Spatially varying relationships between birth registration and influencing factors in Kenya, using a suite of Geographically Weighted Regressions

Abstract

Background

Everyone deserves legal recognition, yet millions of children remain unregistered, with the majority (87%) residing in sub-Saharan Africa and southern Asia. Despite global efforts to improve birth registration coverage, sub-national disparities persist. Across Kenya’s 47 counties, birth registration completeness rates varies from nearly 100% to as low as 12.2%, suggesting local contextual factors are important. This study explores the influence of contextual factors on the spatially heterogeneous rates of birth registration in Kenya.

Methods

We utilized data from the 2022 Kenya Demographic and Health Survey. The association between registered births and its determinants (child factors, health care indicators, maternal, household and geographical factors) was assessed at the cluster level (villages) using four regression models: ordinary least square (OLS) and spatial local regression using Geographically Weighted Regression (GWR-single spatial scale for all predictors), Multiscale GWR (MGWR-each predictor operates at different spatial scale) and Similarity GWR (SGWR-single spatial scale for all predictors) models. Best-fit models were assessed using adjusted R², AICc and Moran’s I (residual spatial autocorrelation). The key difference between GWR and SGWR lies in how spatial dependency is measured between locations.

Results

A total of 1673 survey clusters were analysed. MGWR was the best-fitting model (AICc = 14870.57, adjusted R² = 0.40, Moran’s I = -0.04 (p-value = 0.999)) and identified localised significant relationships for all variables examined. Evidence of spatially varying relationship (local influence) was observed between birth registration, bank account ownership, and unemployment. Regional influence was observed for female-headed households, while other associations maintained a uniform relationship across the study area (global influence).

Conclusion

Determinants of birth registration vary spatially at different geographical scales, necessitating context-specific targeted strategies to boost registration coverage across diverse areas and populations.

1. Introduction

Everyone deserves legal recognition, yet millions of children remain unregistered, with the majority (87%) residing in southern Asia and sub-Saharan Africa (1,2). Birth registration is a fundamental human right that serves as the first step to establishing legal identity and unlocking access to social and health services (3). Complete civil registration and vital statistics (CRVS) systems provide the most reliable source of vital statistical data essential for policymaking and monitoring national and global health and developing effective strategies that support 12 of the 17 Sustainable Development Goals (4). However, only half (49%) of African countries have a complete CRVS system for births compared to over 88% in the developed countries (5). In Kenya, only 12 (26%) out of the 47 counties met the United Nations (UN)’ target of at least 90% birth registration completeness (percentage of registered events over the expected number of events). Consequently, national and international planning agencies default to estimations of vital events based on population sample surveys or intercensal data interpolation (6,7). However, these alternative sources are retrospective, subject to recall bias and sampling errors. Additionally, these surveys are undertaken infrequently, fail to collect data for all indicators of interest and may not be representative of smaller administrative units to facilitate more targeted planning (7,8).

Global, regional, and philanthropic efforts have tried to address birth registration gaps to strengthen the CRVS system (8–10). These initiatives have focused on global frameworks, mobilization of resources, minimum standards and guidelines, including the UN’s Principles and Recommendations for a Vital Statistics System (11) and the World Health Organization (WHO) CRVS Strategic Implementation Plan (12). Regional initiatives such as the UN Economic Commission for Africa, Africa Programme on Accelerated Improvement of Civil Registration and Vital Statistics (13), and the UN Economic and Social Commission for Asia and the Pacific (14) have helped adapt global CRVS standards to regional challenges. Philanthropic efforts such as those supported by Bloomberg Philanthropies Data for Health Initiatives have developed additional technical guides to strengthen CRVS efforts in the low- and lower-middle income countries (LLMICs) (15).

The collective efforts have yielded meaningful improvements in birth registration (16,8), however, regional and national disparities persist, delaying global efforts towards achieving universal birth registration coverage by 2030 (17). In most countries, low coverage of birth registration has been attributed to the incomplete CRVS system (8,18). Africa, for example, has the largest proportion of global births (27%), yet only 3% of these births occur in countries with complete CRVS systems (2,19,20). Further, more pronounced disparities exist between African countries, with registration rates ranging from nearly universal in Sierra Leone (92.9%) to very low in Ethiopia (2.7%) (21). This emphasizes the need for context-specific assessment of birth registration to better adapt interventions and guidelines by country.

