A bilateral sugar trade agreement was established between Mexico and the USA in 2008 under the North American Free Trade Agreement (NAFTA). This agreement removed restrictions on the flow of sugar between these countries^{(Reference Zahniser and Link1)}. By 2013, Mexico became responsible for 66 % (equivalent to two million short tonnes in raw value) (see Fig. 1) of the total sugar imports to the USA, which led to a sharp decline in domestic sugar prices and an increase in the average annual consumer surplus to nearly $1·67 billion^{(Reference Sinclair and Countryman2)}. Diabetes prevalence in the USA went up instantaneously after 2008^{(Reference Uppal, Chehal and Fernandes3)}. Between 2008 and 2012, crude diabetes prevalence increased from 7·3 to 10·1 % (see the Health Care Cost Institute, 2013). In this paper, we study the impact of the NAFTA sugar provision on diabetes prevalence and evaluate which states in the USA were most affected by the policy.

Fig. 1 Trend of US sugar import from Mexico

Governments around the world have had increased interest in global public health in recent years. Many health experts and economists have turned their attention to the link between sugar consumption and diseases like obesity and type 2 diabetes. For instance, type 2 diabetes, dental caries and obesity are all associated with the intake of sugar-sweetened beverages (SSB)^{(Reference Hu and Malik4)}. Furthermore, the increase in the consumption of foods with high sugar content has been shown to increase age-standardised BMI by 0·033 kg/m^{2(Reference De Vogli, Kouvonen and Gimeno5)}. As a result, the WHO⁽⁶⁾ put out an initiative to reduce added-sugar intake with the goal of improving global public health. Governments throughout the world have responded by establishing programmes like sugar taxes to lower the amount of sugar in food and drinks^{(Reference Stanner and Spiro7)}. Policymakers in the UK, for example, introduced a two-tier soft drink industry levy to reduce sugar intake, successfully reducing childhood obesity^{(8,Reference Rogers, Cummins and Forde9)} . Studies have found SSB taxes to reduce purchases of taxed sugar beverages in Mexico^{(Reference Colchero, Popkin and Rivera10)}, Berkeley, California^{(Reference Silver, Ng and Ryan-Ibarra11)} and the Netherlands^{(Reference Kyriakos, Driezen and Fong12)}, a finding that has been supported by several meta-analyses^{(Reference Andreyeva, Marple and Marinello13)}. Sugar taxes have also been shown to reduce diseases associated with SSB intake, such as obesity and type 2 diabetes^{(Reference Miljkovic, Nganje and de Chastenet14–Reference Fernandez and Raine16)}.

Despite the widely recognised advantages of free trade agreements, which include economic development, lower government spending, technology transfer, enforcing competition among firms, enhancing productivity and reducing prices and markups^{(Reference Chen, Imbs and Scott17,Reference Gonzalez-Garcia and Yang18)} , their possible unintended effects on public health are often ignored. The removal or reduction of sugar trade barriers between partner countries increases imports while decreasing commodity prices, essentially having the opposite effect of a sugar tax. Four pathways were identified by^{(Reference McNamara19)} through which trade openness affects health: higher imports and immigration, agricultural commodities trade, labour markets and structural adjustment measures. Prior work has shown that the unhindered trade in commodities like cigarettes, alcohol and ultra-processed foods poses risks to public health^{(Reference An, Guan and Liu20–Reference Adu, Li and Sawadgo22)}. For example, a 10 % rise in trade openness generates a 0·8 % increase in obesity rates^{(Reference An, Guan and Liu20)}. Since sugar trade liberalisation behaves as a subsidy, making more sugar available for consumption in the recipient country, this poses a risk to public health^{(Reference Cernat, Gerard, Guinea and Isella21,Reference Adu, Li and Sawadgo22)} . As a result, considerable evaluation and steps are required to offset any adverse health effects from trade agreements.

