Defining and characterizing Aflatoxin contamination risk areas for corn in Georgia, USA: Adjusting for collinearity and spatial correlation

Abstract

Aflatoxin is a carcinogenic toxin to humans and animals produced by mold fungi in staple crops. Surveys of Aflatoxin are expensive, and the results are usually not available for implementing within season mitigation strategies. Identification of high and low risk areas and years is essential to reduce the number of samples analyzed for Aflatoxin concentration. Previously a risk factors approach was developed to determine county level Aflatoxin contamination risk in southern Georgia, but Aflatoxin concentrations and risk factor data were not analyzed simultaneously and all risk factors had equal weight which is unrealistic. In the current paper we propose a regression approach to overcome these problems. Spatial Poisson profile regression identified clusters of counties which have similar Aflatoxin risk and risk factor profiles, whilst explicitly taking into account multicollinearity in the risk factor data and spatial autocorrelation in the Aflatoxin data. This approach allows examination of the utility of different highly correlated variables including remotely sensed data that could give information at the sub-county level. The results identify plausible clusters compared to previous work but also give the relative importance of the risk factors associated with those clusters. The approach also helps show that some factors like well-drained soil behave differently from expectations and irrigation data is not useful.

Introduction

Aflatoxin is a carcinogenic toxin produced by the mold fungi Aspergillus flavus and Aspergillus parasiticus. As these fungi can infect crops that are some of the dominant staples in different parts of the world, infection is considered a worldwide problem Brenneman et al., 1993, Liu and Wu, 2010, Wang et al., 2010. Ingestion of infected grain/food can cause esophageal and liver cancer in humans, livestock and poultry Ghasemi-Kebria et al., 2013, Liu and Wu, 2010. Consequently, the U.S. Food and Drug Administration (FDA) has set a limit of 20 ppb, total Aflatoxin, for use of corn, peanut, cottonseed meal, and other feeds/feed ingredients intended for animal consumption, particularly by immature animals. There is also a limit of 100 ppb restricting use of corn and peanut products intended for breeding beef cattle, swine, or mature poultry (FDA, 2015). However, in less developed countries where such administrative standards do not exist, there are 16–31 times more deaths from liver cancer and this has been attributed, at least in part, to Aflatoxin contamination of food (Liu and Wu, 2010).

Even in areas where governmental standards for maximum contamination levels exist, like the USA, food scares still occur Garland and Reagor, 2001, Newman et al., 2007 because testing is limited as it is time-consuming, expensive and requires several grain samples (Papadoyannis et al., 1990). There is also the problem that different methods of determining Aflatoxin have differing accuracies, detection limits and advantages and disadvantages (Wacoo et al., 2014). These challenges of accurate Aflatoxin assessment make regular monitoring infeasible with current technology, which results in the scarce availability of data in the study region, southern Georgia (GA), as well as inconsistent data availability over the study period.

An important way to reduce the risk of contaminated food going undetected and keeping the cost of Aflatoxin measurement to a minimum is to identify years and locations at the highest risk of Aflatoxin contamination. Also, as Aflatoxin levels are usually determined at harvest, no in-season mitigation strategies are possible. However, if years and areas with different levels of risk are identified, strategies that prevent, reduce or manage Aflatoxin levels in crops can be employed at key periods in the growing season. During planting, the seeding rate can be varied between different risk zones, or more resistant corn hybrids can be planted in the high risk areas. During crop growth, irrigation, pest and nutrient management strategies could be altered between the different risk zones. Finally, at harvest, areas with different levels of risk can be harvested separately.

Previous studies have shown that infection of corn by A. flavus or A. parasiticus is linked with high temperatures, drought and high net evaporation Guo et al., 2008, Horn et al., 2014, Payne and Widstrom, 1992 so it is associated with particular climatic areas (Abbas et al., 2007) and soil types (Palumbo et al., 2010). In the southern states of the USA, summer crop corn is highly susceptible to Aflatoxin contamination (Widstrom et al., 1996) due to high temperatures and rainfall variability, along with light textured soils that compound crop water stress. Lack of irrigation infrastructure in some areas also means that crop water stress cannot be easily relieved (Brenneman et al., 1993). Using logistic regression, Salvacion et al. (2011) showed that in southern GA, Aflatoxin risk level changed based on deviations of June maximum temperature and rainfall levels from climatic normals. Damianidis et al. (2015) also found that the risk of contamination changes with corn hybrid, soil type and the weather conditions before and after the mid-silk growth stage (usually in June in south eastern USA). In addition, when the Agricultural Reference Index for Drought (ARID) (Fraisse et al., 2006) was included in determining Aflatoxin risk, a 0.1 increase of in-field drought, as quantified by ARID, around the mid-silk period, increased the probability of Aflatoxin exceeding the 20 ppb FDA threshold (Damianidis et al., 2015).

