- Open Access
- Open Peer Review
This article has Open Peer Review reports available.
Estimating global mortality from potentially foodborne diseases: an analysis using vital registration data
© Hanson et al; licensee BioMed Central Ltd. 2012
Received: 11 March 2011
Accepted: 16 March 2012
Published: 16 March 2012
Foodborne diseases (FBD) comprise a large part of the global mortality burden, yet the true extent of their impact remains unknown. The present study utilizes multiple regression with the first attempt to use nonhealth variables to predict potentially FBD mortality at the country level.
Vital registration (VR) data were used to build a multiple regression model incorporating nonhealth variables in addition to traditionally used health indicators. This model was subsequently used to predict FBD mortality rates for all countries of the World Health Organization classifications AmrA, AmrB, EurA, and EurB.
Statistical modeling strongly supported the inclusion of nonhealth variables in a multiple regression model as predictors of potentially FBD mortality. Six variables were included in the final model: percent irrigated land, average calorie supply from animal products, meat production in metric tons, adult literacy rate, adult HIV/AIDS prevalence, and percent of deaths under age 5 caused by diarrheal disease. Interestingly, nonhealth variables were not only more robust predictors of mortality than health variables but also remained significant when adding additional health variables into the analysis. Mortality rate predictions from our model ranged from 0.26 deaths per 100,000 (Netherlands) to 15.65 deaths per 100,000 (Honduras). Reported mortality rates of potentially FBD from VR data lie within the 95% prediction interval for the majority of countries (37/39) where comparison was possible.
Nonhealth variables appear to be strong predictors of potentially FBD mortality at the country level and may be a powerful tool in the effort to estimate the global mortality burden of FBD.
The views expressed in this document are solely those of the authors and do not represent the views of the World Health Organization.
Foodborne diseases (FBD) encompass a wide spectrum of illnesses that manifest after the ingestion of contaminated foods and food products. They can be caused by a variety of microbial pathogens, chemicals, and parasites that contaminate food at different points in the food production and preparation process. Notwithstanding the morbidity and disability resulting from foodborne diseases, the ingestion of contaminated food can lead to death. Diarrheal diseases alone, a considerable proportion of which are foodborne, kill 1.5 million children every year worldwide . Although most of these diarrheal deaths occur in poor countries, foodborne diseases are neither limited to developing countries nor to children. It is estimated that in the United States, foodborne diseases result in 37.2 million illnesses, 228,744 hospitalizations, and 2,612 deaths each year .
The full extent of the burden and cost of unsafe food is currently still unknown, but its impact on global health, trade, and development is considered to be immense. Recognizing the current data gap, the World Health Organization (WHO) launched the Initiative to Estimate the Global Burden of Foodborne Diseases. This Initiative seeks to use summary health metrics that combine morbidity, mortality, and disability in the form of the disability-adjusted life year (DALY). As part of this work, cause of death data from vital registration (VR) systems were analyzed to explore mortality from causes that are largely but not exclusively foodborne. The results of these analyses are presented in this paper.
Causes of death of potentially foodborne origin
Cause of death
Bacterial intestinal infections
Bacterial foodborne intoxications
Protozoal intestinal diseases
Viral intestinal infections
Diarrhea and gastroenteritis of presumed infectious origin
Acute Hepatitis A
This study included all countries that reported at least 30 potentially FBD deaths per year and for which the WHO estimation of population coverage of VR data was at least 70%. These criteria enabled an analysis of 48 of the 96 countries included in the VR database (Table 2). For the data of these 48 countries, ICD codes were standardized to the most recent version, ICD-10, and any ill-defined or clearly miscoded CoDs were redistributed according to established practice .
Publicly available and validated databases, largely those from United Nations agencies, were accessed online and used to collect a wide range of explanatory variables. In addition to traditionally used health indicators, variables related to agriculture, environment, food consumption, environmental pollution, demographics, and trade were also collected and explored. Nonhealth indicators were selected based on their assumed relationship to food production, population behavior, food safety, and/or possible transmission routes of FBD. (A full list of variables collected is available upon request.)
