Identifying age-common and age-specific factors of Plasmodium infection in Nigerian children under five: Application of a cluster-aware multistage selection framework to the 2018 Nigeria Demographic and Health Survey

Abstract

Despite developmental heterogeneity, children under five are often analyzed as a single group, obscuring age-related differences in Plasmodium infection. Using the 2018 Nigeria Demographic and Health Survey data, we analyzed 2,914 younger children (6–23.9 months) and 5,553 older children (24–59 months) to identify age-common and age-specific factors. We applied a Cluster-aware Multistage Selection (CMS) framework integrating penalized regression, interaction testing, and bidirectional selection while accounting for sampling weights, clustering, and stratification. Plasmodium prevalence was 37.3% in younger children and 49.4% in older children. Each 10-percentage-point increase in community-level livestock and agricultural land ownership was associated with 2.5% (PR = 1.025; 95% CI: 1.011–1.039) and 4.6% (PR = 1.046; 95% CI: 1.029–1.062) higher prevalence, respectively. Socioeconomic disadvantage was associated with higher prevalence. Children in the lowest wealth quintile had 80.7% higher prevalence than those in the richest quintile (PR = 1.807; 95% CI: 1.467–2.224), and children whose mothers had no formal education had 55.8% higher prevalence than those with higher education (PR = 1.558; 95% CI: 1.225–1.980). Severe maternal anemia was associated with 22.5% higher prevalence (PR = 1.225; 95% CI: 1.012–1.483). Stunting and household insecticide-treated net (ITN) ownership were associated with 8.0% (PR = 1.080; 95% CI: 1.024–1.139) and 10.4% (PR = 1.104; 95% CI: 1.017–1.198) higher prevalence, respectively. Conversely, sleeping under an ITN (PR = 0.916; 95% CI: 0.858–0.977), breastfeeding (PR = 0.845; 95% CI: 0.764–0.934), maternal internet use (PR = 0.638; 95% CI: 0.475–0.857), and overweight-for-height (PR = 0.761; 95% CI: 0.631–0.917) were associated with lower prevalence. Age-specific associations were observed only in younger children, including paternal lack of education (PR = 1.301; 95% CI: 1.025–1.652) and rural residence (PR = 1.462; 95% CI: 1.266–1.688). These findings support age-tailored Plasmodium prevention and highlight the utility of CMS for complex population data.

Introduction

Malaria remains a major cause of morbidity and mortality in Sub-Saharan Africa, with Nigeria accounting for 27% of global cases and 31% of malaria deaths in 2022 [1]. Children under five experience a disproportionately high Plasmodium infection burden that leads to malaria, with substantial contributions to severe anemia, neurocognitive impairment, and years of life lost [2,3]. Although usually analyzed as a single age group, children under five are developmentally heterogeneous. Infants and toddlers under two years—the “first 1,000 days”—experience rapid growth, immune immaturity, and nutritional vulnerability, whereas children aged two to five years gain greater independence, mobility, and environmental exposure [4–7]. Understanding age-specific risk factors for Plasmodium infection is critical for optimizing the targeting and effectiveness of prevention strategies.

The Demographic and Health Survey (DHS) provides nationally representative Plasmodium biomarker data through stratified, cluster-based sampling, primarily targeting children under five due to their high Plasmodium infection burden [8]. Most DHS-based Plasmodium infection studies pool children under five or compare age groups only descriptively [9–12]. These analyses often include many covariates to improve explanatory power, but this may increase collinearity, overfitting, and confounding [9,10,12]. Although methods such as purposeful selection and least absolute shrinkage and selection operator (LASSO) can mitigate these issues, they are rarely implemented in ways that fully account for the DHS’s stratified, cluster-based design [13–18].

We aimed to identify age-common and age-specific factors of Plasmodium infection among younger (6–23.9 months) and older (24–59 months) children using the 2018 Nigeria DHS. To do so, we developed the Cluster-aware Multistage Selection (CMS) framework, a three-stage, survey-design–aligned approach that evaluates both main effects and age interactions. Guided by prior literature, we hypothesized that socioeconomic factors would show age-common associations, whereas nutritional, environmental, and caregiver-related factors might vary by age due to differences in developmental vulnerability and exposure patterns [5–7,12,19,20].

Materials and methods

Data sources

We used de-identified data from the 2018 Nigeria DHS, conducted from August 14 to December 29, 2018. The DHS employs stratified two-stage cluster sampling to select approximately 5,000–30,000 households per country [8,15]. Our analysis included 8,447 children aged 6–59 months with available Plasmodium test results, corresponding to the population systematically sampled for Plasmodium testing in the DHS. Data were obtained from the DHS Program website ([href:https://dhsprogram.com/data/]https://dhsprogram.com/data/).

Outcome measure: Plasmodium infection

As the DHS does not measure clinical malaria, we used Plasmodium infection as the outcome. Children were classified as infected if either light microscopy or rapid diagnostic tests (RDT) was positive, reflecting the high sensitivity of combined testing in high-endemic settings [21]. Records with one test missing and the other negative were excluded to improve specificity.