In Kenya specifically, the national birth registration completeness has declined to 77% from 81% in 2022 (22,23). Moreover, 51% of counties fell below the national coverage levels, with some recording a mere 12% completeness rate (23). These subnational disparities suggest underlying, county-specific, localized factors influencing birth registration, which demand further interrogation.

Studies have investigated the factors contributing to disparities in birth registration rates, primarily on a global scale (21,24–27). These studies provide a general understanding of how various determinants influence birth registration. However, they assume that the association of these determinants with birth registration is uniform across different spatial contexts. This assumption, may not always hold, as the processes influencing human behaviour and attitudes toward various social services, including birth registration, can differ across geographical contexts (28).

To explore these nuanced spatial relationships, local spatial regression models (28–30) have been employed in this study. To the best of our knowledge, no study has adopted a local perspective to investigate these inequalities, particularly in Kenya. Here we apply a suite of Geographically Weighted Regression (GWR) models to examine whether the determinants of birth registration vary in their influence across space in Kenya. Specifically, we investigate which determinants show uniform (global) or spatially varying (local) associations, where these associations are strongest or weakest and their direction of influence (positive or negative). Understanding these spatial patterns is useful for more targeted, context-specific interventions to improve birth registration coverage.

2. Methodology

2.1. Study context

The study is conducted in Kenya, located in eastern Africa, with a population of about 47.6 million people as per the latest national census conducted in 2019 and projected to reach 53.3 million in 2025 (31). Nairobi, central and western regions have high population density, while the northern and southern regions are sparsely populated. Kenya has a decentralized system of governance with 47 county governments (32,33). The total fertility rate in Kenya is estimated to be 3.4 births per women, a significant drop from about 7 births per woman in 1989 (34). These rates vary sub nationally with the lowest rates in urban counties such as Nairobi and the highest among semi-arid populations in the Northeast of the country (35,36). Figure 1 shows the 47 counties of Kenya and the estimated live births per one kilometre square (37).

Figure 1. World pop estimated live births per one kilometer grid square across Kenya’s 47 counties namely:

Mombasa [1], Kwale [2], Kilifi [3], Tana River [4], Lamu [5], Taita Taveta [6], Garissa[7], Wajir [8], Mandera [9], Marsabit [10], Isiolo [11], Meru [12], Tharaka-Nithi [13], Embu [14], Kitui [15], Machakos [16], Makueni [17], Nyandarua [18], Nyeri [19], Kirinyaga [20], Murang’a [21], Kiambu [22], Turkana [23], West Pokot [24], Samburu [25], Trans Nzoia [26], Uasin Gishu [27], Elgeyo-Marakwet [28], Nandi [29], Baringo[30],Laikipia [31], Nakuru [32], Narok [33], Kajiado [34],Kericho[35], Bomet [36], Kakamega [37], Vihiga [38], Bungoma[39], Busia [40], Siaya [41], Kisumu [42], Homa Bay [43], Migori [44], Kisii [45], Nyamira [46], Nairobi [47].

2.2. Efforts to improve CRVS in Kenya

In 2019, Kenya, attained its highest national birth registration completeness rate of 89%, a significant improvement from 61% in 2017. Several initiatives that may have led to these improvements include expanding physical access to civil registration services (CRS) through additional registration offices, complemented by mobile registration centres in marginalized areas (38). In collaboration with agencies such as the United Nations High Commission for Refugees, the civil registration department has invested in strategies to provide legal identity to all citizens, including nomadic and border populations (e.g those living along Kenya-Somali border), minorities, refugees, and migrants, who can be left stateless (39,40). Integrating birth registration services into Maternal and Child Health (MCH) services such as at immunization, and promoting health facility deliveries through free maternity services where the birth registration process is automatically initiated by healthcare personnel (41,42), improved registration rates. Mandatory birth certificates for school enrolment may have also served as an incentive for parents to register their children’s births (22,43). Other ongoing initiatives aim to streamline the birth registration process through digitization and automation of civil registration systems to ensure efficient storage, application, and processing of certificates may also have improved documenting birth registration nationwide (44).

