Revisiting the effect of democracy on population health

Abstract

I use a novel dichotomous measure of democracy to simulate a quasi-natural experiment and implement a difference-in-differences analysis to identify the heterogeneous treatment effect of democracy on population health across countries from 1960 to 2010. To counteract potential sources of bias resulting from unparallel and stochastic trends between treated and control units, I adopt a principal components difference-in-differences estimator that exploits factor proxies constructed from control units to account for unobserved trends. The main results indicate that countries that transitioned from non-democracy to democracy are more likely to experience health improvements, compared to countries retaining non-democratic institutions. However, the health-enhancing impact of democratization turns out to be much smaller in size than previously established. I posit that conventional estimates exaggerate the economic significance of the health returns to democratization due to inadequate attention to cross-border spillovers, global common shocks, and worldwide heterogeneity in the democracy-health nexus.

‘Whenever health depends on collective action—whether through public works, the provision of health care or education—politics must play a role’
—Deaton (2013: 98)

1. Introduction

The conventional wisdom is that democratic institutions—including competitive elections, the separation of power among distinct bodies of government, and the protection of civil liberties allowing citizens to scrutinize government policies—augment a government’s decision to deliver public healthcare in a way that improves health status of a majority of low- and middle-income segments of the population. This argument implies that a transition from non-democracy to democracy is central to driving health improvements. To the extent that health capital is a key conduit of long-run economic growth (Knowles and Owen 1995, 1997; Bloom, Canning, and Sevilla 2004), the widely acknowledged health-enhancing impact of democratization provides a partial explanation for why democratic countries, relative to their non-democratic counterparts, are more likely to experience robust economic growth (Papaioannou and Siourounis 2008; Madsen, Raschky, and Skali 2015; Acemoglu et al. 2019; Eberhardt 2022).

The empirical literature on the importance of democratic institutions for health gains offers highly mixed findings (Supplementary Appendix A1). Many studies document evidence supporting a positive association between democracy and population health within a cross-country framework (Besley and Kudamatsu 2006; Annaka and Higashijima 2021; Batinti, Costa-Font, and Hatton 2022), whereas others demonstrate that democracy has a negligible impact on health outcomes across countries (Ross 2006; Gerring, Thacker, and Alfaro 2012). Existing estimates, however, may not reflect causal inference because democracies differ widely from non-democracies in numerous characteristics (i.e. institutions, cultures, and histories) that may affect population health. Kudamatsu (2012) attempts to address this concern by using micro data and establishes empirically the positive influence of democratization on individual health outcomes in sub-Saharan Africa. Subnational evidence, albeit largely immune to concerns about unobserved heterogeneities across countries, does not provide a generalized understanding of the role of democracy in driving global income differences via its influence on national health status. Under the presence of diminishing returns to effective units of labour, individual-level estimates presumably exaggerate the economy-wide growth impact of democratization attributed to improved aggregate health capital (Acemoglu and Johnson 2007).

Quantifying the health effect of democratization is difficult due to concerns about unobserved confounders, reverse causation, and measurement errors in democracy indicators. Given that previous studies are far from conclusive, the impact of democratic institutions on population health remains an open empirical research question. This article offers new estimates of the influence of democratization on health improvements across countries between 1960 and 2010. It goes beyond previous research by explicitly accounting for heterogeneity in the democracy-health nexus and the presence of global common shocks. I use a novel dichotomous measure of democracy to simulate an experiment delivering quasi-random variations in political development. These distinguishing features, as argued below, are important for securing reliable inference on the relationship between democracy and population health across countries. I find that countries that experienced democratic transitions have better health outcomes than those who remained non-democracies throughout the period 1960–2010. However, my preferred estimates indicate that the size of the health-enhancing effect of democratization is considerably smaller than previously thought.

Existing estimates of the health returns to democratization are obtained under strong parameter homogeneity and cross-sectional independence assumptions. Deviation from these restrictions undermines the validity of statistical inference on the cross-country democracy-health relationship. Anecdotal evidence shows that several non-democratic countries have developed efficient public health systems and achieved large health improvements. Some democracies characterized by prevalent corruption and lack of state capacity have proved difficult to improve population health via democratization. It is therefore argued that the impact of democracy on health outcomes exhibits large heterogeneity across countries. Furthermore, previous research has predominantly ignored global common shocks (i.e. climate change, epidemiological events, and the global economic cycle) that affect countries differentially. My main contribution is that, differently from the existing empirical literature, I consider such heterogeneity by allowing parameter estimates to vary across units in the panel. Furthermore, I model the impact of democratization on population health under the presence of unobserved common factors and international spillovers of knowledge, technology, and institutions. I establish empirically that the economic significance of the health-enhancing effect of democratization reduces substantially when accounting for parameter heterogeneity and unobserved common factors.

Another contribution of this study is to exploit a novel dichotomous measure of democracy developed by Acemoglu et al. (2019) to account for the possibility that commonly adopted measures of democracy reflect spurious changes in political regimes (Acemoglu et al. 2019). Furthermore, existing democracy indices merely capture the level of political freedom rather than the incidents of democratization (Papaioannou and Siourounis 2008). This implies that previous research remains largely silent about the health impact of democratic transitions. Therefore, adopting the dichotomous democracy index of Acemoglu et al. (2019) is important for understanding whether health outcomes increase, decrease, or remain unchanged in response to democratization. I use this index to simulate a quasi-natural experiment that identifies the heterogeneous treatment effect of democracy on population health. In particular, I compare the average change in health status of treated countries that experienced democratic transitions with that of control countries who retained non-democratic institutions between 1960 and 2010. To account for unparallel trends between the treated and control groups, I follow Chan and Kwok (2022) to use a principal components difference-in-differences estimator. This approach involves constructing common proxies from a panel of the control group, which are used as extra covariates in the regression analysis for the treated group, in line with the identification approach adopted in Eberhardt (2022). Under the assumption that common factors capture a wide range of unobserved time-varying heterogeneities (Pesaran 2006), the empirical strategy enables me to counteract the aforementioned sources of bias.

I also add to the existing literature by documenting further evidence from a generalized synthetic control analysis and three waves of democratization from 1789 to 2015. Specifically, I impute counterfactuals that reflect country-specific trajectories of life expectancy at birth among countries that experienced democratic transitions in the absence of democratization for the years when the treatment effect happened. These counterfactual scenarios are compared with the actual data, yielding estimates of the average treatment effect of democracy on population health. The results lend credence to the positive average treatment effect of democratization on national health status. Additionally, I construct a panel of 169 countries spanning the period 1789-2015 and use the multiplicative polyarchy index available in the Varieties of Democracy dataset as an alternative measure of democracy. Exploiting variations in political institutions during three waves of democratization, I distinguish the long-run effect from the short-run impact of democratization on population health.

The rest of this study is organized as follows. Section 2 discusses the heterogeneous impact of democratization on population health. Section 3 describes data and key variables. Identification methods and main results are presented in Sections 4 and 5, respectively. Section 6 contains additional evidence. Section 7 concludes the study.

2. How does democratization affect population health?

Conventional explanations of the health-enhancing impact of democratization primarily draw on the idea that democratic institutions are central to effective provision of public healthcare services. As depicted in Fig. 1, there have been large increases in life expectancy at birth and rapid declines in mortality worldwide since the end of World War II (Becker, Philipson, and Soares 2005; Cutler, Deaton, and Lleras-Muney 2006). The period of unprecedented health improvements commencing in the 1940s, commonly known as the international epidemiological transition, was mainly triggered by the development and international diffusion of effective medical interventions related to the treatment of infectious diseases using vaccines and antibiotics (Acemoglu and Johnson 2007). These widespread health gains were also attributed to significant decreases in malnutrition, and advances in infrastructure leading to better access to clean water supply and improved sanitation facilities (Besley and Kudamatsu 2006). Implementing modern health interventions and developing infrastructure to a large extent depend on the involvement of governments to address inherent market failures associated with public goods provision. It is therefore hypothesized that democracies, relative to their non-democratic counterparts, are more likely to experience health improvements because democratic institutions augment a government’s decision to provide public healthcare in a way that improves population health.

Figure 1

Democracy and population health (1789–2015).

Notes: This figure depicts the evolution of democratic institutions and health outcomes averaged for different world regions classified by the World Bank, including East Asia and Pacific (EAP), Europe and Central Asia (ECA), Latin America and Caribbean (LAC), Middle East and North Africa (MENA), North America (NA), South Asia (SA), and sub-Saharan Africa (SSA), between 1789 and 2015. Democracy is captured by the multiplicative polyarchy index obtained from the Varieties of Democracy dataset. Health outcomes include life expectancy at birth and infant mortality rate. See Section 3 for variables’ description and data sources.

Open in new tab Download slide

It has been established that a government’s incentive to provide public goods is embedded in a political regime (Fujiwara 2015). This viewpoint rhymes well with key predictions of the political economy models developed by Romer (1975) and Roberts (1977) emphasizing the role of elections in shaping the redistribution of resources. These major contributions to understanding the driving forces of public goods provision reveal that increased political participation of low-income voters may translate into better government spending in healthcare (Fujiwara 2015), which is particularly beneficial to impoverished groups of the population and hence promotes national health status. An important distinguishing feature of democracies lies in the presence of electoral competition, free media, and a system of checks and balances (Dahl 1989). These institutional constraints limit the power of the executive and improve the political accountability of governments, thereby enhancing the quality of public goods provision. Additionally, the development of a well-functioning public health system in response to the demand of a majority of low- and middle-income voters is important for obtaining political support in democracies; this arguably affects the security of tenure of effective leaders and political stability. Democratic institutions therefore augment the adoption of pro-health policies that benefit a plurality of the voting electorate. By contrast, in non-democratic regimes, where political power is typically concentrated in a narrow segment of the population, governments tend to gain political support by spending on cash transfers that disproportionately benefit powerful elites and entrenched groups (Meltzer and Richard 1981; Kammas and Sarantides 2019; Vu 2022). Besley and Kudamatsu (2006) also posit that democracies have advantages over non-democracies when it comes to selecting competent and incorruptible political leaders, resulting in better implementation of health policies and hence fostering national health status.¹

While the hypothesized positive influence of democracy on population health has gained support in many empirical studies (see, i.e., Besley and Kudamatsu 2006; Kudamatsu 2012; Annaka and Higashijima 2021; Batinti, Costa-Font, and Hatton 2022), there exists evidence that democracy has a negligible impact on health outcomes (Ross 2006).² As articulated in Olson (1982), governments influenced by special interest groups may not provide public goods based on the interest of the median voter. In democratic regimes characterized by the domination of influential elites and ethnic groups with clientelistic support from the middle- and upper-income classes, governments tend to deliver welfare services in a way that deprives the low-income group (Keefer and Khemani 2005; Powell-Jackson et al. 2011; Chapman 2018; Batinti, Costa-Font, and Hatton 2022). If this is true, democratic transitions eventually fail to ensure wider access to healthcare and hence undermine population health. International experience demonstrates that several non-democratic countries, such as China and Cuba, can focus on achieving better health outcomes due to their enhanced coercive abilities. In contrast, Bangladesh, Indonesia, Kyrgyzstan, or Nepal, albeit classified as democracies in 2010, still suffer from pervasive corruption and the absence of effective state institutions, which remain major impediments to improving health conditions.

