Spatial Distribution and Risk Factors Associated with Contraceptive Use among Women of Reproductive Age in Zambia

2.1. Source of Data

The current study utilized cross-sectional data from a sample of 13,683 respondents in the 2018 Zambia Demographic and Health Survey (ZDHS). The data in the study is the sixth in a series of Demographic and Health Surveys in Zambia; the prior studies were held in 1992, 1996, 2001-02, 2007, and 2013-14. The data collection was conducted from 18^th July, 2018 to 24^th January, 2019 using a stratified two-stage sample design. In the first stage of sampling, clusters of enumeration areas were selected. A total of 545 clusters were constructed within each sampling stratum, with enumeration areas selected using probability proportional to size (PPS) based on the design frame. The second stage of the cluster sampling design involved the selection of households within each cluster for the survey through systematic sampling. A household listing procedure was conducted using the sampling frame of selected clusters, with an average of 133 households found in each cluster. A fixed number of 25 households were selected from each cluster using an equal probability systematic selection approach. Further to this sample, all women aged 15-49 who stayed the night before the survey at the household as permanent residents or visitors were eligible to be interviewed. From the interviewed households, 14,189 women aged 15-49 were identified as eligible for individual questionnaires. A total of 13,683 women were given the questionnaires, yielding a response rate of 96%. More details on the data collection and sampling methods are available in a study conducted previously [13].

2.2. Response Variable

The primary variable of interest in the study was contraceptive use status in women. A woman is said to be a contraceptive user if she utilizes at least one of the available contraceptive methods (modern or traditional). Thus, the binary response variable indicates whether a woman is using any contraceptive method. This variable is coded as “1” and “0” for a contraceptive user and a non-contraceptive user, respectively.

2.3. Explanatory Variables

The explanatory variables selected in the study were categorized into variable characteristics or factors, such as socio-economic, geographical, environmental, or demographic characteristics. The variable selection was based on an extensive review of existing research on contraceptive use to identify variables that had been previously studied, aligning with the theoretical framework, and conducting exploratory data analysis to identify correlations between variables. The socio-economic factors included religion (Catholic, Protestant, Muslim, Other), household wealth index (poorest, poorer, middle, richer, richest), woman’s employment status (employed, unemployed), heard of family planning (yes, no), visited by a fieldworker in the last 24 months (yes, no), discussion about family planning at a health facility (yes, no), knowledge of sexually transmitted infections (yes, no), and know about contraception side effects (yes, no). The geographical factors included region (Central, Copperbelt, Eastern, Luapula, Lusaka, Muchinga, Northern, North-Western, Southern, Western) and place of residence (urban and rural). The demographic factors included age at first cohabitation (continuous), age at first sex (<15 years, 15-24 years, 25-34 years, 35+ years), age of a partner (continuous), age of a woman (continuous), decision maker for using contraceptives (mainly respondent, mainly husband/partner, joint decision, other), decision maker on woman’s healthcare (respondent alone, respondent and partner, respondent and other person, partner, someone else, other), education level of a partner (no education, primary, secondary, higher, do not know), education level of a woman (no education, primary, secondary, higher), family size (continuous), marital status of a woman (never in a union, married/living with a partner, widowed, divorced, no longer living together/separated), number of living children (0, 1-2, 3-4, 5+), partner’s desire for more children (both wants the same, partner wants more, partner wants fewer, do not know), preferred waiting time for the birth of another child (<12 months, 1 year, 2 years, 3 years, 4 years, 5 years, 6+ years, immeasurable, do not know), and woman’s desire for more children (never had sex, wants within 2 years, wants within 2+ years, unsure timing, undecided, wants no more, sterilized/declared infecund). The above variables were chosen based on the literature [1, 3-5, 8, 9, 11, 14-21].

2.4. Statistical Methods

The current study employed a quantitative analysis approach, and data were analyzed using spatial generalized linear mixed models to account for both fixed and random effects, as well as spatial variability.

