 Research
 Open access
 Published:
Optimizing spatiotemporal correlation structures for modeling food security in Africa: a simulationbased investigation
BMC Bioinformatics volume 25, Article number: 168 (2024)
Abstract
This study investigates the impact of spatio temporal correlation using four spatiotemporal models: SpatioTemporal Poisson Linear Trend Model (SPLTM), Poisson Temporal Model (TMS), SpatioTemporal Poisson Anova Model (SPAM), and SpatioTemporal Poisson Separable Model (STSM) concerning food security and nutrition in Africa. Evaluating model goodness of fit using the Watanabe Akaike Information Criterion (WAIC) and assessing bias through root mean square error and mean absolute error values revealed a consistent monotonic pattern. SPLTM consistently demonstrates a propensity for overestimating food security, while TMS exhibits a diverse bias profile, shifting between overestimation and underestimation based on varying correlation settings. SPAM emerges as a beacon of reliability, showcasing minimal bias and WAIC across diverse scenarios, while STSM consistently underestimates food security, particularly in regions marked by low to moderate spatiotemporal correlation. SPAM consistently outperforms other models, making it a top choice for modeling food security and nutrition dynamics in Africa. This research highlights the impact of spatial and temporal correlations on food security and nutrition patterns and provides guidance for model selection and refinement. Researchers are encouraged to meticulously evaluate the biases and goodness of fit characteristics of models, ensuring their alignment with the specific attributes of their data and research goals. This knowledge empowers researchers to select models that offer reliability and consistency, enhancing the applicability of their findings.
Introduction
Food security is a critical concern, particularly in the context of Africa, where numerous countries face persistent challenges in ensuring adequate access to safe, nutritious, and affordable food for their populations. Understanding the complex dynamics of food security is essential for policymakers, researchers, and practitioners to develop effective interventions and strategies to address this pressing issue.
Traditionally, food security analysis has focused on temporal aspects, examining trends and changes in food availability, access, and utilization over time [1]. Nicholson, Stephens [1] reiterated the overemphasis on availability and accessibility indicators. In their study, they put forth four recommendations: avoiding confusion between "food availability" and "food security," consolidating food access indicators, assessing stable outcomes for food security indicators, and establishing empirical data linking results from agricultural systems models to outcomes related to food access. This emphasizes the necessity of including a broader spectrum of variables when evaluating food security and nutrition, particularly within the African context. However, it is increasingly recognized that food security is not solely determined by temporal factors but is also influenced by spatial dynamics [2]. Cooper, Brown [3] conducted a review of factors associated with food security, analyzing 16,152 abstracts from 3297 publications spanning the years 1975 to 2018. The review revealed that a majority of the studied publications focused solely on spatial effects rather than considering spatiotemporal impacts. This highlights the necessity for exploring spatiotemporal modeling.
To capture spatial–temporal variations and better comprehend the intricacies of food security in Africa, there is a growing need for spatiotemporal modeling approaches. Spatiotemporal models integrate both the spatial and temporal dimensions of data, enabling a more comprehensive understanding of the underlying processes and patterns that shape food security dynamics [4]. Spatiotemporal modeling offers several advantages as it explores spatial dependencies, assesses temporal dependencies, and captures persistence, trends, and cyclical patterns in food security outcomes. Moreover, it provides a framework to investigate interactions between spatial and temporal factors, recognizing complex feedback loops and dynamics within food systems [5].
Over the years, Bayesianbased spatiotemporal modeling has emerged as a significant area of interest due to its versatile applications across various disciplines, including ecology and epidemiology [6]. Aswi, Cramb [7] conducted a substantive case study on Dengue Incidence, exploring a range of Bayesian spatiotemporal models and evaluating their goodness of fit using the Watanabe Akaike Information Criterion (WAIC). Similar research has also been conducted by Ibeji, Mwambi [8], Wahyuni and Syam [9], and Yoo and Wikle [10], showcasing the growing interest in Bayesian spatiotemporal modeling in different contexts.
Previous studies [1, 3] have employed spatiotemporal modeling from a frequentist perspective, driven by limitations such as the absence of subjectivity, challenges in handling prior information, and issues with nonrobustness. Additionally, the difficulties in establishing efficient and reliable computing techniques for Intrinsic Conditional Autoregressive (ICAR) models, as outlined by Ver Hoef, Peterson [11], might have prompted the adoption of a spatiotemporal perspective, as seen in works by Fenta, Zewotir [12]and Sahu [13]. Are all spatiotemporal models robust to spatial and temporal correlations[14]? Hence, our goal was to conduct a comparative investigation into the impact of spatiotemporal correlation on four models [13] by evaluating their goodness of fit and biases in modeling food security in Africa. The Bayesian framework allows us to incorporate subjective insights, dealing adeptly with prior information and ensuring robustness—a notable improvement over the limitations identified in frequentist methodologies [11]
To assess goodness of fit, we employed the WatanabeAkaike Information Criterion (WAIC) for model comparison. This criterion, based on the entire posterior distribution, strikes a balance between model fit and complexity, addressing concerns of both underfitting and overfitting. WAIC is specifically designed to estimate outofsample predictive accuracy, making it a versatile and valuable tool for researchers working with various types of Bayesian models [14]. Furthermore, we examined the extent of bias within this category of models through the utilization of Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) measurements using Monte Carlo Simulations.
This comparison helps to understand how each model performs and whether certain models are more robust to variations in the spatiotemporal correlations. By achieving these objectives, the study sought to provide valuable insights into the suitability and performance of various spatiotemporal models in the context of food security analysis, especially considering the impact of different spatiotemporal correlations. The findings would contribute to a better understanding of the strengths and limitations of different modeling approaches and aid in making informed decisions when analyzing food security data in Africa.
Our methodology uniquely combines dimensionality reduction through PCA, Bayesian modeling, and simulationbased exploration, addressing the complexities of spatiotemporal correlation structures in the context of food security. This enables the exploration of diverse policy scenarios, ranging from weak to strong spatiotemporal correlation structures. By simulating these scenarios, we not only account for uncertainties but also offer a comprehensive understanding of potential impacts on the spatiotemporal dynamics of food security in Africa.
