Agricultural Support and Food Import Dependency in Developing Countries: Evidence from Continuous Treatment Effect Models

Abstract

In this article, we analyze the relationship between agricultural support measures and food import dependency for a 52-country sample from 1985 to 2017 using databases from the World Bank, the Center for Systemic Peace, and the Groningen Center for Growth and Development. Using an instrumental variable framework, we apply a continuous treatment effect and control for endogeneity to describe the extent of food import dependency in response to domestic support for agriculture. Our results suggest heterogeneous associations for aggregate food import dependency at different levels of political aid intensity, while our analysis further reveals nonlinear dose–response patterns, suggesting that moderate levels of agricultural support are associated with lower food import dependency, whereas very high support intensities are not systematically associated with additional reductions. Although estimates of dose–response functions confirm that countries providing moderate support to agriculture tend to exhibit lower levels of agri-food import dependency, these findings should nevertheless be interpreted cautiously given the potential limitations regarding instrument validity and data availability. The primary contribution of our study is the explicit modeling of heterogeneous treatment intensity effects and potential endogeneity. These findings should therefore be interpreted as conditional empirical associations rather than definitive causal effects.

1. Introduction

1.1. Context

Classical international trade theory does not predict that trade openness is mechanically associated with lower food trade deficits. Rather, the theory of comparative advantage emphasizes welfare and income gains from trade, which may reduce the share of income spent on food and improve access to food through imports. However, in economies with limited agricultural resources, trade liberalization may be associated with higher food import dependency, especially when free trade replaces restrictive agricultural import policies. Consequently, the relationship between trade, agricultural policies, and food import dependency remains theoretically ambiguous [1,2], motivating our empirical investigation into how the effect of trade openness on food import dependency is related to domestic agricultural support policies and their intensity.

Importantly, food import dependency is not interpreted here as an inherently undesirable outcome, but rather as an equilibrium outcome reflecting countries’ structural characteristics, policy choices, and exposure to international markets. Food import dependency should not be interpreted as food insecurity. Food trade has long been recognized as a key mechanism for smoothing consumption and mitigating food insecurity, particularly in the presence of weather-related supply shocks. In addition, by integrating local markets into broader trading networks, food imports are associated with lower exposure to local yield shocks. Burgess and Donaldson [3] have shown that the dramatic reduction in transport costs due to the development of railroads in India substantially lowered famine incidence and weakened the link between local weather shocks and food availability. In this sense, food import dependency reflects exposure to international markets rather than food insecurity per se. While such exposure can enhance food security by stabilizing supplies, it may also increase vulnerability to external price shocks, foreign exchange constraints, or policy disruptions. This dual role of food imports underscores the importance of examining how domestic agricultural support policies condition the relationship between trade openness and food import dependency.

Food import dependency is also shaped by deep structural factors such as factor endowments, geography, and climatic constraints. Landlocked countries, small island states, and countries located in temperate zones face inherent limitations in producing certain food products, irrespective of agricultural support policies. In such contexts, some degree of food import dependency is unavoidable unless agricultural support generates substantial technological change and productivity gains. Despite the stabilizing role of food trade, several authors have expressed skepticism toward agricultural development strategies that prioritize export promotion in developing countries. Such strategies, when focused on a limited set of export-oriented crops, may increase dependence on food imports and heighten vulnerability to external shocks, with potential adverse implications for food security [1,2,4]. In this context, agricultural support policies may be associated with differing patterns of import dependency and may also create conditions under which export revenues support domestic food production.

However, other authors have emphasized that export earnings may relax foreign exchange constraints and finance investments in domestic food production, thereby mitigating food import dependency [5,6]. Lamb [6] concludes that a country’s export earnings could enable it either to import more food products or to import intermediate goods needed to increase domestic food crop productivity. In the latter case, local agricultural product promotion follows from the promotion of export crops and would help to limit dependence on food imports. In the same vein, other authors [5] have explored how the development of export crops could finance the creation of institutions to support agriculture (e.g., through investments in infrastructure, extension services, agricultural research, or subsidies).

The literature highlights how changes in agricultural support and trade policies, often motivated by domestic price stabilization objectives, can increase countries’ exposure to international food markets. In particular, policies of food price insulation, which are generally explained by governments’ desire to reduce domestic price volatility or by loss aversion, are well documented in the literature, as are their unintended consequences in terms of magnifying price volatility at the world level. Larochez-Dupraz and Huchet-Bourdon [7,8], for instance, focus on how changes in the nominal rate of assistance (NRA) affect food import dependency rather than food import volumes per se.

However, agricultural support measures can be heterogeneous between countries and between regions, which can affect trade results, including food imports [9]. Moreover, different farming systems imply different levels of agricultural support [9].

Based on these conclusions, one can question the role that support measures play in the food import dependencies of developing countries. Although an extensive amount of scholarship has examined agricultural support, trade outcomes, and food security, relatively fewer studies have jointly examined heterogeneous agricultural support intensity and food import dependency while explicitly accounting for treatment heterogeneity and potential endogeneity within a continuous treatment framework [7,8]. Our study contributes to this literature by examining how the intensity of agricultural support policies is associated with food import dependency across developing countries, recognizing that such dependency reflects both policy choices and structural constraints. Our contribution does not lie in establishing the existence of a relationship between agricultural support and food import dependency per se, but rather in refining the empirical analysis by explicitly modeling heterogeneous treatment intensity effects and potential endogeneity.

To achieve this, we estimate a dose–response function that accounts for potential endogeneity using the continuous treatment model proposed by Cerulli et al. [10]. Compared with standard treatment effect approaches [11], this framework relies on a stepwise instrumental variable regression procedure [10] designed to mitigate potential endogeneity biases in the estimation of the relationship between agricultural support measures and food import dependency in the countries studied [12]. In addition, the continuous treatment framework explicitly allows treatment effects to vary across support intensities and accommodates nonlinear dose–response relationships, whereas conventional panel IV approaches primarily estimate average marginal effects. However, one major challenge in implementing endogenous treatment models lies in the identification of valid instruments [13]. To address this issue, we draw on the theoretical and empirical literature on the determinants of agricultural support, which provides several potential instruments for addressing endogeneity concerns [14,15,16,17,18].

1.2. Research Gap

Despite the extensive literature on agricultural protection, trade policy, and food security, relatively few studies have examined how varying intensities of agricultural support are associated with food import dependency in developing countries while explicitly accounting for treatment heterogeneity and potential endogeneity. Existing studies often rely on aggregate support indicators, such as the nominal rate of assistance or producer subsidy equivalents [7,9,15,19,20,21], and generally use standard panel data approaches that may not adequately capture nonlinear responses to support intensity or the endogenous nature of agricultural policy decisions.

While previous contributions, including those by Larochez–Dupraz and Huchet–Bourdon [7], have primarily focused on the relationship between aggregate support measures and food imports using conventional random-effects models, agricultural support policies are inherently heterogeneous across countries and regions [9], and different policy instruments may affect food imports through distinct mechanisms [22,23,24]. Consequently, aggregate indicators may mask important variations in the relationship between agricultural support intensity and food import dependency.

Moreover, because agricultural support decisions are nonrandom and may be influenced by political, institutional, and structural factors [25,26,27], standard estimation approaches may suffer from endogeneity bias. Although the literature recognizes these challenges, relatively few empirical studies explicitly address both treatment heterogeneity and potential endogeneity simultaneously when examining food import dependency.

This study contributes to the literature by applying a continuous treatment effect framework combined with instrumental variables, following Cerulli [28] and Baum and Cerulli [29], to analyze the heterogeneous associations between agricultural support intensity and food import dependency in developing countries. By explicitly accounting for nonlinear responses and potential endogeneity concerns, we provide a more nuanced assessment of the relationship between agricultural support and food import dependency.

More importantly, relatively little empirical evidence exists regarding whether different intensities of agricultural support are associated with distinct food import dependency patterns across regions and support levels. In this study, we address this gap by explicitly modeling heterogeneous and nonlinear treatment intensity effects.

