Study design and setting

This study involved a secondary analysis of data from the 2018 Zambia Demographic and Health Survey (ZDHS). The ZDHS is a nationally representative survey conducted by the Zambia Statistics Agency (ZamStats) in collaboration with the Ministry of Health (MOH) and ICF International, using a stratified two-stage cluster sampling design. Comprehensive details on the sampling methodology design, data collection procedures, and survey instrument are provided in the ZDHS final report [14]. The primary objective of the 2018 ZDHS was to collect up-to-date and comprehensive data on vital demographic and health indicators to inform national policy and planning.

Sampling and data measurement of ZDHS

The ZDHS employed a two-stage sampling design to ensure representativeness at national, provincial, and urban–rural levels. In the first stage, enumeration areas (EAs) were selected as clusters using probability proportional to size within each stratum, with a total of 545clusters included. In the second stage, a household listing was conducted within each cluster, identifying an average of 133 households per cluster. From these, 25 households per cluster were systematically sampled, yielding a total of 13,625 households. The detailed methodology of the ZDHS has been well-documented in prior publications [14].

Anemia testing was done on all women of reproductive age (15–49 years old) from whom consent was obtained [14]. The missing data constituted only 4% and was, therefore, negligible since 96% of the selected women for the interviews responded.

Explanatory variables (Risk Factors)

The study incorporated both individual-level and community-level variables available in the ZDHS dataset. Previous studies have attributed the prevalence of anemia to both individual and community-level factors.

Individual-level factors encompassed demographic and health-related characteristics such as age, religion, marital status, educational attainment, wealth index, and gravidity. Community-level factors included variables such as residence type (urban or rural), geographic region, water source, and type of toilet facilities, which reflect populations living under similar environmental or infrastructural conditions. The classification of variables as individual- or community-level was guided by the assumption of independence between levels.

This study used secondary data that containing no personal identifier; therefore, the requirement for individual informed consent was waived by the University of Zambia Biomedical Research Ethics Committee (UNZABREC), which approved the study (Approval Number: 6300−2025).

Statistical analysis

Statistical analysis was conducted using Stata 17 to generate descriptive statistics and address the research questions under investigation. Frequencies and percentages were used to summarize both individual- and community-level variables. Given that some regions with smaller populations were oversampled while others with larger populations were underrepresented, weighted frequencies and percentages were calculated to account for population size differences across regions. The detailed weighting methodology is outlined in the ZDHS 2018 report [14].

To identify factors associated with anemia among women, bivariate logistic regression analyses were first conducted to select variables for inclusion in both the multivariate standard logistic regression and mixed-effects multilevel logistic regression models. Missing observations for the anemia and HIV variables, as well as inconclusive responses in the HIV variable, were excluded, reducing the sample from 13,625–13,055 observations.

To select between the multivariate standard logistic regression and the mixed-effects multivariable multilevel logistic regression model, a log-likelihood ratio test was conducted. A model with a higher log-likelihood value was preferred [15].

Standard logistic model

(1)

Mixed-effects multilevel model

A mixed effects multilevel model is a standard logistic regression model, amended to have random effects for each EA. Defining , we have

(2)

for j = 1,..., 545 EAs, with i = 1,..., women in EA j.

Model selection

A log-likelihood ratio test was conducted to select between the standard logistic regression and the mixed-effects multilevel logistic regression model as shown in Table 1. As a rule of thumb can be seen that a model with a higher log-likelihood value is preferred [15].

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Standard logistic and Mixed effects multilevel regressions.

https://doi.org/10.1371/journal.pone.0330400.t001

Mixed-effects logistic regressions

We specified 4 distinct models, aiming to identify the risk factors at different levels, model 1 included no predictors to assess community-level variation in anemia prevalence; model 2 added individual-level factors; model 3 included community-level factors; and Model 4 combined both individual- and community-level factors. In more detail, the differences between the models are summarized in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Specification of estimated models.

https://doi.org/10.1371/journal.pone.0330400.t002

The fixed-effects results were expressed as odds ratios (ORs) with 95% confidence intervals (CIs), while adjusted odds ratios (AORs) with 95% CIs were calculated to identify independent factors associated with anemia. Statistical significance was determined using a p-value threshold of <0.05. To assess multicollinearity, variance inflation factors (VIFs) were computed, with variables considered to have low collinearity if their VIF values were below [16].