The simulationbased approach not only enhances our comprehension of intricate spatiotemporal correlation structures but also equips policymakers with a potent tool for scenario analysis. Policymakers can now evaluate the potential impact of various interventions, spanning from weak to strong spatiotemporal correlation structures. This insight proves invaluable for devising evidencebased policies that effectively address the dynamic challenges of food security in Africa. Given the intrinsic connection between food security, nutritional wellbeing, and community health, the simulation of diverse scenarios offers a nuanced understanding of potential outcomes. This, in turn, assists public health professionals in designing targeted interventions to enhance the nutritional landscape in the region. Our approach establishes a robust framework for modeling complex spatiotemporal correlation structures.
This paper is structured as follows: In the next section, we describe briefly the methodology of the spatiotemporal interaction models. In Sect. "Simulation study", we describe the Monte Carlo simulation scheme that was used to examine the performance of the four considered models. In particular, we report on goodness of fit (WAIC) and BIAS (RMSE and MAE). The simulation results are presented in Sect. "Results". Finally, in Sect. "Discussion", we summarize our findings and give final remarks.
Methodology and simulation
DATA
The Food and Agriculture Organization (FAO) of the United Nations plays a crucial role in supplying data and information to support the attainment of Sustainable Development Goal 2. This goal aims to eradicate hunger, food insecurity, and malnutrition on a global scale. The comprehensive dataset on food and nutrition security can be accessed on the FAO website, where you can also find metadata specifically pertaining to Africa. This metadata comprises variable definitions, data sources, years, and units. To address missing values within the dataset, we employed the missForest, which is random forest algorithm[15]. This research covered a 20year timeframe, ranging from 2000 to 2019. The study gathered a total of 1080 observations, representing data from all 54 countries on the continent. This dataset encompassed 42 variables, each dedicated to specific aspects of food security and nutrition in accordance with the FAO's definitions. The FAO employs the Food Insecurity Experience Scale (FIES) as a metric for measuring food insecurity. In order to maintain consistency and facilitate comparisons between countries, the FIES Survey Module is implemented among nationally representative samples of the adult population. This standardized approach ensures that data on food insecurity is collected in a uniform manner across different nations [16].
Afridi, Jabbar [16] highlighted the significance of utilizing a convergence of evidence strategy, which involves the use of multiple metrics, to identify the key factors associated with food security and nutrition. The initial set of 40 variables sourced from the FAO dataset underwent Principal Components Analysis (PCA). This method was employed to mitigate potential data loss and tackle the challenge of multicollinearity. It also aimed to overcome limitations associated with the conventional selection of based solely on experts' judgment regarding covariates or factors influencing food security and nutrition. PCA offers a more datadriven and objective approach to identifying relevant covariates (explanatory variables) for food security and nutrition. In the end, a total of ten factors were chosen as explanatory variables in the study. These factors included nutrient intake, average food supply, consumption status, childcare, caloric losses, environment, undernourishment, food or nutritional stability, adequate dietary supply, and newborn feeding practices. Together, these ten factors accounted for approximately 74.6% of the total variance present in the dataset [17].
The results of the Kaiser–Meyer–Olkin and Bartlett's Sphericity test values (with a KMO value of 0.729 and a p value of 0.00) indicated that PCA was a suitable approach, as suggested by [18]. Principal component extraction considered eigenvalues larger than one [19]. To ensure that each major component offered distinct information, we applied a VariMaxorthogonal rotation method, in line with Kaiser [19].
The first principal component (PC1) essentially represents nutrient intake and accounts for 19.26% of the total variation in the data. The second principal component (PC2) is associated with factors like food supply, food production, and dietary energy and protein supply. PC2 explains 14.96% of the variation and quantifies the average food supply in Africa. PC3 reflects African consumption levels based on indicators like Gross Domestic Product (GDP) per capita, childhood stunting and overweight, and adult obesity prevalence. It accounts for 10.96% of the variation.PC4, explaining 5.71% of the variance, pertains to childcare factors. PC5 (5.51%) relates to calorie loss, PC6 (4.47%) to environmental influences, PC7 (3.72%) to undernourishment, PC8 (3.43%) to food stability, PC9 (3.39%) to dietary adequacy, and PC10 (3.19%) to infant feeding practices.
The modeling formulation employed in this study is from the spatiotemporal framework. This framework integrates a spatial conditional autoregressive (CAR) prior and an autoregressive (AR) process. By incorporating both spatial and temporal interconnections, this method effectively captures the underlying patterns in the data. The CAR prior is employed to represent the spatial component, taking into account the spatial autocorrelation between neighboring locations. On the other hand, the AR process is utilized to model the temporal component, capturing the temporal autocorrelation over time. By combining these two components, the model can effectively analyze the spatiotemporal dynamics of the data [20].
In Bayesian statistics, when specifying a prior distribution for random variables exhibiting spatial autocorrelation, a common strategy involves combining a uniform prior that encompasses a wide range of values for the intercept (mean) with the intrinsic conditional autoregressive (ICAR) distribution. This approach allows for flexibility and accommodates the spatial autocorrelation structure in the data. Leroux, Lei [21] proposed a remarkable modification to conditional autoregressive (CAR) models that we incorporated into our work. Their suggestion involves using the equation \({\text{Q}}\left({\text{W}}, \rho \right)= \rho \left({{\varvec{W}}}_{{\varvec{d}}}  \mathbf{W}\right)+ \left(1  \rho \right)\mathbf{I}\), Where, \({{\varvec{W}}}_{{\varvec{d}}}\) represents a diagonal matrix of the weighted average value derived adjacency, and W is the ndimensional adjacency matrix. I represent the identity matrix of size\(N\times N\), and 1 is an \(N\times 1\) vector of ones. The parameter ρ, which controls the spatial correlation strength, takes on values between 0 and 1, inclusive. This results in a CAR distribution \({\varphi }_{i}\rho , {k}^{2},{\varphi }_{j},j\ne i \sim N\left(\frac{\rho \sum_{j=1}^{n}{w}_{ij}{\varphi }_{j}}{\rho \sum_{j=1}^{n}{w}_{ij}+1\rho } ,\frac{{k}^{2}}{\uprho \sum_{j=1}^{n}{w}_{ij}+1\rho }\right)\). In this context, \({w}_{ij}\) refers to the element in the nth dimensional adjacency matrix W corresponding to the ith row and jth column. The random effects have a spatial variance of \({k}^{2}\) and autocorrelation of \(\rho\)
The spatial weight matrix related to the units \(i\) and \(j\) is represented by each item \({w}_{ij} \in\) W. The constituent of \({w}_{ij}\) is (\((i,j)\), which remains the neighborhood matrix with \(54\times 54\) dimension. The matrix's nonzero entries reveal whether the two places are neighbors. Typically, the weighted matrix is written as:
The dataset's spatial autocorrelation is verified by applying Moran's I = \(\frac{n}{{s}_{0}}\frac{\sum_{ij}({{w}_{ij}(x}_{i}\mu )({x}_{j}\mu ))}{{\sum }_{i}{({x}_{i}\mu )}^{2}}\), 'n' represents the count of points under investigation,\({x}_{i}{x}_{j}\) denote the observed values at two distinct points, \(\mu\) denote the expected value of '\(x\) and \({w}_{ij}\) signify the spatial weight element. Moran's I range [− 1,1], large values of the relevant metrics are close to other large value clusters when the value is 1, while large values are close to low values when the value of the Moran's I is 1[12, 13]
In the model,\({Y}_{it}\) represented the number of severe food insecurity individuals in a country \(i\) (\(i=\mathrm{1,2},..., 54)\) during year \(t(t=\mathrm{1,2},\dots , 20)\) and \({n}_{it}\) denoted the population size of country i at time t. The design matrix X represented the food security and nutrition covariates components) are represented by the design matrix \({\varvec{X}}\), and \({\varvec{\beta}}\) was a vector comprising related fixed effects parameters:
By introducing the spatiotemporal random effect \({\psi }_{it}\) to Eq. (1), we obtained the following expression:
To capture the various spatiotemporal features (spatiotemporal interactions) of the multiple aspects of the spatiotemporal random effect \({( v}_{it})\), we decomposed it, leading to the four models in the sequel.
Spatiotmporal poisson linear trend model (SPLTM)
where \(\overline{t }\) is calculated as \(\frac{T+1}{2}\) Here \({\omega }_{1}\) and \({\omega }_{2}\) represent the overall intercept and slope parameters, respectively, and are assigned a flat prior distribution. On the other hand, the incremental trend(slope) and intercept parameters for the ith region, denoted as \({a}_{i}\) and \({b}_{i}\)
respectively, are assigned a conditional autoregressive (CAR) prior distribution with different values of ρ and \({k}^{2}.\) Specifically, the parameters parameters \({{\varvec{a}}=(a}_{1},\dots ,{a}_{n})\) and \({{\varvec{b}}=(b}_{i},\dots , {b}_{n})\) follow the CAR prior distributions \({\varvec{a}}\sim CAR({\varvec{a}}{\rho }_{int},{{k}^{2}}_{int},{\varvec{W}})\) and \({\varvec{b}}\sim CAR({\varvec{b}}{\rho }_{slo},{{k}^{2}}_{slo},{\varvec{W}})\) respectively. Similarly, \({{k}^{2}}_{slo},{{and k}^{2}}_{int}\) represent the variance parameters. The parameters \({\rho }_{slo}{, {\text{and}} \rho }_{int}\) are assigned independent uniform prior distributions in the unit interval (0, 1), indicating that their values can range between 0 and 1. On the other hand, the variance parameters, \({{k}^{2}}_{slo},{{and k}^{2}}_{int}\) follow the inverse gamma prior distribution. In this model, important trends for severe food insecurity are accounted for and their level of significance is accessed.
Spatiotemporal poisson anova model ( SPAM).
Knorr‐Held [22] proposed a method for analyzing data that incorporates the interaction between space and time by using a model that factors spatial and temporal main effects based on the Analysis of Variance. Equation 4 is derived from Eq. 2 to represent the spatial and temporal effect with attraction.
The three sets of parameters indicated by \({a}_{i}\), \({b}_{t}\) and \({c}_{it}\) in Eq. 4 are all considered as random effects, each having its own unique probability distribution.
\({c}_{it }\sim N\left(0,{{k}^{2}}_{I}\right),i=1,\dots ,n, t=1,\dots ,T\),
Here Z is the T × T matrix that represents temporal adjacency, where each element \({z}_{ij}\) takes a value of 1 if the absolute difference between i and j is equal to 1, and 0 otherwise. It is assumed that the interaction effect \({(c}_{it })\) is independent for all values of i and t.The parameters \({\rho }_{s}\) and \({\rho }_{T}\) are assigned uniform priors with a range of values between 0 and 1, similar to before(Sect. "DATA"), while the variance parameters \({{k}^{2}}_{s}\),\({{k}^{2}}_{T}\), and \({{k}^{2}}_{I}\) are assigned inverse gamma priors.
Spatiotemporal poisson separable model(STSM)
The separable model, which includes an overall time trend with temporalspecific spatial effects, is an alternative model to the one presented in Sect. "Spatiotemporal poisson anova model ( SPAM).". In this case, independent conditional autoregressive (CAR) models are assigned to the spatial effects \({{\varvec{a}}=(a}_{1t},\dots ,{a}_{nt})\) for each \(t=1,\dots ,T\) and to b = (\({{\text{b}}}_{1}\),…, \({{\text{b}}}_{T}\)), as shown in the following equation.
Z refers to the same definition as previously provided in Sect. "Spatiotemporal poisson anova model (SPAM).". The variance parameters \({{k}^{2}}_{t}\), where t ranges from 1 to T, and \({k}^{2}\), are assigned inverse gamma prior distributions, similar to the previous model. The parameters \({\rho }_{s}\) and \({\rho }_{T}\) are assigned independent uniform prior distributions with a range of values between 0 and 1.
Poisson temporal model for spatiotemporal effect (TMS)
The postulation is made that a particular scenario within the framework of the separable model, as discussed in Sect. "Spatiotemporal poisson separable model(STSM)", corresponds to a temporal autoregressive model with a lag of one. In this special case, the parameter \({b}_{t}\) is equal to zero for all values of t, resulting in the simplification as \({v}_{it}={a}_{it}\), here \({b}_{t}=0\) aimed at the entire \(t\) and
The precision matrix Q(W, \({\varvec{\rho}}\) S) with spatial dependence is defined in Eq. (1), where the temporal autocorrelation is caused by the mean \(\rho T,{{\varvec{a}}}_{t1}\). The prior distributions are assumed to remain unchanged.
The chain models
After decomposing the spatiotemporal random effect \({v}_{it}\) and adding it to the linear predictor \({\eta }_{it}\) as shown in Eqs. 2, four different models were obtained: STLTM, SPAM, STSM, and TMS, represented by Eqs. 6–9, respectively:
Markov chain monte carlo: hamiltonian monte carlo
To construct a Markov chain that converges to the target distribution, Markov Chain Monte Carlo (MCMC) techniques are employed. A Markov chain is constructed by iteratively sampling \({\mathbf{q}}^{(h+1)}\) from the conditional distribution \(p\left(\mathbf{q}{\mathbf{q}}^{h}\right)\) at each time point\(h>0\). The process begins with an initial value\({\mathbf{q}}^{(0)}\).Importantly, the value \({\mathbf{q}}^{(h+1)}\) at each step depends solely on the current value\({\mathbf{q}}^{(k)}\), and is independent of its more distant past values\({\mathbf{q}}^{(h1)}\),\({\mathbf{q}}^{(h2)}\),…, …, \({\mathbf{q}}^{(0)}\) . This specific property, where each step in the chain only depends on the current state and not on previous states beyond the immediate predecessor, is a fundamental requirement for the chain \({\mathbf{q}}^{(h)}\) to be considered a firstorder Markov Chain.
The primary challenge in Bayesian modeling convergence lies in constructing a Markov chain that possesses essential properties such as stationarity, irreducibility, aperiodicity, and ergodicity, while also having a stationary distribution equal to the target posterior distribution \(\pi \left({\varvec{q}}{\varvec{y}}\right)\). Although there are alternative methods like the Metropolis–Hastings algorithm and the Gibbs sampler, the Hamiltonian Monte Carlo (HMC) is widely regarded as the superior choice. This is because HMC excels in exploring regions of high posterior probability and concurrently employs a Metropolis–Hastings step to address the issue of oversampling peaks while neglecting low probability areas, For a comprehensive understanding of HMC, numerous excellent resources are available, including Lambert [23]. So, executing Hamiltonian Monte Carlo (HMC) for a parameter vector \({\varvec{\theta}}\) requires following the procedures delineated in Lambert's 2018 publication, which are presented in a sequel.