1.3. Some Stylized Facts

Figure A2, Figure A3 and Figure A4 (Appendix B) reveal that, over the periods 1985–1994, 1995–2004, and 2005–2017, the Africa region had, on average, the lowest levels of agricultural support compared to other studied regions of the world, while also having the highest levels of food import dependency during the same periods. The figures also show that, although the average level of agricultural support in this region of the world increased over the different periods, such support has not always improved the incomes of agricultural producers (the average values of nominal rates of assistance to agriculture have remained negative over all these periods in the Africa region).

In the Latin America and Asia regions, although some individual countries exhibit patterns similar to those observed in Africa, when considering the countries in these regions of the world overall, on average, the rates of food import dependency remained considerably lower than those observed in Africa for the periods 1985–1994, 1995–2004, and 2005–2017 (see Figure A2, Figure A3 and Figure A4—World Regions and Latin American Countries); Similarly, the Asia region has remained relatively less dependent on food imports than the Latin American region over these periods.

In the Europe and Oceania regions, agricultural assistance measures remained positive for most countries over the three periods of 1985–1994, 1995–2004, and 2005–2017 (see Figure A2, Figure A3 and Figure A4—World Regions, Europe, and Oceania), indicating that support for agriculture consolidated the income of producers. However, at the same time, the import dependency of these countries remained low compared to countries in other regions of the world (see Figure A2, Figure A3 and Figure A4).

These findings suggest a substantial heterogeneity in agricultural support across both regions and countries. To examine how agricultural support intensity is associated with food import dependency, we adopt an endogenous treatment approach based on instrumental variable techniques, one that assumes that macroeconomic and sociopolitical factors partly explain cross-country differences in agricultural support intensity.

These regional patterns should nevertheless be interpreted within broader structural transformation processes. The substantial development literature has documented that developing economies often implicitly or explicitly tax agriculture during early stages of industrialization, while agricultural protection tends to increase as economies diversify and industrial sectors mature. Differences in land endowments, demographic pressures, export specialization, and institutional capacity may therefore partly explain the observed heterogeneity in agricultural support and food import dependency across regions.

1.4. Research Objective and Contribution

In this study, we examine the relationship between agricultural support intensity and food import dependency in developing countries. More specifically, we investigate whether different levels of agricultural support are associated with heterogeneous patterns of food import dependency using a continuous treatment effect framework that accounts for potential endogeneity.

Our study contributes to the literature in three main ways that extend existing empirical approaches beyond standard average-effect estimation: First, we explicitly examine whether agricultural support intensity is associated with nonlinear dose–response relationships rather than assuming homogeneous marginal effects across support levels. Second, we highlight substantial regional heterogeneity in estimated treatment effects, suggesting that the relationship between agricultural support and food import dependency differs across structural and institutional contexts. Third, we highlight the importance of considering both structural constraints and policy heterogeneity when examining food import dependency across developing countries. Unlike conventional panel data approaches focusing on average marginal effects, our analysis explicitly examines nonlinear and heterogeneous treatment intensity associations.

Our study addresses the following research question:

How does the intensity of agricultural support relate to food import dependency in developing countries?

The primary contribution of this study therefore lies not in identifying a previously unknown relationship between agricultural support and food import dependency, but rather in refining the empirical assessment of this relationship through the explicit modeling of treatment intensity heterogeneity and potential endogeneity.

The remainder of this article is organized as follows: Section 2 reviews the literature on agricultural support and food import dependency. Section 3 presents the materials, variables, and empirical methodology. Section 4 reports our empirical results and discusses the heterogeneous treatment effects across regions. Section 5 concludes the article.

2. Literature Review

2.1. Determinants of Agricultural Support

In this section, we present the political economy and institutional foundations motivating both the empirical specification and the selection of explanatory and instrumental variables used in the continuous treatment framework.

Theoretical contributions to the field of political economy related to institutions and democracy [27,30], the role of constitutions [31], political competition [26], and electoral institutions [32] have deepened research on the determinants of support in agriculture. For Olper [27], farmers’ voices can be better heard in an electoral democracy where interest groups are free to compete for political rent, while authoritarian regimes, which are better able to discourage the rent-seeking activities of interest groups, tax or do not support their agricultural sectors [27]. These arguments suggest that agricultural protection tends to be higher in democratic settings [25,33,34].

However, the problem with this view of democracy is the likelihood that governments will adopt ineffective policies that benefit specific interest groups [25,33,34]. Indeed, an important difference between democratic and authoritarian regimes is the degree of external influence [27]. In a well-functioning democracy, public policies are defined under the influence of various groups that sometimes have competing interests [25,33,34]. According to Olper [27], divergences in the interests of different groups can undermine the effectiveness of defined public policies, while, in a less democratic environment, the government is less affected by such interest groups [27]. Thus, one might suggest that agricultural policy transfers should decrease in a democracy [27,30]. Thus, theoretically, the net effect of democracy on agricultural protection is uncertain, a result that, in general, is consistent with existing empirical evidence (e.g., see [19]).

Based on these theoretical contributions from the literature, Fałkowskia and Olper [26] confirm the existence of a positive link between political competition and agricultural protection. For these authors, political platforms cannot neglect the preferences of the majority [26]. In countries where the agricultural population is the majority, Fałkowskia and Olper [26] find that political competition increases the chances of the implementation of public policies favoring agricultural interests. Furthermore, a second explanation of political competition’s effect on agricultural protection can be found in the literature on the median voter theorem [35]. The assumption of this theorem is that, in countries where the agricultural population constitutes the majority, the median income likely closely follows agricultural income [26].

Therefore, winning an election may require accommodating the preferences of agricultural voters. Nevertheless, political competition can also have a positive impact on agricultural protection in countries where the agricultural electorate is only marginal [26]. In these contexts, parties with a high level of political competition are encouraged to attract voters outside their traditional electoral base [32]. Political platforms can thus respond to the wishes of relatively small groups of pivotal voters [26].

However, political competition could also increase the level of agricultural protection without reference to the size of the agricultural electorate [26]. Scholarship on the process of political reversal has shown that, depending on the risk of losing elections, political elites can be encouraged to adopt socially ineffective policies. In addition, Persson et al. [36] have shown that government spending is higher in coalition governments than in a one-party government. This suggests that a positive relationship exists between political competition and public spending, which thus affects agricultural protectionism [26]. As we have shown previously, a positive link between political competition and agricultural policy could be established in several respects.

Olson [37] finds that a shift from taxes to agricultural subsidies during economic development is linked to a reduction in the stowaway problem associated with farmers’ collective actions. If, in addition, transport, communication infrastructures, and education develop as the economy develops, organizational efficiency increases, which increases the difficulties of collective action in rural areas [37]. However, Becker [38] adds that the influence of each group depends on how well the group expends resources to produce pressure. It is therefore not difficult to imagine that, in the context of inequalities in income ownership, agricultural policies reflect the expectations of the most vulnerable groups of people [38].

The various developments discussed above suggest a variety of instruments for agricultural support measures. For the purposes of this study, we have chosen the human capital index, income inequality, degree of democracy, regime sustainability, openness, and competitiveness of executive recruitment as determinants of governmental decisions to provide agriculture support measures. These variables were selected because the literature suggests that political institutions, inequality, human capital accumulation, and governance structures influence governments’ incentives and capacities to support agriculture, while also shaping broader economic and trade dynamics [26,31,37,38,39].

Beyond political and institutional determinants, the central role of factor endowments—particularly land endowments—in shaping agricultural protection policies is emphasized in the literature. Countries with limited arable land relative to their population, such as Japan or Korea, tend to protect agriculture, reflecting both concerns about food availability and the political economy of protection instruments. By contrast, land-abundant countries, such as Argentina, have historically taxed agriculture, as protecting large export-oriented sectors would require costly export subsidies and impose substantial fiscal burdens. Even high-income net exporters face strong constraints in protecting agriculture due to the budgetary costs associated with financing support for exports.

In addition to explaining cross-country differences in the level of agricultural support, the literature also examines short-term variations in protection. A large body of work on food price insulation has shown that protection rates tend to vary inversely with world prices, as governments adjust trade policies to reduce domestic price volatility or in response to loss aversion. Consistent with this view, Bastos et al. [40] document a direct link between domestic output shocks and trade policy responses, finding that adverse rainfall shocks are associated with reductions in import tariffs.