It follows from the extant literature that the role played by democratic institutions in achieving better health outcomes is far from uniform across countries. This is a plausible assumption because countries differ widely from one another in various dimensions of contemporary or historically persistent socio-economic development. Previous research reveals that worldwide heterogeneity in the health returns to democratization depends on various country-specific characteristics, including, but not limited to, interpersonal population diversity, genetic distance to the world frontiers of modern health technologies and political institutions, statehood experience, and the timing of the Neolithic revolution.

Gerring, Thacker, and Alfaro (2012), in particular, find that older autonomous democracies, relative to their newly established counterparts, are more likely to experience the positive effect of democratization on population health. The underlying idea is that long-standing democracies are typically endowed with better governance quality, which is conducive to the provision of public healthcare (Gerring, Thacker, and Alfaro 2012). Hansen (2013) shows that countries with greater genetic distance to the USA tend to experience barriers to the international diffusion of modern health technologies. This is because genetic distance—a proxy for countries’ dissimilarities in cultures, ancestry, and historical legacies—impedes the cross-border exchange of technologies, knowledge, and institutions (Spolaore and Wacziarg 2009). For this reason, countries with greater genetic distance to the world frontiers of modern health technologies and democratic institutions tend to be characterized by the persistence of poor health and underdeveloped political institutions, making it hard to improve population health via democratization. Other studies reveal that prehistorically determined population diversity is detrimental to establishing inclusive political institutions (Galor and Klemp 2017) or public goods provision (Ashraf and Galor 2013; Arbatlı et al. 2020). Accordingly, diverse countries characterized by large heterogeneity in individual preferences for the provision of public goods and government policies may find it difficult to reconcile such diverse preferences. This hampers the ability to achieve better health outcomes by developing inclusive democratic institutions. It has been established that state history (Vu et al. 2022) and the timing of the Neolithic revolution (Galor and Moav 2007) are associated with better national health status.³

3. Data and descriptive statistics

To estimate the effect of democratization on health outcomes, I use a binary indicator of democracy provided by Acemoglu et al. (2019). The construction of this index exploits existing democracy indices available in the Freedom House and Polity IV datasets and closely follows Papaioannou and Siourounis (2008). Specifically, Acemoglu et al. (2019) create a dummy variable taking a value of 1 if a country in a given year is coded as democratic and 0 otherwise. They consider a country democratic if it is classified as ‘free’ or ‘partially free’ in the Freedom House dataset and records a positive Polity IV score.⁴ This dichotomous measure of democracy is available for 183 countries from 1960 to 2010, for which I could obtain data on population health. The percentage of democracies increased significantly from 31.5% in 1960 to 64.1% in 2010, reflecting widespread worldwide diffusion of democratic institutions during the third wave of democratization and the 1990s that witnessed a transition to representative rule in various former socialist countries (Papaioannou and Siourounis 2008; Acemoglu et al. 2019).

A key advantage of using the dichotomous democracy indicator is to capture transitions in political regimes over time.⁵ Democratization or a transition from non-democracy to democracy is highly correlated with the probability of implementing various political institutions related to electoral democracies worldwide, including the aforementioned institutional constraints on executive power (Acemoglu et al. 2019). There were 121 democratization events and 71 reversals from democracy to non-democracy between 1960 and 2010; reversals could be driven by a coup or fraud elections overthrowing representative rule (Papaioannou and Siourounis 2008).⁶ I divide all the countries with available data into three different groups, including (1) always democracies (44 countries for which the dichotomous democracy score is always 1 from 1960 to 2010), (2) never democracies (45 countries for which the dichotomous democracy score is always 0 from 1960 to 2010), and (3) transitioned democracies (94 countries that experienced democratization or democratic reversals from 1960 to 2010). For the group of transitioned democracies, the average number of democratization events (ranging between 0 and 4) is approximately 1.3 while the average number of reversals (ranging between 0 and 3) is nearly 0.8. Out of 94 transitioned democracies, 4 countries (including Gambia, Myanmar, Somalia, and Venezuela) only experienced democratic reversals and did not transition to democracy, 43 countries made a permanent democratic transition (with one and only one event of democratization and no democratic reversals), and 47 countries experienced at least one incident of democratization and/or at least one democratic reversal.

The main outcome variable is life expectancy at birth, which corresponds to the number of years a newborn infant would live if prevailing patterns of mortality at the time of its birth were to stay the same throughout its life. Higher values of life expectancy at birth reflect better national health status. I also exploit infant mortality rate as an alternative measure of population health. It captures the number of infants (per 1,000 live births) dying before reaching one year of age, with lower values indicating health gains. I match these variables to the democracy dataset of Acemoglu et al. (2019), yielding an unbalanced panel of 183 countries spanning the period 1960–2010. I use data for eighty-four countries in the group of transitioned democracies and thirty-nine countries belonging to the group of never democracies in the main analysis, dictated by data availability (see Supplementary Appendix A2 for the list of countries). Given the importance of economic development for achieving better health outcomes, I use GDP per capita and its quadratic term and trade openness as main control variables. Data on population health and other covariates were obtained from the World Bank’s Development Indicators. Supplementary Appendix Table A3 contains descriptive statistics of key variables for transitioned democracies and never democracies. Supplementary Appendix Table A4 shows pairwise correlations between key variables for the sample of transitioned democracies.

Furthermore, I construct another unbalanced panel of 169 countries from 1789 to 2015. I use the multiplicative polyarchy index available in the Varieties of Democracy (V-Dem) dataset (Coppedge et al. 2023) as an alternative (continuous) measure of democracy. It reflects the electoral principle of democracy, such as freedom of association, clean elections, freedom of expression, elected officials, and suffrage (Coppedge et al. 2023). This index takes values ranging between 0 and 1, with higher scores corresponding to better democratic institutions. I collected data on life expectancy at birth between 1789 and 2015 from various sources, including Riley (2005) and the United Nations’ World Population Prospects. Data on infant mortality rate before and after 1960 were obtained from Mitchell (2013) and the World Bank’s Development Indicators, respectively. Data on key control variables, including GDP per capita and trade openness, were obtained from the Maddison Project and the Historical Bilateral Trade and Gravity dataset of Fouquin and Hugot (2016), respectively. Figure 1 depicts the evolution of political institutions and health outcomes averaged for various world regions between 1789 and 2015. It shows substantial improvements in health status and political institutions across countries over years, particularly during the third wave of democratization.⁷

4. Empirical strategy

4.1 Potential threats to identification

Attempts to estimate the influence of democratization on population health face several challenges. An important issue relates to the argument that democratic transitions are not randomly assigned across countries (Acemoglu et al. 2019). Democracies differ from non-democracies in numerous characteristics that may simultaneously drive political development and health improvements. For this reason, estimates of the health returns to democratization can be confounded by a third omitted variable. For example, countries with higher levels of economic development, where people typically enjoy better health status, are endowed with better resources for investments in political development. Additionally, health improvements may affect the nature of political institutions. This is because individuals who live longer and healthier tend to believe that they have controls over their lives (Knowles and Owen 2010), potentially leading to higher demands for the political accountability of governments. Moreover, commonly adopted democracy indices can be measured with errors (Papaioannou and Siourounis 2008; Acemoglu et al. 2019). As such, variations in existing democracy scores may be spurious proxies for actual changes in political regimes. Indeed, obtaining an internationally comparable measure of democracy over years is challenging because ‘[democracy] has meant different things to different people at different times and places’ (Dahl 2000: 3). Therefore, achieving reliable statistical inference on the impact of democratization on population health is contingent on addressing plausible concerns about omitted variable bias, reverse causality, and measurement errors in democracy indicators.

Conventional regression models, as those in Besley and Kudamatsu (2006), may provide invalid statistical inference on the health effect of democratization under the presence of cross-country heterogeneity. This argument is in line with Pesaran and Smith (1995) documenting that homogeneous parameter estimators (i.e. OLS, fixed-effects, 2SLS, and GMM) offer biased and inconsistent estimates if strong parameter homogeneity assumptions are violated.⁸ In addition, health knowledge, technologies, and institutions inevitably transcend national borders, especially within an increasingly interconnected world (Aidt, Albornoz, and Hauk 2021). Under the presence of such spillovers or global common shocks (i.e. the global economic cycle, climate change, and global epidemiological events), the requirement of cross-sectionally independent errors, which underlies much of the panel data literature, can be violated. As suggested by Pesaran (2006), failure to properly account for cross-sectional dependence yields biased estimates and spurious inference. Therefore, identifying the health effect of democratization requires attention to parameter heterogeneity and unobserved common factors.

Against this background, I attempt to estimate the heterogeneous treatment effect of democracy on population health under the presence of global common shocks. I exploit variations in the timing of democratization to simulate a quasi-randomized experiment and implement a difference-in-differences analysis. I use transitioned democracies as treated units, while countries coded as never democracies belong to the group of control units. The identifying assumption is that, in the absence of the treatment effect, the average outcomes for the treated and control countries would have followed parallel trends over time. The average treatment effect of democracy on population health can be measured by the difference between the average change in health outcomes experienced by the treated group and that experienced by the control group. Under the assumption of parallel trends, systematic differences between the treated and control countries can be accounted for by deducting group-specific averages of the outcome variable of interest. In this regard, double differencing permits ruling out the possibility that the democracy-health nexus can be confounded by permanent differences between the treated and control groups or the presence of common time trends unrelated to democratization. By simulating a natural experiment delivering quasi-random variations in political development, I therefore counteract potential sources of bias induced by unobserved confounders, reverse causation, or measurement errors in the democracy index.

A major challenge with my identification strategy relates to non-parallel trends under the presence of unobserved time-varying heterogeneities between countries. To address this problem, I use the principal components difference-in-differences estimator (PCDID) of Chan and Kwok (2022). However, I rely on the common correlated effects (CCE) estimator of Pesaran (2006), rather than principal component analysis (PCA), to construct common proxies that help account for unparallel and stochastic trends, following Eberhardt (2022). Drawing on Pesaran and Smith (1995), my empirical strategy also permits relaxing the highly restrictive assumption of slope homogeneity. This article, differently from the related empirical literature, explicitly allows the impact of democratization on population health to vary, in sign and economic and statistical significance, across countries.

4.2 The PCDID approach under a common factor framework

The PCDID estimator of Chan and Kwok (2022) provides an intuitive approach to estimating specific parameters that identify treatment effects based on factor-augmented regressions. The implementation follows a two-stage regression procedure. In the first stage, Chan and Kwok (2022) propose using a data-driven method, such as PCA, to construct factor proxies based on data from the control group. The second stage involves including these factor proxies in the regression for the treated group to account for endogeneity concerns related to non-parallel trends and endogenous selection into treatment.

A key distinguishing feature of the PCDID estimator is to apply a data-driven approach to the construction of factor proxies by using data from control units. These factor proxies capture unobserved common shocks and their heterogeneous effects across countries by using regression residuals from the control panel, consistent with Bai (2009). A challenge with the implementation of PCA in this context relates to the unbalanced nature of panel data for never democracies. This is because missing observations need to be filled in via a linear projection (regression) method (Stock and Watson 2002; Eberhardt 2022). An alternative way of constructing factor proxies draws upon the common factor framework developed by Pesaran (2006). The basic idea is to use cross-sectional averages of the observables to capture a wide range of unobserved time-varying heterogeneities and their differential impacts across countries (see Supplementary Appendix A3). It is noteworthy that attempts to measure and control for numerous confounding factors in standard regression models are very difficult partly due to the unobserved (or observed but noisily measured) nature of many confounding characteristics.⁹ In this regard, the CCE estimator of Pesaran (2006) offers an intuitive dimensionality-reducing approach to capturing numerous unobserved common factors (Chudik and Pesaran 2015b).