However, since 2019, birth registration completeness has declined by about 12% points, falling to 77% in 2023 (36) and significant disparities among counties persist (22,23,43). It is worth noting the reversal of the gains at the national level, especially after 2021, might be influenced by the COVID-19 pandemic that led to inaccessibility of registration centres and delays in submitting registration form as a result of lockdowns and curfews (22). Irrespective of the period, counties with the lowest birth registration rates are predominantly located in Arid and Semi-Arid Lands, where the childbearing rate is notably high, at 200 births per 1,000 women (36,45). This situation underscores the ongoing challenges faced in achieving equitable birth registration across different regions of Kenya.

2.3. Methodology Overview

Figure 2 illustrates our workflow, which consists of five main steps. In the first step, we assembled data from various sources, extracted and recoded variables to define the outcome (proportion of registered births) and its determinants (informed by literature) at to the cluster level. Second, we ran an ordinary least square (OLS) model to understand the association between birth registration and its determinants at the global scale under the assumption of uniform relationship across space. The OLS model coefficients provided a baseline for comparing with local spatial regression (46). This model violated key assumptions of OLS regression, including independence of residuals, necessitating the application of three local regression models in third step to explore spatial heterogeneity in relationships across regions (28–30).

Figure 2. Overview of approaches used in this study.

The GWR model that allows the coefficients to vary spatially was first implemented but there was significant positive residual spatial autocorrelation. Therefore, both the Multiscale GWR (MGWR) that allows the coefficients to vary at different spatial scales and the Similarity GWR (SGWR) which considers similarities in both geographical location and data attributes were tested in the fifth step. The final step involved assessment of the model fit based on Adjusted R², Corrected Akaike Information Criterion (AICc) and residual spatial autocorrelation.

2.4. Data

2.4.1. Registration Coverage Data

Data on birth registration were obtained from the 2022 Kenya Demographic and Health Survey (KDHS) (45), a nationally representative cross-sectional survey, conducted from 17^th February to 31^st July 2022. The survey was implemented in a total of 1691 clusters. These clusters are georeferenced and geographically randomly displaced to protect the confidentiality of the respondents. Specifically, urban clusters are displaced up to 2 kilometres and rural clusters up to 5 kilometres, with a further 1% of randomly selected clusters displaced up to 10 kilometres (47). The displacement is restricted so that the points stay within the counties.

We used information from the children’s (under 5) and household modules for living children who are usual residents of their households. This age group was selected to minimize recall bias and align with the target demographic group for all the independent variables analysed.

2.4.2. Dependent variable

The outcome variable for this study was the proportion of children under the age of 3 years registered with the civil authority. Birth registration is the official recording of a child’s birth by the government, initiated by the primary guardian to establish legal identity. In Kenya, registration must occur within six months, with late registration requiring additional verification and fees. A child was considered registered if the respondent (women 15-49 years) stated that the birth was registered regardless of whether a birth certificate was available. At the cluster level, we aggregated the individual level data to produce weighted numbers of all children that were surveyed (denominator) and among them those registered (numerator), from which we derived empirical proportions of children registered (Figure 3).

Figure 3. Spatial distribution of birth registration for children under 3 years across clusters in Kenya.

The blue dots indicate high birth registration rate predominantly in the central, western and coastal regions while the red dots indicate low birth registration rates primarily in the north eastern regions.

2.4.3. Independent variables

The independent variables included in the analysis are detailed in Table 1. These variables were informed by literature (21,25,48–56) and cover a range of demographic and socio-economic conditions which may have direct or proximate relationships with registration rates in Kenya. The variables extracted from the survey were categorized as child factors, including demographic characteristics related to children, indicators reflecting access and utilization of health care services, maternal characteristics and behaviour, and household factors. Among the various ethnic and religious groups in Kenya, we selected the Kikuyu ethnic group (the largest in Kenya) as a proxy for ethnicity and Catholicism as a proxy for religion, based on sample size, geographic distribution, and relevance to the civil registration process. Weighted proportions of each independent variable were computed at cluster level where the numerator represented weighted number of individuals of interest (e.,g educated mothers, children delivered in hospitals), and the denominator represented the total weighted number of relevant target group (all mothers, children). Geographical factors (travel time & urbanicity) were obtained from externally modelled surfaces, as they were not available as part of the KDHS. Specifically, data on urbanization was extracted from the 2023 Global Human Settlement Raster Layer (GHSL) and reclassified into urban areas (cities, large settlements, dense and semi-dense towns and suburbs) and rural areas (villages, dispersed areas and very low-density areas) (57,58). The average travel time to civil registration offices (CROs) and health facilities was computed in hours based on a cost friction surface (59–61).