2.2. Agricultural Support and Import Dependency

Multiple agricultural support instruments have been used by various countries both independently and in combination [22], and have had varying effects on food imports [24] (p. 48). These instruments differ fundamentally in the channels through which they affect food imports. Border measures such as tariffs influence food imports through a dual mechanism by simultaneously encouraging domestic production and discouraging domestic consumption. In contrast, domestic support measures primarily operate through production incentives, without directly restraining demand. As a result, instruments of similar ad valorem magnitudes may have markedly different effects on food imports. Against this analytical background, the literature commonly distinguishes three broad categories of policy instruments that may distort trade [22,23,24]: First are the instruments aimed at limiting imports (e.g., import quotas, customs duties, variable levies, or nontariff barriers). Second are measures to encourage exports (e.g., export subsidies, variable refunds, or export credits). Third are the so-called internal support measures aimed directly at increasing the incomes of producers (e.g., production subsidies, subsidies for the use of factors of production, subsidies for domestic consumption, support for producer prices through public purchases, or guaranteed prices). Import restriction measures protect local producers against competing imports [23].

According to many authors, the use of these instruments can induce local production exceeding the levels that would exist at market prices, to the detriment of international producers and exporters [41]. For example, customs duties raise consumer prices for imported products, increasing government revenues and encouraging domestic producers to increase their production of import substitutes [23]. Thus, customs duties are a form of incentive to develop local production and limit dependence on imports [22]. Because tariffs simultaneously reduce domestic demand and increase domestic supply, their expected impact on food imports is generally larger than that of domestic support measures of comparable magnitude. By contrast, import quotas operate through a different mechanism: unlike customs duties, the income from the differential between the sale price of an imported product with and without protection goes in whole or in part to license holders [39]. These license holders form what are called quota rents which, to some extent, can be captured by the state when licenses are sold or auctioned [39].

Furthermore, although nontariff barriers, for example sanitary and phytosanitary measures applied to imports, are not in themselves instruments of trade protection, they can become so [23,42]. This is often the case when they are used in such a way as to serve as a shield for national producers against international competition [42]. Indeed, it is not uncommon for states to adopt such measures—not so much to prevent health risks [23] but rather in response to the activism of certain lobbies [42]. Regarding export support instruments, export subsidies increase effective producer prices and producer incomes, but at a fiscal cost to governments [43,44]. By encouraging exports, such subsidies may also create incentives for imports, unless they are paired with border measures—such as import tariffs or export restrictions—designed to prevent arbitrage or “round-tripping.”

The exchange rate is also a significant trade determinant [43]. Indeed, all other things being equal, an exchange rate reduction can lead to an increase in exports and a decrease in imports [43]. It is important to remember, however, that because depreciation can raise the prices of exported and imported goods, it tends to have an inflationary effect [43]. As a result, a fear of fueling internal inflation often prevents authorities from resorting to depreciation when faced with an unrealized inflation situation, despite its potentially positive impact on the trade balance [43].

Finally, agricultural subsidies can increase local production and reduce competing imports. For example, input subsidies and investment subsidies aim to reduce production costs by lowering the costs of inputs and investments, respectively [45,46]. Consequently, they can stimulate local production [45] while simultaneously helping local products to be more competitive than foreign products [45]. Again, however, these are measures that affect product prices and/or producer income. Output price subsidies affect food imports exclusively through their supply-side effects. By contrast, input subsidies—when applied to inputs supplied elastically—may generate a stronger production response than output price subsidies of equivalent value, as a smaller share of the policy benefit accrues to fixed factors. In addition, as discussed above, the effects of these measures on import demand are influenced by the price elasticity and income elasticity of import demand [47].

Increasingly, the literature on the effects of agricultural support measures on food imports uses aggregate support measures, such as the nominal or effective rate of protection [21], the equivalent of production subsidies [9,48], or the nominal or effective assistance rate [1,7,15]. In their work, Larochez–Dupraz and Huchet–Bourdon [7] analyzed the effects of the nominal assistance rate for agriculture on the imports of 39 developing countries over the period 2005–2010. Using two random-effects panel data models, these authors estimated the effects of the nominal assistance rate on food market prices, and then the effects of the latter on food imports [7]. Their results show that changes in the nominal rate of assistance move inversely with world food prices, consistent with price insulation behavior [7]. However, such adjustments do not necessarily imply a stabilization of food import volumes or values, as price and quantity effects may offset each other [7].

Although interesting, these results could be biased, as they do not consider the potential presence of endogeneity in the adopted measure of domestic support for agriculture. Indeed, the decision of whether to support agriculture is a nonrandom decision that could be influenced by unobserved factors which could also influence food imports. In addition, it should be noted that the aggregates of nominal assistance rates are not able to accurately capture the relative contributions to the reduction in food imports of the various policy instruments used in their construction [49]. This is particularly true in cases where certain policies, such as import taxes, could have negative effects on those imports, while other policies, such as export subsidies, could have positive effects on imports [49]. Likewise, if the import-competing and exportable sectors are each subject to trade taxes, the nominal rates of assistance obtained can be close to zero, even without zero food import flows [49]. In this case, using a random-effects model to estimate the effects of the nominal assistance rate on imports might not show the contribution of zero values to the nominal assistance rate.

Taken together, these considerations highlight that aggregate measures such as the nominal rate of assistance combine policy instruments with heterogeneous and sometimes offsetting effects on food imports. This complexity, combined with the endogenous nature of agricultural support policies, motivates our empirical approach that explicitly accounts for heterogeneous responses to support intensity. To this end, we estimate a dose–response function for the nominal rate of assistance on food imports.

3. Materials and Methods

Figure 1 summarizes the overall research design and empirical strategy adopted in this study, from data collection and variable construction to endogeneity assessment, dose–response estimation, regional heterogeneity analysis, and the interpretation of treatment effects.

3.1. Data Sources, Sample and Variables

We used annual data from 52 developing countries (

Table A3. List of developing countries used in the study.

04 World Regions
Africa	Asia	Europe and Oceania	Latin America
52 Countries
Benin	Bangladesh	Bulgaria	Argentina
Burkina Faso	China	Czech Republic	Brazil
Cameroon	India	Hungary	Chile
Chad	Indonesia	Latvia	Colombia
Cote d’Ivoire	Malaysia	Lithuania	Dominican Republic
Egypt	Pakistan	Russian Federation	Ecuador
Ethiopia	Philippines	Slovakia	Mexico
Ghana	Sri Lanka	Slovenia	Nicaragua
Kenya	Thailand	Turkey
Madagascar	Vietnam	Ukraine
Mali		Ireland
Morocco		Portugal
Mozambique		New Zealand
Nigeria
Senegal
South Africa
Tanzania
Togo
Uganda
Zambia
Zimbabwe

Source: Authors, 2021.

in Appendix E), derived from different sources (see

Table 1. Variables.

	Variables	Meanings and References	Sources
Dependent variable
	lfidi	Logarithm of the food import dependency index (in %) [7]	The method of calculation and the sources of the data used are specified in the following paragraphs.
Treatment variables
	cna	Nominal assistance coefficient [12]	Calculated from nominal rate data. These are taken from the World Bank database (see details in the following paragraphs).
	t	Binary variable that takes the value 1 if cna > 0 and 0 if cna = 0 [12]	Calculations of authors
Covariates
	lcons	Logarithm of the annual growth rate of household consumption expenditure per capita (in %) [17,42]	World Bank (https://data.worldbank.org/indicator/NE.CON.PRVT.PC.KD.ZG (accessed on 10 June 2020))
	ltar	Logarithm of average tariff rates applied (in %) [12]	World Bank (https://data.worldbank.org/indicator/TM.TAX.MRCH.WM.AR.ZS (accessed on 10 June 2020))
	lpop	Logarithm of total population (in 100 million people) [12]	World Bank (https://data.worldbank.org/indicator/SP.POP.TOTL (accessed on 10 June 2020))
	lagri	Logarithm of share of agricultural production in GDP (in %) [12]	World Bank (https://data.worldbank.org/indicator/SP.POP.TOTL (accessed on 10 June 2020))
Instruments
cna	hc	Human capital index (in %) [50]	Growth and Development Centre of Groningen (www.ggdc.net/pwt (accessed on 10 June 2020))
	gini	Gini index [51]	World Bank (http://iresearch.worldbank.org/PovcalNet/index.htm (accessed on 10 June 2020))
t	democ	Democracy index [31]	Center for Systemic Peace (https://www.systemicpeace.org/inscrdata.html (accessed on 10 June 2020))
	durable	Indicator of the sustainability of schemes [31]	Center for Systemic Peace (https://www.systemicpeace.org/inscrdata.html)
	xropen	Opening indicator for executive recruitment [26]	Center for Systemic Peace (https://www.systemicpeace.org/inscrdata.html)
	xrcomp	Competitiveness indicator for executive recruitment [26]	Center for Systemic Peace (https://www.systemicpeace.org/inscrdata.html)
	hc	Human capital index (en %) [50]	Growth and Development Centre of Groningen (www.ggdc.net/pwt)
	gini	Gini index [51]	World Bank (http://iresearch.worldbank.org/PovcalNet/index.htm)

Source: Authors.

