Study design

We conducted an ecological study using municipal-level data for Mexican adults from 2005 to 2021 to identify heterogeneity within large geographic areas by examining homogeneous sub-areas. Small-area estimation is a statistical approach employed to predict characteristics of interest, such as mortality rates, utilizing existing data. This method involves substantial stratification based on geographic and temporal factors, especially for small geographical units. By avoiding the need for primary data collection within each unit, it streamlines the estimation process [7]. In the context of Mexico, the smallest geographical and administrative units are the municipalities, whose areas range from 2.2 to 53,104 Km², and population density ranges from 0.14 to 19,436 inhabitants/ Km² [10]. Adopting the municipality as the small-area unit for our analysis, we used available data on population sociodemographics, access to healthcare services, and mortality records at the municipal level. This data, sourced from censuses and public institutions covering the period from 2005 to 2021, facilitated the estimation of municipal predictors for NCDs mortality rates, stratified by sex and quinquennium of age. Following the SSA approach, we initially designated the municipality as the primary area of interest, encompassing a total of 2,454 municipalities in Mexico for the year 2005. We processed microdata from 2005 to 2021 and fitted a negative binomial model to estimate the relationship between observed municipal factors and NCDs mortality. Subsequently, we predicted the standardized MR for each municipality in each year. Additionally, we generated annual state-level graphs and municipal-level maps based on both crude and standardized MRs.

Data sources

We collected data from official public open-access microdata at the municipal-level from 2005 to 2021. The municipal-level was defined in accordance with Mexico’s 2005 geographical division encompassing a total of 2,454 municipalities in 32 states [6]. Databases were obtained from the National Population Council (CONAPO) [11], National Autonomous University of Mexico (UNAM) [10], National Institute of Statistics and Geography (INEGI) [6], National Council for the Evaluation of Social Development Policy (CONEVAL) [12], National Directory of Economic Units (DENUE) [6], and the General Direction of Health Information (DGIS) of the Ministry of Health [13].

Study outcome

Our dependent variable was the number of deaths from NCDs. The single variable that encompassed the total deaths from NCDs included deaths from: T2D, CVD, CKD, cardiovascular diseases (CVD), hypertension, dyslipidemia and obesity. The data pertaining to adult deaths (individuals aged ≥20 years) attributed to NCDs were obtained from DGIS [13] for the period spanning from 2005 to 2021. The NCDs considered in this study were categorized in accordance with the International Statistical Classification of Diseases, 10th edition (ICD-10). Our dependent variable was a composite variable encompassing the total deaths from NCDs, including T2D (codes E10-E14), CVD (codes I20-I25 and I63-I64), and CKD (codes N18 and N19), along with deaths related to the following NCD risk factors: hypertension (codes I10, I15, O10, R030), dyslipidemia (code E78), and obesity (code E66). These variables were standardized as annual municipal rates, stratified by sex and quinquennium of age.