Table 1. Description of variables included in the global and local spatial analysis.

Categories	Variables (cluster level)	Description
	Dependent variable	Proportion of children under 3 years registered.
	Independent variables
Child factors	Child’s age	Proportion of children under 1 years of age.
	Sex of the child	Proportion of female children.
	Birth order	Proportion of 4th born and higher children.
	Birth interval	Proportion of children with 8-23 months birth interval.
Health care utilization indicators	Birth attendant	Proportion of children attended to by skilled health personnel.
	Place of birth	Proportion of children born at health facility.
	Immunization status	Proportion of children who are fully immunized.
Maternal factors	Mother’s age	Proportion of women aged 35 and above.
	Child mortality experience	Proportion of women with history of child loss.
	Total children born ever (Parity)	Proportion of women with 4 or more children.
	Maternal education	Proportion of women with no education.
	Maternal occupation	Proportion of women not working.
	Frequency of access to media	Proportion of women who access media (either radio, newspaper or TV) at least once per week in the cluster.
Household factors	Household head	Proportion of households with females as household heads.
	Bank account ownership	Proportion of households with at least one bank account holder.
	Media availability	Proportion of households that have all three media devices (TV, Phone, Radio).
	Wealth index*	Proportion of poorest households.
	Religion	Proportion of households practicing Catholic faith.
	Ethnicity	Proportion of households belonging to Kikuyu (largest ethnic group in Kenya) ethnic group per cluster.
Geographical factors**	Urbanicity	Cluster level variable that defines whether a cluster is urban or rural (57,58).
	T ravel time to CROS	Average travel time to civil registration offices adjusted for displacement.
	Travel time to health facilities (in hours)	Average travel time to health facilities (62).

^*A composite measure of the cumulative living standard of a household, computed using the principal component analysis (PCA) using data on household’s ownership of selected set of assets, such as televisions, bicycles, and cars; dwelling characteristics such as flooring material etc. Households are divided into five wealth quintiles (poorest, poor, medium, richer, richest) based on their wealth scores (63).

^**These variables were not extracted from the survey but sourced from externally modelled raster surfaces. To account for coordinate displacement, averages of travel time were taken within 2 or 5km buffers and majority classification was used to classify the cluster as urban or rural based on the GHSL layer

2.5. Spatial Modelling Approaches

2.5.1. Global regression analysis

To understand the broad association between birth registration and the determinants at the cluster level, we used an OLS linear model (Equation 1). This model assumes a consistent relationship between the dependent and independent variables across all data points and yields a single parameter for each association. The OLS model (combination of different determinants), that best explained the variability in birth registration, was selected based on the lowest AICc using backward stepwise regression implemented in R via StepReg package (version 1.6.1).

$y={\beta }_0+{\displaystyle \sum }_{k=1}^p{\beta }_k{x}_k+\epsilon$ Equation 1. OLS linear regression model

Where y is the dependent variable, x_k: k-th independent variable, ranging from 1 to p (number of independent variables), β₀: the intercept, β_k: regression coefficient for k-th independent variable, ε: The error term.

OLS assumptions

We conducted several statistical tests to assess whether assumptions of the OLS linear model were violated including multicollinearity among the independent variables, assessed using the variance inflation factor (VIF) cut off of 10 (64). The details are provided in the Supplementary file and Figures S1 and S2.