The PCDID approach builds upon the econometric literature on policy evaluation methods. Yet, it departs from conventional quasi-experimental estimation methods, including canonical difference-in-differences or synthetic controls analyses, by estimating factor-augmented regressions to address plausible concerns about non-parallel trends and the correlation between unobservables and treatment effects. Previous studies show that two-way fixed-effects regressions may yield estimates of weighted sums of the average treatment effects, where some of the weights can be negative (de Chaisemartin and D’Haultfœuille 2020; Goodman-Bacon 2021; Athey and Imbens 2022; de Chaisemartin and D’Haultfœuille 2023). Because of the negative weights, the results do not reflect causal inference on policy intervention; this problem arises from the possibility that treatment effects may vary across groups and over time. Therefore, inadequate attention to heterogeneity in treatment effects may invalidate statistical inference on canonical difference-in-differences estimates. To address this concern, Chan and Kwok (2022) propose using the mean-group-type (MG) estimator of Pesaran and Smith (1995) that allows the treatment effect to vary across countries. Overall, I rely on the PCDID-MG estimators under the common factor framework of Pesaran (2006) to estimate the heterogeneous treatment effect of democracy on population health. I first construct cross-sectional averages of the observables based on data from the control countries (never democracies) and exploit these factor proxies to capture differential trends between the control and treated groups. Then, I estimate factor-augmented regressions using data from the treated units (transitioned democracies), where factor proxies are allowed to enter the regressions as extra covariates. Below, I provide a semi-technical explanation of the main identification strategy, following Chan and Kwok (2022).

Let the potential health outcome of country

i

in year

t

be denoted by

{H E}_{i t} (1)

and

{H E}_{i t} (0)

with and without treatment, respectively. The treatment effect of democracy on population health is given by

Δ_{i t} : = {H E}_{i t} (1) - {H E}_{i t} (0)

⁠.

D_{i t}

denotes the dichotomous measure of democracy. Let each country

i

belonging to the treated group (⁠

i \in E

⁠) have

T_{0 i}

pre-treatment periods and

T_{1 i}

post-treatment periods. Hence,

D_{i t} = 1_{\{i \in E\}} 1_{\{t > T_{0 i}\}}

⁠. The observed outcome of treatment for a treated country

i

in post-intervention period

t > T_{0 i}

is given by:

{H E}_{i t} = D_{i t} {H E}_{i t} (1) + (1 - D_{i t}) {H E}_{i t} (0) = Δ_{i t} 1_{\{i \in E\}} 1_{\{t > T_{0 i}\}} + {H E}_{i t} (0),

(1)

where

{H E}_{i t} (0) = ζ_{i} + λ_{i} x_{i t} + μ_{i}^{'} f_{t} + {\tilde{ϵ}}_{i t}

⁠;

Δ_{i t}

denotes the treatment effect of democracy on population health that differs across countries and over time;

x_{i t}

is a vector of observed confounding factors with country-specific parameters (⁠

λ_{i}

⁠);

μ_{i}^{'} f_{t}

is an

ℓ \times 1

vector of unobserved common factors (⁠

f_{t}

⁠) with country-specific factor loadings (⁠

μ_{i}^{'}

⁠); and

{\tilde{ϵ}}_{i t}

is the error term. The treatment effect is decomposed into

Δ_{i t} = {\bar{Δ}}_{i} + {\tilde{Δ}}_{i t}

⁠, where

{\bar{Δ}}_{i} : = E (Δ_{i t}| t > T_{0 i})

identifies the country-specific impact of democratization on population health averaged over post-intervention periods;

{\tilde{Δ}}_{i t}

is the cross-sectionally demeaned, time-varying idiosyncratic component of

Δ_{i t}

⁠. It is assumed that

E ({\tilde{Δ}}_{i t}| t > T_{0 i}) = 0

for each treated country

i \in E

⁠. Inserting

Δ_{i t} = {\bar{Δ}}_{i} + {\tilde{Δ}}_{i t}

into Equation (1), I obtain the following reduced-form model:

{H E}_{i t} = {\bar{Δ}}_{i} 1_{\{i \in E\}} 1_{\{t > T_{0 i}\}} + λ_{i} x_{i t} + ζ_{i} + μ_{i}^{'} f_{t} + ϵ_{i t},

(2)

where

ϵ_{i t} : = {\tilde{Δ}}_{i t} 1_{\{i \in E\}} 1_{\{t > T_{0 i}\}} + {\tilde{ϵ}}_{i t}

is the composite idiosyncratic error term in the reduced-form model; it has zero mean following the earlier decomposition. If factor loadings do not vary across countries (⁠

μ_{i}^{'} = μ^{'}

⁠), Equation (2) illustrates the conventional two-way fixed-effects estimator. This corresponds to the inclusion of year fixed effects in the regression to account for unobserved year-specific factors. However, this approach requires that the confounding impact of unobserved common factors is homogeneous across countries. Given that there exist large and persistent differences between countries in many aspects of socio-economic and political development, it is plausible that global common shocks affect countries differentially.¹⁰ This motivates using the CCE estimator of Pesaran (2006) to account for the country-specific confounding impact of unobserved common factors (see Supplementary Appendix A3).

Chan and Kwok (2022) demonstrate that ${\bar{Δ}}_{i}$ can be estimated consistently under the presence of heterogeneous treatment effects by using the MG estimator of Pesaran and Smith (1995). This yields mean estimates by averaging all the country-specific coefficients; $\hat{Δ} = N^{- 1} \sum_{i = 1}^{N} {\bar{Δ}}_{i}$ ⁠, where $N$ is the number of countries in the panel. The MG estimates are derived from a two-stage regression procedure. In the first stage, I estimate the treatment effect of democracy on population health for each country, allowing for slope heterogeneity. The country-specific coefficients are then averaged to obtain mean estimates. This approach further helps to mitigate plausible concerns about the correlation between unobserved confounding factors and the treatment effect under the presence of parameter heterogeneity. Specifically, homogeneous parameter models yield biased and inconsistent estimates of a heterogeneous relationship because ${(\bar{Δ}}_{i} - Δ) D_{i t}$ will be included in the error term of the model. In general, my empirical strategy based on the PCDID MG estimators is consistent with recent studies emphasizing the importance of allowing for heterogeneity in treatment effects (see, i.e., de Chaisemartin and D’Haultfœuille 2020; Callaway and Sant’Anna 2021; Goodman-Bacon 2021; Athey and Imbens 2022; de Chaisemartin and D’Haultfœuille 2023).

4.3 The baseline model

Consistent with the above empirical framework, I capture factor proxies by cross-sectional averages of the observables using data from never democracies. I then estimate factor-augmented models using a panel of transitioned democracies (⁠

i \in E

⁠), where factor proxies are allowed to enter the regression analysis as extra covariates. The baseline model can be specified as follows:

{H E}_{i t} = α_{i} + β_{i} D_{i t} + γ_{i} X_{i t} + λ_{1 i} {\bar{H E}}_{t} + λ_{2 i} {\bar{X}}_{t} + ε_{i t},

(3)

where

{H E}_{i t}

is health outcomes for country

i \in E

in year

t

⁠;

D_{i t}

is the dichotomous democracy measure of Acemoglu et al. (2019);

X_{i t}

is a set of control variables, including log of GDP per capita and its quadratic term and log of trade openness;

{\bar{H E}}_{t}

and

{\bar{X}}_{t}

denote cross-sectional averages of the observables (constructed by using data from control countries) with country-specific factor loadings (⁠

λ_{1 i}

and

λ_{2 i}

⁠); and

ε_{i t}

is the error term.

β_{i}

identifies the country-specific effect of democratization on population health from 1960 to 2010.

A possible concern is that the health returns to democratization may exhibit time lags. Thus, I estimate the following cross-sectionally augmented distributed lag (CS-DL) model:

{H E}_{i t} = α_{i} + β_{i} D_{i t} + \sum_{j = 0}^{p} μ_{i j} {Δ D}_{i, t - j} + \sum_{j = 0}^{p} {λ_{i j} \bar{H E}}_{t - j} + \sum_{j = 0}^{p} {ω_{i j} \bar{X}}_{t - j} + ε_{i t},

(4)

where

Δ

denotes the first-difference operator;

β_{i}

identifies the long-run effect of democratization on population health while the short-run impact is captured by

μ_{i j}

⁠. As suggested by Chudik and Pesaran (2015a) and Chudik et al. (2016), estimating a dynamic model based on the CCE estimator of Pesaran (2006) necessitates controlling for a sufficient number of lags of cross-sectional averages of the observables. Following Chudik and Pesaran (2015a) and Chudik et al. (2016), the rule-of-thumb lag length is given by

p = int (T^{1 / 3}) = 3

⁠, where

T

is the number of years in the panel.

My empirical strategy rests upon the assumption that common proxies capture a wide range of unobserved heterogeneities between the treated and control groups, which underlies the interactive effects structure (⁠ $μ_{i}^{'} f_{t}$ ⁠) of Pesaran (2006) and Bai (2009). It is therefore possible to control for non-parallel and stochastic trends by including common factors in the regression. However, common factors capture not only unobserved common shocks but also idiosyncratic errors of control units. This implies that estimates of the average treatment effect can be biased and inconsistent due to potential correlation between idiosyncratic errors of treated and control units. Therefore, the validity of the PCDID estimator also depends on the asymptotic condition that $\sqrt{T} / N_{C} \to 0$ as $T, N_{C} \to \infty$ ⁠, where $T$ is the length of the sample period and $N_{C}$ is the number of control units; in my sample, $\sqrt{T} / N_{C} = \sqrt{51} / 45 = 0.16$ ⁠. This requirement ensures that the asymptotic bias can be removed (Chan and Kwok 2022). Hence, interpreting the empirical estimates presented in the following section requires attention to the caveat that any deviation from these requirements may invalidate statistical inference on the heterogeneous treatment effect of democracy on population health. It is noteworthy that the baseline control variables are endogenous in regression models explaining the variation in health outcomes across countries and over years. However, attempts to addressing concerns about endogeneity bias in several extra covariates are very difficult. As discussed below, I first observe the stability of the empirical estimates derived from specifications with and without control variables to check whether the health returns to democratization can be attributed to conventional correlates of health improvements. Additionally, I compare the MG and PCDID MG estimates to draw inference on the heterogeneous treatment effect of democracy on population health when explicitly accounting for parameter heterogeneity and unobserved common factors.

5. Empirical estimates

5.1 Main results

5.1.1 Static estimates

Estimates of Equation (3) are reported in Table 1. These results capture the effect of democratization on life expectancy at birth (Panel A) and infant mortality rate (Panel B) across transitioned democracies from 1960 to 2010. I first use the MG estimator of Pesaran and Smith (1995) to estimate the baseline model, as shown in Columns (1) and (2). The specifications from Columns (3) and (4) are augmented with common proxies, constructed by using data for never democracies, to account for unparallel trends between the treated and control groups. The lowest root mean squared error (RMSE) shown in Column (4) indicates that the PCDID MG estimators fit the data better. This is consistent with the argument that allowing for parameter heterogeneity and unobserved common factors is important for securing reliable inference on the cross-country democracy-health nexus.