), from the period 1985–2017. Table 1 presents the variables, definitions, references, and data sources used in the analysis.

Table 2. Descriptive statistics.

Variable	Description (Units)	Mean	Std. Dev.	Min	Max
fidi	Food import dependency ratio (in %)	0.2108	0.2864	0.1328	4.9876
cons	Growth rate of per capita consumption expenditure (in %)	0.6542	0.1281	0.2748	1.3842
tar	Average of the weighted tariff rates per imported product (in %)	9.6665	7.4791	0.0440	56.3600
pop	Total population (in millions of inhabitants)	85.1634	226.5825	1.9434	1421.0220
cna	Nominal assistance coefficient	38.8640	12.0160	0	100
t	Binary variable which takes the value 1 if cna> 0 and 0 if cna = 0	0.9435	0.2310	0	1
hc	Human capital index	2.1810	0.6620	1.0220	3.7940
gini	Gini index	41.0970	9.1680	19.1720	64.80
agripib	Production of the agricultural sector (in % of GDP)	0.2975	2.1740	4.92 × 10−6	57.0475
poprur	Rural population (in millions of inhabitants)	48.8789	147.0594	0.5185	889.2167
democ	Democracy index	5.2930	3.6750	0	10
durable	Sustainability index of regimes	18.5820	22.8380	0	140
xrcomp	Competitiveness index for executive recruitment	2.0520	1.0840	0	3
xropen	Opening index of executive recruitment	3.4580	1.3410	0	4
N	Sample size	1716

Source: Authors.

reports the descriptive statistics for the main variables. Appendix A describes how the agricultural support indicator and the food import dependency index are calculated. The validity of the exclusion restrictions and instruments is assessed in the Results section through a series of endogeneity and overidentification tests.

Table 2 presents the descriptive statistics for the variable data used in this study.

Although the selected instruments are grounded in the political economy literature on agricultural support determination, some institutional variables, such as democracy, inequality, human capital, or regime stability, may also influence food import dependency through channels other than agricultural support. Consequently, the exclusion restriction may not be perfectly satisfied, which constitutes an important limitation of the present study. Accordingly, the exclusion restriction should be interpreted as approximate rather than strict.

Our analysis ends in 2017, as harmonized and internationally comparable data on agricultural support measures were not consistently available beyond that year for the full sample of countries included in this study. The empirical analysis therefore reflects relationships observed prior to major global disruptions affecting food systems after 2017, including the COVID-19 pandemic, supply-chain disruptions, and subsequent food price shocks. Consequently, our findings should be interpreted as reflecting the pre-2018 international policy environment.

3.2. Estimation Strategy

In practice, our analysis estimates both the model with exogenous treatment (i.e., where selection of the treatment level depends only on observable factors, as in Equation (7)) and the model with endogenous treatment (i.e., where selection of the treatment level depends on both observable and unobservable factors, as in Equations (8)–(10)). Because we postulated from the outset that treatment is endogenous, we then draw from the work of Heckman [52], Durbin [53], Wu [54], and Hausman [44] to confirm or refute our hypothesis regarding the endogeneity of our treatment variable.

We test the validity of our instruments using both Durbin and Wu–Hausman’s endogeneity tests [55] and Basmann’s [40] and Sargan’s [56] chi-squared tests, supplemented with a correlation test between the variables. Consequently, the identification strategy should be interpreted as relying on an approximate rather than strict exclusion restriction.

Compared to the estimation approach proposed by Hirano and Imbens [11], our estimation approach, which was inspired by the work of Cerulli [28], does not require a full normality assumption. We model the outcome variable for both zero and nonzero values of the treatment variable. Moreover, we can account for both the importance of null values in the treatment variable and the endogeneity of the latter under reasonable assumptions.

Finally, it should be added that, for both the ordinary least squares and the instrumental variable estimation, we model the dose–response function by approximating it using a third-degree polynomial. A third-degree polynomial is used to allow for flexible non-linearities in the dose–response relationship, while avoiding overfitting that may arise with higher-order polynomials. This choice is consistent with prior applications in the continuous treatment literature (Cerulli [28]; Baum and Cerulli [29]). Because trade-weighted tariffs may understate protection levels in the presence of prohibitive tariffs, we acknowledge potential endogeneity concerns arising from the construction of weighted average tariffs, and the estimated effects should be interpreted cautiously. Although alternative tariff specifications could provide additional robustness assessment, comparable data limitations prevented their implementation within the scope of the present study. More broadly, additional robustness analyses, including alternative instrument sets, alternative agricultural support indicators, and sensitivity analyses based on alternative model specifications, were beyond the scope of the available data and should be prioritized in future research.

All statistical analyses were performed using Stata/SE 16.0 (StataCorp LLC, College Station, TX, USA). Continuous treatment effect models were estimated following the approach proposed by Cerulli [28] and implemented through the corresponding Stata routines.

3.3. Econometric Framework

Our econometric framework is inspired by Baum and Cerulli [29] and Cerulli [28]. We let the treatment variable be the “nominal rate of assistance (See Appendix A)” transformed into the “nominal coefficient of assistance” (cna) and the outcome be the “food import dependency index (See Appendix B)” (fidi). The cna corresponds to a monotonic transformation of the nominal rate of assistance (NRA), designed to express agricultural support as a positive coefficient bounded between 0 and 100 for estimation purposes. We consider two different and exclusive outcomes: one referring to a country i at time T when it is treated, ${fidi}_{1iT}$, and the other referring to the same country when it is not treated, ${fidi}_{0iT}$. We denote $t_{iT}$ as a treatment indicator, which takes the value 1 for treated countries and 0 for untreated countries, where $x_{1iT}$ = ($x_{1iT}$, $x_{2iT}$, $x_{3iT}$, …, $x_{MiT}$) is a line vector of M exogenous and observable characteristics for country i (i = 1,… N). N represents the number of countries participating in the experiment. Here, $N_1$ is the number of treated countries, and $N_0$ is the number of untreated countries; thus, N = $N_1$ + $N_0$.

We define two separate functions, $g_1$($x_{iT}$) and $g_0$($x_{iT}$), as the responses of country i to the vector of variables $x_{iT}$ when country i is treated and untreated, respectively, at time T. ${\mu }_1$ and ${\mu }_0$ are two scalars, and $e_1$ and $e_0$ are two random variables with an unconditional mean of zero and constant variance. Finally, we define the variable ${cna}_{iT}$, which takes continuous values in the range [0;100], as the continuous processing indicator, and h(${cna}_{iT}$) is a general derivable function of ${cna}_{iT}$. In the following section, to simplify the notation, we will discard the indices i and T when defining quantities and relations.