Calculation of mortality rates
For each of the 48 countries, mortality figures for the 19 ICD-10 codes representing potentially FBD were combined and averaged across all available years (2000-2005). Mortality rates per 100,000 were calculated using 2005 population estimates from the 2006 revision of The World Population Prospects, a publication of the UN Department of Economic and Social Affairs. Population coverage of vital events registration differed among countries; for countries with less than 100% population coverage of data, mortality rates were proportionally increased such that final mortality rates approximated 100% population coverage.
Univariate linear regressions were performed to assess the ability of each explanatory variable to independently predict potentially FBD mortality. Each explanatory variable was regressed against the log mortality rate (log transformations were applied to satisfy linear assumptions in the regression; no outliers were noted). Variables were retained for further analyses if a relationship to mortality was indicated (α = 0.1) or if expert knowledge merited the variable's retention.
With a sample size of 48 countries and a large number of potential explanatory variables, it was necessary to reduce overall dimensionality during model building. In cases such as this, automated stepwise regression is sometimes used as a tool. However, an automated approach was not employed in this study due to its well-documented flaws  and the lack of available, complete explanatory data from all countries included in this analysis. Rather, a pragmatic approach was applied; model building proceeded via a modified nonautomated backward stepwise process. Beginning with a full model that contained all potential variables, variables were subsequently removed and replaced individually based on a consideration of standard regression outputs (t-tests, adjusted R2, and AIC metrics), colinearity, interaction, logic, and expert knowledge.
With a limited number of exceptions, the set of variables initially considered for inclusion in our multiple regression model were those satisfying two criteria: (1) the variable was identified as a potentially significant predictor of potentially FBD mortality based on univariate regressions (p-value < 0.1 α-level) and/or subject matter knowledge and (2) at least 40 out of the 48 countries were represented in the variable's dataset. The application of these criteria to the initial list of 91 variables narrowed the focus to 40 variables, 18 of which were traditionally used health indicators and 22 of which were designated nonhealth variables. (Additional file 1: Table A1.)
A multiple regression model was built in three steps to further aid in dimension reduction. Step 1: Multiple regression modeling was limited to the 18 traditionally used health indicators as predictors of potentially FBD mortality. Nine of these variables were identified as most likely predictive, based on t-tests using a cutoff value of α = 0.1, and retained for further modeling. Step 2: Similarly, multiple regression modeling was next limited to the 22 nonhealth variables. Eight of these variables were identified as predictive and retained for further modeling. Step 3: The 17 variables identified from the first two steps (nine traditionally used variables and eight nonhealth variables) were compiled for consideration in the final model. Modeling again proceeded according to the nonautomated approach as described above. After reduction, six variables were retained in the final model (see Additional file 1: Table A2). Residuals were generated and examined for any evidence of outliers or heteroscedasticity.
Countries included in the analysis by WHO regions and mortality level classifications
Eastern Mediterranean region
Serbia and Montenegro
Southeast Asia region
Western Pacific region
All statistical analyses and model building were performed using STATA 9 or STATA 10.