Table 1

Open in new tab

The heterogeneous treatment effect of democracy on population health.

	MG	MG	PCDID MG	PCDID MG
	(1)	(2)	(3)	(4)
Panel A. Dependent variable is log of life expectancy at birth
Democracy	0.066***	0.014***	0.008***	0.007***
Democracy	(0.011)	(0.004)	(0.003)	(0.002)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Common proxies	No	No	Yes	Yes
Treated group: transitioned democracies
Observations	3,108	3,108	3,108	3,108
No. of countries	84	84	84	84
No. of democratization events	101	101	101	101
No. of democratic reversals	59	59	71	59
Control group: never democracies
Observations	1,297	1,297	1,297	1,297
No. of countries	39	39	39	39
RMSE	0.072	0.042	0.027	0.020
Panel B. Dependent variable is log of infant mortality rate
Democracy	−0.363***	−0.101***	−0.016	−0.009*
Democracy	(0.047)	(0.022)	(0.010)	(0.005)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Common proxies	No	No	Yes	Yes
Treatment group: transitioned democracies
Observations	3,043	3,043	3,043	3,043
No. of countries	84	84	84	84
No. of democratization events	101	101	101	101
No. of democratic reversals	59	59	59	59
Control group: never democracies
Observations	1,243	1,243	1,243	1,243
No. of countries	39	39	39	39
RMSE	0.344	0.135	0.064	0.039

	MG	MG	PCDID MG	PCDID MG
	(1)	(2)	(3)	(4)
Panel A. Dependent variable is log of life expectancy at birth
Democracy	0.066***	0.014***	0.008***	0.007***
Democracy	(0.011)	(0.004)	(0.003)	(0.002)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Common proxies	No	No	Yes	Yes
Treated group: transitioned democracies
Observations	3,108	3,108	3,108	3,108
No. of countries	84	84	84	84
No. of democratization events	101	101	101	101
No. of democratic reversals	59	59	71	59
Control group: never democracies
Observations	1,297	1,297	1,297	1,297
No. of countries	39	39	39	39
RMSE	0.072	0.042	0.027	0.020
Panel B. Dependent variable is log of infant mortality rate
Democracy	−0.363***	−0.101***	−0.016	−0.009*
Democracy	(0.047)	(0.022)	(0.010)	(0.005)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Common proxies	No	No	Yes	Yes
Treatment group: transitioned democracies
Observations	3,043	3,043	3,043	3,043
No. of countries	84	84	84	84
No. of democratization events	101	101	101	101
No. of democratic reversals	59	59	59	59
Control group: never democracies
Observations	1,243	1,243	1,243	1,243
No. of countries	39	39	39	39
RMSE	0.344	0.135	0.064	0.039

Notes: This table shows outlier-robust mean estimates of the heterogeneous treatment effect of democracy on health outcomes. A constant, not reported to conserve space, is incorporated in all the regressions. Standard errors are reported in parentheses. RMSE is the root mean squared error. *** and * denote statistical significance at the 1% and 10% levels, respectively. Source: Author’s estimations.

Table 1

Open in new tab

The heterogeneous treatment effect of democracy on population health.

	MG	MG	PCDID MG	PCDID MG
	(1)	(2)	(3)	(4)
Panel A. Dependent variable is log of life expectancy at birth
Democracy	0.066***	0.014***	0.008***	0.007***
Democracy	(0.011)	(0.004)	(0.003)	(0.002)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Common proxies	No	No	Yes	Yes
Treated group: transitioned democracies
Observations	3,108	3,108	3,108	3,108
No. of countries	84	84	84	84
No. of democratization events	101	101	101	101
No. of democratic reversals	59	59	71	59
Control group: never democracies
Observations	1,297	1,297	1,297	1,297
No. of countries	39	39	39	39
RMSE	0.072	0.042	0.027	0.020
Panel B. Dependent variable is log of infant mortality rate
Democracy	−0.363***	−0.101***	−0.016	−0.009*
Democracy	(0.047)	(0.022)	(0.010)	(0.005)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Common proxies	No	No	Yes	Yes
Treatment group: transitioned democracies
Observations	3,043	3,043	3,043	3,043
No. of countries	84	84	84	84
No. of democratization events	101	101	101	101
No. of democratic reversals	59	59	59	59
Control group: never democracies
Observations	1,243	1,243	1,243	1,243
No. of countries	39	39	39	39
RMSE	0.344	0.135	0.064	0.039

	MG	MG	PCDID MG	PCDID MG
	(1)	(2)	(3)	(4)
Panel A. Dependent variable is log of life expectancy at birth
Democracy	0.066***	0.014***	0.008***	0.007***
Democracy	(0.011)	(0.004)	(0.003)	(0.002)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Common proxies	No	No	Yes	Yes
Treated group: transitioned democracies
Observations	3,108	3,108	3,108	3,108
No. of countries	84	84	84	84
No. of democratization events	101	101	101	101
No. of democratic reversals	59	59	71	59
Control group: never democracies
Observations	1,297	1,297	1,297	1,297
No. of countries	39	39	39	39
RMSE	0.072	0.042	0.027	0.020
Panel B. Dependent variable is log of infant mortality rate
Democracy	−0.363***	−0.101***	−0.016	−0.009*
Democracy	(0.047)	(0.022)	(0.010)	(0.005)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Common proxies	No	No	Yes	Yes
Treatment group: transitioned democracies
Observations	3,043	3,043	3,043	3,043
No. of countries	84	84	84	84
No. of democratization events	101	101	101	101
No. of democratic reversals	59	59	59	59
Control group: never democracies
Observations	1,243	1,243	1,243	1,243
No. of countries	39	39	39	39
RMSE	0.344	0.135	0.064	0.039

The dichotomous measure of democracy enters all the regressions with a positive and statistically significant coefficient (Panel A). This provides empirical support for the conventional view that democracies, relative to their non-democratic counterparts, are more likely to experience health improvements. Specifically, I find that people in countries that transformed from non-democracy to democracy on average live 6.6% longer than those in countries that retained non-democratic institutions, holding other things equal (Column 1, Panel A). The estimated impact reduces to 1.4% when I add observed covariates to Column (2). Results are consistent with previous studies documenting the statistical and economic significance of the health-enhancing impact of democratization. When I explicitly account for potential deviation from the parallel trends requirement, the positive influence of democratization on population health becomes much smaller in size (Columns 3 and 4). In particular, the size of the estimated coefficients reported in Columns (3) and (4) is less than 50% of the MG estimates. My preferred PCDID MG estimates suggest that the magnitude of the health returns to democratization is significantly smaller than previously thought.

The main results indicate that failure to control for unobserved time-varying heterogeneities between the treated and control groups in the presence of global common shocks and parameter heterogeneity considerably inflates the economic significance of the health-enhancing effect of democratization. This finding remains robust to using infant mortality as an alternative health outcome (Panel B). Specifically, I find that the economic and statistical significance of the mortality-reducing impact of democratization decreases substantially when I explicitly account for unparallel and stochastic trends between the treated and control countries.¹¹Figure 2 depicts the country-specific impact of democratization on population health. It demonstrates that the role played by democratic transitions in shaping national health status is highly heterogeneous across countries. Overall, I provide evidence suggesting that the health-enhancing impact of democratization is much smaller in size than that implied by conventional estimates. This partially stems from the presence of heterogeneity in the cross-country democracy-health relationship and unobserved common factors.

Figure 2

The heterogeneous treatment effect of democracy on population health, static estimates.

Notes: This figure depicts country-specific point estimates and 95% confidence intervals of the effect of democratization on life expectancy at birth, derived from the specification from Table 1, Column (4).

Open in new tab Download slide

5.1.2 Dynamic estimates

A potential concern is that a transition from non-democracy to democracy is unlikely to immediately translate into health improvements. It is argued that a government’s decision to develop a well-functioning public health system required for achieving better health outcomes is a lengthy process (Annaka and Higashijima 2021). This implies that democratization may have a dynamic impact on population health. Therefore, I estimate Equation (4) to capture the long-run and short-run impacts of democratization on population health, as reported in Table 2. For ease of comparison, I first adopt the MG estimator to implement the regression analysis for the treated group (transitioned democracies). Then, I include common proxies in the model specification as additional covariates. To allow for the dynamic impact of unobserved common factors, I augment the regression analysis with between 1 and 3 lags of common proxies.

Table 2

Open in new tab

The heterogeneous treatment effect of democracy on population health, dynamic CS-DL estimates.

Lag length of common proxies	$p = 1$		$p = 2$		$p = 3$
	(1)	(2)	(3)	(4)	(5)	(6)
	MG	PCDID MG	MG	PCDID MG	MG	PCDID MG
Panel A. Dependent variable is log of life expectancy at birth
Democracy	0.094***	0.012***	0.096***	0.011**	0.098***	0.012**
	(0.012)	(0.004)	(0.013)	(0.004)	(0.013)	(0.005)
$Δ$ Democracy	−0.037***	−0.007**	−0.041***	−0.006	−0.045***	−0.004
	(0.006)	(0.003)	(0.007)	(0.004)	(0.008)	(0.004)
$Δ$ Democracy_t-1	−0.030***	−0.008***	−0.035***	−0.008***	−0.041***	−0.008**
	(0.006)	(0.002)	(0.006)	(0.003)	(0.007)	(0.003)
$Δ$ Democracy_t-2			−0.029***	−0.008***	−0.035***	−0.008***
			(0.005)	(0.003)	(0.006)	(0.002)
$Δ$ Democracy_t-3					−0.028***	−0.009***
					(0.005)	(0.003)
Common proxies	No	Yes	No	Yes	No	Yes
Treated group: transitioned democracies
Observations	4,408	4,408	4,314	4,314	4,220	4,220
No. of countries	94	94	94	94	94	94
No. of democratization events	120	120	120	120	119	119
No. of democratic reversals	68	68	65	65	64	64
Control group: never democracies
Observations	2,038	2,038	1,994	1,994	1,936	1,936
No. of countries	45	45	45	45	44	44
RMSE	0.076	0.035	0.074	0.031	0.072	0.028

Panel B. Dependent variable is log of infant mortality rate
Democracy	−0.520***	−0.026	−0.544***	−0.026	−0.552***	−0.031
	(0.052)	(0.017)	(0.055)	(0.019)	(0.058)	(0.020)
$Δ$ Democracy	0.281***	0.031**	0.316***	0.031**	0.329***	0.033*
	(0.033)	(0.013)	(0.036)	(0.015)	(0.039)	(0.017)
$Δ$ Democracy_t-1	0.234***	0.027***	0.273***	0.032**	0.297***	0.026*
	(0.028)	(0.010)	(0.032)	(0.013)	(0.037)	(0.016)
$Δ$ Democracy_t-2			0.224***	0.019*	0.257***	0.026*
			(0.027)	(0.010)	(0.032)	(0.013)
$Δ$ Democracy_t-3					0.220***	0.012
					(0.027)	(0.008)
Common proxies	No	Yes	No	Yes	No	Yes
Treated group: transitioned democracies
Observations	3,951	3,951	3,894	3,894	3,834	3,797
No. of countries	94	94	94	94	94	92
No. of democratization events	119	119	119	119	118	116
No. of democratic reversals	64	64	62	62	61	60
Control group: never democracies
Observations	1,789	1,789	1,764	1,764	1,724	1,724
No. of countries	45	45	45	45	44	44
RMSE	0.307	0.059	0.295	0.050	0.282	0.041

