Causal Effects of Petrocaribe on Sustainable Development: A Synthetic Control Analysis

We examine the causal effects of the energy subsidy programme PetroCaribe in the three dimensions of sustainable development: economic, social and environmental. We use the synthetic control method to construct a counterfactual and compare it to the outcomes of the beneficiary countries and thus estimate the magnitude and direction of the PetroCaribe effect. PetroCaribe had a positive effect on economic growth in most of the beneficiary countries; however, this economic boost was not followed by an improvement in social development. Environmentally, PetroCaribe did not negatively or positively impact the environmental quality of the member countries, in the sense that we do not find a significant effect on the trend of emissions per capita.


IntroductIon
The PetroCaribe programme, initiated by the late President Chavez of Venezuela, sold oil below market price to political allies. This paper is the first to study the implications for the beneficiaries on their economic growth, energy use, and societies.
Oil prices exhibited unprecedented volatility at the beginning of the 2000s, with an upward trend during [2003][2004][2005][2006][2007][2008]. Prices rose from US$30 in 2003 to a historic high of US$147 in 2008. Overall, the extent of the adverse effects of high and volatile oil prices depend on whether a country is an oil exporter or importer, its level of development and on the governmental capability to face oil shocks (Monaldi, 2015;Yépez-García and Dana, 2012). In particular, in oil importing countries that are highly reliant on oil for has been argued that PetroCaribe has had a significant impact on helping member countries deal with the rise in crude oil and food prices. Without the subsidy, the rising costs would have meant a catastrophe for many countries, especially in those with high poverty rates and energy deficit (Benzi and Zapata, 2012;Trinkunas, 2014). Sardinas et al. (2009) highlight the positive impact on the urban development of the Cuban city of Cienfuegos as a result of the improved performance of the Camilo Cienfuegos refinery, a PetroCaribe project focused on infrastructure investment. However, the energy agreement is subject to the same criticisms, in general, of energy subsidies. For some authors, PetroCaribe represents an uneconomical energy practice with limited social and economic benefits and rather an increased debt for the participating countries. The high dependence on a single source of subsidized oil, has sustained the dependence on fuel for power generation, discouraging the transition to alternative, more efficient and less expensive feedstock for electricity. Environmentally, the discouragement in the investment of renewable sources exacerbates the use of fossil fuels, jeopardizing the regional efforts to reduce carbon dioxide emissions (Di Bella et al., 2015;Goldwyn and Gill, 2014;Johnston, 2014;Lacayo, 2013). Source: (PDVSA, 2014(PDVSA, , 2015.

methodology
As mentioned in Section 1, we evaluate the impact of PetroCaribe. Impact evaluation techniques compare outcomes for treated unit with counterfactual baselines to estimate what would have happened without an intervention. The counterfactual is never observed but is estimated using outcomes in similar units, with similar characteristics. A common strategy to estimate such interventions is the difference-in-differences model (DiD). The DiD compares the difference before and after the intervention in the outcome of a treated unit and the control group to determine the net impact. However, the main drawback of the DiD is its key assumption of parallel trends, i.e. it is assumed that in the absence of the intervention, the treatment and control group would have had the same trend across time. To overcome the aforementioned issue, we use the synthetic control method (SCM) developed by Abadie and Gardeazabal (2003) and further developed by Abadie et al. (2010Abadie et al. ( , 2015. The SCM relaxes the parallel trend assumption and constructs a synthetic control match for the treated unit by using untreated units in the control group in such a way that the synthetic counterfactual has a similar behavior to the actual treated unit before the intervention. Since the pioneering work of Abadie and Gardeazabal (2003) who use this approach to analyze the economic effect of terrorism in the Basque country, the SCM has been used in economics and other social science to analyze a wide range of interventions. For example, Abadie et al. (2010), study the effects of a tobacco prevention legislation in California in 1988 on tobacco consumption. Hope (2016) investigates the effect of the Economic Monetary Union on the account balance. Billmeier and Nannicini (2013) estimate the effect of trade liberalization on economic growth. Sills et al. (2015) employ this method in investigating the impact of a local policy initiative to limit deforestation in the Brazilian Amazon. Grier and Maynard (2016) evaluate the impact of the president Hugo Chavez on the Venezuelan economy.