Our primary objective in this study is to analyse the potential causal effects of NAFTA’s unrestricted sugar trade agreement on diabetes prevalence across the fifty US states. While free trade agreements provide enormous benefits, it is critical to recognise their unintended implications on public health. To our knowledge, only two studies^{(Reference Adu, Li and Sawadgo22,Reference Baggio and Chong23)} have examined the direct health implications associated with trade agreements. In^{(Reference Baggio and Chong23)} ‘selected countries’, free trade agreements with the USA and their impact on obesity in the USA are studied. Whereas^{(Reference Baggio and Chong23)} investigates free trade agreements more broadly, our study is distinct by focusing specifically on the potential causal impact of NAFTA’s sugar agreement. Our study is similar to^{(Reference Adu, Li and Sawadgo22)} which investigates the causal impact of the NAFTA sugar trade agreement on diabetes prevalence using the synthetic control method. However, whereas^{(Reference Adu, Li and Sawadgo22)} focused on the USA as a whole, we employ the difference-in-difference (DD) and event-study approaches to examine the causal impact of the NAFTA sugar trade agreement on diabetes prevalence in each of the fifty US states. In doing so, we explore which populations were the most adversely affected by the policy since sugar consumption patterns and economic and sociodemographic characteristics vary from state to state.

To achieve our main objective, we use a DD research design – a quasi-experimental method commonly employed to evaluate the effects of policies or programmes following^{(Reference Wooldridge24,Reference Shahid, Li and Gao25)} and others. The DD estimate allows us to compare the differences in outcomes before and after the treatment (difference one) between a group exposed to the treatment and a control group (difference two). Additionally, we conducted a panel event study, where the ‘event’ was the date of implementation of the NAFTA sugar agreement following^{(Reference Shahid, Li and Gao25)}. Our findings provide compelling evidence supporting concerns that the unrestricted sugar trade agreement under NAFTA poses a risk to public health. We observed an increase in diabetes prevalence ranging from 0·54 to 2·3 % across various states within the USA following the implementation of the trade agreement. We find that states with a higher percentage of their population below the poverty level, a greater percentage of the Black population and a lower percentage of high school graduates were associated with greater increases in diabetes prevalence as a result of the NAFTA sugar trade agreement.

Methods

To identify the causal effect of the NAFTA sugar agreement on diabetes prevalence, we employ a DD model with fixed effects and an event study. The primary limitation to using DD analysis and randomised control trials in macroeconomic research is the challenge of meeting key assumptions, such as the ‘parallel trend assumption’. This assumption, which requires that the trends in the treatment and control groups would have been similar in the absence of the treatment, is difficult to establish at the macroeconomic level. To address these concerns, we rigorously tested the parallel trend assumption to ensure the validity of our estimates following^{(Reference Ryan, Kontopantelis and Linden36)}. The panel event-study method establishes causal relationships by comparing outcomes before and after an event. It effectively controls time-invariant heterogeneity and time trends, allowing for a flexible time specification. However, it relies on the assumption of common trends and can be sensitive to the choice of time periods. Additionally, its findings may have limited generalisability and require substantial data. Despite these limitations, it provides valuable insights into the dynamic effects of events on outcomes^{(Reference Freyaldenhoven, Hansen and Shapiro37)}. In our research design, a treated state adopted the NAFTA sugar agreement in the USA, while a control region (i.e. the control country includes six Organisation for Economic Co-operation and Development countries, including Australia, China, Norway, Japan, Switzerland and the UK) adopted no such trade policy. Intuitively, the idea is to compare the difference in average diabetes prevalence between treated and control regions before and after the treatment. The data analysis was performed using Stata 18 software and packages such as ‘did’, ‘hdid’ and ‘eventdd’, as well as Stata’s built-in regression functionality for linear regression.

The difference-in-differences analysis

We estimate the following equation:

(1) $${Y_{st}} = {\alpha _0} + {\beta _0}{\rm{Trea}}{{\rm{t}}_s} + {\beta _1}{\rm{Pos}}{{\rm{t}}_t} + {\beta _2}{\rm{D}}{{\rm{D}}_{st}} + {\beta _3}{\omega _{st}} + {\gamma _s} + {\tau _t} + {\varepsilon _{st}}$$