Lag length of common proxies	$p = 1$		$p = 2$		$p = 3$
	(1)	(2)	(3)	(4)	(5)	(6)
	MG	PCDID MG	MG	PCDID MG	MG	PCDID MG
Panel A. Dependent variable is log of life expectancy at birth
Democracy	0.094***	0.012***	0.096***	0.011**	0.098***	0.012**
	(0.012)	(0.004)	(0.013)	(0.004)	(0.013)	(0.005)
$Δ$ Democracy	−0.037***	−0.007**	−0.041***	−0.006	−0.045***	−0.004
	(0.006)	(0.003)	(0.007)	(0.004)	(0.008)	(0.004)
$Δ$ Democracy_t-1	−0.030***	−0.008***	−0.035***	−0.008***	−0.041***	−0.008**
	(0.006)	(0.002)	(0.006)	(0.003)	(0.007)	(0.003)
$Δ$ Democracy_t-2			−0.029***	−0.008***	−0.035***	−0.008***
			(0.005)	(0.003)	(0.006)	(0.002)
$Δ$ Democracy_t-3					−0.028***	−0.009***
					(0.005)	(0.003)
Common proxies	No	Yes	No	Yes	No	Yes
Treated group: transitioned democracies
Observations	4,408	4,408	4,314	4,314	4,220	4,220
No. of countries	94	94	94	94	94	94
No. of democratization events	120	120	120	120	119	119
No. of democratic reversals	68	68	65	65	64	64
Control group: never democracies
Observations	2,038	2,038	1,994	1,994	1,936	1,936
No. of countries	45	45	45	45	44	44
RMSE	0.076	0.035	0.074	0.031	0.072	0.028

Panel B. Dependent variable is log of infant mortality rate
Democracy	−0.520***	−0.026	−0.544***	−0.026	−0.552***	−0.031
	(0.052)	(0.017)	(0.055)	(0.019)	(0.058)	(0.020)
$Δ$ Democracy	0.281***	0.031**	0.316***	0.031**	0.329***	0.033*
	(0.033)	(0.013)	(0.036)	(0.015)	(0.039)	(0.017)
$Δ$ Democracy_t-1	0.234***	0.027***	0.273***	0.032**	0.297***	0.026*
	(0.028)	(0.010)	(0.032)	(0.013)	(0.037)	(0.016)
$Δ$ Democracy_t-2			0.224***	0.019*	0.257***	0.026*
			(0.027)	(0.010)	(0.032)	(0.013)
$Δ$ Democracy_t-3					0.220***	0.012
					(0.027)	(0.008)
Common proxies	No	Yes	No	Yes	No	Yes
Treated group: transitioned democracies
Observations	3,951	3,951	3,894	3,894	3,834	3,797
No. of countries	94	94	94	94	94	92
No. of democratization events	119	119	119	119	118	116
No. of democratic reversals	64	64	62	62	61	60
Control group: never democracies
Observations	1,789	1,789	1,764	1,764	1,724	1,724
No. of countries	45	45	45	45	44	44
RMSE	0.307	0.059	0.295	0.050	0.282	0.041

Notes: This table shows outlier-robust mean estimates of the short- and long-run heterogeneous treatment effects of democracy on health outcomes. Standard errors are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. Source: Author’s estimations.

Table 2

Open in new tab

The heterogeneous treatment effect of democracy on population health, dynamic CS-DL estimates.

Lag length of common proxies	$p = 1$		$p = 2$		$p = 3$
	(1)	(2)	(3)	(4)	(5)	(6)
	MG	PCDID MG	MG	PCDID MG	MG	PCDID MG
Panel A. Dependent variable is log of life expectancy at birth
Democracy	0.094***	0.012***	0.096***	0.011**	0.098***	0.012**
	(0.012)	(0.004)	(0.013)	(0.004)	(0.013)	(0.005)
$Δ$ Democracy	−0.037***	−0.007**	−0.041***	−0.006	−0.045***	−0.004
	(0.006)	(0.003)	(0.007)	(0.004)	(0.008)	(0.004)
$Δ$ Democracy_t-1	−0.030***	−0.008***	−0.035***	−0.008***	−0.041***	−0.008**
	(0.006)	(0.002)	(0.006)	(0.003)	(0.007)	(0.003)
$Δ$ Democracy_t-2			−0.029***	−0.008***	−0.035***	−0.008***
			(0.005)	(0.003)	(0.006)	(0.002)
$Δ$ Democracy_t-3					−0.028***	−0.009***
					(0.005)	(0.003)
Common proxies	No	Yes	No	Yes	No	Yes
Treated group: transitioned democracies
Observations	4,408	4,408	4,314	4,314	4,220	4,220
No. of countries	94	94	94	94	94	94
No. of democratization events	120	120	120	120	119	119
No. of democratic reversals	68	68	65	65	64	64
Control group: never democracies
Observations	2,038	2,038	1,994	1,994	1,936	1,936
No. of countries	45	45	45	45	44	44
RMSE	0.076	0.035	0.074	0.031	0.072	0.028

Panel B. Dependent variable is log of infant mortality rate
Democracy	−0.520***	−0.026	−0.544***	−0.026	−0.552***	−0.031
	(0.052)	(0.017)	(0.055)	(0.019)	(0.058)	(0.020)
$Δ$ Democracy	0.281***	0.031**	0.316***	0.031**	0.329***	0.033*
	(0.033)	(0.013)	(0.036)	(0.015)	(0.039)	(0.017)
$Δ$ Democracy_t-1	0.234***	0.027***	0.273***	0.032**	0.297***	0.026*
	(0.028)	(0.010)	(0.032)	(0.013)	(0.037)	(0.016)
$Δ$ Democracy_t-2			0.224***	0.019*	0.257***	0.026*
			(0.027)	(0.010)	(0.032)	(0.013)
$Δ$ Democracy_t-3					0.220***	0.012
					(0.027)	(0.008)
Common proxies	No	Yes	No	Yes	No	Yes
Treated group: transitioned democracies
Observations	3,951	3,951	3,894	3,894	3,834	3,797
No. of countries	94	94	94	94	94	92
No. of democratization events	119	119	119	119	118	116
No. of democratic reversals	64	64	62	62	61	60
Control group: never democracies
Observations	1,789	1,789	1,764	1,764	1,724	1,724
No. of countries	45	45	45	45	44	44
RMSE	0.307	0.059	0.295	0.050	0.282	0.041

Lag length of common proxies	$p = 1$		$p = 2$		$p = 3$
	(1)	(2)	(3)	(4)	(5)	(6)
	MG	PCDID MG	MG	PCDID MG	MG	PCDID MG
Panel A. Dependent variable is log of life expectancy at birth
Democracy	0.094***	0.012***	0.096***	0.011**	0.098***	0.012**
	(0.012)	(0.004)	(0.013)	(0.004)	(0.013)	(0.005)
$Δ$ Democracy	−0.037***	−0.007**	−0.041***	−0.006	−0.045***	−0.004
	(0.006)	(0.003)	(0.007)	(0.004)	(0.008)	(0.004)
$Δ$ Democracy_t-1	−0.030***	−0.008***	−0.035***	−0.008***	−0.041***	−0.008**
	(0.006)	(0.002)	(0.006)	(0.003)	(0.007)	(0.003)
$Δ$ Democracy_t-2			−0.029***	−0.008***	−0.035***	−0.008***
			(0.005)	(0.003)	(0.006)	(0.002)
$Δ$ Democracy_t-3					−0.028***	−0.009***
					(0.005)	(0.003)
Common proxies	No	Yes	No	Yes	No	Yes
Treated group: transitioned democracies
Observations	4,408	4,408	4,314	4,314	4,220	4,220
No. of countries	94	94	94	94	94	94
No. of democratization events	120	120	120	120	119	119
No. of democratic reversals	68	68	65	65	64	64
Control group: never democracies
Observations	2,038	2,038	1,994	1,994	1,936	1,936
No. of countries	45	45	45	45	44	44
RMSE	0.076	0.035	0.074	0.031	0.072	0.028

Panel B. Dependent variable is log of infant mortality rate
Democracy	−0.520***	−0.026	−0.544***	−0.026	−0.552***	−0.031
	(0.052)	(0.017)	(0.055)	(0.019)	(0.058)	(0.020)
$Δ$ Democracy	0.281***	0.031**	0.316***	0.031**	0.329***	0.033*
	(0.033)	(0.013)	(0.036)	(0.015)	(0.039)	(0.017)
$Δ$ Democracy_t-1	0.234***	0.027***	0.273***	0.032**	0.297***	0.026*
	(0.028)	(0.010)	(0.032)	(0.013)	(0.037)	(0.016)
$Δ$ Democracy_t-2			0.224***	0.019*	0.257***	0.026*
			(0.027)	(0.010)	(0.032)	(0.013)
$Δ$ Democracy_t-3					0.220***	0.012
					(0.027)	(0.008)
Common proxies	No	Yes	No	Yes	No	Yes
Treated group: transitioned democracies
Observations	3,951	3,951	3,894	3,894	3,834	3,797
No. of countries	94	94	94	94	94	92
No. of democratization events	119	119	119	119	118	116
No. of democratic reversals	64	64	62	62	61	60
Control group: never democracies
Observations	1,789	1,789	1,764	1,764	1,724	1,724
No. of countries	45	45	45	45	44	44
RMSE	0.307	0.059	0.295	0.050	0.282	0.041

As shown in Table 2, I find that democratization has a positive and statistically significant long-run impact on life expectancy at birth (Panel A). However, the size of the long-run health-enhancing effect of democratization reduces substantially when I attempt to rule out concerns about unobserved heterogeneities between the treated and control units by using the PCDID approach. There also exists large heterogeneity in the estimated long-run coefficients on democracy across countries, as depicted in Fig. 3. I also observe large reductions in the magnitude and statistical precision of the estimated long-run effect of democratization on infant mortality when accounting for non-parallel and stochastic trends (Panel B). These dynamic estimates provide further support for the main results documenting that previous studies overestimate the economic significance of the health returns to democratization. Table 2 also shows that democratization has a negative influence on population health in the short run. This is consistent with Annaka and Higashijima (2021) suggesting that the health-enhancing effect of democratization typically appears with time lags. Overall, the dynamic estimates lend additional support to the long-run health-enhancing effect of democratization albeit with much smaller magnitude.

Figure 3

The heterogeneous treatment effect of democracy on population health, dynamic estimates.

Open in new tab Download slide

5.2 Heterogeneity patterns

Having established heterogeneity in the cross-country democracy-health relationship, I now explore whether some deep-rooted institutional and human characteristics fundamentally drive worldwide differences in the health returns to democratization, as discussed in Section 2. Figure 4 depicts the relationship between the estimated country-specific long-run coefficients on democracy and several fundamental characteristics. Fitting a linear trend of long-run democracy coefficients to genetic distance to the USA indicates that countries with greater genetic distance to the world frontier of modern health technologies and political institutions are less likely to experience the health-enhancing effect of democratization. Furthermore, the contribution of democratization to health improvements is negatively associated with interpersonal population diversity. This reveals that highly diverse countries may find it difficult to foster health improvements by developing democratic institutions. I also find that statehood experience and the timing of the Neolithic Revolution are positively correlated with the estimated long-run coefficients on democracy. This provides partial support for the earlier argument that countries with greater statehood experience and an earlier adoption of sedentary agriculture tend to experience the health-enhancing impact of democratization. Although the confidence intervals of these linear regression lines contain zero in most cases, the results depicted in Fig. 4 offer a partial explanation for worldwide heterogeneity in the impact of democratization on population health.