Following Abadie and Gardeazabal (2003) and Abadie et al. (2010Abadie et al. ( , 2015, let us assume that we observe countries i = 1, …, N + J. Countries 1 to N are exposed to the intervention (here, are signatories of the PetroCaribe programme) at time T 0 + 1 and the remaining J countries form the donor pool from which the synthetic control countries are created. Let Y PC it be the outcome variable observed for country i, member of PetroCaribe (PC) at time t. Similarly, let Y NP it be the outcome variable observed for country i, not member of PetroCaribe (NP), at time t.
The outcome variable for any country i at time t can be written as: (1) where it is the effect of the intervention for country i at time t, and S it is a binary indicator variable that takes the value one if the intervention has taken place and value zero otherwise. Assuming a single signatory (i.e. N = 1), the effect of PetroCaribe for country 1 (i.e. i = 1 and t ≥ T 0 ) in equation (1) can be defined as: In equation (2) the only observed variable is Y PC 1t , hence the counterfactual Y NP 1t can be estimated as follows: where t is a vector of common time-specific factors constant across countries; t is a vector of unknown parameters; Z i is a (r × 1) vector of observed covariates not affected by the intervention, which can be either time-invariant or time-varying; t is a (1 × F) vector of unobserved common factors, i is a (F × 1) vector of unknown unit specific factors, and it are idiosyncratic error terms with zero mean. Let us define a synthetic control unit as a weighted average of countries in the donor pool. That is, it can be represented by a (J × 1) vector of weights W = w 2 , … , w J + 1 � such that w j ≥ 0 for j = 2, ⋯ ,J + 1 and ∑ J + 1 j = 2 w j = 1. Each value of the vector W represents a potential synthetic control for a PetroCaribe country, for which its outcome variable is defined by: Suppose there is a vector of weights w * 2 , … , w * J + 1 � such that: i.e. the weighted average of the pre-treatment outcomes of the control perfectly matches the pre-treatment outcomes of the treated country and the weighted average of the covariates of the control perfectly replicates the covariates of the treated country. Then, the estimated treatment effect for the treated country can be estimated as: Conditions in equation (5) hold exactly only if Y 1t , Z 1 belongs to the convex hull of Y 21 , … , Y 2T 0 , Z � 2 , … , Y J + 11 , … , Y J + 1T 0 , Z � J + 1 , i.e.
(2) 1t = Y PC 1t − Y NP 1t . (3) there should exist some combination of untreated units that exactly match the treated country before the treatment. Usually, is not possible to estimate a perfect synthetic control because there are no weights w * j for condition (5) to hold exactly. Thus, in practice, W * is estimated in a non-parametric fashion and is selected such that (5) holds approximately. W * is selected by minimizing the distance between the vector of characteristics (covariates and pre-treatment outcomes) of the signatory countries (X 1 ) and the weighted matrix that contains the same characteristics of each potential donor pool (X 0 W ) in the pre-treatment period.
Formally, let the vector K = k 1 , … , k T 0 � define a linear combination Let us consider N of such linear combinations be define by the vectors K 1 , … , K N . X 1 is a (k × 1) vector defined as: containing k covariates and pre-treatment outcomes of the signatory country. Similarly, X 0 is a (k × J) matrix that contains the same variables for each country in the donor pool. The differences between the pre-treatment characteristic of the PetroCaribe countries and a synthetic control is given by the vector ‖X 1 − X 0 W ‖. The vector W * is chosen so that it minimizes: where W is a weighting vector that measures the relative importance of each control country in the construction of the synthetic control, and V is a (k × k) symmetric and positive definite diagonal matrix that reflects the relative importance of each covariate and pre-treatment outcome. The choice of V influences the root mean square error of the estimator (RMSPE). Abadie and Gardeazabal (2003) suggest to choose a V such that the RMSPE of the outcome variable is minimized for the pre-intervention period: While the choice of the covariates Z i can be justified by selecting those variables that better explain the outcome variable, there is no consensus on the optimal set of pre-treatment outcomes (Y K j ) that need to be included as predictors. Abadie et al. (2010) suggest as an obvious solution to use the values of the outcome variable for all the pre-treatment years, Bohn et al. (2014), Gobillon and Magnac (2016) use this approach. However, Kaul et al. (2017) show that including all pre-treatment outcomes as predictor leads to all other predictors receiving zero weights. Another very common specification is to use the average of the pre-treatment outcome, Abadie and Gardeazabal (2003), Abadie et al. (2015), Kleven et al. (2013) among others, Electronic copy available at: https://ssrn.com/abstract=3613934 perform their analysis with this linear combination. Bove et al. (2014) select four out of ten pre-treatment period to