in which ${Y_{st}}$ is a continuous variable indicating crude diabetes prevalence in state $s$ at year $t$ , ${\rm{Treat}}$ is a binary variable indicating whether the state (or country) had adopted NAFTA, ${\rm{Post}}$ is a binary variable taking a value of 1 in the post-NAFTA period (and 0 otherwise) and DD is the interaction between ${\rm{Treat}}$ and ${\rm{Post}}$ . Our parameter of interest is ${\beta _2}$ . Finally, ${\gamma _s}$ and ${\tau _t}$ are, respectively, state and year fixed effects. Including year and country fixed effects is crucial for controlling time-specific and country-specific factors affecting diabetes prevalence. Year fixed effects account for time trends, while country fixed effects control for country-specific characteristics. This reduces omitted variable bias, improves comparability across countries and over time and provides more accurate estimates of variable relationships. However, our analysis captures unaccounted time-varying factors by using fixed effects at the country and time level. For example, differences in diabetes screening across countries or over time are captured through the fixed effects. However, if unobserved factors vary significantly across different states and these variations have different temporal trends, our model may not fully account for their impacts. This limitation suggests that the fixed effects model could underestimate the influence of these unobserved variables, which may vary not only by state but also over time. Future research might use models such as mixed-effects models or the inclusion of additional observable covariates to better control these factors. The variable ${\omega _{st}}$ captures variables such as the population aged above 65 years, the log of raw sugar import (metric tonnes) and the percentage of the female population, which have been found to affect diabetes prevalence in prior studies^{(Reference Adu, Li and Sawadgo22)}. Studies such as those by^{(Reference Adu, Li and Sawadgo22,Reference Blouin, Chopra and Van der Hoeven38,Reference Legge, Sanders and McCoy39)} have shown that an increase in sugar imports results in higher consumption of sugar and SSB, leading to a rise in diabetes and other diseases. Research also indicates that the prevalence of diabetes varies by age group, with older adults having the highest prevalence among all age groups^{(Reference Selvin, Coresh and Brancati40–Reference Lin, Thompson and Cheng42)}. Additionally, studies have shown that diabetes prevalence varies by gender, with a higher prevalence among females than males^{(Reference Boyle, Honeycutt and Narayan43)}. To overcome possible serial correlation problems, we used a robust se following^{(Reference Antón, Muñoz de Bustillo and Fernández Macías44)}. The key identifying assumption of DD analyses is that of common trends between treated states and the control countries in the absence of the treatment. We tested the parallel trend assumption following^{(Reference Shahid, Li and Gao25)} and removed states that did not meet the assumption from the analysis. We employed a Wald test, comparing against a null hypothesis of zero, to evaluate whether the linear trends were parallel before the treatment^{(Reference Shahid, Li and Gao25)}.

Event-study analysis

In addition to the DD analysis described above, we conducted a panel event study, with the ‘event’ being the date of implementation of the NAFTA sugar trade agreement in a particular state. We estimated the following equation:

(2) $${Y_{st}} = {\alpha _1} + \mathop \sum \nolimits_{j = 2}^J {\delta _j}{\left( {{\rm{Lag}}\;j} \right)_{st}} + \mathop \sum \nolimits_{k = 2}^K {\lambda _j}{\left( {{\rm{Lead}}\;k} \right)_{st}} + {\mu _t} + {\psi _s} + {\varepsilon _{st}}$$

where ${\psi _s}$ and ${\mu _t}$ are binary variables for state and year and ${\varepsilon _{st}}$ is the unobserved error term. Further, ${\rm{La}}{{\rm{g}}_j}$ and ${\rm{Lea}}{{\rm{d}}_k}$ are two binary variables indicating the number of years until the implementation of the NAFTA sugar agreement in state $s$ . Formally, we defined ${\rm{La}}{{\rm{g}}_j}$ and ${\rm{Lea}}{{\rm{d}}_k}\;$ according to equations (3)–(6):

(3)

(4) $${\left( {{\rm{Lag}}\;j} \right)_{st}} = 1\left[ {t = {\rm{Even}}{{\rm{t}}_s} - j} \right]\;{\rm{for}}\;j \in \left\{ {1, \ldots, J - 1} \right\},$$

(5) $${\left( {{\rm{Lead}}\;k} \right)_{st}} = 1\left[ {t = {\rm{Even}}{{\rm{t}}_s} + k} \right]\;{\rm{for}}\;k \in \left\{ {1, \ldots, K - 1} \right\},$$

(6) $${\left( {{\rm{Lead}}\;k} \right)_{st}} = 1\left[ {t \ge{\rm{Even}}{{\rm{t}}_s} + K} \right].$$

where ${\rm{Even}}{{\rm{t}}_s}$ is a variable indicating the year ${\rm{t}}$ in which the NAFTA sugar agreement was implemented in state $s$ . The first ${\rm{Lag}}$ was omitted to capture the baseline difference between the treated and control state.