Figure 4

The deep determinants of heterogeneity.

Notes: This figure illustrates the relationship between the long-run country-specific coefficients (plotted in Figure 3) and deep-rooted institutional and human characteristics. It shows a fitted linear regression line for the relationship between democracy coefficients and several fundamental factors, including genetic distance to the USA (Spolaore and Wacziarg 2009), ancestry-adjusted predicted genetic diversity (Ashraf and Galor 2013), ancestry-adjusted state history (Borcan, Olsson, and Putterman 2018), and the timing of the Neolithic Revolution (Putterman 2006).

Open in new tab Download slide

5.3 Robustness

As mentioned previously, the benchmark models are augmented with controls for the level of economic development and trade openness to mitigate concerns about omitted variable bias. The basic intuition is that better economic performance can be simultaneously associated with the development of inclusive political institutions and an enhanced ability to deliver public healthcare services (Besley and Kudamatsu 2006; Madsen, Raschky, and Skali 2015; Eberhardt 2022), thus confounding the main results. Previous studies also suggest that trade liberalization is linked to improved health outcomes and democratization (Owen and Wu 2007; Vu 2020; Bharati, Farhad, and Jetter 2023). Thus, accounting for these potential confounders helps rule out alternative explanations for health improvements. The existing literature indicates that preferences for the provision of public goods (including healthcare services) differ widely between rural and urban areas. This argument is consistent with Rajkumar and Swaroop (2008) documenting a positive association between the rate of urbanization and human development. Additionally, the conventional view is that the building of well-functioning states is central to efficient provision of public healthcare services, which helps improve population health (Moon and Dixon 1985; Vu et al. 2022). It follows from these arguments that higher rates of urbanization and state capacity (i.e. fiscal and legal capabilities) predict health improvements. Glaeser and Steinberg (2017) reveal that urbanization may affect democratization via shaping demands for inclusive political institutions; moreover, urban concentration is determined by democratic transitions (Davis and Henderson 2003). There is also evidence suggesting that state capacity is interrelated with and jointly determined by both political institutions and national health status (Hanson 2015). Hence, urbanization and state capacity can be common causes or effects of both democratization and population health.

As reported in Supplementary Appendix Table A5, I allow urbanization and state capacity to enter the baseline regressions as extra covariates. Accordingly, I find large reductions in the statistical and economic significance of the health-enhancing impact of democratization. There are several explanations for this finding. First, these additional covariates can be regarded as ‘bad controls’ in regression models explaining the variation in health outcomes across countries over years (Angrist and Pischke 2009, 2014; Cinelli, Forney, and Pearl 2022). This is because urbanization and state capacity can be affected by or have an influence on both democratization and population health, as discussed above.¹² Moreover, interpreting Supplementary Appendix Table A5 results critically requires attention to endogeneity concerns related to the inclusion of urbanization and state capacity in the regression analysis. Second, accounting for urbanization and state capacity also considerably constrains the feasible sample size, making it difficult to obtain a generalized understanding of the democracy-health nexus across the globe. Finally, substantial reductions in the statistical precision of the baseline estimates can be attributed to a high correlation among control variables. For example, this is reflected in the sizeable correlation of 0.82 between urbanization and log of GDP per capita (Supplementary Appendix Table A4).¹³ Overall, I do not include urbanization and state capacity in the baseline regressions primarily due to concerns about using ‘bad controls’, following Angrist and Pischke (2009 2014), Acharya, Blackwell and Sen (2016), and Cinelli, Forney, and Pearl (2022). It is worth re-emphasizing that this article primarily relies on the common factor framework of Pesaran (2006) as a dimensionality-reducing approach to accounting for a wide range of confounding factors related to international spillovers and unobserved time-varying heterogeneities (Chudik and Pesaran 2015b).

A potential concern is that reversals from democracy to non-democracy could have led to better health outcomes if political reforms are triggered by poor development outcomes. Existing research also documents that several non-democratic regimes, characterized by improved coercive abilities, have achieved significant health improvements. Therefore, I follow Papaioannou and Siourounis (2008) to replicate the main analysis but exclude four countries that only experienced a reverse transition from democracy to non-democracy. I also restrict the analysis to treated countries with one and only one event of democratization from 1960 to 2010 and no democratic reversals after a transition. Additionally, I re-estimate the baseline model but employ a sample of transitioned democracies with more than one incidents of democratization during the period 1960–2010. My findings, however, are robust to these variations, as shown in Supplementary Appendix Table A7.

To further address concerns about heterogeneity in the treatment effect of democracy on population health over time and across countries, I follow Callaway and Sant’Anna (2021) to estimate the group-time average treatment effect. In line with Callaway and Sant’Anna (2021), I use a panel of treated countries comprising only transitioned democracies that made a permanent democratic transition; these treated units did not reverse after democratization. I also use never democracies as the control group, consistent with the main analysis. I compare the average change in population health experienced by the treated group to the average change experienced by the control group. An important distinguishing feature of Callaway and Sant’Anna’s (2021) method is to identify the average treatment effect for each treatment cohort, defined by the year when a country first transitioned from non-democracy to democracy. This estimation design, therefore, allows for heterogeneity and dynamics in treatment effects. Consistent with the core findings, Supplementary Appendix Fig. A1 shows that the treatment effect of democracy on population health exhibits large heterogeneity, in size and statistical significance, across countries experiencing democratization in different years.

6. Extensions

6.1 Evidence from a generalized synthetic control analysis

I now rely on the generalized synthetic control method (GSCM) of Xu (2017) to estimate the impact of democratization on population health. Drawing on the synthetic control method (SCM) developed by Abadie, Diamond, and Hainmueller (2010, 2015), the GSCM provides a data-driven method that allows estimating treatment effects under the presence of non-parallel trends driven by unobserved time-varying heterogeneities across units in the panel. Specifically, the SCM permits constructing counterfactuals for treated units by exploiting the pre-intervention covariates and outcomes of treated units to reweigh control units. This allows matching control units to treated units to account for unobserved time-varying confounders. The GSCM differs from the SCM by using the interactive fixed-effects model of Bai (2009) to accommodate unobserved time-varying confounding factors semi-parametrically, thus improving the plausibility of synthetic control analyses. The GSCM also allows estimating the average treatment effect under the existence of multiple treated units receiving the treatment at different periods (Xu 2017).

Let

{H E}_{i t}

be the health outcomes of country

i

in year

t

⁠. Assume

T

and

C

⁠, respectively, denote treated (transitioned democracies) and control (never democracies) countries. The number of countries included in the GSCM analysis can be decomposed into

N = N_{t r} + N_{c o}

⁠, where

N_{t r}

and

N_{c o}

are the number of treated and control countries, respectively. The health outcomes of a control unit can be expressed as follows:

{H E}_{i} = X_{i} β + F λ_{i} + ε_{i}, where i \in 1, 2, \dots N_{c o} .

(5)

The health outcomes of the counterfactual after stacking all control units together can be written as:

{H E}_{C O} = X_{C O} β + F Λ_{C O}^{'} + ε_{C O},

(6)

where

{H E}_{C O} = [{H E}_{1}, {H E}_{2}, \dots, {H E}_{N_{C O}}]

and

ε_{C O} = [ε_{1}, ε_{2}, \dots, ε_{N_{C O}}]

are

(T \times N_{C O})

matrices;

X_{C O}

is a three-dimensional

(T \times N_{C O} \times p)

matrix; and

Λ_{C O}^{'} = {[λ_{1}, λ_{2}, \dots, λ_{N_{C O}}]}^{'}

is a

(N_{C O} \times r)

matrix. Consistent with Bai (2003, 2009), the identification of

β

⁠,

F

⁠, and

Λ_{C O}

requires that all factors are normalized and they are orthogonal to one another (Xu 2017).

The construction of a synthetic counterfactual for each treated unit follows a three-step procedure. First, data from control units are exploited to estimate an interactive fixed effects model, yielding a fixed number of latent factors and covariate parameter estimates (⁠

\hat{β}, \hat{F}

⁠, and

{\hat{Λ}}_{C O}

⁠):

(\hat{β}, \hat{F}, {\hat{Λ}}_{C O}) = \underset{\tilde{β} \tilde{F} {\tilde{Λ}}_{C O}}{argmin} \sum_{i \in C} {({H E}_{i} - X_{i} \tilde{β} - \tilde{F} {\tilde{λ}}_{i})}^{'} ({H E}_{i} - X_{i} \tilde{β} - \tilde{F} {\tilde{λ}}_{i})

s.t.

{\tilde{F}}^{'} \tilde{F} / T = I_{r}

and

{\tilde{Λ}}_{C O}^{'} {\tilde{Λ}}_{C O} = diagonal

⁠.

The second step involves using these latent factors and estimates to calculate factor loadings for each treated country by minimizing the root mean squared predicted error of the pre-intervention periods:

{\hat{λ}}_{i} = \underset{{\tilde{λ}}_{i}}{argmin} {({H E}_{i}^{0} - X_{i}^{0} \hat{β} - {\hat{F}}^{0} {\tilde{λ}}_{i})}^{'} ({H E}_{i}^{0} - X_{i}^{0} \hat{β} - {\hat{F}}^{0} {\tilde{λ}}_{i})

= {({\hat{F}}^{0'} {\hat{F}}^{0})}^{- 1} {\hat{F}}^{0'} ({H E}_{i}^{0} - X_{i}^{0} \hat{β}), i \in T,

where

\hat{β}

and

{\hat{F}}^{0}

are obtained via the above first-step regression and subscripts ‘0’s indicate pre-intervention periods. In the final step, the counterfactual of the outcome variable is imputed for each treated country, using the estimated latent factors and factor loadings:

{\hat{H E}}_{i t} (0) = x_{i t}^{'} \hat{β} + {\hat{λ}}_{i}^{'} {\hat{f}}_{t}, i \in T, t > T_{0} .

The treatment effect can be derived from comparing the imputed counterfactuals to the observed outcome. The estimated treatment effect of democracy on health outcomes in year

t

of post-intervention periods (⁠

t > T_{0}

⁠) can be given by:

{\hat{ATT}}_{t} = \frac{1}{N_{t r}} \sum_{i \in T} [{H E}_{i t} (1) - {\hat{H E}}_{i t} (0)], t > T_{0} .

[7]

Consistent with the benchmark analysis, I use the dichotomous measure of democracy of Acemoglu et al. (2019) to estimate the time-varying impact of democratization on life expectancy at birth. Given that the advent of democracy occurred in different years across countries, I standardized time for democratic transitions around the year of democratization, denoted as ‘year 0’. As such, pre- and post-treatment years are represented by negative and positive values, respectively. All the regressions are augmented with log of GDP per capita and its squared term, log of trade openness, and country and year fixed effects. Following Xu (2017), I rely on a nonparametric bootstrapping procedure with 1,000 replications to calculate standard errors.