The final predictive model incorporating both health and nonhealth indicators of potentially FBD mortality
Irrigated land (percent)
A 1% increase in irrigated land predicts a 8.9% decrease in FBD mortality
Calorie supply from animal products, average (kcal/person)
An increase of 10 kcals per person predicts a 2% decrease in FBD mortality
Meat production (metric tons per 1,000 people)
An increase of 10 metric tons per 1,000 people predicts a 6% increase in FBD mortality
Literacy rate, adults (percent)
A 1% increase in literacy predicts a 4% decrease in FBD mortality
HIV/AIDS infection prevalence, adults (percent)
A 1% increase in HIV/AIDS predicts a 10% increase in FBD mortality
Deaths caused by diarrheal diseases, age < 5 (percent)
A 1% increase in diarrheal deaths under age 5 predicts a 6% increase in FBD mortality
From the original list of traditionally used health indicators, the fifteen not included in the final model were each added, in turn, as a seventh variable to the model. None of these fifteen variables merited inclusion based on statistical metrics. The effect of the addition of these variables on the original six variables in the model was varied. Consistently, the significance of the three food-related variables remained high, regardless of the addition of an additional heath variable. The one exception was the inclusion of Urban population to the model, which rendered only Meat production nonsignificant. Statistical metrics were considered for a model that included Urban population and excluded Meat production. However, doing so did not appear to improve the fit of model. The two health variables showed greater variability. The significance of adult HIV/AIDS prevalence and percent of deaths under age 5 caused by diarrheal diseases changed depending on which additional health covariate was added to the model. In no case did the alternative model provide statistical metrics that would suggest the rejection of the original model. This finding is supported by the predictive validity check (Additional file 2: Table A3) from which the final model proposed in Table 3 prevailed as the best predictive model in terms of both in-sample and out-of-sample predictive validity.
A country-level development indicator variable was also generated and considered for inclusion in the final model. Countries were divided based on WHO mortality-level classifications (see Table 2). Countries classified as A or B (lowest mortality; 42/48) were grouped, as were countries classified as C, D, or E (highest mortality; 6/48). This variable was not significant (p-value = 0.987, α-level) and was excluded from the model. Interaction terms between the indicator and each of the six covariates were also considered; there was no evidence for significance of these terms within the model.
FBD mortality rates as predicted by the final multiple regression model
Bosnia and Herzegovina
Serbia and Montenegro
Foodborne diseases continue to be a major contributor to morbidity and mortality worldwide . The wide range of illnesses and the presence of multiple routes of transmission make computing epidemiological estimates of the burden of FBD a challenging task. Regardless, reliable mortality estimates are crucial to the World Health Organization's strategy to reduce this burden. The present study, part of WHO's Initiative to Estimate the Global Burden of Foodborne Disease, sought to assess of the utility of nonhealth-related variables as predictors of potentially FBD mortality. Through the model building process, numerous insights were gained.
The most significant finding from this analysis is that so-called nonhealth variables improve the estimation of FBD mortality. Specifically, some nonhealth variables have proved to be more prominent predictors of potentially FBD mortality than some of the traditionally used health indictors. As demonstrated by this analysis, only two of the six variables included in the final model were directly related to health. The four nonhealth variables proved to be robust predictors of potentially FBD mortality, remaining significant even with the addition of a number of other traditionally used health indicators. Health indicators do not therefore appear to be confounding variables in this model. As such, we propose that nonhealth variables are likely providing unique and previously unmodeled information in FBD estimations.
While health indicators will always remain strong predictors of overall health outcomes, food-related variables may provide more specific insights into estimating outcomes of potentially FBD. Model building for this analysis included testing nonhealth variables from a variety of categories that were both food and nonfood related. However, all of the nonhealth variables that remained in the final model could be specifically linked to food production or consumption.
Furthermore, the nonhealth variables that appear in the final model make intuitive sense. For example, increased meat production predicts increased FBD mortality, which could be related to unsafe handling practices. A higher per capita average calorie supply from animal products predicts reduced mortality and could be an indication of improved nutrition. Likewise, increased percent irrigated land predicts reduced mortality, perhaps due a higher availability of safe food supplies. Case studies regarding these and other predictive variables may aid in the confirmation and discovery of contributors to the global mortality burden of FBD.
This analysis also provides support for the use of nonhealth variables in predicting potentially FBD mortality in countries lacking VR data. However, nonhealth models should not be used indiscriminately. In this analysis, 42 of the 48 countries were classified in the two lowest mortality levels (A or B). Additionally, the geographic distribution of these countries was skewed heavily in favor of the Americas and Europe regions. We suspect this grouping of countries to be somewhat self-selected. That is, countries with better reporting of FBD tend to be more developed and have demonstrated a lower incidence of FBD. The predictive capabilities of our model are therefore strongest for countries at a similar development level and with comparable FBD incidence. As such, we focused on predicting potentially FBD mortality in AmrA, AmrB, EurA, and EurB region countries.