Given the above notations, we assume a specific country category generation process for the two exclusive potential outcomes, ${ivsa}_1$ and ${ivsa}_0$, as follows:

$$\left\{\begin{array}{c}t=1 : { fidi}_1 = {\mu }_1+ g_1\left(x\right)+h\left(cna\right)+e_1\\ t=0 : { fidi}_0 = {\mu }_0+ g_0\left(x\right)+e_0\end{array}\right.,$$

(1)

where the function h(cna) is only nonzero in the processed state. Given this, we can also define the causal parameters of interest. Indeed, by defining the treatment effect as the difference TE = (${fidi}_1-{fidi}_0$), we define the treatment effect parameters of interest as the population average treatment effects (ATEs) conditional on x and f. In other words,

$$\left\{\begin{array}{c}ATE(x;cna)= E({fidi}_1-{fidi}_0|x,cna)\\ ATE(x;cna>0)= E({fidi}_1-{fidi}_0|x,cna>0)\\ ATE(x;cna=0)= E({fidi}_1-{fidi}_0|x,cna=0)\end{array}\right.,$$

(2)

where ATE denotes the effect of the overall average treatment, ATET represents the effect of the average treatment on treated countries, and ATENT represents the effect on untreated countries. Assuming a linear parametric form of the parameters $g_1\left(x\right)$ and $g_0\left(x\right)$, we can write the average treatment effect conditional on x and cna as follows:

$$ATE\left(x;cna;t\right)= t\left[\mu +x\delta +h\left(cna\right)\right]+(1-t)\left[\mu +x\delta \right],$$

(3)

where $g_1\left(x\right)=x{\delta }_1$ and $g_0\left(x\right)=x{\delta }_0$, $\mu = {\mu }_1- {\mu }_0$, and $\delta = {\delta }_1- {\delta }_0$. Then, the unconditional ATE related to Model (4.1) is equal to

$$ATE=p(t=1)\left[\mu +{\overline{x}}_{cna>0}\delta +{\overline{h}}_{cna>0}\right]+p(t=0)\left[\mu +{\overline{x}}_{cna=0}\delta \right],$$

(4)

where p() is a probability, and ${\overline{h}}_{cna>0}$ is the average of the response function taken over cna > 0. Insofar as, by the law of iterated expectation, we have ATE = $p\left(t=1\right)\times ATET+ p\left(t=0\right)\times ATENT$, we can derive the following from the previous formula:

$$\left\{\begin{array}{c}ATE = p(t=1)\left[\mu +{\overline{x}}_{cna>0}\delta +{\overline{h}}_{cna>0}\right]+p(t=0)\left[\mu +{\overline{x}}_{cna=0}\delta \right]\\ ATET = \mu +{\overline{x}}_{cna>0}\delta +{\overline{h}}_{cna>0}\\ ATENT = \mu +{\overline{x}}_{cna=0}\delta \end{array}\right.$$

(5)

where the dose–response function is obtained by averaging the ATE (x, cna) over x:

$$ATE(cna) = \left\{\begin{array}{c}ATET+\left[h\left(cna\right)-{\overline{h}}_{cna>0}\right]\\ ATENT si cna=0\end{array}\right.,$$

(6)

which is a function of the treatment intensity cna. Estimating Equation (6) under different identification assumptions is the main objective of the following sections.

3.3.1. Regression Model Under Exogenous Treatment

In view of the previous developments, starting from the definitions and assumptions established above and considering in particular the form of the potential outcomes in Model (1), we adopt the following potential outcome equation [57]: ${fidi}_{iT}$ = ${fidi}_{0iT}$ + t (${fidi}_{1iT}$ − ${fidi}_{0iT}$). Hence, the model to be estimated takes the following form:

$${fidi}_{iT}= {\mu }_{0T}+ t_{iT}ATE+ x_{iT}{\delta }_{0T}+ t_{iT}\left[x_{iT}- {\overline{x}}_T\right]{\delta }_T +t_{iT}\left[{h(cna}_{iT})- {\overline{h}}_T\right]+ {\eta }_{iT}$$

(7)

where ${\eta }_{iT}= e_{0iT}+t_{iT}(e_{1iT}- e_{0iT})$.

In practice, Equation (7) provides the basic model for estimating all our parameters (${\mu }_{0T}$, ${\mu }_{1T}$, ${\delta }_{0T}$, ${\delta }_{1T}$ and ATE) by ordinary least squares, as well as all remaining ATEs. A semiparametric or parametric approach can be used if a parametric or nonparametric form of the function h(cna) is assumed. However, to obtain a consistent estimate of the basic parameters, we must consider the likely existence of endogeneity in the treatment variable, which leads us to the regression model under endogenous treatment.

3.3.2. Regression Model Under Endogenous Treatment

Assuming endogeneity in our treatment variable, Equation (7) can no longer be estimated by ordinary least squares. However, an estimation using instrumental variables can be implemented to help address potential endogeneity bias [58]. To this end, it is sufficient to express the previous Model (7) in a semi-structural form. The model to be estimated then becomes

$${fidi}_{iT}={\mu }_{0T}+t_{iT}ATE+x_{iT}{\delta }_{0T}+t_{iT}\left[x_{iT}-{\overline{x}}_T\right]{\delta }_T+t_{iT}{H}_{1iT}+\mathrm{b}{t}_{iT}{H}_{2iT}+{ct}_{iT}{H}_{3iT}+{\eta }_{iT},$$

(8)

$${t^{\ast }}_{iT}=x_{t,iT}{\beta }_{t,T}+{ϵ}_{t,iT},$$

(9)

$${{cna}^{\prime }}_{iT}=x_{cna,iT}{\beta }_{cna,T}+{ϵ}_{cna,iT},$$

(10)

where $H_{1iT}$ = ${cna}_{iT}$ − E(${cna}_{iT}$), $H_{2iT}$ = ${cna}_{iT}^2$ − E(${cna}_{iT}^2$), and $H_{3iT}$ = ${cna}_{iT}^3$ − E(${cna}_{iT}^3$). Here, ${t^{\ast }}_{iT}$ represents the unobservable latent counterpart of the binary variable $t_{iT}$, $x_{t,iT}$ and $x_{cna,iT}$ are two sets of exogenous regressors $,$ and ${ϵ}_{t,iT}$, ${ϵ}_{cna,iT}$, and ${\eta }_{iT}$ are error terms. Equation (9) is the selection equation, which defines the regression explaining the net benefit indicator t, and the covariate vector $x_{t,iT}$ records the selection criteria used by countries to define the treated and untreated groups. On the other hand, Equation (10) is the treatment-level equation, which defines how the treatment level of countries is determined. The vector of covariates $x_{cna,iT}$ captures the exogenous variables considered to determine the treatment level. In Equation (8), the variables $t_{iT}$, $H_{1iT}$, $H_{2iT}$, and $H_{3iT}$ are endogenous, the last three being functions of the endogenous treatment cna.

4. Results

This section presents the main empirical findings from the continuous treatment effect models. We first examine the evidence supporting treatment endogeneity and the relevance of the identification strategy. We then analyze the estimated treatment effects, the nonlinear dose–response relationship, and the heterogeneity of the association between agricultural support and food import dependency across regions. Additional descriptive figures and supporting analyses are provided in the Appendix A, Appendix B, Appendix C, Appendix D and Appendix E. These appendices contain supplementary results and supporting materials referenced throughout the manuscript.

4.1. Agricultural Support Is Associated with Lower Food Import Dependency

We begin by estimating the association of the nominal assistance rate with the food import dependency using the OLS regression. The results are presented in

Table A1. Dose–response model estimates with exogenous treatment.

Dependent Variable: lfidi	lfidi
t	−0.468 ***
	(−4.33)
lcons	2.692 ***
	(17.55)
ltar	0.310 *
	(2.54)
lpop	0.746 ***
	(4.30)
lagripib	0.164 ***
	(3.63)
_ws_ltar	−0.088
	(−0.71)
_ws_lpop	−0.926 ***
	(−5.33)
_ws_lagripib	−0.185 ***
	(−4.05)
Tw_1	0.0514
	(1.68)
Tw_2	−0.000
	(−1.58)
Tw_3	0.000
	(1.35)
cons	−3.542 ***
	(−9.36)
R2	0.317
N	1716

Note: All the variables are specified in logarithms. T statistics in parentheses: * p < 0.05; ** p < 0.01; *** p < 0.001. Source: Authors, 2021.

in Appendix D. Because we suspect that endogeneity bias exists in the estimation of the treatment effects of the nominal assistance rate on food import dependency, we estimate this relationship and use the dose–response regression with instrumental variables.

