct mechanism linking aid and oil revenues to demands from citizens for greater political accountability. We report the effects of randomly assigned treatments identifying aid funds compared to oil money on behavior of citizens in six survey and lab experiments in Ghana and Uganda. We find no differences in accountability pressures when subjects are randomly assigned to aid or oil conditions in any experiment, including a survey-based field experiment in Uganda that employed very strong information treatments on the extent of aid and oil funds. Aid may well have governance effects through the indirect route of donor oversight, but the results presented here suggest no evidence that aid, compared to oil, directly induces greater accountability demands among citizens.
Paper available here.
Government accountability is severely lacking in many developing countries, yet we know relatively little about the causal dynamics that produce citizen demands for greater responsiveness. We argue that psychological ownership over public money drives governance expectations. We offer a new theory of ownership and accountability and apply it in sub-Saharan Africa. Results from a series of lab-in-the-field experiments in Ghana and Uganda and from a nationally representative survey-based field experiment in Uganda demonstrate that higher feelings of ownership over public revenues significantly increase citizens’ demands on leaders. Furthermore, simple interventions can significantly increase feelings of revenue ownership over oil and aid windfalls, producing accountability pressures indistinguishable from taxes.
Paper available here.
This study tests whether citizens will more readily demand accountability from governments for taxes than for non-tax revenue from oil or aid. Two identical experiments on large, representative subject pools in Ghana and Uganda probe the effects of different revenue types on citizens' actions to monitor government spending. A similar experiment on more than 500 members of parliament from the two countries examined their views toward these revenue sources. Roughly half of all citizens willingly sign petitions and donate money in order to scrutinize all three sources. However, neither Ghanaians nor Ugandans are more likely to take action for tax revenues than for oil or aid when the money is delivered directly to the government. Members of parliament in both countries likewise saw no difference among the three sources. Results also suggest no differences among taxes, oil and aid in citizens' perceptions of misappropriation risk or public-goods provision. However, citizens do differentiate more between revenue delivered directly to the government compared to money given to an NGO. Findings are robust to numerous alternative specifications and subgroup partitions including taxpayers vs. non-taxpayers. Focusing on individual citizens and elites, we show little evidence that taxes strengthen citizens' demands for accountability in two developing countries or that MPs perceive differences in control or public benefits across revenue sources.
Paper available here.
The regression discontinuity (RD) design has become increasingly popular among social scientists. One prominent application is the study of close elections. We explicate several methodological misunderstandings widespread across disciplines by revisiting the controversy concerning the validity of RD design when applied to close elections. Although many researchers invoke the local or as-if-random assumption near the threshold, it is more stringent than the required continuity assumption. We show that this seemingly subtle point determines the appropriateness of various statistical methods and changes our understanding of how sorting invalidates the design. When multiple-testing problems are also addressed, we find that evidence for sorting in US House elections is substantially weaker and highly dependent on estimation methods. Finally, we caution that despite the temptation to improve the external validity, the extrapolation of RD estimates away from the threshold sacrifices the design’s advantage in internal validity.
Paper available here.
The strong, negative relationship between ethnolinguistic diversity and economic outcomes is one of the most robust findings in the empirical political economy literature, and many of its most important entries come from analysis of sub-Saharan Africa. In this chapter, we identify several distinct mechanisms by which ethnolinguistic diversity affects economic performance in Africa, including, among others, an increased likelihood of armed conflict, the underprovision of public goods, and political dysfunction and instability. We then provide a two-part explanation for why Africa in particular is home to such high levels of ethno-linguistic diversity. Drawing on work in anthropology, political science, economics and history, our explanation emphasizes Africa's unique role in early human evolution and the arbitrary borders of its modern states. We conclude by offering suggestions for fruitful future research and discuss the policy implications of existing work.
Paper available here.
Existing work on enumerator effects has focused largely on survey contexts where enumerators have been shown to activate social desirability bias and affect non-response rates. In experimental settings, however, enumerators do far more than administer surveys; they lead discussion groups, implement complex experimental protocols, and deliver key pieces of information to subjects. Yet relatively little attention has been paid to the consequences of enumerator-induced variation in the delivery and efficacy of a treatment. In this article, we evaluate the inferential consequences and substantive magnitude of enumerator-induced treatment effect heterogeneity. We first show that, contrary to conventional wisdom, enumerator effects have inferential implications under many common experimental designs. We then use a combination of original and publicly available data from 12 experiments—including survey, lab-in-the-field and large-scale field interventions—to provide what is to our knowledge the first systematic evidence of the size and distribution of enumerator effects in experimental studies. We conclude by proposing a split-pot design and accompanying treatment assignment algorithm designed to reduce enumerator effects.
Currently under revision, preliminary draft available upon request.
Why do we observe vote buying under the secret ballot? Prevailing wisdom suggests that candidates use vote buying to signal their willingness to provide desirable goods post-election. In contrast, using a combination of observational and experimental evidence from two large, original datasets in Uganda and Ghana, I find that offers of cash or gifts before an election have an unambiguously negative effect on candidate reputation and the reputation of closely aligned political actors. I argue that we observe widespread vote buying in low-credibility environments because voters attempt to maximize the value of current and future elections by demanding offers as the price of their turnout. These results suggest that elections in low-credibility environments may cause candidates to sacrifice long-term reputation for short-term electoral gain, and may lead voters to reward one-time offers of cash or gifts at the expense of long-term public goods provision.
Currently under revision, draft available on request.
Existing research in clientelism has paid special attention to the provision of cash and gifts before elections. Yet candidate strategies also frequently include promises of to provide public goods after the election, even in states where resource and capacity constraints make such provision unlikely. What do candidates promise, and how often are those promises fulfilled? This paper uses a combination of survey and experimental data in Ghana and Uganda to evaluate the causes and consequences of making promises in a low-credibility environment. First, I show that promises of ex post provision are by far the most common strategy observed by voters, surpassing in their frequency all other remaining strategies, including vote buying and the provision of quasi-public goods such as scholarships. I then demonstrate that promise-making produces anchoring effects among voters, as voters price the expectation of future provision into their evaluations of candidates. Given that over-promising creates such strong anchoring effects, why do politicians continue to do it? I argue that the prevalence of over-promising is due to a time-inconsistency problem facing candidates during the campaign period. Using a conjoint experiment, I demonstrate that politicians are unlikely to suffer electoral costs for over-promising, and in some cases can actually benefit from doing so. This dynamic gives candidates few incentives to reign in their promise-making during the pre-election period, particularly when the financial returns to holding office are very large.
Currently under revision, draft available upon request.
Contact
Y2E2 Building Room 364
473 Via Ortega Way
Stanford, CA 94305
brandon.delacuesta
@stanford.edu