Independent covariates

We selected 56 candidate covariates (excluding age) based on a conceptual directed acyclic graph (DAG; Fig 1), comprising 13 community-level and 43 individual-level variables. The DAG was specified to represent hypothesized antecedent factors; thus we did not depict potential reverse pathways from Plasmodium infection to covariates. We calculated community-level covariates as the proportion of children with each characteristic within each DHS primary sampling unit (PSU) and assigned these values to all individuals in that cluster. We used PSUs as the community unit because each corresponds to a census enumeration area, a well-defined geographic proxy for local communities in national surveys [8]. We extracted individual-level covariates from children’s DHS responses including 8 sociodemographic, 11 child and maternal nutrition, 14 household structure and physical-environment, and 10 preventive and health-related behavior variables. We excluded variables with extensive missingness (≥10%) or non-informative responses unless specified by DHS guidelines. We modeled age both as a continuous and a dichotomous variable (6–23.9 vs. 24–59 months). Covariate definitions are provided in S1 Table.

Fig 1. Directed acyclic graph (DAG) of factors associated with Plasmodium infection among children aged 6–59 months in Nigeria, 2018 Demographic and Health Survey.

Conceptual DAG summarizing known and hypothesized pathways linking sociodemographic, environmental, nutritional, and behavioral factors to Plasmodium infection among children under five in Nigeria. The diagram integrates evidence from epidemiology, environmental health, child nutrition, and health-seeking behavior research and motivates identification of age-common and age-specific associations in the present analysis [5,9–11,14,19,20]. Abbreviations: DAG = directed acyclic graph; WASH = water, sanitation and hygiene; ITN = insecticide-treated net.

Statistical analysis

We conducted all analyses in R (version 4.5.1) within RStudio. We implemented the Cluster-aware Multistage Selection (CMS) framework, a three-stage, survey-aligned variable-selection approach incorporating DHS sampling weights, PSU clustering, and stratification to identify age-common and age-specific factors. With no a priori thresholds, we tested multiple tuning combinations and selected the most accurate and parsimonious configuration.

1. Step 1: survey-weighted LASSO regression covariate screening.

We fit a survey-weighted LASSO including all covariates and their interactions with the age-band indicator (6–23.9 vs. 24–59 months), leaving age and DAG-defined structural covariates unpenalized. Binary variables contributed one main and one interaction term; multi-category variables contributed these terms for each dummy indicator (a binary variable representing each response). We selected a parent covariate if any associated coefficient was non-zero at the chosen penalty parameter λ. To assess stability, we repeated LASSO across 200 stratified PSU-bootstrap replicates, advancing covariates selected above a tunable proportion threshold. Details on LASSO and bootstrapping are provided in S1 Text.

2. Step 2: Age-common and age-specific candidate identification.

All retained covariates were entered into a survey-weighted generalized linear model including their main and interaction terms. We screened age-interaction blocks using Wald tests with tunable p-value thresholds and re-evaluated them across the same 200 stratified PSU-bootstrap replicates. We classified covariates passing the tunable threshold as candidate age-specific factors; others as age-common factors.

Step 3: Bidirectional selection for model enrichment.

We fitted a survey-weighted quasi-Poisson model with a log link including main-effect terms for all candidate age-common covariates and both main and interaction terms for candidate age-specific covariates. Backward elimination began from the full model. At each iteration, we marked terms not meeting any retention criterion for removal, and removed the term whose exclusion resulted in the smallest increase (or largest decrease) in the simplified quasi-likelihood information criterion (QICu; Pan 2001), ensuring that the terms contributing least to model fit were eliminated first [22].

Following backward elimination, we conducted a forward reassessment to re-evaluate previously excluded terms, including Step 1 main and interaction terms and Step 2 interaction terms of candidate age-common covariates. We added each candidate term individually, and among those satisfying any retention criterion, we reintroduced the term yielding the greatest improvement (or smallest decrease) in QICu. QICu ties in both directions were resolved by p-value, then alphabetically.

Retention criteria were: 1) statistical significance of Wald block tests; 2) improved or minimally worsened model fit based on dQICu; 3) evidence of confounding, defined as notable mean or maximum changes in other covariates’ PRs when a term was removed. Thresholds for each criterion (p-value, dQICu tolerance, PR-change cutoffs) were tunable. Main-effect components of retained interaction terms and all DAG-defined structural covariates were always preserved in the final model.

3. Model diagnostics and selection of tuning parameters.

We evaluated all combinations of tunable thresholds across Steps 1–3. In step 1 we compared the minimum‐error and one–standard error rules for selecting the LASSO penalty λ, and bootstrap selection thresholds of 30% vs. 60%. In step 2 we evaluated age-interaction screening p-value cutoffs of 0.20, 0.15, and 0.10, alongside the same 30% and 60% bootstrap thresholds. Finally, in step 3 we assessed retention p-value thresholds of 0.15, 0.10, and 0.05 (always stricter than the paired Step 2 threshold), dQICu tolerances of −2, 0, and 2, and confounding thresholds based on relative mean and maximum PR changes of 10%/20% or 20%/40%.

For each configuration, we computed Pearson dispersion, QICu, calibration correlation, Rao–Scott F statistic, and total parameters (main and interaction terms counted separately). We used Brier scores obtained from cross-validation preserving DHS stratification and clustering as the primary ranking metric [23]. We grouped models into 0.1% bins above the minimum Brier score, with the k-th bin defined as $Brie{r}_{min}(\text{1}+\text{0}.\text{0}\text{0}\text{1}·k)$. We then ranked models by Brier‐score bins and, within bins, by parsimony based on the number of parameters.

We summarized retention patterns in a heatmap showing each covariate’s selection frequency and effect direction.

4. Inference.

We used the top-ranked configuration to fit the final survey-weighted quasi-Poisson model. We assessed model fit using residual–fitted plots, and calibration by comparing survey-weighted observed prevalence with predicted probabilities across deciles.