We start by testing the endogeneity of the treatment variable (the nominal assistance coefficient) and the validity of our instruments. The results of the endogeneity tests are presented in

Table 3. Statistical evidence supporting treatment endogeneity and instrument assessment.

Tests	Score—Fischer	p-Value
Endogeneity test
Durbin	(Score) chi2(2) = 15.353	(p = 0.0005)
Wu–Hausman	F (2, 1707) = 7.70503	(p = 0.0005)
Test for overidentification of restrictions
Sargan	(Score) chi2(6) = 305.811	(p = 0.0000)
Basmann	(Score) chi2(6) = 369.31	(p = 0.0000)

Source: Authors’ calculations based on World Bank and FAO data.

.

The different tests presented in Table 3 show significant results at the 1% statistical level. The null hypothesis of exogeneity of the nominal assistance coefficient is rejected by the Durbin and Wu–Hausman test. While the Durbin and Wu–Hausman tests support the presence of endogeneity in the treatment variable, the Basmann [40] and Sargan [47] overidentification tests reject the null hypothesis of full instrument validity, suggesting that some instruments may not fully satisfy the exclusion restriction. Consequently, the instrumental variable estimates should be interpreted cautiously and viewed as suggestive rather than definitive causal evidence. Accordingly, all subsequent interpretations of treatment effects and policy implications should be understood as conditional on the proposed identification strategy and the approximate validity of the selected instruments. All these results seem to be consistent with the correlation table shown in

Table A2. Correlation table.

Variables	lfidi	lcons	ltar	lpop	cna2	Treatment	hc	gini	crisis08	lpoprur	democ	durable	xrcomp	xropen
lfidi	1.000
lcons	0.488 (0.000)	1.000
ltar	0.272 (0.000)	0.299 (0.000)	1.000
lpop	−0.172 (0.000)	−0.079 (0.001)	0.337 (0.000)	1.000
cna2	−0.071 (0.003)	0.022 (0.371)	−0.173 (0.000)	0.056 (0.020)	1.000
treatment	0.018 (0.454)	0.180 (0.000)	−0.006 (0.807)	0.150 (0.000)	0.792 (0.000)	1.000
hc	−0.535 (0.000)	−0.517 (0.000)	−0.613 (0.000)	−0.199 (0.000)	0.120 (0.000)	−0.068 (0.005)	1.000
gini	0.012 (0.607)	0.164 (0.000)	0.176 (0.000)	0.066 (0.006)	0.007 (0.782)	0.142 (0.000)	−0.234 (0.000)	1.000
crisis08	−0.012 (0.613)	0.002 (0.918)	−0.091 (0.000)	0.014 (0.559)	−0.061 (0.011)	−0.045 (0.062)	0.038 (0.116)	−0.008 (0.740)	1.000
lpoprur	0.462 (0.000)	0.344 (0.000)	0.397 (0.000)	0.118 (0.000)	−0.067 (0.006)	0.039 (0.104)	−0.645 (0.000)	−0.213 (0.000)	−0.019 (0.435)	1.000
democ	−0.352 (0.000)	−0.264 (0.000)	−0.524 (0.000)	−0.218 (0.000)	0.186 (0.000)	0.015 (0.542)	0.579 (0.000)	0.010 (0.683)	0.033 (0.171)	−0.535 (0.000)	1.000
durable	−0.140 (0.000)	−0.210 (0.000)	−0.170 (0.000)	−0.000 (0.990)	0.205 (0.000)	0.093 (0.000)	0.258 (0.000)	0.036 (0.140)	0.017 (0.476)	−0.260 (0.000)	0.131 (0.000)	1.000
xrcomp	−0.335 (0.000)	−0.253 (0.000)	−0.408 (0.000)	−0.055 (0.023)	0.138 (0.000)	0.021 (0.384)	0.519 (0.000)	0.046 (0.055)	0.042 (0.084)	−0.479 (0.000)	0.906 (0.000)	0.158 (0.000)	1.000
xropen	−0.235 (0.000)	−0.217 (0.000)	−0.234 (0.000)	0.082 (0.001)	0.032 (0.181)	0.014 (0.563)	0.344 (0.000)	0.026 (0.283)	0.036 (0.136)	−0.294 (0.000)	0.524 (0.000)	0.164 (0.000)	0.739 (0.000)	1.000

Source: Authors, 2021.

in Appendix E, where a relative correlation between the instruments used and the treatment variables can be observed.

In addition to the above results, which confirm the endogeneity of our treatment variable and the relevance and approximate validity of the instruments used, the Mills ratio proposed by Heckman [52] is calculated in the instrumental variable estimation. The significance of the Mills ratio was assessed using the chi-squared test to evaluate potential selection bias.

Table 4. Determinants of agricultural support participation and support intensity.

Variables	Description	T	cna
Lcons	Per capita consumption expenditure	4.506 ***	−9.555 ***
		(0.62)	(3.02)
Ltar	Average tariff rate	−0.440 ***	−0.637
		(0.13)	(0.487)
Lpop	Total population	0.424 ***	−0.166
		(0.13)	(0.32)
Lagripib	Share of agricultural value added in GDP	−0.039	0.322 ***
		(0.04)	(0.11)
Hc	Human capital index	0.261 ***	1.213 **
		(0.18)	(0.58)
Gini	Gini index	0.039 ***	−0.120 ***
		(0.01)	(0.03)
crisis08	2008 financial crisis indicator	−1.062 ***	−1.189
		(0.26)	(1.67)
Lpoprur	Rural population	0.764 ***
		(0.22)
Democ	Democracy index	0.232 ***
		(0.06)
Durable	Regime durability index	0.018 ***
		(0.00)
Xrcomp	Competitiveness indicator for executive recruitment	−0.508 **
		(0.20)
Xropen	Opening index of executive recruitment	0.024
		(0.07)
_cons	Constant	−4.851 ***	50.888 ***
		(1.39)	(3.41)
N		1716

Note: Variables are specified according to the empirical specification described in Table 1. T statistics in parentheses * p < 0.05; ** p < 0.01; *** p < 0.001. Source: Authors, 2021.

presents the results of the Heckman selection model [52] and shows that the chi² test is significant at the 1% statistical level, suggesting that the correction of the selection bias is justified. Although untreated observations represent a relatively small share of the sample, the Heckman selection framework remains informative for assessing potential selection bias associated with zero-support observations. Taken together, these results suggest that accounting for endogeneity through an instrumental variable framework may provide more informative estimates than ordinary least squares [52], although concerns regarding instrument validity remain.

Table 5. Estimated association between agricultural support and food import dependency under endogenous treatment.

Dependent Variable	lfidi
t	−2.827 ***
	(−2.09)
_ws_ltar	0.679
	(1.23)
_ws_lpop	−3.436 ***
	(−5.22)
_ws_lagripib	−1.833 ***
	(−4.90)
Tw_1	0.845
	(1.01)
Tw_2	−0.0156
	(−0.83)
Tw_3	0.000
	(0.60)
lcons	2.501 ***
	(7.48)
ltar	−0.413
	(−0.80)
lpop	3.229 ***
	(5.04)
lagripib	1.87 ***
	(4.71)
_cons	1.290
	(0.77)
N	1716

Note: All the variables are specified in logarithms. T statistics in parentheses: * p < 0.05; ** p < 0.01; *** p < 0.001. Source: Authors’ calculations based on World Bank and FAO data.

presents the results of the instrumental variable estimation.

The estimated average treatment effect remains negative (−2.827) and significant, and it is higher in absolute value than that observed in the exogenous treatment above. While the dose–response curve (Figure 2) shows a similar pattern to that obtained in the ordinary least squares estimation model, its derivative has a less convex shape (although the minimum is still between doses 60 and 70).

Table 4 shows that the entrenchment of democracy, the sustainability of regimes, and an increase in the share of the rural population among the total population each have significant effects at the 1% statistical level and positive effects on the likelihood of interventions being applied to agriculture in the developing countries studied. However, an increase in the degree of competition in the process of executive recruitment, although statistically significant at the 5% level, is associated with a lower likelihood for such interventions. While the former results are consistent with those expected, the latter result is not.