Effect heterogeneity over time

To assess how the effects of the NAFTA sugar agreement vary with time, we follow^{(Reference De Giorgi, Geldsetzer and Michalik45)} to estimate a dynamic model where $\beta $ can vary across years:

(7) $${Y_{st}} = {\alpha _2} + {\beta _0}{\rm{Trea}}{{\rm{t}}_s} + \mathop \sum \nolimits_{t = 9}^{17} {\beta _{1t}}\left[ {{\rm{yea}}{{\rm{r}}_{st\left( {t = 2008} \right)}} - {\rm{year}}{{\left( {T = 1} \right)}_s}} \right] + \mathop \sum \nolimits_{t = 9}^{17} {\beta _{2t}}{\rm{Trea}}{{\rm{t}}_s}\left[ {{\rm{yea}}{{\rm{r}}_{st\left( {t = 2008} \right)}} - {\rm{year}}{{\left( {T = 1} \right)}_s}} \right] + {\gamma _s} + {\tau _t} + {\varepsilon _{st}}$$

${\rm{Trea}}{{\rm{t}}_s}$$ is a binary variable equal to 1 if State $s$ is ever treated (i.e. part of the NAFTA sugar agreement). Then, ${\left[ {{\rm{year}} - {\rm{year}}\left( {T = 1} \right)} \right]_c}$ is the difference between the observation year and the first year of implementation of the NAFTA sugar trade agreement in State $s$ . The parameters of interest are the ${\beta _{2t}}$ , which represents the mean difference in the crude diabetes prevalence in a specific year $t$ . We also control for State $\left( {{\gamma _s}} \right)$ and year $\left( {{\tau _t}} \right)$ fixed effects.

Data

Our analysis is conducted using a panel dataset that includes data at both the state and country levels from 2000 to 2016. The dataset contains 1054 observations, covering the fifty US states and six countries from the Organisation for Economic Co-operation and Development each year between 2000 and 2016. Data on crude diabetes prevalence that comprises type 1 and type 2 diabetes both at the state and country levels were obtained from the US Diabetes Surveillance website maintained by the Centers for Disease Control and Prevention and the NCD Risk Factor Collaboration (NCD-RisC). Data on sugar imports were sourced from the Food and Agriculture Organization of the United Nations Statistics Office (FAOSTAT). We obtained data on the percentage of poverty (all ages) from the 2009 US Census Bureau, Small Area Estimates Branch. The percentage of the Black population was obtained from the 2009 Census Bureau. Gender-related data were acquired from Kaiser Family Foundation (KFF) estimates, which relied on the 2008–2021 American Community Survey’s 1-Year Estimates. Additionally, data on educational attainment (high school diploma or more) were sourced from the US Census Bureau’s 2009 American Community Survey. Although a stronger strategy would concentrate solely on type 2 diabetes, which is affected by dietary modifications, we faced a lack of precise data regarding type 2 diabetes, particularly for the control countries. Considering the worldwide escalation in diabetes occurrence primarily linked to type 2 diabetes^{(Reference Zimmet, Magliano and Herman46)}, we hold the viewpoint that incorporating the combined prevalence of type 1 and type 2 diabetes will not impact the fundamental argument of this research. Figures 2 and 3 illustrate the diabetes trend for the states and countries used in our analysis.

Fig. 2 Trends of diabetes prevalence by states

Fig. 3 Trends of diabetes prevalence by control countries

Results

In our study, we investigate the causal impact of NAFTA’s sugar trade agreement on diabetes prevalence in the fifty US states. Overall, we find that the NAFTA sugar trade agreement has significant positive impacts on diabetes prevalence in most of the fifty states (see Fig. 4 and Table 1). Figure 4 displays the results of the event-study analysis, while Table 1 shows the estimated average treatment effects from the DD analysis. Across most states, both the DD and event analysis methods revealed significant effects of the NAFTA sugar trade agreement on diabetes prevalence. Because the NAFTA sugar trade agreement was a national policy, one would expect the states to experience similar effects. However, the impact on the fifty states varies. The estimated effect varies in statistical significance and magnitude from 2·28 % (P < 0·001) (95 % CI from 1·54 to 3·02) in Alabama to 0·50 % (P < 0·1) (95 % CI from −0·07 to 1·07) in Iowa (Table 1).

Fig. 4 Event-study results

Table 1 Difference-in-differences (DD) estimated results by state

Note: ***P < 0.01, **P < 0.05, *P < 0.1.