For each covariate, the exponentiated main-effect coefficient represented the prevalence ratio (PR) among younger children. When an age interaction was present, the PR among older children was obtained by multiplying the exponentiated main-effect and interaction coefficients. Therefore, if the interaction coefficient was zero, the PR was identical across age groups, and the covariate was treated as age-common.

Results

We included a total of 2,914 younger children and 5,533 older children in the analysis. Among younger children, 37.3% tested positive for Plasmodium infection, compared with 49.4% of older children (Fig 2). Overall, 45.1% of children under five were positive (S1 Fig).

Fig 2. Prevalence of Plasmodium infection among younger (6–23.9 months) and older (24–59 months) children in Nigeria, 2018 Demographic and Health Survey.

Weighted prevalence of Plasmodium infection among children aged 6–59 months, stratified by younger and older age groups. Children testing positive by either rapid diagnostic test or light microscopy were classified as Plasmodium-positive; those negative on both tests were classified as Plasmodium-negative. Estimates incorporate sampling weights, clustering, and geographic stratification.

Weighted frequencies and proportions for individual-level covariates are presented in Table 1A, and median community-level proportions with their interquartile ranges (25th–75th percentiles) are shown in Table 1B. Full covariate definitions and recoding procedures are provided in S1 Table.

Table 1. A. Weighted distributions of individual-level covariates among younger (6–23.9 months) and older (24–59 months) children in Nigeria, 2018 Demographic and Health Survey. Weighted frequencies and proportions of each covariate in each age group are presented. Weights were calculated with clustering and geographical stratification according to the Demographic and Health Survey (DHS) methodologies. Each covariate was defined following the Demographic and Health Surveys (DHS) Standard Recode Manual for DHS-7 and the Guide to DHS Statistics, with additional refinements applied at the researchers’ discretion. Anthropometric indicators were coded according to World Health Organization definitions. Full details of covariate definitions and recoding procedures are provided in S1 Table. BMI = body mass index; ITN = insecticide-treated nets. 1B. Distributions of community-level covariates among younger (6–23.9 months) and older (24–59 months) children in Nigeria, 2018 Demographic and Health Survey. Distributions of community-level covariates across primary sampling units (PSU), each corresponding to a census enumeration area and serving as the operational community unit in this analysis. For each PSU, the proportion of children aged 6–59 months exhibiting each characteristic was calculated; median and interquartile ranges of these PSU-level proportions are reported.