Figure 5 illustrates the time-varying effect of democratization on life expectancy at birth. Panel A depicts the estimated coefficients and 95% confidence intervals of the average treatment effect; the full estimates are reported in Appendix Table A8. Panel B displays the observed life expectancy and the counterfactuals for the treated group. Specifically, the counterfactuals capture the estimated health status in the absence of democratization for the years of democratization. Following democratization (represented by the vertical line at year 0), the difference between the observed life expectancy at birth (averaged across all the treated countries) and a given counterfactual reflects the health effect of democratization. Accordingly, the GSCM counterfactuals are good matches to the actual data in pre-intervention years. This indicates that the counterfactuals are good proxies for the health status that would have been achieved in treated countries in the post-intervention period in the absence of democratization. The results demonstrate that democratization results in an increase in life expectancy at birth and hence widens the gap between the observed health improvements and their artificial counterfactuals after a democratic transition.

Figure 5

The ATT effect of democracy on population health.

Notes: This figure illustrates dynamics in the estimated ATT effect of democracy on life expectancy at birth based on the generalized synthetic control method of Xu (2017).

Open in new tab Download slide

A steady divergence between the actual data and the counterfactuals in the post-treatment period indicates that the health returns to democratization become larger in size over time. As shown in Supplementary Appendix Table A8, the health-enhancing effect of democratization is statistically significant from the sixth year after democratization to the end of the post-intervention period. These findings are consistent with the argument that the health returns to democratization may exhibit time lags. The underlying idea is that democratic institutions, by improving the political accountability of governments, tend to gradually translate into better health outcomes over time due to greater investments in the public provision of healthcare (Annaka and Higashijima 2021). Overall, additional evidence from a generalized synthetic control analysis provides support for the positive treatment effect of democratization on population health.

6.2 Evidence from three waves of democratization

A potential concern is that the main results merely reflect the health effect of democratic transitions during the third wave of democratization. Additionally, distinguishing the long-run effect from the short-run impact critically requires obtaining an internationally comparable proxy for democracy spanning a prolonged period. This sub-section complements the baseline analysis by documenting evidence of the impact of democratic institutions on population health between 1789 and 2015. For this purpose, I exploit variations in political institutions captured by the V-Dem’s multiplicative polyarchy index. This allows me to isolate the health returns to long-lasting democratization from the influence of transitory shocks to democratic institutions on population health. The empirical analysis also explicitly allows for parameter heterogeneity and cross-sectional dependence.

6.2.1 Static estimates

I first estimate the following static model for up to 167 countries from 1789 to 2015:

{H E}_{i t} = α_{i} + β_{i} {MPI}_{i t} + γ_{i} X_{i t} + δ_{i} f_{t} + ε_{i t},

(8)

where

{H E}_{i t}

denotes log of life expectancy at birth or log of infant mortality rate for country

i

in year

t

⁠;

{MPI}_{i t}

is the V-Dem’s multiplicative polyarchy index;

β_{i}

identifies the heterogeneous effect of democratic institutions on health outcomes across countries;

X_{i t}

is a set of control variables including log of GDP per capita and its squared term, and log of trade openness;

f_{t}

is a vector of common proxies with country-specific factor loadings (⁠

δ_{i}

⁠); and

ε_{i t}

is the error term.

Estimates of Equation (8) are reported in Table 3. These outlier-robust mean estimates capture the effect of democracy on population health from 1789 to 2015. To account for endogeneity concerns arising from the correlation between unobserved common factors and the observables, I rely on the CCE estimator of Pesaran (2006) in Columns (3) and (4). Results indicate that democracy has a positive and negative influence on life expectancy at birth and infant mortality rate, respectively (Columns 1 and 2); these impacts are precisely estimated at the 1% or 5% level of statistical significance. When I explicitly account for cross-sectional dependence in Columns (3) and (4), the estimated coefficients on MPI retain their signs but reduce substantially in statistical and economic significance. This suggests that conventional estimates, which ignore or fail to properly accommodate parameter heterogeneity and cross-sectionally dependent errors, exaggerate the importance of democracy for population health, consistent with the core results.

Table 3

Open in new tab

The heterogeneous effect of democracy on population health between 1789 and 2015.

	MG	MG	CCE MG	CCE MG
	(1)	(2)	(3)	(4)
Panel A. Dependent variable is log of life expectancy at birth
MPI	0.524***	0.036**	0.095***	0.010
	(0.044)	(0.016)	(0.023)	(0.007)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Observations	9,940	9,940	9,940	9,940
No. of countries	155	155	155	155
RMSE	0.120	0.050	0.059	0.028
Panel B. Dependent variable is log of infant mortality rate
MPI	−2.578***	−0.362***	−0.269***	−0.007
	(0.192)	(0.064)	(0.070)	(0.018)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Observations	9,411	9,411	9,411	9,411
No. of countries	155	155	155	155
RMSE	0.487	0.151	0.155	0.056

	MG	MG	CCE MG	CCE MG
	(1)	(2)	(3)	(4)
Panel A. Dependent variable is log of life expectancy at birth
MPI	0.524***	0.036**	0.095***	0.010
	(0.044)	(0.016)	(0.023)	(0.007)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Observations	9,940	9,940	9,940	9,940
No. of countries	155	155	155	155
RMSE	0.120	0.050	0.059	0.028
Panel B. Dependent variable is log of infant mortality rate
MPI	−2.578***	−0.362***	−0.269***	−0.007
	(0.192)	(0.064)	(0.070)	(0.018)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Observations	9,411	9,411	9,411	9,411
No. of countries	155	155	155	155
RMSE	0.487	0.151	0.155	0.056

Notes: This table shows outlier-robust mean estimates of the influence of democratic institutions on health outcomes across countries between 1789 and 2015. Standard errors are reported in parentheses. *** and ** denote statistical significance at the 1% and 5% levels, respectively. Source: Author’s estimations.

Table 3

Open in new tab

The heterogeneous effect of democracy on population health between 1789 and 2015.

	MG	MG	CCE MG	CCE MG
	(1)	(2)	(3)	(4)
Panel A. Dependent variable is log of life expectancy at birth
MPI	0.524***	0.036**	0.095***	0.010
	(0.044)	(0.016)	(0.023)	(0.007)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Observations	9,940	9,940	9,940	9,940
No. of countries	155	155	155	155
RMSE	0.120	0.050	0.059	0.028
Panel B. Dependent variable is log of infant mortality rate
MPI	−2.578***	−0.362***	−0.269***	−0.007
	(0.192)	(0.064)	(0.070)	(0.018)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Observations	9,411	9,411	9,411	9,411
No. of countries	155	155	155	155
RMSE	0.487	0.151	0.155	0.056

	MG	MG	CCE MG	CCE MG
	(1)	(2)	(3)	(4)
Panel A. Dependent variable is log of life expectancy at birth
MPI	0.524***	0.036**	0.095***	0.010
	(0.044)	(0.016)	(0.023)	(0.007)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Observations	9,940	9,940	9,940	9,940
No. of countries	155	155	155	155
RMSE	0.120	0.050	0.059	0.028
Panel B. Dependent variable is log of infant mortality rate
MPI	−2.578***	−0.362***	−0.269***	−0.007
	(0.192)	(0.064)	(0.070)	(0.018)
Control variables
Log of GDP per capita	No	Yes	No	Yes
Log of GDP per capita squared	No	Yes	No	Yes
Trade openness	No	Yes	No	Yes
Observations	9,411	9,411	9,411	9,411
No. of countries	155	155	155	155
RMSE	0.487	0.151	0.155	0.056

6.2.2 Dynamic estimates

To distinguish between the short- and long-run impacts of democratic institutions on population health, I set up a dynamic error correction model (ECM) as follows:

{Δ H E}_{i t} = β_{0 i} + β_{1 i} {H E}_{i, t - 1} + β_{2 i} {MPI}_{i, t - 1} + β_{3 i} X_{i, t - 1} + β_{4 i} {Δ MPI}_{i t} + β_{5 i} {Δ X}_{i t} + \sum_{j = 0}^{p} λ_{i j} {Δ f}_{t - j} + μ_{i} f_{t - 1} + ε_{i t},

(9)

where

{Δ H E}_{i t}

⁠,

{Δ MPI}_{i t}

⁠, and

{Δ X}_{i t}

⁠, respectively, denote health outcomes, the V-Dem’s multiplicative polyarchy index, and control variables in first differences for country

i

in year

t

⁠;

{H E}_{i, t - 1}

⁠,

{MPI}_{i, t - 1}

⁠, and

X_{i, t - 1}

are lagged variables of the observables;

β_{4 i}

identifies the short-run impact of democracy on health outcomes while the long-run impact is captured by

- (β_{2 i} / β_{1 i})

⁠. I also augment the regression analysis with cross-sectional averages of the observables to account for cross-sectional dependence, as represented in the second line of Equation (9). Following Chudik and Pesaran (2015a) and Chudik et al. (2016), I include several lags of common proxies; the conventional lag length is given by

p = int (T^{\frac{1}{3}}) = 3

⁠.

Table 4 contains the dynamic ECM estimates. For ease of comparison, I first estimate Equation (9) by using the MG estimator (Column 1). In Columns (2)–(5), I augment all the regressions with cross-sectional averages of the observables and vary the choice of lag length of common factors. The MG long-run estimates presented in Panel A suggest that democratic institutions have a positive impact on life expectancy at birth. I also observe substantial decreases in the statistical precision and magnitude of the long-run effect of democracy on life expectancy at birth when I explicitly account for parameter heterogeneity and cross-sectional dependence. This finding remains intact when I re-estimate Equation (9) but use infant mortality as an alternative outcome variable (Panel B). The short-run coefficients have expected signs but are imprecisely estimated in most cases. This provides suggestive evidence of time lags in the contribution of democratic institutions to achieving better health outcomes, consistent with the results reported in Table 2.

Table 4

Open in new tab

The heterogeneous effect of democracy on population health between 1789 and 2015, dynamic ECM estimates.

		$p = 0$	$p = 1$	$p = 2$	$p = 3$
		Lag length of common proxies
	(1)	(2)	(3)	(4)	(5)
	MG	CCE MG	CCE MG	CCE MG	CCE MG
Panel A. Dependent variable is log of life expectancy at birth in first differences
Long-run coefficient
MPI	0.229***	0.019	0.016	0.028**	0.025**
	(0.061)	(0.014)	(0.013)	(0.013)	(0.010)
Short-run coefficient
MPI	0.001	0.003	0.003	0.006**	0.007***
	(0.002)	(0.002)	(0.002)	(0.002)	(0.003)
Error-correction coefficient
Log of life expectancy at birth lagged	−0.032***	−0.174***	−0.177***	−0.187***	−0.216***
	(0.003)	(0.016)	(0.017)	(0.019)	(0.022)
Observations	9,674	9,674	9,545	9,362	9,179
# of countries	167	167	166	166	166
RMSE	0.034	0.025	0.023	0.021	0.020

Panel B. Dependent variable is log of infant mortality rate in first differences
Long-run coefficient
MPI	17.574	−0.472**	−0.514***	−0.368**	−0.252*
	(20.699)	0.196	0.189	(0.150)	0.142
Short-run coefficient
MPI	−0.012	−0.004	−0.005	−0.006	−0.004
	(0.008)	(0 .006)	(0.006)	(0.005)	(0.006)
Error-correction coefficient
Log of infant mortality rate lagged	0.002	−0.036***	−0.038***	−0.040***	−0.042***
	(0.002)	(0.007)	(0.007)	(0.008)	(0.009)
Observations	10,645	10,645	10,635	10,430	10,227
No. of countries	167	167	166	166	166
RMSE	0.055	0.045	0.045	0.043	0.042