Predictor variables

We employed several sociodemographic and access to healthcare services predictor variables derived from diverse sources to comprehensively assess municipal-level factors. The predictor variables were selected for being available at the municipal level and being considered factors associated with a higher mortality risk from NCDs at the individual level [6, 14–24]. In the negative binomial model, the cases of deaths due to NCDs per year, municipality, age group, and gender were the dependent variable. The analysis units consisted of age and sex categories within the 2,454 municipalities per year Sociodemographic and municipal environmental predictor variables were transformed into annual proportions, densities, and rates per 100,000 inhabitants. The exposure in the model was the adult population of the municipality per year, age group, and gender, allowing for the adjustment of mortality rates based on the population. The population size of municipalities was based on CONAPO projections [11]. Municipal geographic area (in square kilometers) was obtained from estimations provided by UNAM to estimate population density (Total population/Km²) [10, 14, 15]. We collected demographic information, including the proportions of indigenous speakers aged ≥5 years, illiteracy in population aged ≥15 years, from INEGI counts (2005) and census data (2010 and 2020) [6, 22, 23]. Information on the population affiliated with social security services and proportion of households with television was also obtained from INEGI [6, 16, 24]. Poverty estimations at the municipal level were retrieved for the years 2010, 2015 and 2020 from CONEVAL, a governmental institution that assesses and classifies poverty according to income and the presence of the following social deprivations: education, healthcare, social security, housing services, or access to food. The three categories of poverty from CONEVAL’s evaluation are: a) Poverty: population whose income is below the value of the well-being line and who experiences at least one social deprivation, b) Extreme poverty: population experiencing three or more social deprivations and whose income is below the minimum well-being line, c) Moderate poverty: it refers to individuals who, while being poor, do not fall into the category of extreme poverty. The incidence of moderate poverty is calculated by determining the difference between the incidence of the population in poverty and that of the population in extreme poverty. For our study, we used estimations of moderate poverty at the municipal level [12, 18]. We evaluated the ultra-processed food, alcohol & tobacco retail environment at municipal-level using microdata from INEGI, which has been annually registering economic units’ activities, size (base on employed personnel), and locations in the country since 2010 [6, 17]. Each economic units in the DENUE database was coded according to the criteria of the North American Industrial Classification System (NAICS) criteria [6, 25]. We characterized the ultra-processed food, alcohol & tobacco retail environment using codes associated with higher households purchases of ultra-processed-products (UPP) reported by Hernández M, et al. [25] that includes convenience stores, discount stores, small grocery stores, and stores specialized in selling candies, popsicles, soft drinks, and other food and beverages. Additionally, codes for stores selling alcohol & tobacco were incorporated into the classification of ultra-processed food, alcohol & tobacco retail environment. To assess the ultra-processed food, alcohol & tobacco retail environment, we calculated the store density by dividing the number of stores by the total population per 100,000 inhabitants at the municipal level.

Statistical methods

To address missing values in predictor variables not measured between census years (2006–2009, 2011–2019, and 2021), we employed linear and logistic regression models adjusted by age and sex. To standardize the predictor variables, we transformed them into proportions, densities, or rates per 100,000 inhabitants, using population projections from CONAPO at the municipal level, stratified by age group and sex as the denominator. Furthermore, to examine potential non-linear associations, we categorized predictors variables into quintiles over the study period. To estimate the municipal factors associated with the NCDs MR, we utilized negative binomial regression models. These models were adjusted for sex, age group, predictor variables (categorized in quintiles) and year. The dependent variable consisted of the number to NCDs-related deaths by year, sex, age group, and municipality of residence. We assessed multicollinearity among covariates using the diagnostic tool of Variance Inflation Factor (VIF), incorporating variables with values <5 in the statistical models [26]. The final model selection was guided by achieving the best goodness of fit, determined by the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). We adjusted the models using predictor variables in both continuous and categorical formats. The expected MR per 100,000 inhabitants was calculated using the predicted deaths from the negative binomial regression model and was subsequently age-standardized according to Mexico’s 2015 age structure. Additionally, recognizing the impact of excess mortality during the COVID-19 pandemic [27], we conducted a separate model to analyze municipal-level factors from 2005 to 2019. All analyses and maps were developed with Stata 17 (College Station, TX).

Ethics committee approval was not required for this study which analyzed publicly available datasets. Patient consent for publication was not required.

Equation of the model:

Where:

NCDs deaths: Mortality from NCDs in adults ≥20 years old (Dependent variable)

j: municipality

t: year (2005–2021)

n: inhabitants in the municipality ≥20 years old, by age group and gender.

β₀: intercept

D = i | χ_1jt: D stands for a dummy variable that equals one for category i in χ_1jt. The first category in χ_1jt is the reference group.

Covariates: χ_1jt: sex, χ_2jt: age group (quinquennium in adults ≥20 years), χ_3jt: proportion of males, χ_4jt: proportion of indigenous population, χ_5jt: proportion of illiteracy, χ_6jt: proportion of moderate poverty, χ_7jt: population density, χ_8jt: density of ultra-processed food, alcohol & tobacco retail stores, χ_9jt: proportion of households with television, χ_10jt: density of hospitals, χ_11jt: rate of hospital discharges from NCDs, χ_12jt: proportion of the population affiliated to Mexican Social Security Institute, χ_13jt: proportion of the population without social security, χ_14jt: crude mortality rate from causes other than NCDs, χ_15jt: year, χ_16jt: state of the country.