analyze the impact of civil war on GDP. Montalvo (2011) uses only the last two pre-treatment values. Based on the previous discussion, we test five different placebo specifications that differ only in the linear combination of lagged outcome variable used as predictor. For each specification, we compute the SCM on each country j in the donor pool as treated 4 : 1. The average of all pre-treatment outcomes: X j = 3. The first and last period of the pre-treatment: 4. The first, half and last period of the pre-treatment: 5. The first, two half and last two periods: We calculate the post-treatment RMSPE of each specification. As suggested by Ferman et al. (2016), since the control countries did not experience the intervention, we ideal specification will be the one with the lowest RMSPE in the post-treatment period: Once the proper specification is selected, it is vital that the weighted synthetic outcomes match the outcomes for the treated country in the pre-treatment period. To assess whether the synthetic country is a good counterfactual, we estimate the R 2 statistic, the coefficient of determinatin or the fraction of variance explained. This is essentially one minus the pre-treatment MSE normalized by the variance of the treated country: R 2 can range from minus infinite to 1. An R 2 of 1 indicates a perfect match. If R 2 = 0 then the estimated synthetic is no more accurate than the average of the observed data, and a negative R 2 occurs when the mean of the observed data is a better counterfactual than the estimated synthetic control. Best fit is a matter of judgment (Sills et al., 2015) that in this case hinges on the outcome of interest. 4 As explained is Section 3, we construct country-specific donor pool for each PetroCaribe country and its outcome of interest.Thus, specifications (iv) and (v) vary from country to country depending on the outcome and donor pool.
To assess statistical significance, we conduct a series of placebo tests closely following Abadie et al. (2015). The first placebo test, known as inspace placebo, consists in iteratively applying the SCM on each country of the donor pool as if it was treated. Since the control country did not receive any intervention, we should not expect a treatment effect. If the placebo studies exhibits a treatment effect of similar magnitude to the one estimated for the actual treated country, we conclude that this treatment effect is driven entirely by chance and that the analysis does not provide a convincing evidence of a treatment effect.
However, we take into consideration that some control countries in the placebo experiments can have a bad pre-treatment fit with the consequent large RMSPE, casting doubt on their reliability. In order to avoid misleading conclusions, we drop placebo runs with a pre-treatment RMSPEs that are at last 1.5 times higher than that of the PetroCaribe country.
Since this visual analysis involves some amount of subjectivity, we additionally estimate the post-treatment RMSPE to the pre-treatment RMSPE ratios: This scale-free measure allows to estimate the extremity of the impact of the placebo experiments. The empirical distribution of the ratios allows to compute pseudo p-values as follows: The pseudo p-values constructed in this context imply that if the treatment were to be assigned at random, then the probability of getting a ratio at least as large as the one estimated for the PetroCaribe country is 1/J + 1 (Abadie et al., 2010). Note that the pseudo p-values necessarily depend on the number of control countries, meaning that some values cannot be significant at conventional levels (one-tailed test), which does not imply the absence of an effect.
As a second validation check, we test the sensitivity of the baseline model to the countries in the control pool. The so-called leave-one-out test consists in iteratively apply the baseline SCM omitting in each iteration one of the countries that received a positive weight in the baseline specification. 5 (11) This allows assessing whether one of the control countries is driving the results. If the synthetic control follows a similar trajectory, then it is less likely that the results are biased to the inclusion of any single control country.

data and SPecIfIcatIon
The analysis focuses on the effect of PetroCaribe on four outcomes of interest: (i) economic development, represented by GDP per capita; (ii) social development, represented by the Human Development Index; (iii) CO 2 per capita emissions; and iv) electricity use per capita. For covariates, we include the lags of the outcome variable and a set of standard predictors with a stable relationship with the outcome variable. These variables, displayed in Table 2, were selected by their predictive power based on empirical literature (Abadie and Gardeazabal, 2003;Almer and Winkler, 2012;Eren et al., 2014;Larivière and Lafrance, 1999) and availability. The period under consideration for economic growth and social development is 1990 to 2014, while for per capita CO 2 emissions and electricity use is 1980 to 2013. Descriptive statistics are shown in Appendix C. Sources and definitions are provided in Appendix G.