Covariate name	Younger children N (%)	Older children N (%)
Socioeconomic and demographic
Female sex	1403 (47.2%)	2703 (49.1%)
Household wealth quintile
Richest	592 (19.9%)	999 (18.1%)
Richer	597 (20.1%)	1118 (20.3%)
Middle	614 (20.7%)	1126 (20.4%)
Poorer	563 (18.9%)	1118 (20.3%)
Poorest	605 (20.4%)	1146 (20.8%)
Education level of the mother
Higher	288 (9.7%)	485 (8.8%)
Secondary	1097 (36.9%)	1800 (32.7%)
Primary	449 (15.1%)	942 (17.1%)
None	1138 (38.3%)	2279 (41.4%)
Education level of the father
Higher	452 (15.2%)	783 (14.2%)
Secondary	1102 (37.1%)	1954 (35.5%)
Primary	411 (13.8%)	803 (14.6%)
None	1007 (33.9%)	1966 (35.7%)
Occupation of the mother
Office work	1574 (53%)	3050 (55.4%)
Manual labor	124 (4.2%)	238 (4.3%)
None	1274 (42.9%)	2218 (40.3%)
Occupation of the father
Office work	1208 (40.6%)	2191 (39.8%)
Manual labor	1524 (51.3%)	2900 (52.7%)
None	240 (8.1%)	415 (7.5%)
Area of residence
North Central	398 (13.4%)	782 (14.2%)
North East	463 (15.6%)	827 (15%)
North West	871 (29.3%)	1699 (30.9%)
South South or South East	739 (24.9%)	1211 (22%)
South West	502 (16.9%)	988 (17.9%)
Household lives in rural regions	1634 (55%)	3111 (56.5%)
Child & Maternal nutrition
Fulfills minimum dietary diversity	677 (22.8%)	0 (0%)
Fulfills minimum meal frequency	1230 (41.4%)	0 (0%)
Fulfills minimum acceptable diet	314 (10.6%)	0 (0%)
Given iron supplement within 7 days	548 (18.4%)	1038 (18.8%)
Child given vitamin A supplement within 6 months	1381 (46.5%)	2569 (46.6%)
Anemia level of the mother
Normal	1279 (43%)	2278 (41.4%)
Mild	797 (26.8%)	1483 (26.9%)
Moderate	862 (29%)	1643 (29.8%)
Severe	34 (1.1%)	102 (1.8%)
BMI status of the mother
Normal	1858 (62.5%)	3395 (61.7%)
Very thin	79 (2.7%)	148 (2.7%)
Thin	248 (8.3%)	349 (6.3%)
Overweight	527 (17.7%)	1055 (19.2%)
Obese	260 (8.8%)	560 (10.2%)
Height-for-age
With stunting	981 (33%)	2349 (42.7%)
Normal	1991 (67%)	3157 (57.3%)
Weight-for-height
With wasting	349 (11.8%)	229 (4.2%)
Normal	2543 (85.6%)	5166 (93.8%)
Overweight	80 (2.7%)	112 (2%)
Weight-for-age
Underweight	682 (23%)	1248 (22.7%)
Normal	2261 (76.1%)	4238 (77%)
Overweight	28 (1%)	20 (0.4%)
Currently breastfed	1984 (66.8%)	90 (1.6%)
Household structure and physical environment
Number of household members
2-4 members	907 (30.5%)	1230 (22.3%)
5-7 members	1143 (38.5%)	2364 (42.9%)
8 or more members	922 (31%)	1912 (34.7%)
Access to electricity	1698 (57.1%)	3061 (55.6%)
House has finished flooring material	2117 (71.2%)	3889 (70.6%)
2 or more rooms for sleeping	2101 (70.7%)	4057 (73.7%)
Cooks inside	2012 (67.7%)	3725 (67.6%)
Owns agricultural land	1870 (62.9%)	3541 (64.3%)
Owns livestock	1438 (48.4%)	2955 (53.7%)
Access to improved sanitation facility	1649 (55.5%)	2875 (52.2%)
Shares toilet with others	896 (30.1%)	1661 (30.2%)
Place of handwashing
At home: fixed	736 (24.8%)	1398 (25.4%)
At home: mobile	1673 (56.3%)	3092 (56.1%)
Not at home	563 (18.9%)	1016 (18.5%)
Has soap and water when washing hands	884 (29.8%)	1618 (29.4%)
Ash, mud, sand present at handwashing place	30 (1%)	60 (1.1%)
Access to improved water source	1865 (62.7%)	3466 (62.9%)
Owns water source	861 (29%)	1530 (27.8%)
Prevention and health-related behaviors
Household owns ITN	2077 (69.9%)	3789 (68.8%)
Household owns more than 1 ITN per 2 members	566 (19%)	1042 (18.9%)
Slept under an ITN last night	1631 (54.9%)	2754 (50%)
Mother uses the internet at least once a week	281 (9.5%)	455 (8.3%)
Mother exposed to mass media at least once a week	1278 (43%)	2328 (42.3%)
Visited health facility in the last 12 months	1700 (57.2%)	2856 (51.9%)
Mother had problems in accessing healthcare	1578 (53.1%)	2918 (53%)
Appropriate disposal of children's stool	1650 (55.5%)	1269 (23%)
Mother received drugs for intestinal parasites during pregnancy	513 (17.3%)	459 (8.3%)
Place of delivery
Own or other's home	1619 (54.5%)	3103 (56.3%)
Government facility	844 (28.4%)	1570 (28.5%)
Private facility	510 (17.1%)	833 (15.1%)
1B
Covariate	Median proportion (25th-75th percentile)
Socioeconomic and demographic
Household is poor or poorer	20.0% (0.0%–80.0%)
Mother received primary education or higher	88.9% (28.6%–100.0%)
Father received primary education or higher	80.0% (50.0%–100.0%)
Mother is employed	60.0% (33.3%–90.0%)
Father is employed	100.0% (85.7%–100.0%)
Household structure and physical environment
Household owns livestock	50.0% (0.0%–83.3%)
Household owns agricultural land	75.0% (28.6%–100.0%)
House has finished flooring material	100.0% (50.0%–100.0%)
Access to electricity	71.4% (0.0%–100.0%)
Access to improved sanitation facility	57.1% (0.0%–100.0%)
Access to improved water source	75.0% (33.3%–100.0%)
Prevention and health-related behaviors
Mother exposed to mass media at least once a week	40.0% (9.1%–75.0%)
Household owns ITN	72.7% (42.9%–100.0%)

A total of 144 tuning-parameter combinations were evaluated (S2 Table). Covariate-selection patterns were highly stable across models, with consistent retention of significant covariates, no reversals in effect direction, and non-significant covariates appearing mainly in lower-ranked models, indicating robustness of the selection procedure (S2 Fig). The highest-ranking model used the minimum-error rule for λ in Step 1, a stratified PSU-level bootstrap threshold of 30% in Steps 1 and 2, and a Step 2 screening threshold of p < 0.20. In Step 3, covariates were retained if they met at least one of the following: p < 0.05, model improvement by dQICu tolerance ≤2, or confounding PR changes ≥20% (mean) and ≥40% (maximum) among all other covariates. This model showed moderate overdispersion (Pearson dispersion = 3.75), no systematic residual–fitted patterns, and excellent calibration (calibration coefficient = 0.989), with close agreement between observed and predicted prevalence across risk deciles (S3 Fig). It was also the second-most parsimonious model, with 46 parameters.

Using these tuning parameters, 56 covariates entered Step 1, of which 35 were retained (S3 Table). In Step 2, 5 covariates were classified as age-common and 30 as candidate age-specific factors. After bidirectional selection in Step 3, 21 covariates remained, of which 17 were age-common and 4 were age-specific. All retained variables were selected during backward elimination, and none were reintroduced during forward reassessment.

PRs from the final model are shown in Table 2. Community-level PRs reflect prevalence changes per 10-percentage-point increase in PSU-level exposure, and individual-level PRs compare children with versus without the characteristic.

Table 2. Prevalence ratios for age-common and age-specific factors associated with Plasmodium infection among children aged 6–59 months in Nigeria. Prevalence ratios (PRs) and 95% confidence intervals (CIs) estimated from a survey-weighted generalized linear model using a three-stage, survey-aligned variable-selection framework to identify age-common and age-specific associations. Younger children were aged 6–23.9 months and older children 24–59 months. Age-common factors exhibit consistent associations across both age groups, whereas age-specific factors differ between them. Covariates associated with higher prevalence are shown in red; those associated with lower prevalence are shown in blue. Abbreviations: PR = prevalence ratio; CI = confidence interval; PSU = primary sampling unit; BMI = body mass index; ITN = insecticide-treated net.