1.
Choose an initial location, \({\mathbf{q}}^{(0)}\), randomly from an initial proposal distribution.

2.
During iteration h, generate an initial momentum (m) randomly from a proposal distribution, such as \({\varvec{m}}\sim N({\varvec{\mu}},\Sigma\)).

3.
Progress the current point\({\varvec{q}}\), m to \({{\varvec{q}}}^{*}\), \({\mathbf{m}}^{*}\) by executing L steps of the leapfrog algorithm, as implemented by the NoUTurn Sampler (NUTS) [24]. The pair \({{\varvec{q}}}^{*}\), \({\mathbf{m}}^{*}\) represents the updated position of the parameter and its momentum after completing the L leapfrog steps.

4.
Compute the Metropolis acceptance probability, denoted as \(\alpha ({\varvec{q}},g),\) for the proposal \(g=({{\varvec{q}}}^{*}\),\({\mathbf{m}}^{*})\) and current point \({\varvec{q}}=\left({\varvec{q}},{\varvec{m}}\right)\) based on the target density \(\pi \left({\varvec{q}},{\varvec{m}}{\varvec{y}}\right)\) where
$${\text{log}}\pi \left({\varvec{q}},{\varvec{m}}{\varvec{y}}\right)\propto {\text{log}}\left(\pi \left({\varvec{q}}{\varvec{y}}\right)\right)+\frac{1}{2}{{\varvec{m}}}{\prime}{\Sigma }^{1}{\varvec{m}}.$$ 
5.
Generate a random number u from a uniform distribution U(0,1), and set \({\mathbf{q}}^{(k+1)}={{\varvec{q}}}^{*}\) if \(u<\) \(\alpha ({\varvec{q}},g)\) orthewise, set \({\mathbf{q}}^{(h+1)}={\varvec{q}}\)
We employed the MCMC algorithm that generates 10,000 samples from the posterior distribution with N = 120,000, burn.in = 20,000, and thin = 10.
Simulation study
The main aim of our study was to explore how the spatial and temporal correlation structures impact the spatiotemporal dynamics of food security and nutrition in Africa. To achieve this, we conducted a Monte Carlo simulation involving 1,000 random samples (replications) with varying spatial and temporal correlations (0.05, 0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75, 0.85, 0.95) for longitudinal twenty repeated measures in each country. Every aspect of the analysis, which encompassed tasks such as data generation, simulation, and the estimation of model parameters, was executed utilizing the R programming language (R 4.3.1). For each scenario or combination of parameter values, we executed a total of 1000 simulation runs.
The combination of PCA and VIF allows for a more holistic evaluation of multicollinearity, validating the efficacy of the PCA dimensionality reduction while ensuring the dependability of regression model outcomes. This integrated approach fortifies the robustness of statistical analyses, contributing to a more comprehensive understanding of the interrelationships among variables, as evidenced in prior studies [25, 26]. From Fig. 1, we observed that all ten components had a VIF value of 1. The numerical value of 1 for VIF observed for these 10 components confirms that there is no inflation in the percentage of variance (i.e., the standard error squared) for each coefficient. Therefore, the application of PCA is justified.
Datagenerating process
During the data generation process (response variable), we specified four versions of spatiotemporal models to investigate the effects of spatiotemporal correlation using randomly generated (artificial) dependent variables and the 10 explanatory variables (covariates) of food security in Africa identified when the 40 indicators from FAO was subjected to PCA. The spatial units were the 54 countries of Africa. These ten components (covariates) were obtained from the 40 variables provided by the FAO dataset after subjecting them to Principal Component Analysis (PCA) as described in Sect. "Methodology and simulation".
We devised the following iterative algorithm, which was utilized to induce the desired spatiotemporal correlation when generating the values of the dependent variables.
The first version.

a.
Define the spatial correlation (0.05, 0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75, 0.85, 0.95) and temporal correlation value (0.05, 0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75, 0.85, 0.95) combinations, one hundred pairs.

b.
Use information in (a) to generate two correlation matrices: the 'spatial correlation matrix' and 'temporal correlation matrix'. These matrices have the specified correlation values and dimensions based on 'num_countries' and 'num_observations', respectively.

c.
Use the information in (b) to generate the noise using multivariate normal distributions with means of 0 and covariance matrices 'spatial_correlation_matrix' and 'temporal_correlation_matrix', in(b). Here the spatial and temporal data is generated by adding the spatial noise to the coordinates and replicating the temporal noise.