Treated Countries And Donor Pool
Although PetroCaribe officially has 18 members, three countries, the Bahamas, Guatemala and St. Lucia never entered into a bilateral agreement and are thus omitted from the analysis. For the remaining countries, we impose the following conditions: (i) the treatment needs to be sustained through a significant period, otherwise, if the post-treatment period is short, the SCM cannot estimate any real treatment effect. Four countries do not meet this condition. Belize and Honduras had interruptions in their supply. 6 Suriname and El Salvador joined at a later date, 2012 and 2014 respectively. (ii) The treated country cannot be an outlier in the dataset. Recalling that countries with extreme values of observed characteristics are unlikely to satisfy condition (5) in Section 2, in such case, the SCM cannot give a correct prediction. In this regard, Haiti was excluded. Being the poorest country in Latin America and the Caribbean, and one of the poorest in the world, US$1,737 in 2014, its outcomes of interest lie in the extremes, which make it difficult to build a donor pool with countries of similar characteristics. (iii) Countries do not have to be exposed to other significant shocks during the treatment period. Two countries do not satisfy this condition. Haiti suffered losses equivalent to 113 percent of GDP as a result of an earthquake that struck the country in 2010 (ECLAC, 2014), three years after joining PetroCaribe. In Jamaica, high fluctuation in its GDP, CO 2 emissions and energy consumption are a major results of the closure of three of four bauxite and alumina plants in 2008; bauxite industry is the largest contributor to its GDP.
Regarding the treatment date, for some beneficiaries the delivery of oil was not made immediately after the signing of the agreement, but it was delayed a few years. Therefore, the treatment date is established as the year in which the countries received the first cargo of oil. Table 3 shows the treatment date considered in the analysis.
Taking into consideration the heterogeneous characteristics of the PetroCaribe beneficiaries, we build a country-specific donor pool for each outcome of interest. The potential donor pool is restricted to the following conditions: (i) the countries need to remain unexposed to the intervention through the period under study; (ii) to avoid interpolation bias, which might occur by interpolating across countries with different observed characteristics, the donor pool is limited to countries that closely resemble the PetroCaribe members. As a first filter, we selected countries that belong to the same income level as the PetroCaribe countries, according to the World Bank classification (World Bank, 2017). Later, we choose as donor pool those countries which values of the outcome variable lie within the range of the 50 percent of the value of the outcome of interest of the PetroCaribe country. This is a crucial step in the construction of the synthetic country, since if the control countries are not sufficiently similar, any difference in the outcome of the two sets may simply reflect disparities in their characteristics (Abadie et al., 2015). The donor pool as well as the descriptive statistics are shown in Appendix C.

reSultS
As mentioned in Section 2, the first step in the analysis involves the choice of the appropriate specification, i.e. the one that minimizes the RMSPE for each country and outcome of interest. For the sake of brevity, the results of each specification are shown in Appendix A. Control country and covariate weights are displayed in Appendix B. Robustness is discussed in the context of the main findings. The results of the placebo test are displayed in Appendix D, E and F.

Economic Growth
As mentioned in Section 4, Haiti and Jamaica are removed from main analysis because they did not satisfy the conditions to carry out an adequate analysis. On one hand, the extreme low values of Haiti compared to the donor pool, and the exogenous significant shock in Jamaica. For illustrative purposes, both circumstances are reflected in the low values of the pretreatment fit shown in Table 4. For Antigua and Barbuda, the pre-treatment fit is weak, with a low value of 0.186. Moreover, the results are not robust to any falsification test performed (see appendixes D, E and F), thus, Antigua and Barbuda is also dropped from the main analysis. Fig. 1 illustrates the synthetic control estimates for the PetroCaribe countries with a good pre-treatment fit, as mentioned in Table 4.
In four countries, PetroCaribe significantly boosted economic development. The largest effect can be seen in Cuba, with an average gain of 26.29 percent in GDP per capita and a gain of 41.55 percent in 2014. The results are highly robust to the placebo test, with a pseudo p-value of 0.066. In Dominica, the pre-treatment fit of 0.938 is nearly perfect. The average gain in its per capita GDP due to PetroCaribe is 11.19 percent and in 2014 per capita GDP is 14.19 percent higher than it would have been without the agreement. Guyana experienced an average gain in the post-treatment period of 11.19 percent, while in the Dominican Republic, the gain was 7.39 percent. All the results are robust to the placebo test, which is reflected in the pseudo p-value, and to the leave-one-out test, i.e. the positive effect of PetroCaribe in is not driven by any control country in the donor pool.