2.5. Model Structure

The model form of the spatial generalized linear mixed model is given by:

(1)

The spatial autocorrelation is assumed to depend on the distance instead of the location of responses, so that the covariance between observations y(s_i) and y(s_j) at different locations is given by:

(2)

Where c(.) is a function of the distance h = s_i - s_j between two locations [22, 23]. It is assumed that this covariance has a constant mean that is second-order stationary, i.e., E[y(s_i)]=E[y(s_j)] Furthermore, to include the location s_i, it is assumed that (i) y(s_i|γ) is conditionally independent of any location with mean E[y(s_i|γ)] = μ(s_i) and (ii) The distribution of s is defined by γ [24]. Consider a generalized linear model (GLM) case classically formulated as g(μ) = Xβ, where X is a design matrix. To include spatial processes, assume (i) and (ii) [25]. Thus, a model with spatial random effects is given as:

(3)

Where X and Z are the design matrices, and the random effect at the location s_i is normally distributed with mean=0 and variance = γ(θ). Moreover, the spatial correlation parametrized by θ is γ(θ), and the error terms are also normally distributed with mean = 0 and variance = σ_ε²I, where I is the identity matrix. Literature has suggested that several functions could be used to describe spatial dependency. However, in geostatistics, spatial correlation is generally described by three main functions: (i) Semivariogram (also known as variogram), (ii) Correlogram, and (iii) Covariance [25]. The variogram is the most commonly used function in geostatistics for fitting spatial correlation models, and the current study utilizes this function.

2.6. Spatial Autocorrelation Analysis

Accounting for spatial variability requires investigating the autocorrelation among residuals. The purpose of investigating residual autocorrelation is to enable the researcher to account for autocorrelation, maintain the normality assumption, and remove any structural errors from the final model. Moran’s I coefficient and Geary’s contiguity ratio [26, 27] are the two standard techniques applied when investigating the presence of spatial autocorrelation. Moran’s I coefficient evaluates the distribution of the given data and generates a single value output ranging from -1 to +1 to evaluate spatial autocorrelation. The value around -1 suggests the scattered distribution of data, values around 0 indicate a random data distribution, and values around +1 indicate a clustered data distribution [26, 28].

Geary’s contiguity ratio statistic is another measure of spatial autocorrelation that examines the similarity of the pair of attributes and produces an output value ranging from 0 to 2, where an output of 0 suggests maximum autocorrelation, 1 suggests the absence of spatial autocorrelation, and a value of 2 suggests a strong negative autocorrelation [27]. In addition to these statistics, the scatter plot and the structure of the empirical semivariogram can be constructed and examined.

2.7. Spatial Prediction

Spatial prediction (also known as spatial forecasting or spatial modeling) is a process of prediction that is used in spatial analysis and involves using data from known locations to predict the value of a variable at unobserved locations (S_i, for i = 1, 2, 3,..., n). In research, it is standard practice to measure one or more quantities at certain geographical locations to predict the observed values in unsampled locations [29]. Geographic information systems (GIS) are typically used for spatial analysis as they offer tools for generating feature statistics and performing geoprocessing tasks like data interpolation [24]. Kriging has long been used in spatial prediction to forecast one variable at a time using data from the same kind of variable. Kriging provides the best interpolation technique based on regression against observed values of neighboring spatially covariant-weighted data points. It offers several benefits, such as mitigating the impact of data clustering, granting individual points within a cluster a lower weight than isolated data points, providing a variable estimate, guaranteeing the availability of estimation error, which serves as a foundation for stochasticity, and enabling the simulation of potential outcomes [30]. Ordinary kriging was used in this study to predict variables in unsampled locations. The observed statistical features of the data were combined using the ordinary kriging approach. The partial sill corresponds to the most considerable variability in the absence of spatial dependence, and the strength of the statistical correlation is measured as a function of distance, where spatial correlation declines [31].

