DataPresentation of the global burden of disease database

We extracted our outcome variables and some of our covariates from the GBD database [26]. This data source has been described as the most comprehensive health database worldwide. It provides estimates of 369 diseases and injuries for 204 countries and territories, using more than 86,000 data sources from 1990 to 2019. The GBD also provides a number of covariates measuring socioeconomic, demographic, health system access, climate, and food consumption indicators.

Primary data sources include a comprehensive catalogue of health-related data such as surveys, studies from the scientific literature, censuses, registries, and other administrative data. To be included in the GBD study, input sources have to comply with guidelines provided by GATHER which are the gold standard for observational studies [28]. Following GATHER ensures that adequate information is available to assess the quality of the source data, in particular that study samples are representative of the general population and have clinical thresholds established by international classifications.

All available data are standardized, mapped to the GBD cause list, stratified by age and sex, corrected for miscoded causes and redistributed to appropriate causes, aggregated over various sources, and finally pooled into a single database [26].

The number of sources available in the GBD database to provide estimates for each condition was as follows: 2569 for OUD; 455 for low back pain; 3944 for alcohol use disorders; 197 for anxiety disorders; 147 for bipolar disorder; 461 for cannabis use disorders; 528 for depressive disorders; 1949 for eating disorders; and 202 for schizophrenia [29]. Note that OUD, alcohol use disorders, and eating disorders have sources for fatal in addition to non-fatal outcomes, which the other disorders have not.

These relatively large numbers are mitigated by two caveats. First, available data may be of poor quality, for instance if it does not use the preferred case definition or an appropriate measurement method. Using a 5 stars classification system, the GBD provides a grading of data quality for each country, which depends on data availability, completeness, detail of mortality data and percentage of deaths coded to ill-defined codes or highly aggregated causes [26]. In the current study, we only included data from locations whose data quality was rated with at least 3 stars (N = 87 countries from the original 204 locations where DALYs were estimated; Table 1).

Second, the availability of the data sources is unequally distributed geographically and temporally. To generate robust cause-specific estimates by age, sex, year, and location, the GBD uses powerful statistical modelling techniques that incorporate external data (e.g. from other locations and over time), and enforce the consistency between epidemiological parameters. A detailled description of these methods is provided elsewhere [26, 30,31,32].

Outcome

Our outcome variable was the burden of OUD 2019. Subsequent analyses involved the burden of low back pain and that of other adult mental and substance use disorders for the same year. These were extracted from estimates of the GBD study 2019 as the 2019 age-standardized Disability-Adjusted Life Years rates per 100,000 inhabitants (hereby 2019 DALYs) [26]. Mental and substance use disorders included in the study were: alcohol use disorders, anxiety disorders, bipolar disorder, cannabis use disorders, depressive disorders, eating disorders, schizophrenia.

Briefly, for a specific year, country and disability, DALYs are a measure of overall disease burden, expressed as the sum of the number of estimated Years Lived with Disability (YLD) and early death (Years of Life Lost, YLL). YLD is measured according to the formula YLD = Prevalence x Disability Weights. Disability weights are based on population surveys to lay descriptions of sequelae highlighting major functional consequences and symptoms [26]. Disability weights are held invariant between age and sex groups, as well as locations and over time. Disability Weights are measured on a scale from 0 (full health) to 1 (death). Importantly, in the GBD study, each disability is collectively exhaustive and exclusive of any comorbidity.

Outcomes were log-transformed to allow interpretation of coefficients as percentage differences, and to account for possible non-gaussian distributions.

Exposure

The Globalization Index was extracted from the KOF Swiss Economic Institute [33]. The KOF is a “Konjunkturforschungsstelle” which in German means “Business Cycle Research Center”. The Globalization Index measures de facto and de jure economic, social and political dimensions of globalization for the period 1970 to 2019 on a scale of 1 to 100. As of 2019, it is available for 195 countries. To calculate the Index value, different variables are used and are aggregated using statistically determined weights (obtained from principal component analysis) [33].

The sub-​segment of economic globalization comprises:

de facto financial globalization: foreign direct investment, portfolio investment, international debt, international reserves, international income payments. International debt in particular is defined as the sum of inward and outward stocks of international portfolio debt securities and international bank loans and deposits (as a percentage of the GDP);

de jure financial globalization: investment restrictions, capital account openness, and international investment agreements.