In contrast, Grenada, Nicaragua, St. Vincent and the Grenadines and St. Kitts and Nevis, did not experience a higher per capita GDP than they would have had without PetroCaribe. Grenada, which received the first Vincent and the Grenadines synthetic Vincent and the Grenadines shipment of oil in 2007, experienced during the post-treatment period a per capita GDP that was 5.43 percent below its synthetic counterfactual. As can be seen in Fig. 1, Grenada experienced a decrease in its per capita GDP in 2008, while its synthetic counterfactual continued with the growing trend. The gap narrows towards the end of the post-treatment period, with a gap of −5.3 percent in 2014. In Nicaragua, the SCM estimated an average decrease in the post-treatment period of −9.3 percent in per capita GDP. As in Grenada, Nicaragua experienced a decrease in its per capita GDP in 2008 while the trend in its synthetic continued upwards. In both countries, results are robust to the placebo test.
St. Vincent and the Grenadines and St. Kitts and Nevis also show an average decrease in their per capita GDP, −1.87 percent and −4.03 percent, respectively. However, the results are not robust. The pseudo p-value in both countries indicate that the probability to obtain a placebo country with an effect higher or equal to that experienced in the treated country is fifty percent, concluding that the effect of PetroCaribe in both countries is not statistically significant.

Social Development
The impact of PetroCaribe on social development is estimated only in six countries due to data availability. As can be seen in Table 5, the SCM achieved a good pre-treatment fit in all countries. Fig. 2 illustrates the effect of PetroCaribe on the Human Development Index.
Only Cuba experienced a positive effect. Twelve years after the Agreement, the HDI is 0.16 percent points higher than its synthetic counterpart. In Guyana and Haiti, PetroCaribe is not reflected in an increase in their HDI. The results are fairly robust to the placebo tests. Jamaica did not experience a higher HDI in comparison with its counterfactual, however, the results are not statistically significant, therefore, we can not drive conclusions about the real effects of PetroCaribe in the country. Finally, in Nicaragua and the Dominican Republic, PetroCaribe had no discernible effect, the divergence from their respective counterfactuals is small.The results in both countries are statistically significant.

Carbon Dioxide Emissions Per Capita
The synthetic control method could estimate a good match in the pretreatment period for all the countries except for Cuba. As can be seen in Table 6, the pre-treatment fit in Cuba is −0.379, thus, we exclude this country from the main analysis. Fig. 3 plots the trajectories of the PetroCaribe countries and their estimated synthetic counterfactual. First, let us focus on the case of Antigua and Barbuda and Guyana, the only two countries that experienced higher  year Dominican Republic synthetic Dominican Republic treatment period. The pseudo p-value of 0.083 gives us confidence in our results, as does the robustness seen in the leave-one-out test. Guyana has an average increase of 7.88 percent and at the end of the treatment period, the CO 2 per capita emissions are 14.3 percent higher than that of its synthetic counterfactual. However, the placebo test shows that 4 of the 10 control countries have a higher pre/post-RSME than that of Guyana. As such, we cannot conclude that PetroCaribe increased emissions in this country. In Grenada, the path of the treated is slightly higher than the synthetic counterfactual, 5.03 percent on average. With a pseudo p-value of 0.461 and a highly robust leave-one-out test, the results for Grenada are statistically significant. We next move to the countries where PetroCaribe had a negative or close to zero effect in their CO 2 per capita emissions, i.e. have a lower level of emissions compared to their counterfactual. Dominica has, on average, 13.06 percent less emissions than what would have had without PetroCaribe. At the end of the post-treatment period, the emissions are 17.72 percent lower than those of its synthetic. In Nicaragua, the estimated effect of the agreement at the end of the treatment is a difference of −11.32 percent in comparison with its counterfactual. St. Vincent and the Grenadines has an estimated effect of 6.62 percent fewer emissions than its counterfactual. The effects are statistically significant in all these countries.  Finally, the Dominican Republic experienced an average decline of 5.79 during the treatment period, while St. Kitts and Nevis has a small difference of −4.17 percent compared to its synthetic counterfactual. The effects, however, are not statistically robust to the placebo test, nor to the leave-one-out test. Therefore, we cannot be confident about the true effect.