2.8. Analysis

The analysis in this study was carried out using the Proc GLIMMIX procedure from SAS Enterprise by fitting a generalized linear mixed model to the data. Several covariance structures were considered, namely linear, spherical, exponential, Gaussian, powered exponential, and Matern. ArcGIS Pro was used to produce a smooth map showing the contraceptive use prevalence patterns at the district level for each region of Zambia.

3.1. Descriptive Analysis

Table 1 gives a summary of the descriptive statistics of the categorical background characteristics of the respondents. The frequencies and percentages of women using and not using contraceptives are presented in Table 1. Out of the 13,683 women from the sample, 35.18% used at least one type of contraceptive method, and 64.82% were non-users. It was observed that the prevalence of using contraceptives among women was highest in the Eastern region, with 4.96% of users, and the lowest in the Western region, with 2.20% of users. As presented in Table 1, 14.14% of women from urban areas used contraceptives, whereas 21.04% of women from rural areas used contraceptives. It was observed that 2.62% of the women using contraceptives had no education, 19.65% had primary education, 13.46% had secondary education, and 2.15% had higher education. Furthermore, 5.94% of women using contraceptives were Catholic, 28.71% were Protestant, 0.23% were Muslim, and 0.31% were from other religions.

Table 2 shows the minimum, median, and maximum values of the continuous background characteristics used in the study. It was observed that the minimum age of cohabitation for women in the study was 10 years, the median was 19 years, and the maximum age was 44 years. Moreover, the minimum age of the partner was 15 years, the median was 37 years, and the maximum was 95 years. Moreover, the minimum age of a woman was 15 years, the median was 27 years, and the maximum was 49 years. Furthermore, the minimum family size of a woman was 0, the median was 2, and the maximum was 14.

The constructed empirical semivariogram is displayed in Fig. (1). It was observed that the distance between clusters increased, and the distribution of the semivariogram was not uniform. Thus, inferences could be made about an indication of surface trend.

Fig. (2). displays the scatter plot of the contraceptive use data observed in Zambia. The plots suggest a distribution that is not indicative of a uniform distribution, but rather an indication of the random spread of the response. Therefore, direct inferences about the existence of surface trends in the data can be made.

The current study analyzed various explanatory variables to investigate their impact on women’s contraceptive use using a spatial generalized linear mixed model. The results are presented in Table 3. The level of significance used in the analysis was 5%. Women who had their first sexual intercourse between the ages of 15 and 24 were found to be 1.208 times more likely to use contraceptives compared to those who initiated sexual activity before the age of 15 (P-value = 0.0006). Similarly, women who had their first sexual intercourse between the ages of 25 and 34 were 1.605 times more likely to use contraceptives compared to those who initiated sexual activity before the age of 15 (P-value = 0.0304). Additionally, for each one-year increase in a woman’s age, the likelihood of contraceptive use decreases by a factor of 0.948 (P-value < 0.0001). A woman without any education was 0.527 times less likely to use contraceptives compared to a woman with higher education (P-value < 0.0001). A woman with only primary education was 0.738 times less likely to use contraceptives than a woman with higher education (P-value = 0.0123). Furthermore, for each additional person in a woman's family, the chance of her using contraceptives increases by 1.096 (P-value < 0.0001).

A woman from the poorest household was 0.704 times less likely to use contraceptives than a woman from the richest household (P-value = 0.0017). Similarly, a woman from a poorer household was 0.800 times less likely to use contraceptives compared to a woman from the richest household (P-value = 0.0331). Additionally, a woman who has never been in a union was 0.640 times less likely to use contraceptives than a woman who was married or living with a partner (P-value = 0.0022). A widowed woman was 0.183 times less likely to use contraceptives than a woman who was married or living with a partner (P-value = 0.0002). On the other hand, a woman with 1-2 living children was 5.963 times more likely to use contraceptives than a woman with no children (P-value 0.0065). A woman with 3-4 living children was 11.897 times more likely to use contraceptives than a woman with no children (P-value <0.0001). A woman with 5+ living children was 14.152 times more likely to use contraceptives compared to a woman with no children (P-value <0.0001).