2.5.2. Local spatial analysis

Presence of spatially correlated residuals not adequately captured by the global OLS model violates the OLS assumption that residuals are independent and identically distributed (28). This suggest that the underlying processes may be spatially heterogenous and influenced by local contextual factors. To explore this spatial heterogeneity in relationships, we applied local regression that allow coefficients to vary from one location to another. These models are fitted at each location using data whose influence diminishes with distance (Tobler’s first law of geography) (28). We therefore fitted three local models:

Geographically Weighted Regression

The GWR (Equation 2) allowed relationships to vary from one location to another at the same spatial scale (single bandwidth) (65,66).

$y_i={\beta }_{i0}+{\displaystyle \sum }_{k=1}^p{\beta }_{ik}{x}_{ik}+{\epsilon }_i$ Equation 2. Standard GWR model

Where y_i is the dependent variable at location i: x_ik is the k-th independent variable at location i;p is the number of independent variables;β_i0 is the intercept parameter at location i; β_ik is the local regression coefficient for the k-th independent variable at location i; and ε_i is the random error at location i.

The local parameters are estimated by weighted least squares (Equation 3) (65).

${\widehat{\beta }}_i={\left(X^T{W}_{i}X\right)}^{-1}{X}^T{W}_{i}y$ Equation 3.GWR coefficient estimator

Where X is matrix of independent variables, including a column of intercepts represented as 1s; y is the dependent variable vector; for a set of p independent variables ${\widehat{\beta }}_i={\left({\beta }_{i0},\dots {\beta }_{ip}\right)}^T$ is the vector of p +1 local regression coefficients; and W_i is the diagonal matrix representing each observed data’s geographical weighting for regression point i.

The local weighting scheme (W_i) is calibrated using a kernel function, that determines how quickly the influence of nearby points on a regression point decreases with distance and a bandwidth that controls the extent of weighting. This study used an adaptive bi-square kernel (Equation 4) with a nearest-neighbour measure of proximity, allowing the bandwidth h_j (distance within which observations are considered) to vary based on the local density of the points. This approach is particularly suitable for data like the KDHS where sampled points may not be evenly distributed.

$w_{ij}=\{\begin{array}{cc}{\left[1-{\left(\frac{d_{ij}}{h_j}\right)}^2\right]}^2,& if{d}_{ij}<h_j\\ 0,\hfill & if{d}_{ij}\ge h_{ij}\end{array}$ Equation 4. Bi-square kernel weight estimator

Where w_ij is the bi-square kernel weight for observation i at distance d_ij from the regression point j. h_j represents the adaptive bandwidth, defined as the distance to the k-th nearest neighbor for regression point j. The weights decrease with distance d_ij from the regression point to zero beyond the adaptive bandwidth h_j.

An optimal bandwidth was selected based on the lowest (AICc), which corrects AIC’s bias in small sample sizes. Points with significant associations were identified where the t-value was larger than ±1.96 at the 95% confidence interval following standard MGWR practice, as local models do not provide adjusted t- or p-values.

Multiscale Geographically Weighted Regression

MGWR (Equation 5) extends GWR by allowing relationships between variables to vary across different spatial scales (local to global), using distinct bandwidths for each explanatory variable (28,30) as detailed in the Supplementary file.

$y_i={\beta }_{bwi0}{x}_{i0}+{\displaystyle \sum }_{k=1}^p{\beta }_{bwik}{x}_{ik}+{\epsilon }_i$ Equation 5. MGWR model

Where yi is the dependent variable at location i:xik is the k-th independent variable at location i;p is the number of independent variables; β_bwik0x_i0 is the intercept parameter at location i; β_bwik is the local regression coefficient for the k-th independent variable at location i; and εi is the random error at location i. The equation is similar to GWR (Equation 3) but with bwk included in the β_bwik indicating bandwidths used for calibrating each k-th conditional relationship._bwik

The MGWR results were categorized based on the number of neighbouring points considered when estimating local parameters (bandwidths) that indicate their scale of influence (global, regional, local) (28,67). A variable was categorized as having global influence if its bandwidth exceeded 75% of the global bandwidth (0.75 * 1673 = 1254). Conversely, variables with bandwidths less than 25% of the global bandwidth (< 418) were deemed to exert local influence. Variables with bandwidths ranging between 418 and 1254 were categorized as having regional influence. Additionally, we evaluated the impact of each variable (level of influence) within the study area, calculated as the percentage of significant clusters associated with each variable.