Covariate name	Age pattern	Younger children PR (95% CI)	Older children PR (95% CI)
Community-level factors (per 10-percentage-point increase)
Socioeconomic and demographic
Household is poor or poorer	Age-common	1.013 (1.000–1.026)
Mother received primary education or higher	Age-common	1.000 (0.982–1.018)
Household structure and physical environment
Household owns livestock	Age-common	1.025 (1.011–1.039)
Household owns agricultural land	Age-common	1.046 (1.029–1.062)
Individual-level factors
Socioeconomic and demographic
Household wealth quintile	Age-common
Richest		Reference
Richer		1.668 (1.384–2.012)
Middle		1.833 (1.515–2.218)
Poorer		1.753 (1.432–2.147)
Poorest		1.807 (1.467–2.224)
Education level of the mother	Age-common
Higher		Reference
Secondary		1.303 (1.045–1.624)
Primary		1.544 (1.228–1.941)
None		1.558 (1.225–1.980)
Education level of the father	Age-specific
Higher		Reference	Reference
Secondary		1.095 (0.858–1.397)	0.944 (0.567–1.570)
Primary		1.173 (0.895–1.537)	0.972 (0.557–1.696)
None		1.301 (1.025–1.652)	0.966 (0.592–1.575)
Area of residence	Age-common
North Central		Reference
North East		0.763 (0.677–0.860)
North West		0.967 (0.868–1.077)
South South or South East		1.078 (0.953–1.219)
South West		1.559 (1.367–1.778)
Household lives in rural regions	Age-specific	1.462 (1.266–1.688)	1.263 (0.950–1.679)
Child & Maternal nutrition
Given iron supplement within 7 days	Age-specific	0.879 (0.757–1.021)	1.078 (0.795–1.462)
Child given vitamin A supplement within 6 months	Age-common	0.936 (0.880–0.995)
Anemia level of the mother	Age-common
Normal		Reference
Mild		1.132 (1.060–1.208)
Moderate		1.144 (1.077–1.215)
Severe		1.225 (1.012–1.483)
BMI status of the mother	Age-common
Normal		Reference
Very thin		0.975 (0.869–1.095)
Thin		0.956 (0.874–1.046)
Overweight		0.896 (0.825–0.974)
Obese		0.788 (0.669–0.928)
Height-for-age	Age-common
With stunting		1.080 (1.024–1.139)
Normal		Reference
Weight-for-height	Age-common
With wasting		1.047 (0.953–1.150)
Normal		Reference
Overweight		0.761 (0.631–0.917)
Currently breastfed	Age-common	0.845 (0.764–0.934)
Household structure and physical environment
Place of handwashing	Age-common
At home: fixed		Reference
At home: mobile		1.005 (0.931–1.084)
Not at home		1.091 (0.996–1.196)
Prevention and health-related behaviors
Household owns ITN	Age-common	1.104 (1.017–1.198)
Slept under an ITN last night	Age-common	0.916 (0.858–0.977)
Mother uses the internet at least once a week	Age-common	0.638 (0.475–0.857)
Appropriate disposal of children's stool	Age-specific	0.942 (0.850–1.044)	1.073 (0.866–1.330)

Sixteen covariates had at least one response category significantly associated with Plasmodium infection (Table 2). Two community-level covariates showed age-common effects. A 10-percentage-point increase in community-level household livestock ownership was associated with a 2.5% increase in prevalence (PR = 1.025; 95% CI: 1.011–1.039), and a similar increase in community-level agricultural land ownership was associated with a 4.6% increase (PR = 1.046; 95% CI: 1.029–1.062).

Of the 14 individual-level covariates, 12 exhibited age-common effects and 2 were age-specific. Among age-common covariates associated with increased Plasmodium prevalence, the largest increases were observed for socioeconomic disadvantage. Children in any wealth quintile below the richest had prevalence increases of 66.8–80.7% (PR = 1.668–1.807; all p < 0.05), followed by children whose mothers had less than higher education, who had increases of 30.3–55.8% (PR = 1.303–1.558; all p < 0.05). Maternal anemia was associated with modest increases of 13.2–22.5% (PR = 1.132–1.225). Smaller increases were observed for household ITN ownership (PR = 1.104; 95% CI: 1.017–1.198) and child stunting (PR = 1.080; 95% CI: 1.024–1.139).

Several age-common covariates were associated with decreased Plasmodium prevalence. Children whose mothers used the internet at least once a week showed the largest reduction, with 36.2% lower prevalence (PR = 0.638; 95% CI: 0.475–0.857), followed by children who were overweight-for-height, with a 23.9% reduction (PR = 0.761; 95% CI: 0.631–0.917), and by children of mothers with obesity, who had a 21.2% reduction (PR = 0.788; 95% CI: 0.669–0.928). More modest reductions were observed for current breastfeeding with 15.5% lower prevalence (PR = 0.845; 95% CI: 0.764–0.934), maternal overweight with a 10.4% reduction (PR = 0.896; 95% CI: 0.825–0.974), sleeping under an ITN with an 8.4% reduction (PR = 0.916; 95% CI: 0.858–0.977), and recent vitamin A supplementation with a 6.4% reduction (PR = 0.936; 95% CI: 0.880–0.995).