The sub-​segment of social globalization comprises:

Interpersonal contacts:

de facto: international telephone connections, transfers, tourism flows and migration;

de jure: telephone subscriptions, international airports and visa restrictions.

Information flows:

de facto: international patent applications, international students, and high-​technology trade;

de jure: access to television and the internet, press freedom, and international internet connections.

Cultural globalization:

de facto: trade in cultural goods, registrations of international trademark rights, trade in personal services, and the numbers of McDonald’s restaurants and IKEA stores;

de jure: civil rights, gender equality and expenditure on education.

The sub-segment political globalization comprises:

de facto: number of embassies, international non-governmental organizations (NGOs) and participation in UN peacekeeping missions;

de jure: membership of international organizations and international treaties, number of partners in investment treaties.

The KOF provides a comprehensive file that describes the sources and definitions of each sub-segment of globalization (https://ethz.ch/content/dam/ethz/special-interest/dual/kof-dam/documents/Globalization/2018/Globalisation%20Index%202018_1.zip).

We used the KOF Globalization Index (KOFGI) for the year 2018, and further discretized values into 4 different levels (Table 1):

The lowest level of globalization included 23 countries with a globalization index between 41 and 64 (Low; [41, 64]; N = 23 countries);

The middle-low level of globalization included 23 countries with a globalization index between 64 (excluded value) and 72 (Mid-Low; (64,72]; N = 23 countries);

The middle-high level of globalization included 23 countries with a globalization index between 72 (excluded) and 82.5 (Mid-High; (72,82.5]; N = 19 countries);

The highest level of globalization included 23 countries with a globalization index between 82.5 (excluded) and 91 (High; (82.5,91]; N = 22 countries).

Further analysis investigated the above-mentioned sub-indices of globalization, which we also discretized into 4 different levels of increasing magnitude.

CovariatesConfounders

Confounders impact both the exposure (economic, socio-cultural and political aspects of globalization) and outcome (the burden of OUD), and are crucial to include in a statistical model to avoid biases in parameter estimates.

We included the following country-level confounders:

- A country’s level of development. Broadly speaking, development is associated with a high standard of living, a high level of industrialization, and advanced technological infrastructure [34]. Specifically, developed economies are associated with efficient transportation infrastructure, telecommunication and information systems, good labor skills, and political stability, that in turn are thought to attract foreign investors [35, 36], and would also favor economic and socio-cultural globalization.

On the other hand, a country’s level of development has been associated with the rate of opioid analgesic consumption [37, 38]. Likewise, as mentioned above, high-income countries have seen a dramatic increase in the burden of OUD compared to low- and middle-income countries [2], suggesting that socioeconomic status and development may be related to OUD [39].

We extracted the Socio-Demographic Index (SDI) for the year 2018 from the GBD database [40] as an index of the level of development of a country. The SDI is a composite indicator of development status, and is measured as the geometric mean of indices of total fertility rate, mean education for those aged 15 and older, and lag distributed income per capita. Its range is 0–1, with values closer to 1 indicating a higher level of development.

- Unemployment rate. A low rate of unemployment would increase the spending power of individuals and be an indicator of a country’s growing economy as well as social stability. In contrast, a high rate of unemployment would signal a failure to utilize the available labor force, i.e. a stagnating, non-resilient economy. According to these accounts, foreign investors would favor countries with a low unemployment rate – provided that they have highly skilled and available employees [41, 42].

On the other hand, unemployment rates have been related to increased opioid prescribing and misuse, and higher overdose mortality [43, 44].

We extracted the Unemployment rate for the year 2018 (unemployed citizens as a percentage of the labor force) from the World Bank database [45].

- Income inequality. Income inequality might drive decreased globalization. Indeed, in countries with a high level of inequality, the poor and middle classes incur reduced spending power and may take on debt for consumption, increasing economic, social and political instability. Inequality also causes increased rents and decreased productive activity, reducing growth and development [46]. Overall, these might lead to reduced foreign investments, as well as lower rates of social and political aspects of globalization.