d.
Use information (c) to calculate a latent variable lamda = exp(spatial datax + temporal data). Where 'spatial datax' represents the xcoordinate of the spatial data for each country. 'temporal data': This represents the temporal noise generated using multivariate normal distributions.

e.
Generate the observed response variable \(y\) where each count is expected to be around the corresponding 'lambda' value. Where \(y > 0\)

f.
Fit spatiotemporal interaction model for the four (SPLTM, SPAM, STSM, TMS) models specified in Sect. "The Chain Models"

g.
Estimate the parameters, and compute the bias and goodnessoffit measures for evaluation

h.
Perform steps (a) to (g) iteratively for a total of 1,000 repetitions, each time varying the spatial correlation and temporal correlation values. After each iteration, calculate the relevant parameters accordingly.
For model validation:

I.
All other procedures from (a) to (e) remained unchanged, as outlined in the first version.

II.
To enable the computation of validation statistics, designate 70 percent of the data for training and 30 percent for testing.

III.
Fit spatiotemporal interaction model for the SPAM model

IV.
Estimate the parameters, and compute the coverage percentage for validation

V.
Perform steps (I) to (IV) iteratively for a total of 1,000 repetitions, each time varying the spatial correlation and temporal correlation values. After each iteration, calculate the relevant parameters accordingly.
Statistical analyses in simulated datasets
To determine the best predictive model from the four available options, we employed a range of statistical metrics to evaluate and compare their performance. Specifically, we computed the WAIC, RMSE, and MAE for each of these models. The model with the lowest statistical metric values indicates superior predictive ability, striking a balance between model fit and complexity.
Goodness of fit test: WAIC
The optimal fit among the four competing models was evaluated using WAIC [27, 28], To calculate WAIC, the loglikelihood of each data point is computed based on the model's parameters. The total loglikelihood is then adjusted for the variance of these loglikelihoods, with a correction term for effective parameters.
Here, "lpd" represents the log pointwise predictive density, which is calculated as the summation from i equals 1 to n of the natural logarithm of the density function \(p(\left({y}_{i}\theta \right)\).In this context, \(p(\left({y}_{i}\theta \right)\) represents the density of observing the data point \({y}_{i}\) given the model parameters \(\theta\). The effective number of parameters, denoted as "pWAIC," is computed as the summation from i equals 1 to n of the variance of the log predictive densities, \({\text{Var}}[{\text{log}}\left\{p\left(\left({y}_{i}\theta \right)\right\}\right].\) We assessed the goodness of fit by calculating the average modelbased WAIC across the 1000 replications.
BIAS: RMSE and MAE
Let \({\widehat{y}}_{i}\) denote the estimated severe food security value for the \(i\)th observation using a specific method, while \({y}_{i}\) represents the observed value for the same \(i\)th observation. The variable n represents the number of data points in each replication. The RMSE is calculated by taking the square root of the average of the squared differences between the actual values and the estimated parameters generated by the model in each iteration. The RMSE can be expressed as follows:
The MAE is determined by taking the mean of the absolute differences between the actual values and the estimated parameters obtained from the model in each iteration, as described by Eq. 12.
Both RMSE and MAE for each replication are arithmetically averaged throughout the 1000 replication.
Validation
This involved the validation of the best model among the four competing models. We computed the coverage percentage (CP) for the 95% predictive intervals using the formula \({\text{CP}}=100\frac{1}{{\text{k}}}{\sum }_{{\text{i}}=1}^{{\text{k}}}{\text{I}}\left({{\text{L}}}_{{\text{i}}}\le {{\text{y}}}_{{\text{i}}}\le {{\text{U}}}_{{\text{i}}}\right)\), where \({{\text{y}}}_{{\text{i}}}\) represents the observed value for i = 1,…,k and \({({\text{L}}}_{{\text{i}}},\) \({{\text{U}}}_{{\text{i}}})\) represents the 100(1 − α)% predictive interval for predicting \({{\text{y}}}_{{\text{i}}}\). In this calculation, we employed the indicator function I(·). It's important to note that we used 70% of the data as the training set and the remaining 30% as the test set [29, 30].
Results
We present the simulation results based on the types of estimated statistical metrics explained in the earlier section. Our primary focus is on evaluating the performance of spatiotemporal models for food security concerning Africa. These metrics are visually represented in a matrix plot and Table.
For the goodness of fit test, it is represented in Table 1. In Table 1 “SP” represents spatial correlation, “TM” denotes temporal correlation, or the correlation over time. The cells in the table are coloredcoded to easily identify the best and worst results. Red cells indicate higher or worse values, meaning WAIC. Yellow cells point to lower or better values, signifying less WAIC. The color scheme facilitates interpretation by signaling which areas or spatiotemporal correlation pairs exhibited the strongest or weakest WAIC metric.
Matrix plots are a crucial tool for presenting the relationships and patterns discovered in our analysis of the four competing spatiotemporal models. They provided a visual narrative that illustrates how different spatial and temporal correlation structures impact food security and nutrition modeling for Africa. In these matrix plots, the cells are colorcoded to represent the values of the parameters, including RMSE, and MAE, for each of the four models under consideration. lower RMSE and MAE values (deeper colors) indicate better model performance. This visualization method helps us to grasp and compare the biases of these models across various spatial and temporal correlation scenarios.
Goodness of fit assessment
The WAIC analysis provided insights into how sensitive the different models were to the specific combinations of spatial and temporal correlation included in the modelling. Regarding WAIC, Table 1 illustrates that the spatiotemporal Poisson Linear Trend Model (SPLTM) exhibits values ranging from 6000 to 20,000, with the model exhibiting the highest cluster of WAIC values when compared to the four competing models. The highest WAIC value observed was 18,450.05, while the lowest value recorded was 6529.104. Notably, lower WAIC values were observed when both spatial and temporal correlations were weak to moderate, falling within the range of 0.05 to 0.5. This suggests that the model performs better or exhibits more favorable characteristics under these conditions. Conversely, higher WAIC values were observed in scenarios characterized by strong temporal and spatial correlations. This implies that the model's performance does not improve with higher values of the parameters, specifically in cases of elevated spatial and temporal correlation. This behavior indicates that the model is particularly sensitive to increased spatial and temporal correlation, particularly when there are high levels of spatiotemporal correlation.