Similarity and Geographically Weighted Regression

The GWR and MGWR models are based on geographical similarity, where locations close to one another tend to have similar characteristics to those far away. However, even if two points are close, their data characteristics (e.g. registration rates) may differ due to inherent contextual characteristics at these locations (29). For instance, one area could develop at a higher rate than another. The GWR and MGWR may overlook these dissimilarities in data, even in nearby locations, due to the assumption that proximity equals similarity. As illustrated in Equation 6, the SGWR model addresses this limitation by combining the geographically weighted matrix produced in GWR (Equation 3), with a similarity weight matrix. The resulting final matrix (W_GSi) comprises a parameter (Supplementary file, Equation S1) that controls the contribution of each matrix. Details of this model are published elsewhere (29).

${\widehat{\beta }}_i={\left(X^T{W}_{GSi}X\right)}^{-1}{X}^T{W}_{GSiy}$ Equation 6. SGWR coefficient estimator

Model fit assessment

All regression models (OLS, GWR, SGWR and MGWR) were assessed for the goodness of fit using adjusted R² and AICc. Both one- and two-sided Moran’s I tests were conducted on model residuals to verify spatial independence. The best-performing model, defined as the one with the lowest AICc value, highest adjusted R² value and minimal spatial autocorrelation, was retained for further analysis.

The analysis was conducted on a Windows 11 PC with an Intel Core i7 processor and 16 GB RAM. The OLS model was fitted using the Stats package in R (version 4.4.0); GWR and MGWR models were implemented using the GWmodel package (version 2.3-3) in R (64), and the SGWR model was implemented as detailed in Lessani and Li (68). All data preparation and visualization were performed using Stata/SE 18.0 (StataCorp LLC, College Station, TX, USA), R software (version 4.4.0), and ArcGIS Pro (version 3.0.3; Esri, Redlands, CA, USA).

3. Results

3.1. Spatial distribution of birth registration

A total of 1,673 clusters, which included children under 3 years (Figure 3), with a complete set of data across all independent variables were included in the analysis. The average birth registration rate for children under 3 years was 76.8% (range 0.0% to 100.0%), and the median rate was 83.3% (IQR; 36.4). The spatial distribution of children registered exhibited significant clustering, as indicated by Global Moran’s I 0.3 (p-value <0.001). Overall, 51.0% of the clusters had birth registration rates exceeding 80% and were concentrated in the central, western and coastal regions. Only 4.1% of the clusters had registration rates below 20%. The distribution of registration rates across the remaining clusters was as follows: 6.3% had rates between 20%-40%, 13.2% between 40%-60%, and 25.4% between 60%-80%.

3.2. Performance of regression models

Table 2 shows the performance of the three spatial regression models compared to the global model. With respect to the metrics used to evaluate the performance of the models at the cluster level (adjusted R², AICc and Moran’s I), overall, all the spatial regression models outperformed the OLS model. Among the local spatial regression models, MGWR had the best fit, with the lowest AICc of 14870.57 and the highest adjusted R² of 0.40, indicating that 40% of the variation in birth registration was explained by the independent variables included in the model.

Table 2. Performance of OLS and local spatial regression models (GWR, SGWR & MGWR).

Metric	Model	OLS	GWR	SGWR	MGWR
Adjusted R2		0.24	0.33	0.34	0.40
AICc		15135.39	15047.38	15046.98	14870.57
Moran’s, I value (one-sided test)		0.17 (p-value < 0.001)	0.08 (p-value = 0.001)	0.08 (p-value=0.001)	-0.04 (p-value= 0.999)
Moran’s, I value (two-sided test)		0.17 (p-value < 0.001)	0.08 (p-value = 0.001)	0.08 (p-value=0.001)	-0.04 (p-value= 0.02)

MGWR was also the only model with no significant positive residual spatial autocorrelation, Moran’s I -0.04 (p-value= 0.999) under one-sided test. Under two-sided test, the results were consistent showing that MGWR reduced residual spatial autocorrelation more than other models.