Furthermore, we can observe that, for these countries, while an increase in the annual growth rate of household consumption expenditure per capita, in the Gini index, or in the size of the population could be accompanied by a statistically significant increase at the 1% level in the probability that governments will intervene in agriculture, it would also be associated with the lowest levels of such support. While the reduction in support levels is significant in the first two cases, it is not significant in the case of population size.

In contrast, an increase in tariffs would be associated with both the lowest probability that governments would support agriculture (significant at the 1% statistical level) and lower levels of support granted (not significant). The results in Table 4 also suggest that the crisis in 2008 had a significantly negative effect on both the probability of government agricultural support and on the levels of such support, without being significant. Finally, the effects of the human capital index are positive both on the probability of government agricultural support and on the levels of support granted but are significant only in the latter case.

In the results from the instrumental variable model estimation shown in Table 5, agricultural support, the annual growth rate of household consumption expenditure, population size, and the share of agriculture in the GDP have significant effects at the 1% statistical level on food import dependencies. Specifically, if the annual growth rate of consumption expenditure, population size, and the share of agriculture in the GDP each increase by 1%, the dependence on food imports is associated with an increase of 2.06%, 3.23%, and 1.87%, respectively, all other things being equal. Tariffs, on the other hand, have a negative but insignificant effect on food import dependencies. These estimated relationships should nevertheless be interpreted as conditional empirical associations rather than definitive causal effects, given the concerns raised by the overidentification tests regarding the validity of the exclusion restrictions.

4.2. The Association Between Agricultural Support and Food Import Dependency Differs Across Regions

Table 6. Average reduction in food import dependency associated with agricultural support.

Variables	Mean	N
ATE	−2.827	1716
ATET	−2.136	1619
ATENT	−14.363	97

Source: Authors’ calculations based on World Bank and FAO data.

presents statistics on the ATE, the ATET, and the ATENT.

Thus, food import dependency is estimated to be lower by approximately 2.827% on average for all countries if they were to benefit from agricultural support measures, conditional on the covariates used in the estimation of the instrumental variables dose–response model. For beneficiary countries, the reduction averages 2.136%, while for nonbeneficiaries it averages 14.363%. Figure 3 presents the density functions for the ATE, ATET, and ATENT kernel (the kernel density estimate is a non-parametric estimate that allows the visualization of the distributions) and the dose–response function.

Beyond the average negative association between agricultural support and food import dependency, the results reveal a substantial heterogeneity across treatment intensities and regions. In particular, the estimated dose–response patterns suggest that moderate support intensities are associated with lower food import dependency, whereas very high support levels are not systematically associated with proportionally larger reductions. These findings highlight the importance of accounting for nonlinear and heterogeneous treatment intensity associations that may remain insufficiently captured by conventional linear panel data approaches.

An analysis of the distributions of the kernel density estimates in Figure 3 shows a highly similar trend in the evolution of the effects of the ATE(x) and ATET(x) curves, suggesting that the trend in food imports for countries that have agricultural support measures in place is almost the same as that of all the countries taken together. Observation of the dose–response function reveals an inverted U-shaped relationship for lower doses of approximately 70 and a U-shaped relationship for doses above 70. This implies that, for doses smaller than 70, food import dependency appears higher at lower support intensities and tends to decline as support intensity increases up to a threshold. The opposite situation seems to emerge for doses above 70.

This result suggests that the relationship between agricultural support and food import dependency is not monotonic. Moderate support intensities appear to be associated with lower food import dependency, whereas very high support intensities do not necessarily generate proportionally larger associations. These findings highlight the importance of accounting for nonlinear and heterogeneous treatment intensity effects. However, given the limitations identified regarding instrument validity, these estimated patterns should be interpreted as conditional empirical associations rather than definitive causal responses to agricultural support intensity.

While the overall pattern of treatment effects for all countries seems relatively like that of the treated countries, differences may exist between countries and even regions. Indeed, support measures differ from one country to another, as do the intensities of these support measures. It could be argued that, from one region to another, the implementation of support measures varies and, in turn, so do their effects. Based on the results of our dose–response function estimation, we compare the average treatment effects across the regions (See the study regions in Table A3 in Appendix E) under study.

Table 7. Regional heterogeneity in treatment effects.

Variable	Statistics	Africa	Asia	Europe and Oceania	Latin America
ATE(x)	Mean	−1.201	−6.200	−3.411	−1.933
	Std. Dev.	0.12	0.31	0.23	0.18
	N	693	330	429	264
ATET(x)	Mean	−0.644	−6.028	−1.724	−1.664
	Std. Dev.	0.06	0.31	0.09	0.15
	N	663	325	373	258
ATENT(x)	Mean	−13.518	−17.315	−14.645	−13.489
	Std. Dev.	0.06	0.4	0.24	0.01
	N	30	5	56	6

Source: Authors’ calculations based on World Bank and FAO data.

presents the results of the average effects per region, which statistically confirm the fact that the average treatment effects differ (the results of the t-test of the means reject at the 1% statistical threshold the equality of the observed means between the regions) from one region to another. These regional differences further suggest that agricultural support policies operate within heterogeneous structural and institutional environments. Consequently, similar levels of support intensity may be associated with different food import dependency patterns across regions. Furthermore, it can be observed that, in absolute terms, the association between support measures and food import dependency is larger in the regions of Asia, Central Europe, Latin America, and Africa. While this overall trend is similar to the observations for both treated and untreated individuals, it should be noted that, for the latter, agricultural support measures appear to be more strongly associated with lower food import dependency in Africa than in Latin America.

In view of these results, it is necessary to refer to the characteristics of each region based on the covariates used to better understand the effects obtained.

Table 8. Regional characteristics associated with treatment heterogeneity.

Variable	Statistics	Africa	Asia	Europe and Oceania	Latin America
nra	Average	−0.067	0.042	0.263	0.001
	Standard deviation	0.28	0.26	0.48	0.28
cons	Average	0.728	0.618	0.554	0.669
	Standard deviation	0.12	0.12	0.08	0.08
tar	Average	11.549	13.971	3.763	8.938
	Standard deviation	6.16	10.78	2.47	4.44
pop	Average	28,930.490	310,754.300	24,537.780	49,301.250
	Standard deviation	30,488.27	444,426.50	39,840.18	56,537.12
agri_pib	Average	0.406	0.006	0.473	0.084
	Standard deviation	3.22	0.01	1.42	0.38
hc	Average	1.653	2.151	2.946	2.362
	Standard deviation	0.40	0.44	0.45	0.37
gini	Average	42.805	37.746	34.057	52.239
	Standard deviation	7.61	6.28	7.61	5.00
poprur	Average	66.075	64.544	33.935	27.129
	Standard deviation	12.57	13.37	9.27	12.22
democ	Average	2.887	4.594	8.520	7.239
	Standard deviation	3.05	3.49	2.17	2.03
xrcomp	Average	12.266	23.882	26.065	16.375
	Standard deviation	12.90	20.94	35.03	13.48
xropen	Average	1.355	2.109	2.739	2.693
	Standard deviation	1.06	1.02	0.52	0.70
durable	Average	2.860	3.721	3.972	3.865
	Standard deviation	1.75	1.02	0.33	0.73
N		693	330	429	264

Source: Authors’ calculations based on World Bank and FAO data.

provides descriptive statistics by region.

Table 8 shows that the regions with the highest levels of support for agriculture (Asia, Europe, and Oceania) are the ones that tend to exhibit relatively lower levels of food import dependency. Compared to the rest of the regions under study, these regions are also characterized by relatively more advanced democratic processes, relatively lower levels of income inequality, relatively higher levels of human capital index, and relatively high average population levels (overall or rural). In contrast, the regions with lower levels of support for agriculture (Latin America and Africa) are associated with comparatively higher levels of food import dependency.