For spatiotemporal Poisson Anova Model (SPAM), the WAIC values are distributed within the range of 3800 to 5000. The lowest WAIC value observed was 3893.43, while the highest recorded WAIC value was 4917.99 (Table 1). Among the four models compared, SPAM consistently obtained the lowest WAIC scores in the vast majority of simulated scenarios concerning diverse spatial and temporal dependency combinations. This indicates that SPAM excelled in capturing and modeling the underlying spatial and temporal correlations within the data, as determined by the WAIC metric for model selection and performance. In the case of the Poisson Temporal Model (TMS) for the Spatiotemporal correlation effect, as shown in Table 1, the WAIC values also fall within the range of 4000 to 5000 for all 100 pairs of spatiotemporal correlation. The lowest WAIC value was 4060.257, and the highest WAIC value was 4756.981.
It is worth noting an interesting pattern observed in the results: lower WAIC values, indicating better model fit, were seen when the spatial and temporal correlation values included in the simulations ranged from low to moderately strong, specifically when both types of correlations took on values between 0.05 and 0.75. This pattern held for both the SPAM and TMS models. In other words, these two models performed best at capturing the relationships in the data when the spatial and temporal dependencies were in the moderate rather than extremely low or high ranges. This provides useful insight into the conditions under which these types of spatiotemporal models are better able to model correlation structures. WAIC values deteriorated when the data exhibited strong spatiotemporal correlations, as opposed to when these correlations were at lower or moderate levels. This implies that although the SPAM and TMS models exhibited good performance when dealing with modest spatial and temporal dependencies, their capability to effectively capture patterns in the data decreased as the strength of the dependencies across geographical areas and over time increased substantially.
Concerning the performance of the spatiotemporal Poisson Separable Model (STSM), a notable observation is that the performance of the STSM is particularly responsive to variations in both spatial and temporal correlation. The STSM demonstrates relatively strong performance in scenarios characterized by low to moderate levels of spatial and temporal correlation, ranging from 0.05 to 0.55. However, it exhibits higher WAIC values when there is a combination of high spatial correlation with lower temporal correlation. The WAIC values in this case fall within the range of 4100 to 5000, with the highest WAIC value recorded at 4999.791 and the lowest at 4150.11.
Bias evaluation
Figure 2 depicts the distribution of bias levels, as indicated by RMSE, for all four models. Notably, for both the SPAM and TMS models, the values fall within the range of 0.7 to 1.2, and interestingly, both models exhibited similar distribution patterns. The parameter values vary according to different levels of spatiotemporal correlation, which ranged from 0.7 to 1.00. Notably, there was only one observation exceeding 1.00, occurring when the temporal correlation was 0.25 and the spatial correlation was 0.05. According to Fig. 2, the SPLTM exhibits the highest range of RMSE values, spanning from 1.2 to 16.94. In contrast, the STSM shows a narrower range, with RMSE values ranging from 1.8 to 0.89.
Regarding the MAE results presented in Fig. 3, the SPLTM model exhibited the widest range of values, spanning from 8.9 to 0.7. In contrast, the SPAM model displayed narrower intervals, with values ranging from 0.8 to 0.5. For the TMS model, the MAE intervals were 0.99 to 0.5 (Fig. 3). Interestingly, the majority of MAE metrics for SPAM fell within the range of 0.6 to 0.75, except in situations when spatial correlation was 0.65 and temporal correlation was 0.05 (Fig. 3). Similar patterns were observed for cases of low spatial (0.05) and temporal (0.25) correlation, as well as spatial (0.45) and temporal (0.85) correlation. These distributions suggest that the SPAM model may not perform well under such conditions, but it excels when there is a strong correlation in both spatial (0.85) and temporal (0.95) (Fig. 3). For the TMS model, lower MAE metrics were clustered when spatial correlation ranged from 0.45 to 0.95 for nearly all temporal correlations (Fig. 3). Regarding the STSM model, the MAE metrics ranged from 0.65 to 0.85, with the highest values observed in scenarios where both spatial and temporal correlations were very weak (0.05) or when temporal correlation was strong (0.95) with weak spatial correlation (0.05) (Fig. 3).
Coverage percentagebased validation of the selected model: SPAM
Based on the principles of statistical learning, the SPAM model is considered the best among the four competing models. Therefore, it is crucial to validate its performance using the 70% train and 30% test split from a machine learning perspective. This validation will help assess the model's capability for handling unseen data across various spatiotemporal correlation scenarios especially concerning Africa’s food security modelling. We noticed a homogeneous distribution of the coverage percentage across almost all variations in spatiotemporal correlation concerning the SPAM. The coverage percentages were noted to vary between 96.8% and 80.6%. Notably, the SPAM model consistently achieved higher coverage percentages (> 90%) for nearly all levels of spatiotemporal correlation, with one exception: when the temporal correlation was high (0.85) and the spatial correlation was weak (0.05) Fig. 4. It achieved its lowest coverage percentage at 80 (Fig. 4), although there were instances (17 observations) where the coverage percentage ranged from 85 to 90.
Discussion
As we concentrate on evaluating the performance of spatiotemporal models under different spatiotemporal correlation scenarios, we summarize our findings in a table and matrix plots. When necessary, we contextualize our results within the existing literature.
In the table and matrix plot displaying WAIC, RMSE, and MAE values for our four models (SPLTM, SPAM, STSM, and TMS), we have identified intriguing and consistent monotonic patterns that offer valuable insights into their performance and bias characteristics. In essence, these observations underline the sensitivity of all four models to variations in spatiotemporal correlation. On average, each of these models gradually converges towards the true values as spatiotemporal correlation increases, which is reflected by the corresponding increase in WAIC values. These findings strongly support the notion that an augmentation in spatiotemporal correlation levels can indeed impact the performance of these models [31]. Our results reinforced the significance of selecting models that align with the specific patterns evident in the data.