On the other hand, income inequalities have been related to the burden of OUD. For instance, in Canada, people living in lower-income areas have been found to experience higher rates of opioid-related harms [47]; and in the US, areas with greater income inequality have higher rates of overdose deaths [48].

We extracted a measure of income inequality for the year 2018, as measured by the p90p100 index, from the World Inequality Database [49]. This measure represents the ratio of individuals whose income belongs to the top 10% of the population, divided by the entire population. Its range is 0–1, with values closer to 1 indicating a higher level of inequality.

- Urbanization. Cities are described as the “theaters where globalization plays out” [50], where people, capital, information, services and goods converge to enable economic and socio-cultural globalization [50].

On the other hand, urbanization has been linked to OUD. In the US, while all states saw an increase in opioid-related harms [51], some have reported that a larger proportion of people suffering from OUD may originate from rural (vs. metropolitan areas) [51, 52].

We extracted an index of urbanization for the year 2018 from the GBD database (from 0, low degree of urbanity to 1, high degree of urbanity) [53].

- A pre-exposure measure of the outcome. Here, a pre-exposure measure of OUD cannot be considered, stricto sensu, as a confounder as it may not directly affect globalization. However, it is often (if not always) advised to include a pre-exposure outcome value in a statistical model [54,55,56], as it may be related to an underlying source that might also affect the exposure. Examples of such underlying sources may be historical, geographical or cultural factors both related to globalization and the burden of OUD, e.g. typically factors linked to liberalism and consumerism.

We extracted a pre-exposure measure of the outcome (log-transformed 1990 age-standardized DALYs for OUD – but also all other outcomes tested in this study) from the GBD database [26].

- Other important determinants of globalization, especially economic globalization, may be: quality of institutions and economic structures, market size, trade openness, tendency to tax economic actors, or labor cost [35, 57, 58]. However, we reasoned that these were not obviously related to the burden of OUD (other than via other confounders mentioned above and already included in our statistical model).

Factors aiming to improve the precision of the parameter estimates

Factors aiming to improve the precision of the parameter estimates are those factors that are only related to the outcome. We hypothesized such factors to be:

- Children sexual abuse [59]. From the GBD database, we extracted the age-standardized summary of exposure value (SEV) for children sexual abuse for the year 2018, which measures a population exposure to children sexual abuse and takes into account the contribution of that risk to disease burden [53]. SEV is reported on a scale from 0 to 100%; SEV takes the value zero when there is no excess risk for a population and the value 100 when the population is at the highest level of risk.

- Quality of measurement of the outcome. We used the data quality rating system from the GBD database [25].

- Access to and quality of healthcare services [52]. We extracted an index of access to and quality of healthcare services for the year 2018 from the GBD database [53]. Such an index ranges from 0 (low quality of healthcare services) to 100 (high quality).

Mediators

Key mediators of the relationship between globalization and OUD were discussed above as being linked to liberalization, privatization, disengagement from the State and, overall, a free-market, deregulated economy. In turn, these are thought to promote: (1) profit-making clinical practices that over-prescribe opioids; (2) pharmaceutical companies that aggressively marketize opioid products; (3) over-demand and over-consumption of opioids from patients; (4) companies that de-localize industrial activities to countries with cheaper workforce; and (5) disengagement from unions, and more generally public regulations and policies services, that fail to protect citizens from financial and economic losses, but also from occupational issues (including occupational pain and diseases). It is important to note that as these factors may be on the causal path between globalization and OUD, they should not be included as covariates in our statistical model. Indeed, adjusting for these factors would under-estimate the total effect of globalization on OUD.

AnalysisMethodological strategy

Our methodological strategy consisted in the following steps:

1.

For each observation\(i\), we predicted log 2019 DALYs \(\widehat_}\) at each of the four levels of globalization, taking into account our set of covariates. Estimation relied on targeted maximum-likelihood estimation (TMLE) and machine learning (ML) (see estimation strategy below) [27]. Briefly, to reach \(\widehat_}\), observed (log-transformed) outcome values \(DALY_\)were initially estimated based on our set of covariates. Initial predictions were then updated using inverse probabilities of exposure (a.k.a. weights) to the four levels of globalization (which were themselves predicted based on the covariates).

2.