5. Discussion and Policy Implications

5.1. Determinants of the Decision for and/or Intensity of Agricultural Support

Our results suggest that stronger democratic institutions are associated with both the likelihood and intensity of agricultural support. This is consistent with the predictions of many authors [26,27,59,60], who start with the observation that, in these countries, the agricultural sector is more important than the manufacturing sector [59,60] and accounts for most of the economically vulnerable population [26]. Thus, as the democratic process strengthens, it can be expected that this sizeable segment of the global population will also increase pressure on government elites to act in favor of agriculture [26,27].

In addition, the results of this work support the thesis that agricultural support measures can be used to help ensure the sustainability of political systems in developing countries. Agriculture in many developing countries remains characterized by relatively low levels of technological adoption [61], and under these conditions, where countries require a modernization of agricultural means, the survival of policy regimes—far from depending on the detailed management of society or tight governmental control over social processes—instead depends on the outcomes of policy responses sensitive to the forces of change, flexible adjustments of system structures to meet the needs of innovation, and open policy processes that allow for progressive and orderly agricultural development [62]. Much of the Western democratic world made peaceful progress in this way, despite new political philosophies (see “Redistricting: The Key to Politics in the 1980’s” of Heslop [11].

Our results also suggest that strengthening the democratic process or improving the sustainability of regimes in developing countries would play an important role in the decision to support and/or intensity of support for agriculture only if attention is paid to both income inequality and human capital development. Indeed, human capital development would require a well-trained and well-educated population [26,42] who would understand the costs and benefits of agricultural policy [26]. Therefore, policymakers should be sensitive to strengthening human capital, as well as to the impacts of the pressures their electorate may exert in terms of demanding improved agricultural support. The fact that an increase in income inequality is associated with an increase in agricultural support reinforces the work showing that this support can be used as a means of redistributing agricultural income [63]. However, these authors have pointed out that the generic nature of many measures implies that the bulk of agricultural support would go to farm households that do not truly need it [63].

5.2. Agricultural Support and Dependence on Food Imports

Our analysis shows that, in developing countries, agricultural support measures are associated with food import dependency patterns, which may interact with food security outcomes depending on country-specific contexts. Moreover, the magnitude of the observed associations appears to vary according to the intensity of support. These results are consistent with the findings of Anderson et al. [1,2,15,19,20,64,65] and d’Hammoudi et al. [66], who show that agricultural support measures can be associated with lower food import dependency while ensuring the welfare of both producers and consumers. One possible interpretation of these results is that such measures may create or be accompanied by conditions favorable for the promotion of domestic food crops [66].

Indeed, the governments of these countries could contribute by ensuring that the implementation of measures to support export crops does not come at the expense of domestic food production. In doing so, food producers (who typically comprise the poorest portion of the rural population in these countries) could participate in and benefit from domestic production activities as export crop producers do, which, in addition to ensuring their income, could contribute to enhancing food stocks [66]. However, it should be noted that, in the absence of a food stock management mechanism, food producers could suffer losses that would undermine efforts to build up the national food stock [67], as is the case with export taxes or import subsidies.

Moreover, policies aimed solely at reducing food import dependency may entail significant environmental and ecosystem costs if they encourage production in agro-ecologically unsuitable areas, highlighting the need to balance trade, sustainability, and food security objectives.

6. Conclusions

6.1. Research Objective

This study examined how the intensity of agricultural support is associated with food import dependency in developing countries. More specifically, it investigated whether different levels of agricultural support are associated with heterogeneous patterns of food import dependency when treatment heterogeneity and potential endogeneity are explicitly considered within a continuous treatment framework. To address this research question, we applied the endogenous continuous treatment framework proposed by Cerulli [68,69,70] and Baum and Cerulli [29] to a panel of 52 developing countries observed between 1985 and 2017.

6.2. Main Empirical Findings

As in the work of Magrini et al. [12], our results suggest heterogeneous and nonlinear associations between agricultural support intensity and food import dependency. Moderate levels of agricultural support appear to be associated with lower food import dependency, although the magnitude and direction of these associations vary across treatment intensities and regions. Our findings also reveal substantial regional heterogeneity, with the estimated associations varying considerably across Asia, Europe–Oceania, Africa, and Latin America. Rather than establishing the existence of a previously unknown relationship between agricultural support and food import dependency, our study contributes primarily by refining the empirical analysis through the explicit modeling of heterogeneous treatment intensity effects and potential endogeneity within a continuous treatment framework. These findings should be interpreted in light of the identification assumptions underlying the instrumental-variable framework and the limitations highlighted by the overidentification tests.

Importantly, food import dependency should not be interpreted as synonymous with food insecurity. Food imports may contribute to food availability and consumption smoothing, particularly in countries facing climatic, geographic, or structural production constraints, consistent with Burgess and Donaldson [3]. At the same time, excessive dependence on external markets may increase vulnerability to international price volatility, foreign exchange constraints, and trade disruptions. Our results should therefore be interpreted as reflecting patterns of external dependence rather than direct measures of food insecurity.

6.3. Limitations

Several limitations of our study should be acknowledged. First, although the selected instruments are grounded in the political economy literature, the overidentification tests suggest that the exclusion restriction may not be perfectly satisfied, implying that the instrumental-variable estimates should be interpreted cautiously. Consequently, the estimated relationships should be viewed as conditional empirical associations that are consistent with the proposed identification strategy rather than as definitive causal estimates. Second, the use of aggregate agricultural support indicators may mask heterogeneous effects associated with specific policy instruments such as tariffs, subsidies, or export support measures. In addition, some structural geographic factors—such as exchange rates, terms of trade, landlocked status, climatic conditions, or agricultural land endowments—were discussed conceptually but could not be fully incorporated into the empirical specification because of data comparability limitations. Third, the analysis was constrained by the availability of harmonized international data, which limited the study period to 2017. Finally, the use of trade-weighted tariff measures may understate effective protection levels in the presence of prohibitive tariffs. Furthermore, additional robustness analyses based on alternative instrument sets, alternative support indicators, and alternative model specifications were not feasible given current data constraints and therefore remain important avenues for future research.

More generally, this study relies on observational cross-country data rather than an experimental design. While the empirical framework attempts to mitigate endogeneity concerns through instrumental variables and continuous treatment estimation, caution remains warranted when interpreting the estimated relationships. Furthermore, major global shocks occurring after 2017—including the COVID-19 pandemic, supply-chain disruptions, and recent food price crises—may have modified the structural relationships examined in this analysis. Consequently, the estimated relationships should be viewed as conditional empirical associations that are consistent with the proposed identification strategy rather than as definitive causal estimates.

6.4. Policy Implications

Given the observational nature of the data and the limitations associated with instrument validity, the policy implications discussed below should be interpreted with appropriate caution. From a policy perspective, the results suggest that agricultural support policies should be designed in a differentiated manner reflecting countries’ structural conditions, institutional capacities, and regional specificities. Moderate and well-targeted support measures appear to be associated with lower food import dependency in certain contexts, but excessively high or poorly coordinated interventions may be associated with diminishing effectiveness or unintended distortions. The findings also highlight that structural constraints—such as geography, factor endowments, market size, and institutional quality—condition the effectiveness of agricultural support policies. Consequently, uniform policy prescriptions are unlikely to be appropriate across developing countries. The findings therefore do not support a single optimal model of agricultural support applicable to all developing countries, nor should lower food import dependency be interpreted as a direct proxy for improved food security outcomes.

6.5. Future Research

Future research could extend the analysis using more recent post-2017 data, alternative measures of agricultural protection, and disaggregated policy instruments to better distinguish the heterogeneous associations of specific support mechanisms with food import dependency. Further work could also explore dynamic adjustments, country-specific institutional contexts, and alternative identification strategies that could provide more robust empirical evidence. Additional robustness analyses based on alternative instrument sets, alternative support indicators, and sensitivity analyses across model specifications would also help assess the stability of the estimated relationships.

Overall, this study contributes primarily by combining a continuous treatment effect framework with instrumental variables to examine heterogeneous associations between agricultural support intensity and food import dependency in developing countries. Our analysis highlights substantial regional heterogeneity and nonlinear dose–response patterns that may remain insufficiently captured by conventional linear approaches. Future research should prioritize robustness analyses using alternative instrument sets, alternative measures of agricultural support, and complementary identification strategies.