Geographic region showed substantial variability in Plasmodium prevalence. Compared with the North Central region, prevalence was 23.7% lower in the North East (PR = 0.763; 95% CI: 0.667–0.860) and 55.9% higher in the South West (PR = 1.559; 95% CI: 1.367–1.778).

Two covariates demonstrated age-specific effects. Among younger children, having a father with no education versus higher education increased prevalence by 30.1% (PR = 1.301; 95% CI: 1.025–1.652), but no significant association was observed among older children (PR = 0.966; 95% CI: 0.592-1.575). Younger children living in rural areas also had markedly higher prevalence, with a 46.2% increase (PR = 1.462; 95% CI: 1.266–1.688), again with no significant association among older children (PR = 1.263; 95% CI: 0.950-1.679).

Discussion

In this nationally representative analysis of Nigerian children under five, we applied the Cluster-aware Multistage Selection (CMS) framework to identify age-common and age-specific factors of Plasmodium infection while accounting for the stratified and clustered design of the Demographic and Health Survey. Using CMS, we reaffirmed established socioeconomic, nutritional, ecological, and behavioral determinants and revealed age-dependent heterogeneity obscured in pooled under-five analyses.

We found that community-level ecological factors of livestock and agricultural land ownership were consistently associated with higher Plasmodium prevalence. We observed geographic heterogeneity, with lower prevalence in the North East and higher prevalence in the South West relative to the North Central region.

Among age-common individual-level factors, we found that socioeconomic disadvantage showed the strongest associations, with clear gradients by household wealth and maternal education. Nutritional factors showed smaller but consistent associations, with stunting associated with higher prevalence and overweight or obesity, vitamin A supplementation, and breastfeeding associated with lower prevalence. Maternal anemia was associated with higher prevalence, whereas maternal overweight or obesity was associated with lower prevalence. Prevention-related behaviors showed divergent patterns, with ITN ownership associated with higher prevalence but child-level ITN use associated with lower prevalence. Maternal internet use showed the strongest protective association with Plasmodium prevalence.

We identified lack of paternal education and rural residence as age-specific factors; each associated with higher prevalence only among younger children.

Finally, we demonstrated the utility of CMS for high-dimensional population surveys. CMS enabled robust variable selection while accounting for complex survey design, yielding stable and interpretable models for population-level inference.

The associations between community-level livestock ownership, agricultural land use, and higher Plasmodium prevalence align with prior evidence that Plasmodium infection transmission is shaped primarily by shared ecological environments rather than individual household characteristics. Agricultural activity and livestock presence serve as proxies for vector-favorable landscapes that increase exposure risk at the community level [24–26].

We identified geographic heterogeneity that aligns with prior evidence of regional variation in Plasmodium infection burden across Nigeria [27]. Although both the North East and South West are ecologically suitable for Anopheles vectors, the lower prevalence observed in the North East after socioeconomic adjustment likely reflects differences in population density and Plasmodium control coverage, including seasonal malaria chemoprevention (SMC) [26,28–30]. In contrast, persistently higher prevalence in the highly urbanized South West is consistent with evidence that urbanization does not uniformly confer protection, as unplanned growth, poor drainage, and stagnant water can sustain Anopheles breeding even in nominally urban settings [28].

Strong and graded associations between socioeconomic disadvantage and Plasmodium prevalence are well documented [10–12,31,32]. Prior studies suggest that household wealth and maternal education reflect disparities in housing quality, preventive access, health literacy, and care-seeking linked to Plasmodium susceptibility [1,19,31,33,34]. Evidence also supports bidirectional causation, whereby Plasmodium infection may reinforce socioeconomic disadvantage through financial burden [35].

The associations observed for nutritional and maternal health factors align with established links between early-life nutrition, immune competence, and Plasmodium infection susceptibility. Chronic undernutrition may impair immune responses and parasite clearance, while malaria from Plasmodium infection may contribute to growth faltering, suggesting a bidirectional association with Plasmodium infection [36–38]. Breastfeeding and vitamin A supplementation are known to support immune function and have been associated with reduced Plasmodium infection [39,40].

Divergent associations for prevention-related behaviors warrant careful interpretation. Household ITN ownership was associated with higher prevalence, likely reflecting residual confounding and programmatic targeting of ITNs to higher-risk households in endemic areas, rather than a causal effect. In contrast, child-level ITN use was associated with lower prevalence, consistent with established protective efficacy [1,11,41]. This discrepancy highlights a measurement distinction between access and utilization, as ownership does not ensure consistent or correct use [42]. Additionally, reverse causality may contribute, whereby households experiencing prior malaria episodes are more likely to acquire or retain ITNs. Together, these findings emphasize that effective utilization, rather than access alone, is critical for reducing Plasmodium infection risk. Protective association for maternal internet use may reflect improved access to health information, as suggested by prior regional studies from Nigeria [43].

We identified age-specific associations for paternal education and rural residence, contrasting with prior pooled analyses that reported inconsistent or uniformly elevated risks, suggesting masked age-related heterogeneity [24,27,44–47]. Early childhood is a period of heightened caregiver dependence, when caregiver characteristics may disproportionately influence child health [34,43,46]. Therefore, paternal education may be more consequential in early childhood, before caregiving roles shift toward mothers as children age, and younger children in rural settings may experience greater exposure through caregiver proximity; however, these pathways could not be evaluated using DHS data.