i: covariates categories (e.g., i = (0,1,2) for the sex variable as in χ_1jt or i = (1,2,3,4,5) for the quitile-based variable as in χ_2jt)

ε: unobserved random variables (or error term)

μ: unobserved fixed variables

Trends in NCDs mortality

The expected NCDs MR in Mexican adults increased over fifty percent from 2005 to 2021. Between 2015–2017, a trend emerged presenting an average annual decrease of two percentage points in the risk of NCDs mortality, persisting until the onset of the COVID-19 pandemic. This decline in NCDs mortality during this period could potentially be linked to government strategies focus on improving the quality of health care services in the previous years [28]. In 2020–2021, we observed a resurgence in MR compared to the preceding study period, which may be attributed to the advent of the COVID-19 pandemic.

The observed increase in MR across states and municipalities displayed marked heterogeneity, which could be attributed to the distribution of various ecologic health-related indicators [29]. A comprehensive study that assessed health-related sustainable development goals indicators in 1990–2017, including different measurements such as: socio-demographic, marginalization and concentration indexes, financial protection, poverty and health expenditure, revealed substantial disparities in performance among different states in Mexico [29]. For example, states such as Nuevo León and Sinaloa had a better progress towards the health-related indicators, and Oaxaca and Chiapas a worse performance. These disparities were related to socioeconomic factors including poverty and health expenditure [29].

Municipal predictors of NCDs MR

As expected, mortality showed a direct correlation with age. However, from age 50, women had higher mortality than men. These findings aligned with a comprehensive meta-analysis, which attributed the higher NCD-related mortality among women to genetic, biological, gender disparities, cultural, and environmental factors [20]. Since the year 2000, countries in the Americas, excluding Mexico, experienced a decline in the CVD mortality among women; while Mexico experienced stagnation, a trend linked to its evolving socioeconomic development [21]. Consequently, gender disparities in the burden of NCDs has changed in favour of women which are less affected by tobacco and alcohol use but face a higher risk of physical inactivity [30].

Furthermore, our study revealed that municipalities with a higher proportion of indigenous populations had elevated NCDs mortality. Over the past two decades, MR among indigenous groups have been greater than those of non-indigenous populations, primarily attributed to NCDs [6]. While social and cultural determinants of indigenous health have been described, the heterogeneity in living environments [31] and the influence of commercial determinants [22] could offer additional explanations for our findings [22, 31]. A meta-analysis has suggested lower indigenous MR in urban compared to rural areas [22], likely attributable to the health advantages associated with urban living [22]. Regarding the commercial determinants of health, which involve private companies’ impact on public health goals, industries have been known to promote the consumption of UPP, alcohol, and tobacco [31]. Consequently, these unhealthy products have emerged as the primary drivers of health inequalities within indigenous population [31].

Our study also identified an unexpected association between a higher proportion of illiteracy and a lower NCDs mortality risk. These findings contrast with existing evidence linking lower education levels to decreased CVD mortality [23]. However, when analyzing NCDs mortality between 1990–2018, an increase in deaths was observed in countries with middle to low development countries while very-low development countries contributed the least to this mortality [19]. Authors attributed this trend to lower urbanization and dependence on local food in the least developed areas [19]. In this context, illiteracy may be linked to regional development, and a lower NCDs burden might be observed in least developed areas [19]. Another plausible explanation for the inverse correlation between illiteracy and NCDs mortality could be attributed to the positive trend in the life expectancy among illiteracy populations, while remained stable for non-illiterate populations [6, 33]. Further studies are needed to comprehensively analyze this relationship. Our results also exposed a lower NCDs mortality risk in municipalities with a higher proportion of population unaffiliated to health services. This finding could be attributed to undiagnosed morbidity and mortality resulting from NCDs. For example, after the implementation of health insurance programs in U.S.A. between 2007 and 2016, the authors observed an increase in NCDs prevalence trends from pre- to post-implementation programs [24].