Moreover, a woman who did not know their partner's desire for more children was 0.761 times less likely to use contraceptives than a woman who had the same desire for more children with a partner (P-value <0.0001). A woman living in rural areas was 0.819 times less likely to use contraceptive methods compared to a woman living in urban areas (P-value = 0.0076). A woman who preferred to wait 1 year for the birth of another child was 1.190 times more likely to use contraceptives compared to a woman who preferred to wait < 12 months (P-value <0.0001). On the other hand, a woman who preferred to wait 2 years for the birth of another child was 0.937 times less likely to use contraceptives compared to a woman who preferred to wait < 12 months (P-value = 0.0001). Additionally, a woman who preferred to wait a non-numeric period for the birth of another child was 0.920 times less likely to use contraceptives than a woman who preferred to wait < 12 months for the birth of another child (P-value = 0.0003).

As presented in Table 3, a woman from the Luapula region was 0.579 times less likely to use contraceptives than a woman from the North-Western region (P-value < 0.0001). Similarly, a woman from the Northern region was 0.615 times less likely to use contraceptives compared to her counterparts in the North-Western region (P-value = 0.0003). In the Southern region, a woman was 0.723 times less likely to use contraceptives than a woman in the North-Western region (P-value = 0.0114). A woman from the western region was 0.644 times less likely to use contraceptives compared to a woman in the North-Western region (P-value = 0.0012). Additionally, a woman who did not discuss family planning at a health facility was 0.805 times less likely to use contraceptives than a woman who did (P-value < .0001). Lastly, an unemployed woman was 0.797 times less likely to use contraceptives compared to an employed woman (P-value < 0.0001).

3.2. Interaction Effects

The current study examined various two-way interaction effects and found that the number of living children and a woman’s desire for more children, as well as the marital status of a woman and her preferred waiting time for the birth of another child, were significant. The expected probabilities of the joint effects between a woman's desire for more children and the number of living children are shown in Fig. (3). It was observed that a woman with no children had a lower likelihood of using contraceptives regardless of the desire for more children. Furthermore, a woman with no children and who wanted more but was unsure of the timing was observed to have a lower likelihood of using contraceptives. Moreover, a woman who had at least three children was more likely to utilize contraceptives than a woman who had at most 1 or 2 children, regardless of her desire for more children.

Fig. (4) shows the association between the preferred waiting period for the birth of another child and marital status. (Fig. 4) illustrates a high likelihood of contraceptive use among women who have never been married, as well as across all categories of preferred waiting time before the birth of another child. This likelihood of contraceptive use was found to be higher among married women across all categories of preferred waiting time for the birth of another child. Furthermore, widowed women had a lower likelihood of contraceptive use throughout their preferred waiting times for the birth of another child. A woman who has never been in a union, married, or widowed, and who would rather wait a year to have another child, had a higher likelihood of using contraceptives than other women. For a divorced woman, the possibility of contraceptive use throughout the recommended waiting period was also similar. Additionally, a woman who wished to wait three years for the birth of another child and was separated from her partner was more likely to take contraceptives.

3.3. Smooth Map

The outcomes of the spatial prediction process, as determined by interpolation using the ordinary kriging approach, are shown in Fig. (5). This graph illustrates the variation in district and region-specific contraceptive usage prevalence throughout Zambia. Contraceptive use was found to be similar among women who were closer to each other than it was among those who were more distant. In various parts of the same region, the prevalence of using contraceptives was very low, low, moderate, and high. This can be verified by the four distinct colors that this figure displays. It was noted that the prevalence of contraceptive use was very low, and almost non-existent in areas such as Lusaka, Muchinga, Eastern, and Central. Fig. (5) displays the distribution of contraceptive use in Zambia at the district level. Contraceptive use was very low in the North-Western and Northern regions. The Western, Eastern, Southern, and Luapula regions exhibited a low prevalence of contraceptive use, whereas Lusaka, Muchinga, Copperbelt, and Central indicated a high prevalence.