Our findings among children under five align with Plasmodium infection patterns reported in older children and adults. Unlike children under five, there is no DHS-equivalent standardized survey that reliably estimates Plasmodium prevalence in older age groups in sub-Saharan Africa [48]. Global Burden of Disease data analyses across sub-Saharan Africa demonstrated lower Plasmodium prevalence among school-aged children (6–12 years; 17.6%) compared to adolescents and adults (≥12 years; 36.5%), mimicking our finding of higher prevalence among older children (49.4%) compared with younger children (37.3%) [48]. Across age groups, evidence suggests a repeating age-related gradient in Plasmodium prevalence, with lower prevalence among younger children and primary school–aged children and higher prevalence among older children under five and adolescents, indicating that ages two and twelve may represent key transition points in mobility, exposure, and caregiving intensity. Age-common factors identified in our study, including lower socioeconomic status, limited education, and inconsistent ITN use, were also significant determinants in older age groups [48–52].

Notably, some studies have found that rural residence has not been consistently associated with Plasmodium infection beyond age five once specific environmental and behavioral exposures are considered [49,53]. Instead, proximate factors such as household farming activity, stagnant water, proximity to rivers, and vegetation near households appear to drive infection risk, suggesting that the age-specific association observed among younger children in our study may also reflect the impact of age-dependent behaviors or environmental exposures rather than rural residence itself [49,54].

Methodologically, many DHS-based multivariable studies are limited by collinearity, inadequate model validation, and incomplete appreciation of the survey design, compromising stable selection and population-level inference [13,15,16,18]. These challenges likely contribute to heterogeneous findings and underscore the utility of CMS as a rigorous yet flexible approach.

Our study has several limitations. Because the 2024 DHS discontinued Plasmodium infection testing and the 2021 Malaria Indicator Survey did not collect covariates as comprehensive as those in the DHS, we relied on the 2018 DHS data, which may limit the timeliness of our findings. The cross-sectional DHS design supports association analyses but precludes assessment of transmission seasonality and evaluation of malaria control programs, including seasonal malaria chemoprevention. DHS Plasmodium infection testing is restricted to children under five, which precludes assessment of older children and limits inference across the full pediatric age spectrum. Outcome misclassification is possible due to persistent RDT positivity. Residual confounding from unmeasured environmental or behavioral factors may also influence observed associations. CMS relies on researcher-defined, data-informed tuning parameters, which may introduce selection uncertainty, and age was modeled using pre-specified categories rather than flexible nonlinear functions, potentially obscuring more complex age-related patterns. While CMS incorporated stability thresholds and survey-aligned validation to mitigate overfitting, post-selection bias cannot be fully eliminated.

Conclusion

Using the 2018 Nigeria DHS, we applied the Cluster-aware Multistage Selection (CMS) framework to identify age-common and age-specific determinants of Plasmodium infection among children under five. Higher Plasmodium prevalence among younger children in rural and socioeconomically disadvantaged households highlights priority populations for age-targeted Plasmodium control, including household-level prevention and caregiver-focused interventions. CMS provides a rigorous, reproducible approach for inference in clustered, multilevel data and is broadly applicable to complex population-based studies such as DHS. Public health interventions should account for age heterogeneity within under-five populations, and future studies should investigate underlying immunological, behavioral, and socioeconomic mechanisms while validating CMS across settings.

Supporting information

S1 Table. Full list and definitions of covariates evaluated in the Cluster-aware Multistage Selection to identify age-common and age-specific factors associated with Plasmodium infection among children under 5 years in Nigeria, based on the 2018 Nigeria Demographic and Health Survey.

Raw covariates were merged, filtered, or recoded as needed to create more interpretable, generalizable, and epidemiologically meaningful variables. The Demographic and Health Survey (DHS) Standard Recode Manual and the Guide to DHS Statistics (DHS-7) were consulted in constructing these covariates.

(XLSX)

S2 Table. Diagnostic assessments for models used to tune threshold parameters in the Cluster-aware Multistage Selection identifying age-common and age-specific factors associated with Plasmodium infection among children under five in Nigeria, 2018 Nigeria Demographic and Health Survey.

Tuning parameters included the method (minimum-error or 1-standard-error) used to determine the penalty λ term for least absolute shrinkage and selection operator regression in Step 1; the proportion of bootstrap trials required for a covariate to pass Step 1; the p-value cutoff for identifying candidate age-specific covariates in Step 2; the p-value cutoff for selecting significant covariates in Step 3; the cutoff of modified quasi-likelihood information criterion improvement for identifying model-improving covariates in Step 3; and the cutoff of relative prevalence ratio changes for identifying potential confounding covariates in Step 3. For each model, we report the Pearson dispersion statistic, the modified quasi-likelihood information criterion, the calibration correlation coefficient, the Rao–Scott adjusted global model test, the cross-validated Brier score and bins grouped by 0.1% increments of the best score, and the number of estimated parameters. Models were ranked first by their Brier score bins and, within ties, by parsimony. Selected model is highlighted in black. The tuning parameters from the top-ranked model were selected. Abbreviations: F = Rao-Scott F statistic; λ = penalty term used for least absolute shrinkage and selection operator regression; dQIC = difference in the simplified quasi-likelihood information criterion; dPR = change in prevalence ratio; φ = Pearson dispersion statistic; QICu = simplified quasi-likelihood information criterion; r = calibration correlation coefficient.

(XLSX)

S3 Table. Step-by-step results of the Cluster-aware Multistage Selection identifying age-common and age-specific factors associated with Plasmodium infection among children under five in Nigeria, 2018 Nigeria Demographic and Health Survey.