Recently, Kerry et al. (2017b) used a risk factors approach to identify years and areas at high risk of Aflatoxin contamination. They used an additive indicator approach based on key thresholds in variables that previous research Abbas et al., 2007, Brenneman et al., 1993, Palumbo et al., 2010, Salvacion et al., 2011, Widstrom et al., 1996 had identified as important to Aflatoxin contamination risk. This approach, however, had the limiting assumption that each variable has equal weight in determining overall risk. To overcome this problem, we consider regression models and assess the relative importance of different risk factors (explanatory variables or covariates) for predicting Aflatoxin levels (response variable).

While climate information available from weather stations are clearly helpful in determining drought prone areas and years, it is only through the use of interpolated data that information on drought can be extracted for areas that are not close to weather stations. One way to gain more complete spatial and temporal information on actual drought stress in crops is from remotely sensed imagery. Normalized difference vegetation index (NDVI) data from images indicate the degree of crop greenness and have been used to indicate variations in crop health and yield within fields in response to differences in water supply and nutrients Hatfield and Prueger, 2010, Wang et al., 2016. Thermal InfraRed (IR) data have also been used to indicate drought stress Jones et al., 2009, Sepúlveda-Reyes et al., 2016. In a preliminary study based on one high risk year, Kerry et al. (2017a) found an 82% agreement in the cells that were identified as high and low risk when NDVI and Thermal IR were used as opposed to rainfall and maximum temperature data. It is important to further determine if these variables can be used to indicate drought stress in other years and thus, Aflatoxin contamination risk level can be identified within growing seasons and at a finer spatial scale. Thus, our goals include determining if NDVI and Thermal IR data from remotely sensed imagery are useful predictors of Aflatoxin levels and could therefore be used to determine risk at finer scales.

A challenge associated with the representation of the risk of Aflatoxin contamination levels and building a regression model is that the Aflatoxin measurements were obtained from corn grain samples, but their specific location other than the county of origin is unknown, so they are essentially aggregated data represented at the county level. Other information, such as weather, soil, or remotely sensed data are available for different representations, such as point locations for weather stations, polygon maps of soil type and raster data from imagery but all were aggregated to the county level. Regardless of the unit of analysis, aggregating point data to areas is likely to induce spatial autocorrelation. Also, Kerry et al. (2017b) found spatial structure in Poisson variograms of Aflatoxin data showing that spatial autocorrelation is present in these data. The presence of spatial autocorrelation in standard regression models needs to be accounted for, otherwise the uncertainty intervals of errors will be under-estimated.

The final challenge with applying a multiple regression approach to these data is that the known risk factors for Aflatoxin contamination in corn crops are highly correlated causing problems of multicollinearity. For example, the correlation of irrigation and excessively-drained soil class is as high as 0.60 and the two remotely sensed variables (NDVI and Thermal IR) were negatively correlated (−0.71) with one another. The recent development of spatial Profile regression Molitor et al., 2010, Molitor et al., 2011, Papathomas et al., 2011 is a promising alternative to address both the presence of spatial autocorrelation in Aflatoxin contamination risk and collinearity among the explanatory variables. Profile regression is a Bayesian statistical approach that assesses the link between potentially collinear variables and a response variable through cluster membership (Liverani et al., 2016). Its application in environmental and social epidemiology Hastie et al., 2013, Coker et al., 2016 and infectious disease studies (Shekhar et al., 2017) has demonstrated its potential to be useful in the characterization of Aflatoxin contamination risk areas.

In this paper, we identified areas with different levels of Aflatoxin contamination risk by applying Bayesian spatial profile regression to both Aflatoxin measurements and risk factors simultaneously, while accounting for problems of collinearity among risk factors and the spatially correlated structure present in Aflatoxin risk. Profile regression was performed for two time frames, high risk years identified by Kerry et al. (2017b) and the entire study period, respectively. The latter allows comparison with the high risk regions of Aflatoxin contamination derived from Poisson Kriging (Kerry et al., 2017b) purely based on the primary data from corn plant samples. Lastly, we estimated the uncertainty associated with contamination risk for each cluster using both Aflatoxin measurements and risk factors within a Bayesian framework by deriving the posterior distribution of random effects for each cluster. In applying Bayesian spatial profile regression to these data we aimed to achieve two goals. First, to determine if this more statistically principled approach gives different or additional insights compared to previous methodologies used and standard Poisson regression both of which have limiting assumptions given the characteristics of the data. The second goal was to identify which of several highly correlated continuous and dichotomous variables were the most useful in predicting risk. The fact that profile regression deals with multicollinearity in the risk factors made this a possible avenue of investigation. As part of this second aim, remotely sensed data were included in the list of continuous variables to determine if they are useful at predicting risk and could therefore be used in future studies to identify risk at the sub-county scale.