In Mexico, despite improvements in social conditions since 2008, particularly in terms of healthcare access, there has been limited progress in population income [12, 28]. Poverty in this context, comprises factors that encompass the overall well-being of populations, including aspects like food security. This could result in insufficient fulfillment of essential needs, such as access to affordable and nutritious diets [28, 32]. Our results showed a notable association between municipalities with a higher proportion of poverty and an increased risk of NCDs mortality. However, within a food insecurity context, the population is more likely to turn to high-energy products, such as UPP, due to their affordability [32]. This propensity towards the consumption of such products increases their likelihood of obesity and, consequently, NCDs [32]. Regarding ultra-processed food, alcohol & tobacco retail environment, our results are consistent with a study where unhealthy food retail increased all-cause mortality, including NCDs [17]. The UPP contains sugars, saturated fats, sodium, and food additives linked to elevated morbidity and mortality from NCDs [33]. However, another study found limited evidence of associations between food retail and NCDs, possibly due to non-standardized measurements [4].

Population densities can be considered risk features for NCDs [14]. Despite that the high population density appears to be associated with higher NCDs mortality [14], we found that population density was inversely linked to NCDs mortality risk. Studies suggested that better infrastructure (for walking and cycling) in high-density cities might mitigate the density-mortality link [15]. Although infrastructure data at municipal-level is not publicly available in Mexico, a study [34] indicated that the sitting time (including watching television) and physical inactivity increased by more than 40% between 2006 and 2018. Higher NCDs mortality risk was found in our study for in municipalities where most of households have a television. Previous studies have linked prolonged television viewing to increased risk of NCDs and mortality [16].

Our investigation revealed that municipalities with higher hospital density had lower NCDs MR. In Mexico, hospitals are predominantly concentrated in urban areas, with only 3.3% located in rural regions [28]. This distribution of healthcare facilities raises a critical issue because individuals residing in rural areas face a higher mortality risk compared to those in urban, primarily due to healthcare inequalities and deficiencies in infrastructure and medical personnel [12, 35]. Thus, the observed protective effect of hospital density against NCDs MR underscores the persistent rural-urban health disparities that exists within the country. Interestingly, our research uncovers a paradoxical finding in relation to public healthcare services and NCDs mortality. We observed that municipalities with high proportion of affiliation with IMSS had a higher risk of NCDs mortality, even in the model for the period from years 2005 to 2019. IMSS is the largest public healthcare provider in the country, covering approximately 38% of adults, however also faces challenges [36], particularly during COVID pandemic. According to reports [6] in 2021 42.0% (469,542) of deaths in Mexico occurred in public and private hospitals, with 43% of these occurring within the IMSS, and the remainder in other healthcare facilities [6]. Consequently, our findings regarding mortality and IMSS affiliation must be interpreted cautiously recognizing the complex and multifaceted landscape of healthcare services in Mexico.

Estimating age-adjusted and standardized MR offers the advantage of reducing the ’volatility’ in the estimations for municipalities with very small populations. In such cases, we often observe mortality rates in the thousands (e.g., in 2021 municipality Santo Domingo Tlatayápam from Oaxaca, with 4 cases and 102 inhabitants, had a crude MR = 3,922 per 100,000 inhabitants vs an estimated MR = 513.5, CI 95% 487.6–540.8) where a single death can exert a significant impact on the metrics due to the very small population denominator. Therefore, relying solely on crude mortality rates to target preventive and treatment efforts in these municipalities may not necessarily pinpoint the most urgent areas for intervention. Consequently, this article provides insights into identifying priority locations across the country that require targeted interventions to prevent mortality cases related to NCDs and outlines the characteristics of these municipalities.

Strengths and limitations

Our study is novel in using a municipal approach to describe NCDs mortality predictors in Mexico between 2005 and 2021. The methodology can be applied to other diseases. Our study also provides a broader understanding of NCDs mortality heterogeneity throughout the country, providing relevant knowledge on the specific health needs of each state and municipality. We analyzed extensive data across years integrating a relevant number of observations and showing robust and consistent results. Limitations include the ecological design lacking temporal criteria; therefore, there is no causal inference between the risk factors and NCDs mortality. Caution against ecological fallacy suggests interpreting results as group-based instead of individual. The precision of our estimations depended on the quality and the study covariates representativeness. Moreover, since population size was relatively small for some municipalities, the observed rate could differ from the expected due to the random data variation. Finally, measurement errors are usually greater in smaller areas than in large areas, so the estimates’ reliability is more precise in the former. Nevertheless, the covariables in this study at the municipal level produced representative estimations with predictive strength.