Recent estimates from county-level studies of FBD mortality were consistent with predictions from our model. A 2011 study from the United States estimated the number of deaths from foodborne infections at approximately 2,612 each year (90% credible interval 1,723-3,819) . This compares favorably with the predicted number of deaths from our model at 3,058 deaths per year with an upper limit of 19,135 deaths per year (rate of 1.01 per 100,000, upper limit 6.32 per 100,000). In the Netherlands, a national study of the burden of FBD concluded that approximately 80 persons die from foodborne infections every year ; our model again predicts well, estimating 43 deaths per year with an upper limit of 326 persons per year. Given the known problems of miscoding and misclassification inherent to CoD registration data, particularly in diseases that are not frequently observed, we assume that VR has underestimated the true incidence of potentially FBD mortality.
The predictive accuracy of nonhealth variables was supported by the validity of the FBD estimates generated by our model. The results of the predictive validity check in Additional file 2: Table A3 were consistent with our final model choice. Specifically, we found that our final model had good out-of-sample predictive validity (i.e., our model yielded smaller prediction errors than did other candidate models when predicting data points outside the original data set). We suspect that factors such as geographic location, cultural practices, and economic status are likely to have a strong effect on which particular nonhealth indicators are predictive of FBD mortality. Therefore it is hoped that further model building efforts will focus on specific regions or development levels. However, the lack of available data in these areas provides an obstacle to this aim, further underscoring the need for improved data collection and reporting in countries with high mortality rates.
Foodborne diseases are a global problem, causing considerable morbidity and mortality annually. In this study, we have shown the potential value of using nonhealth variables alongside traditionally used health variables to predict potentially foodborne disease mortality at the country level. Our analyses were limited by the paucity of vital registration data in certain regions, but nevertheless demonstrate the predictive strength of nonhealth variables. Moreover, in addition to enhancing the specificity of cause of death modeling, nonhealth predictors could provide practical alternative measures of foodborne disease mortality in countries where traditionally used indicators may not be available or measured. Such analyses could provide valuable insight into the cause and source of disease at the country level, informing subsequent policies and interventions. As such, the expansion of the traditional health model to include nonhealth variables has the potential to be a powerful new tool in disease burden studies, including the current effort by WHO to estimate the global burden of foodborne disease.
The authors would like to acknowledge the contributions of Laura Boehm, Katherine Forsythe, Lauren Guengerich, Kelsey Harr, Brianna Hirst, and Logan Stuck during the initial stages of this work. Additionally, the authors thank Mrs. Doris Ma Fat for providing the WHO vital registration mortality data and the members of WHO Foodborne Disease Burden Epidemiology Reference Group (FERG) for valuable comments and inputs.
- World Health Organization: The global burden of disease: 2004 update. World Health Organization, Geneva, Switzerland; 2008.Google Scholar
- Scallan E, Hoekstra RM, Angulo FJ, Tauxe RV, Widdowson M-A, Roy SL, Jones JL, Griffin PM: Foodborne illnesses acquired in the United States--major pathogens. Emerg Infect Dis 2011,17(1):7-15. 10.3201/eid1701.09-1101p1View ArticlePubMedPubMed CentralGoogle Scholar
- Chapman G, Hansen KS, Jelsma J, Ndhlovu C, Piotii B, Byskov J, Vos T: The burden of disease in Zimbabwe in 1997 as measured by disability-adjusted life years lost. Trop Med Int Health 2006,11(5):660-671. 10.1111/j.1365-3156.2006.01601.xView ArticlePubMedGoogle Scholar
- Judd CM, McClelland GH: Data Analysis: A Model Comparison Approach. New York: Harcourt Brace Jovanovich; 1989.Google Scholar
- National Institute for Public Health and the Environment: Our food, our health: Healthy diet and safe food in the Netherlands. National Institute for Public Health and the Environment (RIVM), Bilthoven, The Netherlands; 2006.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.