Our study highlights that the SPLTM consistently exhibited WAIC values in these comparisons. This recurrent pattern of SPLTM holding the highest WAIC values raises significant concerns about its suitability for modeling food security and nutrition in the African context. In stark contrast, our findings emphasize the remarkable consistency of the SPAM model. SPAM consistently outperformed the other models, as evidenced by its persistently lower WAIC values when compared to SPLTM, STSM, and TMS for almost all scenarios under study. This robust and unwavering performance strongly suggests that SPAM stands out as exceptionally wellsuited for the complex task of modeling the dynamic interplay of food security and nutrition in the African context. The reliability demonstrated by SPAM makes it a top choice for researchers seeking to comprehend and predict the intricate spatiotemporal patterns within this critical domain. These findings align with existing research [13] on the behavior of spatiotemporal interaction models and reinforce the superiority of the SPAM model in this context.
Our analysis of bias using RMSE and MAE unearthed interesting patterns that shed light on the behavior of the four spatiotemporal models. The SPLTM model consistently exhibited relatively larger bias values across a wide spectrum of spatiotemporal correlation values. This consistent pattern of overestimation in food security across diverse spatiotemporal scenarios underscores the model's limitations and its tendency to inflate predictions when necessary. In contrast, the Poisson Temporal Model (TMS) demonstrated relatively more variable bias patterns. In scenarios with low spatial and temporal correlation, it consistently recorded substantial bias values, indicating a potential tendency to overestimate food security. However, in high correlation settings, TMS exhibited a notable shift towards underestimation, showcasing the influence of varying spatiotemporal correlation on its predictive behavior.
The SPAM emerged as the model with minimal bias across most scenarios. This consistent performance highlighted SPAM's reliability in accurately modeling food security dynamics, irrespective of the degree of spatiotemporal correlation. Conversely, the STSM exhibited a notably high bias, consistently underestimating food security. This bias pattern was particularly prominent in areas characterized by low to moderate spatial and temporal correlation, emphasizing the model's limitations in accurately capturing the dynamics of food security in such settings. Our bias evaluation not only highlights the relative performance of these models but also provides researchers with critical insights for model selection based on their specific research questions and the characteristics of their spatiotemporal data. Understanding these bias patterns is paramount to making informed decisions when choosing the most appropriate model for accurate and reliable predictions in the domain of food security.
In light of these findings, we recommend that model selection should be a balanced consideration of both biasness and goodness of fit metrics. A model that exhibits superior goodness of fit but introduces significant bias may not provide accurate predictions, and conversely, a model with low bias but poor goodness of fit may fail to capture underlying spatiotemporal dynamics. Researchers should aim to strike a harmonious equilibrium between these two aspects to select models that align with the specific characteristics of their data and research objectives.
The statistical insights gained from our research have the potential to extend their impact beyond the realm of spatiotemporal modeling. Other disciplines grappling with complex and correlated data, such as epidemiology, environmental science, and economics, can draw upon the methodology and lessons learned in this study. our research fosters interdisciplinary collaboration and opens new avenues for the analysis of intricate spatiotemporal patterns in diverse fields. This crossdisciplinary applicability enhances the relevance and broader impact of our findings.
While the integration of PCA and Bayesian modeling contributes to the efficiency of our methodology, it is crucial to recognize potential computational complexities, particularly with larger datasets. Given that our study focused on 20 temporal observations for each subject (country), future research endeavors could delve into computational optimization techniques. These techniques would be beneficial for handling datasets of varying sizes, ranging from smaller (e.g., 5, 10, 15) to larger (e.g., 25, 30,35) temporal observations. Implementing such strategies would help maintain computational efficiency without compromising model accuracy, addressing scalability concerns associated with different data sizes. The Bayesian modeling approach relies on specific assumptions, and deviations from these assumptions can influence the robustness of the findings. For example, the assumed prior distributions may introduce subjectivity. Future studies could incorporate sensitivity analyses to evaluate the impact of different priors on model outcomes. This would contribute to enhancing the credibility and reliability of our Bayesian approach.
To the best of our understanding, there hasn't been prior research assessing the performance of the four spatiotemporal models (SPLTM, SPAM, STSM, and TMS) in terms of their sensitivity to spatiotemporal correlations using Monte Carlo simulations. It's important to note that, as with any simulation study, our findings are bound by the specific scenarios and data employed in our simulations. Hence, while our analysis of bias and goodness of fit provides valuable insights into the capabilities and constraints of spatiotemporal modeling approaches, it's important to note that these findings might not be directly transferrable to scenarios or contexts that were not explicitly investigated in our study.
Conclusion
The primary objective of our study was to investigate the influence of spatial and temporal correlation structures on the spatiotemporal patterns of food security and nutrition (FSN) in Africa. We conducted this exploration within the spatiotemporal framework, considering four Bayesian Poisson models: SPLTM, SPAM, STSM, and TSM. Monotonic patterns became evident when we subjected the four competing models to varying spatiotemporal correlations. the SPAM model consistently exhibited the lowest RMSE, MAE, and WAIC values across all spatiotemporal correlations. This notable consistency implies a high level of robustness in the SPAM model's performance. These results strongly suggest that SPAM is a reliable and solid choice for capturing the dynamics of food security. The results of our study provide practical guidance for researchers and practitioners engaged in spatiotemporal modeling. By identifying the strengths and weaknesses of each model under varying degrees of spatiotemporal correlation, we offer a valuable framework for informed model selection. Researchers should strive to find a balanced equilibrium between these two factors (goodness of fit and bias metrics) when choosing models that are in line with their data's unique features and research goals. This knowledge empowers researchers to select models that offer reliability and consistency, enhancing the applicability of their findings.
Availability of data and materials
The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.
Abbreviations
 SPLTM:

Spatiotemporal poisson linear trend model
 TMS:

Poisson temporal model
 SPAM:

Spatiotemporal poisson anova model
 STSM:

Spatiotemporal poisson separable model
 WAIC:

Watanabe akaike information criterion
 RMSE:

Root mean square error
 MAE:

Mean absolute error
 MCMC:

Markov chain monte carlo
 FAO:

Food and agriculture organization
 PCA:

Principal components analysis
 PC:

Principal component
 GDP:

Gross domestic product
 CAR:

Conditional autoregressive
 ICAR:

Intrinsic conditional autoregressive
 HMC:

Hamiltonian monte carlo
 SP:

Spatial correlation
 TM:

Temporal correlation
References
Nicholson CF, et al. Food security outcomes in agricultural systems models: current status and recommended improvements. Agric Syst. 2021;188:103028.
Zhai H, et al. Understanding spatiotemporal patterns of land use/land cover change under urbanization in Wuhan, China, 2000–2019. Remote Sens. 2021;13(16):3331.
Cooper MW, et al. Text mining the food security literature reveals substantial spatial bias and thematic broadening over time. Glob Food Sec. 2020;26:100392.
Donatelli C, et al. Statistical detection of spatiotemporal patterns in the salinity field within an intertidal basin. Estuaries Coasts. 2022;45(8):2345–61.
Masolele RN, et al. Spatial and temporal deep learning methods for deriving landuse following deforestation: a pantropical case study using Landsat time series. Remote Sens Environ. 2021;264:112600.
SantosFernandez, E., et al., SSNbayes: An R package for Bayesian spatiotemporal modelling on stream networks. arXiv preprint arXiv:2202.07166, 2022.
Aswi A, Sukarna SC, Mengersen K. Effects of climatic factors on dengue incidence: a comparison of bayesian spatiotemporal models. J Phys Conf Ser. 2021;1863:012050. https://doi.org/10.1088/17426596/1863/1/012050.
Ibeji JU, Mwambi H, Iddrisu AK. Bayesian spatiotemporal modelling and mapping of malaria and anaemia among children between 0 and 59 months in Nigeria. Malar J. 2022;21(1):1–12.
Wahyuni, M.S. and R. Syam. Comparison of Bayesian Spatiotemporal Models of Tuberculosis in Makassar, Indonesia. in Journal of Physics: Conference Series. 2021. IOP Publishing.
Yoo, M. and C.K. Wikle, A bayesian spatiotemporal level set dynamic model and application to fire front propagation. arXiv preprint arXiv:2210.14978, 2022.
Ver Hoef JM, et al. Spatial autoregressive models for statistical inference from ecological data. Ecol Monogr. 2018;88(1):36–59.
Fenta HM, Zewotir T, Muluneh EK. Space–time dynamics regression models to assess variations of composite index for anthropometric failure across the administrative zones in Ethiopia. BMC Public Health. 2022;22(1):1–11.
Sahu SK. Bayesian modeling of spatiotemporal data with R. Boca Raton: Chapman and Hall/CRC; 2021.
Devine, S., Falk, C.F., Fujimoto, K.A., Comparing the Accuracy of Three Predictive Information Criteria for Bayesian Linear Multilevel Model Selection. 2023
Bofa A, Zewotir T. Filling the gap in food and nutrition security data: What imputation method is best for Africa’s food and nutrition security? Lith J Stat. 2022;61:16–31.
Afridi GS, et al. An analysis of food insecurity in pakistan: prevalence of undernourishment (PoU) and food insecurity experience scale (FIES). J Appl Econ Bus Stud. 2021;5(1):175–90.
Bofa A, Zewotir T. The determinants of severe food insecurity in Africa using the longitudinal generalized Poisson mixed model. Lith J Stat. 2023;62:1–19.
Bartlett MS. A note on the multiplying factors for various χ2 approximations. J R Stat Soc Ser B Stat Method. 1954;16(2):296–8. https://doi.org/10.1111/j.25176161.1954.tb00174.x.
Kaiser HF. The application of electronic computers to factor analysis. Educ Psychol Measur. 1960;20(1):141–51.
Jalilian A, Mateu J. A hierarchical spatiotemporal model to analyze relative risk variations of COVID19: a focus on Spain, Italy and Germany. Stoch Env Res Risk Assess. 2021;35:797–812.
Leroux BG, Lei X, Breslow N. Estimation of disease rates in small areas: a new mixed model for spatial dependence. In: Elizabeth Halloran M, Berry D, editors. Statistical models in epidemiology, the environment, and clinical trials. New York, NY: Springer New York; 2000. p. 179–91. https://doi.org/10.1007/9781461212843_4.
KnorrHeld L. Bayesian modelling of inseparable spacetime variation in disease risk. Stat Med. 2000;19(17–18):2555–67.
Lambert, B., A student's guide to Bayesian statistics. A Student's Guide to Bayesian Statistics, 2018: p. 1–520.
Hoffman MD, Gelman A. The NoUTurn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo. J Mach Learn Res. 2014;15(1):1593–623.
Kianfar N, et al. Spatiotemporal modeling of COVID19 prevalence and mortality using artificial neural network algorithms. Spatial Spatiotemporal Epidemiol. 2022;40:100471.
Kianfar N, Mesgari MS. GISbased spatiotemporal analysis and modeling of COVID19 incidence rates in Europe. Spatial Spatiotemporal Epidemiol. 2022;41:100498.
Gelman A, Carlin JB, Stern HS, Dunson DB, Vehtari A, Rubin DB. Bayesian data analysis. Chapman and Hall/CRC; 2013. https://doi.org/10.1201/b16018.
Watanabe S, Opper M. Asymptotic equivalence of Bayes cross validation and widely applicable information criterion in singular learning theory. J Mach Learn Res. 2010;11(12):3571–94.
Cheng FY, et al. Using machine learning to predict ICU transfer in hospitalized COVID19 patients. J Clin Med. 2020;9(6):1668.
Mangalathu S, et al. Datadriven machinelearningbased seismic failure mode identification of reinforced concrete shear walls. Eng Struct. 2020;208:110331.
Perera S. A Bayesian spatial interaction framework for optimal facility location in urban environments. Warwick: University of Warwick; 2022.
Acknowledgements
We gratefully acknowledge the Food and Agriculture Organization (FAO) for providing valuable data on food security, which greatly contributed to the research presented in this paper. Their dedication to collecting and disseminating reliable information is essential in addressing global food security challenges.
Funding
This project work was carried out without any external funding(public, commercial, or notforprofit sectors).
Author information
Authors and Affiliations
Contributions
The authors, AB and TZ, were responsible for the conception and design of the experiments, analysis, and interpretation of the data, and writing of the paper.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
The Suite of Food Security Indicators presents a fundamental collection of indicators focused on food security. This selection is based on the recommendations of experts who convened during the Committee on World Food Security (CFS) Round Table on hunger measurement at FAO headquarters in September 2011. These indicators were chosen to encompass various dimensions of food insecurity. The selection process considered expert insights and the availability of data that permits crossregional and temporal comparisons. The FAO Statistical Quality Assurance Framework's Principle 10 underscores the strict confidentiality of data subject to national policies, particularly regarding individuals, entities, and small aggregates. Such data are exclusively used for statistical purposes or in accordance with legal mandates.
Consent for publication
Not applicable' for that section.
Competing interests
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Bofa, A., Zewotir, T. Optimizing spatiotemporal correlation structures for modeling food security in Africa: a simulationbased investigation. BMC Bioinformatics 25, 168 (2024). https://doi.org/10.1186/s1285902405791w
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s1285902405791w