Variable selection results from the three-stage covariate screening procedure. Community-level factors were defined as primary sampling unit level proportions of individuals exhibiting each covariate, and individual-level factors were measured directly for each child. In Step 1, survey-weighted least absolute shrinkage and selection operator with stratified primary sampling unit bootstrapping retained covariates selected in ≥60% of bootstrap replicates. In Step 2, generalized linear models (GLMs) with bootstrapping identified candidate age-varying covariates when age-interaction terms were significant in ≥60% of replicates; remaining covariates were treated as candidate age-common covariates. In Step 3, backward elimination with forward reassessment retained covariates that showed significant associations with Plasmodium infection, improved model fit by modified quasi-information criterion changes, or acted as confounders based on changes in mean or maximum prevalence ratios. Refer to S1 Text for details. The final age pattern (“age-common” or “age-specific”) and retention status are reported for each covariate. Abbreviations: LASSO = Least Absolute Shrinkage and Selection Operator; GLM = Generalized Linear Model; DAG = Directed Acyclic Graph; BMI = Body Mass Index; ITN = Insecticide-Treated Net; QICu = simplified quasi-likelihood information criterion; dQIC = QICu of reduced model minus QICu of retained model; PR = prevalence ratio; dPR = Relative prevalence ratio change.

(XLSX)

S4 Table. Covariate-level results from the final Cluster-aware Multistage Selection model identifying age-common and age-specific factors associated with Plasmodium infection among children under five in Nigeria, 2018 Nigeria Demographic and Health Survey.

Community-level factors were defined as primary sampling unit level proportions of individuals exhibiting each covariate, and individual-level factors were measured directly for each child. Model statistics are reported for covariates retained in the final survey-weighted quasi-Poisson regression model with a log link following the multi-level, three-stage, survey-design-aware, bidirectional covariate selection procedure. Interaction terms between the age-bin indicator (24–59 months versus 6–23.9 months) and each covariate X are presented as Age × X. Abbreviations: CI = confidence interval; ITN = insecticide-treated nets.

(XLSX)

S1 Fig. Prevalence of Plasmodium infection among younger (6–23.9 months) and older (24–59 months) children in Nigeria, 2018 Demographic and Health Survey.

Weighted prevalence of Plasmodium infection among children aged 6–59 months, stratified by younger and older age groups. Children testing positive by either rapid diagnostic test or light microscopy were classified as malaria-positive; those negative on both tests were classified as malaria-negative. Estimates incorporate sampling weights, clustering, and geographic stratification.

(TIF)

S2 Fig. Covariate retention across 144 tuning configurations in the Cluster-aware Multistage Selection identifying age-common and age-specific factors associated with Plasmodium infection among children under five in Nigeria, 2018 Nigeria Demographic and Health Survey.

The Cluster-aware Multistage Selection framework included several tunable components. Step 1 varied the LASSO penalty-rule for λ (minimum-error vs 1-standard-error). Steps 1–2 varied the stratified PSU-level bootstrap selection threshold (30% vs 60%) and Step 2 used p-value cutoffs of 0.20, 0.15, or 0.05. Step 3 applied stricter p-value thresholds (0.15, 0.10, 0.05), simplified quasi-likelihood information criterion or QICu of reduced model minus retained model tolerances of (−2, 0, or +2), and confounding thresholds based on relative mean/maximum PR changes (20%/40% or 10%/20%). These combinations yielded 144 unique tuning configurations. Models were ranked by Brier-score bins and, within bins, by parsimony. In the plot, models are ordered left to right from highest to lowest ranking. For each model, covariates are shown as retained (X) or excluded; if retained, associations are colored as significant increasing (red), significant decreasing (blue), or non-significant (grey). Results are presented separately for younger (6–23.9 months) and older (24–59 months) children, with age-common and age-specific associations distinguished by shading intensity. Abbreviations: BMI = body mass index; ITN = insecticide-treated nets.

(TIF)

S3 Fig. Residual and calibration plots for the final model selected using the Cluster-aware Multistage Selection to identify age-common and age-specific factors associated with Plasmodium infection among children under five in Nigeria, 2018 Nigeria Demographic and Health Survey.

Model selected from the Cluster-aware Multistage Selection was evaluated using residual–fitted plots, and calibration was assessed by comparing survey-weighted observed prevalence with predicted probabilities across deciles of fitted probabilities.

(TIF)

S1 Text. Statistical methods used in the Cluster-aware Multistage Selection to identify age-common and age-specific factors of Plasmodium infection among children under five in Nigeria, using the 2018 Nigeria Demographic and Health Survey.

(PDF)

Data Availability

This study used de-identified data from the 2018 Nigeria Demographic and Health Survey (NDHS). Researchers may request access through the DHS Program data portal (https://dhsprogram.com/data/available-datasets.cfm) by registering for an account, submitting a research project description, selecting the 2018 Nigeria DHS, and agreeing to the DHS data-use conditions. Following approval, researchers may download the dataset and its accompanying documentation and codebook. The DHS recode files used in this study were NGKR7BFL.DTA (Children’s Recode), NGPR7BFL.DTA (Household Member Recode), and NGHR7BFL.DTA (Household Recode). Because DHS data-use conditions prohibit redistribution, the recode files are not included with this article or in the code repository. No DOI or accession number is assigned to the 2018 Nigeria DHS microdata. The full R code is publicly available at https://github.com/woojae1995/Nigeria_DHS_2018_age_specific_RF.

Funding Statement

The authors received no specific funding for this work.