To ensure that our analyses did not suffer from positivity violations, we verified that each country had some probability (i.e. a positive probability) of being exposed to various levels of globalization given their characteristics [60,61,62]. Probabilities of exposure were predicted using ML algorithms, scaled by the marginal probability of exposure (a.k.a. stabilized) [62], and were trimmed to the 99.9th percentile.

3.

Taking the lowest level of globalization (Low) as the reference, we estimated how increasing levels of globalization were associated with our (updated) predictions of 2019 DALYs (mean log difference: High, Mid-High, Mid-Low vs. Low) using the following linear equation:

$$\widehat_}\equiv _+_Glo_+_Glo_+_Glo_$$

, where:

\(Glo_,Glo_, Glo_\)are dummy regressors which, for each observation \(i\), take either values \(1\) or \(0\) according to whether the observation belongs to the High, Mid-High or Mid-Low globalization level, respectively;

\(}_, _, _\)are the adjusted marginal effect, in percentage, of increasing levels of Globalization (from Low to High, Mid-High and Mid-Low, respectively), on predictions of 2019 DALYs.

Estimation strategyTargeted maximum likelihood estimation

TMLE is a doubly-robust estimator, which relies on the estimation of (1) the outcome regression (where outcome values are predicted based on our set of covariates); and (2) the exposure mechanism (where probabilities of exposure are predicted based on the covariates). Both steps are further integrated by updating (“targeting”) the initial estimation of the predicted outcome based on each country’s inverse probability of exposure (a.k.a. weights) [27]. This aims to incorporate information from the exposure mechanism and create a pseudo-random population of observations with respect to the distribution of the covariates. This also optimizes the bias-variance tradeoff for the given parameter of interest.

Crucially, this makes TMLE a doubly-robust method that will yield unbiased estimates if either the estimated outcome regression or exposure mechanism is consistently estimated [27]. When both the outcome regression and exposure mechanisms are consistently estimated, TMLE is an asymptotically efficient estimator [27]. More technical details on how to obtain targeted maximum likelihood estimates are provided elsewhere [27, 60].

Machine learning

To maximize our chances to consistently estimate the outcome regression and the exposure mechanism, we used data-based estimation techniques, a.k.a. ML algorithms [27]. ML algorithms significantly improve the quality of estimation compared to pre-specified parametric models (typically general linear models) which put strong assumptions on data distributions, are likely mis-specified and prone to confirmation biases. Here, we used an ensemble ML algorithm called SuperLearner [63], which we defined as a weighted linear combination of the following basic learners: Linear model with main effects only, Stepwise regression with a step forward procedure, Linear regression with L1-regularization [64], Multivariate adaptive regression splines (MARS) [65], Random Forest (RF) [66].

We used the default hyperparameters of the SuperLearner R package for the MARS (inc. maximum degree of interaction = 2) and RF algorithms (inc. number of trees = 500; number of variables to possibly split at in each node = 2; minimum node size = 5 for the outcome regression and 1 for the exposure mechanism). Cross-Validated R2 and multi-class Area Under the Receiving Operating Curves were our main indicators of performance. As an example, we report the performances of ML algorithms for the outcome regression and the exposure mechanism when analyzing the association between globalization and the burden of OUD 2019 in Supplementary Table 1 (trimming probabilities of exposure to the 99.9th percentile and not removing outliers).

Main analyses

Using the above methodological strategy, we investigated:

(1)

whether globalization was associated with the burden of OUD 2019;

(2)

whether globalization was associated with the burden of low back pain and that of other mental and substance use disorders for the same year;

(3)

whether sub-categories of globalization were associated with the burden of OUD 2019.

Sensitivity analyses

Finally, to check the robustness of our results, we ran two other sets of analysis where:

we used two other levels of truncation for each country’s probability of exposure (truncating probabilities to the 99th and 97.5th percentiles instead of the 99.9th percentiles).

we removed outlier observations, defined as those with an outcome value inferior to the 1st percentile or superior to the 99th percentile (Supplementary Table 2).

The STROBE for cross-sectional analysis guided the writing of this manuscript [67]. For this country-level analysis that relied on publicly available data, no ethics approval was required. All analyses were performed using R version 4.1 and package lmtp [68].