Summing up, there is little evidence that PetroCaribe led to an increase in per capita CO 2 emissions in the member countries.

Electricity Use Per Capita
In the study of the effect of PetroCaribe in per capita electricity, the SCM was not able to estimate a good pre-treatment match for Antigua and Barbuda, Guyana and the Dominican Republic. These three countries and Jamaica are removed from the main analysis. As can be seen in Table 7 and Fig. 4, in the rest of the countries, the pre-treatment fit is fairly good.
Although all countries show an upward trend in their electricity consumption, only two countries, Nicaragua and St. Kitts and Nevis, increased their electricity consumption after joining PetroCaribe. In Nicaragua, the increase was on average 4.85 percent during the treatment period and in 2014, the last treatment year, the electricity use was 7.11 percent higher than that of its synthetic estimate. The effects, however, are not statistically significant. In St. Kitts and Nevis, the electricity consumption was, on average, 17.12 percent higher than that its counterfactual. The pseudo p-value is 0.083, highly statistically significant. In contrast, electricity use in Cuba and Grenada is less than the electricity use in heir synthetic estimates. For Cuba,   1980 1985 1990 1995 2000 2005 2010 2015 year Vincent and the Grenadines synthetic Vincent and the Grenadines to their synthetic control. Although all show an overall increase in their HDI during the period under analysis, PetroCaribe did not have the positive impact that was expected by its sponsors. These results contradict the conclusion of SELA 2015, p. 20) that PetroCaribe "has made a bigger contribution" on the HDI in the beneficiary countries. Rather, they confirm one criticism of energy subsidies, that they do not always improve the social development of individuals. The Human Development Index is dominated by education and health, which are stock variables that change only slowly over time. PetroCaribe does not have an impact in the short run analyzed here, but it may have in the long run.
PetroCaribe had no effect on per capita CO 2 emissions. Emissions neither increased-as may have been expected from a programme that subsidizes oil-nor fell-the stated intention of the recipient countries. Although some countries show an increase in emissions, the difference with their counterfactual is minimal. We can conclude that PetroCaribe did not result in a worsening of CO 2 per capita emissions. As economic growth accelerated, this implies that PetroCaribe must have reduced the carbon intensity of the recipient economies.
Regarding electricity use, some results are positive and others negative, but only two are statistically significant and economically meaningful. Jamaica saw a large drop, and St. Kitts and Nevis a large increase. However, these outcomes are not strongly supported by the leave-one-out test. Nicaragua, St. Kitts and Nevis and St. Vincent and the Grenadines were beneficiaries of a series of power generation projects, supported by PetroCaribe. Access to cheaper oil for power generation and accelerated growth appears to have been offset by greater efficiency, perhaps in terms of reduced transmission and distribution losses, which are around 20 percent, one of the highest in the world. We cannot draw firm conclusions about the effects of PetroCaribe on electricity use in its member countries.
The policy implication is that an energy subsidy like PetroCaribe can promote economic development in the beneficiary countries without a significant worsening of per capita CO 2 emissions. PetroCaribe can provide the insights and evidence that oil subsidies of this type, in which the savings derived from the oil bill are destined for a series of energy infrastructure along with social development programs are not incongruent with the discourse of sustainable development.
Further research should investigate how a subsidy of this kind impacts the development of renewable energy sources, and whether it acts as a disincentive to the transition towards alternative sources of energy. This is of particular interest for countries that heavily rely on imported fossil fuel for power generation. A deeper analysis into the impact of PetroCaribe on the structure of economic activity and public expenditure would be useful. The analysis here should be repeated when the data allow for an analysis of the impact in the long run. The limitations of the methodology used did not allow us to estimate the impact of PetroCaribe in Haiti, the poorest country among the beneficiaries and the most dependent on Venezuelan oil. We need better counterfactuals for this country. Lack of fit for some countries, should not be interpreted as a lack of effect. Another caveat is that the synthetic control method does not explicitly consider the interactions and spillovers between the treated countries, whose economies are integrated. Furthermore, PetroCaribe is a composite treatment, not just the programme itself but also through its geopolitical realignment. All this is deferred to future research.