		$p = 0$	$p = 1$	$p = 2$	$p = 3$
		Lag length of common proxies
	(1)	(2)	(3)	(4)	(5)
	MG	CCE MG	CCE MG	CCE MG	CCE MG
Panel A. Dependent variable is log of life expectancy at birth in first differences
Long-run coefficient
MPI	0.229***	0.019	0.016	0.028**	0.025**
	(0.061)	(0.014)	(0.013)	(0.013)	(0.010)
Short-run coefficient
MPI	0.001	0.003	0.003	0.006**	0.007***
	(0.002)	(0.002)	(0.002)	(0.002)	(0.003)
Error-correction coefficient
Log of life expectancy at birth lagged	−0.032***	−0.174***	−0.177***	−0.187***	−0.216***
	(0.003)	(0.016)	(0.017)	(0.019)	(0.022)
Observations	9,674	9,674	9,545	9,362	9,179
# of countries	167	167	166	166	166
RMSE	0.034	0.025	0.023	0.021	0.020

Panel B. Dependent variable is log of infant mortality rate in first differences
Long-run coefficient
MPI	17.574	−0.472**	−0.514***	−0.368**	−0.252*
	(20.699)	0.196	0.189	(0.150)	0.142
Short-run coefficient
MPI	−0.012	−0.004	−0.005	−0.006	−0.004
	(0.008)	(0 .006)	(0.006)	(0.005)	(0.006)
Error-correction coefficient
Log of infant mortality rate lagged	0.002	−0.036***	−0.038***	−0.040***	−0.042***
	(0.002)	(0.007)	(0.007)	(0.008)	(0.009)
Observations	10,645	10,645	10,635	10,430	10,227
No. of countries	167	167	166	166	166
RMSE	0.055	0.045	0.045	0.043	0.042

Notes: This table shows outlier-robust mean estimates of the long-run and short-run impacts of democratic institutions on health outcomes across countries between 1789 and 2015. Standard errors are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. Source: Author’s estimations.

Table 4

Open in new tab

The heterogeneous effect of democracy on population health between 1789 and 2015, dynamic ECM estimates.

		$p = 0$	$p = 1$	$p = 2$	$p = 3$
		Lag length of common proxies
	(1)	(2)	(3)	(4)	(5)
	MG	CCE MG	CCE MG	CCE MG	CCE MG
Panel A. Dependent variable is log of life expectancy at birth in first differences
Long-run coefficient
MPI	0.229***	0.019	0.016	0.028**	0.025**
	(0.061)	(0.014)	(0.013)	(0.013)	(0.010)
Short-run coefficient
MPI	0.001	0.003	0.003	0.006**	0.007***
	(0.002)	(0.002)	(0.002)	(0.002)	(0.003)
Error-correction coefficient
Log of life expectancy at birth lagged	−0.032***	−0.174***	−0.177***	−0.187***	−0.216***
	(0.003)	(0.016)	(0.017)	(0.019)	(0.022)
Observations	9,674	9,674	9,545	9,362	9,179
# of countries	167	167	166	166	166
RMSE	0.034	0.025	0.023	0.021	0.020

Panel B. Dependent variable is log of infant mortality rate in first differences
Long-run coefficient
MPI	17.574	−0.472**	−0.514***	−0.368**	−0.252*
	(20.699)	0.196	0.189	(0.150)	0.142
Short-run coefficient
MPI	−0.012	−0.004	−0.005	−0.006	−0.004
	(0.008)	(0 .006)	(0.006)	(0.005)	(0.006)
Error-correction coefficient
Log of infant mortality rate lagged	0.002	−0.036***	−0.038***	−0.040***	−0.042***
	(0.002)	(0.007)	(0.007)	(0.008)	(0.009)
Observations	10,645	10,645	10,635	10,430	10,227
No. of countries	167	167	166	166	166
RMSE	0.055	0.045	0.045	0.043	0.042

		$p = 0$	$p = 1$	$p = 2$	$p = 3$
		Lag length of common proxies
	(1)	(2)	(3)	(4)	(5)
	MG	CCE MG	CCE MG	CCE MG	CCE MG
Panel A. Dependent variable is log of life expectancy at birth in first differences
Long-run coefficient
MPI	0.229***	0.019	0.016	0.028**	0.025**
	(0.061)	(0.014)	(0.013)	(0.013)	(0.010)
Short-run coefficient
MPI	0.001	0.003	0.003	0.006**	0.007***
	(0.002)	(0.002)	(0.002)	(0.002)	(0.003)
Error-correction coefficient
Log of life expectancy at birth lagged	−0.032***	−0.174***	−0.177***	−0.187***	−0.216***
	(0.003)	(0.016)	(0.017)	(0.019)	(0.022)
Observations	9,674	9,674	9,545	9,362	9,179
# of countries	167	167	166	166	166
RMSE	0.034	0.025	0.023	0.021	0.020

Panel B. Dependent variable is log of infant mortality rate in first differences
Long-run coefficient
MPI	17.574	−0.472**	−0.514***	−0.368**	−0.252*
	(20.699)	0.196	0.189	(0.150)	0.142
Short-run coefficient
MPI	−0.012	−0.004	−0.005	−0.006	−0.004
	(0.008)	(0 .006)	(0.006)	(0.005)	(0.006)
Error-correction coefficient
Log of infant mortality rate lagged	0.002	−0.036***	−0.038***	−0.040***	−0.042***
	(0.002)	(0.007)	(0.007)	(0.008)	(0.009)
Observations	10,645	10,645	10,635	10,430	10,227
No. of countries	167	167	166	166	166
RMSE	0.055	0.045	0.045	0.043	0.042

Overall, additional evidence from three waves of democratization indicates that the health-enhancing impact of democracy is much smaller, in sign and statistical significance, than previously established. This is arguably attributed to inadequate attention to parameter heterogeneity and unobserved common factors in existing studies.

7. Conclusion

There is a widespread consensus that democratization augments a government’s decision to provide public healthcare services, thus improving population health. This article offers new estimates of the heterogeneous treatment effect of democracy on national health status. My results provide support for the health-enhancing impact of democratization, but the magnitude of such effect is much smaller than that implied by conventional estimates. My findings indicate that existing studies significantly inflate the economic significance of the contribution of democratization to achieving better health outcomes. This at least partially stems from inadequate attention to worldwide heterogeneity in the democracy-health relationship and the presence of global common shocks. Future research should focus on a more comprehensive analysis of the patterns of worldwide heterogeneity in the health returns to democratization. This promisingly improves our understanding of the democracy-health nexus, which helps inform policies geared towards achieving better health outcomes by developing inclusive political institutions.

Footnotes

Additionally, the advent of democracy may improve national health status through its positive influence on income, educational attainment, and health information (Batinti et al. 2022).

See Supplementary Appendix A1 for a review of related studies.

Vu et al. (2022) document empirically that the early formation and development of historical states above the tribal level help improve population health through conferring present-day countries with improved legal, fiscal, and organizational capabilities. As established in Galor and Moav (2007), countries that made an earlier transition to sedentary agriculture tend to enjoy better health outcomes because of greater exposure to infectious diseases and other environmental hazards.

Acemoglu et al. (2019) rely on the classifications of Cheibub et al. (2010) and Boix, Miller, and Rosato (2013) when data are unavailable in the Freedom House and Polity IV datasets.

Let $D_{i t}$ be the dichotomous measure of democracy; $D_{i t} = 1$ if country $i$ in year $t$ is classified as democracy and $D_{i t} = 0$ denotes a non-democratic regime. The incident of democratization is represented by ${Dem}_{i t} \in \{0,1\}$ ⁠; for a country transitioning from non-democracy to democracy, we have ${Dem}_{i t} = 1$ $(D_{i t} = 1, D_{i t - 1} = 0)$ ⁠. A democratic reversal is given by ${Reversal}_{i t} \in \{0,1\}$ ⁠; for a country transitioning from democracy to non-democracy, we have ${Reversal}_{i t} = 1$ $(D_{i t} = 0, D_{i t - 1} = 1)$ ⁠. For a country retaining democratic or non-democratic institutions, ${Dem}_{i t} = {Reversal}_{i t} = 0$ $(D_{i t} = D_{i t - 1} = 0| D_{i t} = D_{i t - 1} = 1)$ ⁠.

Supplementary Appendix Tables A1 and A2 provide more details of the incidents of democratic transitions.

This is consistent with the earlier argument that unprecedented health gains across the world were mainly driven by the international epidemiological transition starting in the 1940s. Indeed, there were limited improvements in population health in most parts of Africa, the Americas (except the USA), and Asia prior to the introduction of effective modern health interventions in the 1940s (Acemoglu and Johnson 2007). This is in contrast to early improvements in health conditions in Western Europe and the United States from the mid-19th century (Acemoglu and Johnson 2007).

These models carry an implicit assumption that parameter estimates are homogeneous across cross-sectional units in the panel. Specifically, the highly restrictive assumption of parameter homogeneity requires that the underlying relationship between democracy and population health is common across countries.

Due to the unknown relationship between unobserved common factors and the outcome variable, the empirical estimates may be susceptible to bias from model misspecification. Additionally, incorporating a large number of confounding factors in standard regression models significantly constrains the feasible degrees of freedom.

For example, Dell, Jones, and Olken (2012) establish that the negative impact of global climate change on economic growth is heterogeneous across countries depending on the level of economic development.

The MG estimates suggest that democratization helps reduce the rate of infant mortality, an association that is statistically significant at the 1% level (Columns 1 and 2). When I augment the regression analysis with common proxies, the mortality-reducing impact of democratization becomes imprecisely estimated at conventional levels of statistical significance (Columns 3 and 4).

See also Acharya et al. (2016) for discussions on intermediate variable bias induced by the inclusion of ‘bad controls’ in the regression analysis.

In Supplementary Appendix Table A6, I exclude the set of key control variables that are highly correlated with urbanization and state capacity. Results indicate that democracy enters all the regressions with an imprecisely estimated coefficient.

Supplementary material

Supplementary material is available at the Oxford Economic Papers Journal online. These are the data and replication files and the online appendix.

Funding

No funding was received for this study.

Conflict of interest statement. None declared.

Acknowledgements

I am grateful to Anindya Banerjee and anonymous referees for helpful comments that improved the article. The usual disclaimer applies.

References

Abadie

Diamond

Hainmueller

(

2010

) ‘

Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California’s Tobacco Control Program’

Journal of the American Statistical Association

105

493

–

505

Month:	Total Views:
August 2024	520
September 2024	494
October 2024	259
November 2024	262
December 2024	235
January 2025	172
February 2025	141
March 2025	381
April 2025	256

Article Contents

Revisiting the effect of democracy on population health

Abstract

1. Introduction

2. How does democratization affect population health?

3. Data and descriptive statistics

4. Empirical strategy

4.1 Potential threats to identification

4.2 The PCDID approach under a common factor framework

4.3 The baseline model

5. Empirical estimates

5.1 Main results

5.1.1 Static estimates

5.1.2 Dynamic estimates

5.2 Heterogeneity patterns

5.3 Robustness

6. Extensions

6.1 Evidence from a generalized synthetic control analysis

6.2 Evidence from three waves of democratization

6.2.1 Static estimates

6.2.2 Dynamic estimates

7. Conclusion

Footnotes

Supplementary material

Funding

Acknowledgements

References

Supplementary data

Citations

Views

Altmetric

Email alerts

Citing articles via

Latest

Most Read

Most Cited

This Feature Is Available To Subscribers Only