We classified the fifty states by the magnitude and significance of the estimated impact of the NAFTA sugar agreement on diabetes prevalence. We assessed the parallel trend assumption as per Shahid et al. (2022) and excluded states that did not fulfil this assumption from the analysis. Our estimates suggest that two states (Alabama and Arkansas) saw crude diabetes prevalence increase by greater than two percentage points as a result of the NAFTA sugar agreement. Fifteen states had an impact greater than 1 % but less than 2 % (Alaska, Arizona, Florida, Georgia, Indiana, Kansas, Mississippi, Nevada, New Mexico, North Carolina, Ohio, Oregon, South Carolina, Texas and Virginia). Eighteen states saw less than a 1 % impact (California, Connecticut, Illinois, Iowa, Maine, Massachusetts, Michigan, Minnesota, Missouri, Nebraska, New Hampshire, New York, North Dakota, Pennsylvania, Rhode Island, Washington, Wisconsin and Wyoming). Nine states (Colorado, Delaware, Hawaii, Idaho, Maryland, Montana, South Dakota, Utah and Vermont) had no effect on diabetes prevalence from the NAFTA sugar trade agreement. This illustrates that, while most states saw a rise in diabetes prevalence because of the NAFTA sugar trade agreement, some did not. Lastly, six states (Kentucky, Louisiana, New Jersey, Oklahoma, Tennessee and West Virginia) did not meet the parallel trend assumption.

To better understand the rationale behind the disparities in trade policy impact across states, we explore the relationship between selected covariates that have been shown in the health literature to be major factors influencing diabetes prevalence. We use 2008 values of the covariates for the analyses, as this represents the midpoint of our dataset. We concentrated on poverty, educational attainment, percentage of the population that is Black and percentage of the population that is female. All these variables have a statistically significant association (P value < 0·001) with the estimated diabetes prevalence. Higher poverty level, for example, is associated with higher diabetes prevalence, having a greater Black population is associated with diabetes prevalence, having a lower percentage of the population with a high school degree is associated with higher diabetes prevalence and having a higher percent female population is associated with higher diabetes prevalence.

These findings are consistent with previous health research. For example, research has demonstrated that individuals with greater levels of education are more likely to practise preventative healthcare behaviours, such as eating healthier meals, exercising more and preventing type 2 diabetes and obesity^{(Reference Pampel, Krueger and Denney47,Reference Montez and Zajacova48)} . Also, studies have investigated the effect of race on type 2 diabetes prevalence. The difference in type 2 diabetes prevalence across race and ethnic groups includes the prevalence of certain risk factors such as obesity and limited access to healthy foods. For example^{(Reference Divers, Mayer-Davis and Lawrence49)}, found that type 2 diabetes prevalence is higher in the Black population (5·97 %) compared with whites (0·77 %) in the USA from 2014 to 2015. It has also been shown that diabetes incidence has a greater link with poverty level^{(Reference Gaskin, Thorpe and McGinty50)}. Several studies have found that diabetes prevalence varies between males and females^{(Reference Gaskin, Thorpe and McGinty50)}. These findings point to the fact that the state-level differences in crude diabetes prevalence as shown in the current study can be attributed to sociodemographic characteristics such as poverty, race, gender and educational attainment.

Figures 5–7 as well as Table 2 present the results of equation (7). Table 2 shows the effect of the NAFTA sugar agreement for the fifty states across the posttreatment years. We observe that although the average treatment effect on the treated (ATET) was positive in 2008 (the treatment year) with its 95 % CI ranging from –0·17 to 0·30, it lacks statistical significance. This is not unexpected because just like any trade policy, the year of the agreement does not present a substantial impact. This is also reflected in Figs. 5–7, showing a marginal effect for 2008. We obtained statistically significant ATET (0·22 %) in 2009 with a 95 % CI ranging from 0·01 to 0·43. The ATET for 2010 is 0·31 % with 95 % CI ranging from 0·11 to 0·51. The ATET then increased from 0·92 % in 2012 (95 % CI from 0·14 to 1·70) to 1·12 % in 2016 (95 % CI from 0·13 to 2·11).

Fig. 5 Average treatment effect on the treated in pretreatment and posttreatment years

Fig. 6 Average treatment effect on the treated in posttreatment years

Fig. 7 Average treatment effects on the treated over different lengths of exposure to treatment

Table 2 Posttreatment – average treatment effect on the treated (ATET)

Number of observations: 952.

As a robustness check, we used equation (1) to apply the difference-in-difference approach at the aggregate level (see Table 3). Our regression model specifications revealed a statistically significant impact of the NAFTA sugar trade agreement on diabetes prevalence. This result is consistent with all our other findings.

Table 3 Results of difference-in-difference (DD) regression

Robust se in parentheses: ***P < 0·01, **P < 0·05, *P < 0·1.