Random effects, capturing variations between clusters, were assessed using the Intra-Cluster Correlation Coefficient (ICC), Percentage Change in Variance (PCV), and Median Odds Ratio (MOR). The ICC quantified the proportion of total variability attributable to clustering effects, while the MOR converted unexplained cluster-level variance into the OR scale, facilitating the interpretation of heterogeneity. The PCV measured the proportion of total variance explained by individual- and community-level factors. These measures collectively informed the suitability of the model specification. These measures were computed using the following formulas:

(3)

(4)

(5)

Where V represents the estimated variance of clusters, is the variance in the initial model, and is the variance in subsequent models with additional terms. The multi-level analysis was well suited for our study given the correlated nature of our data. It evaluates how variables at different levels influence the dependent variable while addressing biases introduced by clustering [17]. By producing corrected standard errors, confidence intervals, and significance tests, multilevel models yield reliable parameter estimates [18].

The results of the fourth model, selected based on the AIC and BIC, indicated that age group, marital status, sex of the household head, pregnancy status, breastfeeding status, number of children born in the preceding five years, HIV test results, residence, and religion were significantly associated with anemia among women.

Spatial mapping

The ZDHS data is georeferenced, enabling spatial analysis to be carried out. To maintain the confidentiality of respondents, ZDHS has a georeferenced data release policy, which provides two levels of protection. First, data from the same enumeration area (EA) are aggregated to a single-point coordinate. Second, a GPS coordinate displacement process is applied, which masks the true locations. Urban clusters were displaced a distance of up to 2 km (0–2 km), rural clusters up to 5 km (0–5 km), and an additional randomly selected 1% of rural clusters are displaced up to ten kilometers (0–10 km) [14].

A spatial analysis was conducted to examine the geographic distribution of anemia prevalence among women of reproductive age (WRA) in Zambia using geospatial data from the 2018 Zambia Demographic and Health Survey (ZDHS). A shapefile containing Zambia’s administrative boundaries with province-level anemia prevalence estimates (percentages) was obtained from the ZDHS website. The geospatial data were imported into QGIS for spatial processing and visualization. Anemia prevalence was classified into seven categories and displayed using a graduated color scale, with yellow representing the lowest prevalence and red the highest prevalence. Provincial labels were added to enhance interpretability. The resulting map illustrates spatial disparities in anemia prevalence across Zambia’s provinces.

Sociodemographic characteristic

Table 3 presents the descriptive statistics of the study participants. The analysis included data on a total of 13,055 women from 545 clusters across 10 provinces. The median age (IQR) of the participants was 27 (20 –36) years. Over half (56%) of the women were married, and 8% had no formal education. About half (54%) of the women in the study were rural residents, and 71% identified as protestant Christians. About 8% of the women were pregnant at the time of the survey, while 23% were breastfeeding and 14% were HIV positive. Furthermore, nearly three-quarters (73%) of the households were male-headed, 72% had access to improved water sources while 57% had unimproved toilet facilities.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Descriptive statistics and prevalence of anemia among study participants (n = 13,055).

https://doi.org/10.1371/journal.pone.0330400.t003

Table 3 shows the variation of anemia across different Socio-demographic and health-related characteristics among study participants. The estimated anemia prevalence among women of reproductive age was 31%. Significant regional variations were observed, Lusaka and Western had the highest prevalence (36% and 38%, respectively), while Central province had the lowest (24%). Adolescent women (aged 15–19) years and older women (40–49 years) had the highest and equal anemia prevalence (33%). Single women exhibited higher anemia prevalence (34%) compared to married women (28%). The prevalence of anemia among pregnant women was higher (41%) compared to the women who were not pregnant (30%).

Factors such as marital status, HIV positivity, current pregnancy, current breastfeeding, and gravidity were significantly associated with anemia prevalence (p = 0.001), whereas access to medical care, place of residence, and sanitation showed no significant association. Gravidity was excluded from the multivariate standard logistic regression and mixed-effects multilevel logistic regression models to minimize multicollinearity, given its strong correlation with age and marital status in the study population.

The spatial distribution of anemia prevalence in Zambia is shown on a map using a sequential color gradient, with lighter shades indicating lower prevalence and darker shades indicating higher prevalence. The lowest prevalence is observed in Central Province and the highest in Western Province as shown in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Map of Zambia showing the prevalence of anemia at provincial level.

The base layer of the map was obtained from the Database of Global Administrative Areas (GADM) version 4.1 (https://gadmorg/download_country.html). GADM data is freely available for academic and non-commercial use and is compatible with the CC BY 4.0 license. Anemia prevalence data were obtained from the 2018 Zambia Demographic and Health Survey.

https://doi.org/10.1371/journal.pone.0330400.g001

Mixed effect

Table 4 summarizes the multilevel mixed-effects logistic regression estimates for determinants of anemia among women of reproductive age. Women aged 25–29 years old were 16% less likely to be anemic compared with women in the youngest age group (15–19 years old) (AOR = 0.84; 95% CI 0.71 to 0.97). The married women were 0.78 times less likely to be anemic than women who never married (AOR = 0.78; 95% CI 0.68 to 0.90). The odds of anemia were higher in women who were pregnant (AOR = 1.76; 95% CI 1.52 to 2.03) compared with those who were not pregnant. Women who were currently breastfeeding were 15% (AOR = 1.15; 95% CI 1.02 to 1.30) more likely to be anemic. Women who gave birth to one child in the preceding 5 years of the survey were at low risk of having anemia (AOR = 0.84; 95% CI 0.76 to 0.94). In this study, women who were HIV positive had a twofold increased odds of having anemia compared with women classified as HIV negative (AOR = 2.21; 95% CI (1.97 to 2.49). Higher odds of anemia were observed in Copperbelt(AOR = 1.33; 95% CI 1.06 to 1.68), Eastern (AOR = 1.27; 95% CI 1.01 to 1.61), Lusaka (AOR = 1.3; 95% CI 1.03 to 1.66), Luapula (AOR = 1.52; 95% CI 1.21 to 1.90), Northern (AOR = 1.31; 95% CI 1.03 to 1.67), Northwestern (AOR = 1.74; 95% CI 1.36 to 2.22), Southern(AOR = 1.52; 95% CI 1.20 to 1.91) and Western provinces (AOR = 1.99; 95% CI 1.56 to 2.54) compared with Central province.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Multilevel mixed-effects logistic regression of factors associated with anemia among women of reproductive age.

https://doi.org/10.1371/journal.pone.0330400.t004

Random effects

As shown in Table 4, the random-effects model demonstrates significant variation in anemia prevalence across communities (p < 0.001). In other words, the anemia prevalence rate was not similarly distributed across the communities. About 6% of the variance in the odds of anemia in women could be attributed to community-level factors, as calculated by the ICC based on estimated intercept component variance. After adjusting for individual-level and community-level factors, the variation in anemia across communities remained statistically significant. About 5% of the odds of anemia variation across communities was observed in the full model (model 4). In the null model, the MOR for anemia was 1.52, indicating that there was a variance between communities (clustering) (1.52 times larger than the reference (MOR = 1). When both individual and community factors were included in the model, the unexplained community variation in anemia decreased to MOR of 1.48. This indicates that in the full model the effects of clustering are still statistically significant when we considered both individual and community factors.

The fit of the models are compared by relying on Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) and are summarised in Table 5. As can be seen that the model with the lowest value describes the data best [19].

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Estimation results: AIC and BIC.

https://doi.org/10.1371/journal.pone.0330400.t005