This indicates that MGWR effectively accounted for spatial relationship in the data, making it the preferred local model for the results presented here. Boxplots showing the distribution of coefficients across all the models is shown in the Supplementary file, Figure S3. For spatial context, the travel time variable was selected and visualized in the map (Supplementary file, Figure S4).

3.4. Summary of OLS and MGWR results

The backward stepwise process resulted in the removal of 8 variables: the child’s age, sex, and birth interval; maternal factors: birth attendant, child loss history, and total number of children ever born (Parity); and household factors: frequency of access to media and media availability at households.

A summary of OLS and MGWR results for the 15 variables is shown in Table 3. The OLS model shows that nearly all factors were associated with birth registration (p<0.05) except for unemployment of women (mother) and average travel time to health facilities. From this model, an increase in the proportion of children born in health facilities, those fully immunized, the proportion of older women (35+), married women, female-headed households, households with bank account holders, those belonging to Kikuyu ethnic group and catholic faith were associated with high birth registration rates. Conversely, an increase in the proportion of younger children in large families (birth order 4+), uneducated women, poorest households and average travel time to CROs and urban clusters are associated with low birth registration rates.

Table 3. A summary of OLS and MGWR coefficients and proportion of significant clusters.

	OLS		MGWR
Variables	Coef [95% CI] 1	VIF	Mean	Min	Max	Bandwidth†	% of significant Clusters (n=1673)
(Intercept)	0.00 [-0.04, 0.04]	1.75	47.30	2.48	73.04	59	_
Birth order (4+)	-0.08 [-0.14, -0.02]**	2.53	-0.03	-0.04	-0.03	1670	59.8
Children born at a health facility	0.23 [0.17, 0.30]***	1.12	0.21	0.17	0.23	1615	95.2
Children fully immunized	0.06 [0.01, 0.10]*	1.29	0.05	0.02	0.08	1323	80.9
Women aged 35 (+)	0.08 [0.04, 0.13]***	1.29	0.07	0.06	0.07	1670	62.3
Married women	0.06 [0.01, 0.11]*	2.86	0.09	0.05	0.13	1320	96.4
Uneducated women	-0.08 [-0.15, -0.01]*	1.40	-0.04	-0.05	-0.04	1670	5.5
Unemployed women	-0.04 [-0.09, 0.01]	1.21	-0.07	-0.54	0.52	75	54.8
Female-headed households	0.06 [0.01, 0.10]*	1.59	0.06	-0.05	0.13	782	56.0
Households with bank account holders	0.10 [0.05, 0.15]***	2.76	0.10	-0.08	0.28	212	52.3
Poorest households	-0.10 [-0.17, -0.03]**	1.17	-0.09	-0.11	-0.06	1558	42.7
Households belonging to Kikuyu ethnic group.	0.05 [0.00, 0.09]*	1.04	0.06	0.05	0.06	1670	25.2
Households practicing catholic faith.	0.05 [0.00, 0.09]*	2.05	0.05	0.05	0.06	1670	48.5
Average travel time to health facilities (hrs) per cluster	0.06 [0.00, 0.12]	2.10	2.81	1.70	3.98	1663	43.6
Average travel time CROs (hrs) per cluster	-0.06 [-0.12, 0.00]*	1.47	-0.70	-1.11	-0.61	1670	46.0
Urbanicity (whether a cluster is urban or not)	-0.07 [-0.12, -0.02]**	1.75	-1.31	-1.92	-0.06	1659	53.3

*p<0.05; **p<0.01; ***p<0.001

^†Number of neighboring points considered when estimating local parameters

The OLS model residuals were spatially autocorrelated (Moran’s I = 0.17, p < 0.001) and justified using spatial regression models to explore potential spatial heterogeneity in the determinants. Other assumptions violated include non-constant variances in the residuals errors (heteroscedasticity) (Breusch-Pagan test, X² = 37.19, p<0.001) and normality assumption (Jarque-Bera test X² = 443.9, p<0.001). However, multicollinearity was not an issue since all the VIF values for all independent variables included in the model were far below the 10 VIF cut-off.

Majority of the variables (n = 13) had a uniform relationship with birth registration across the study area (global influence). Local influences were observed for the proportion of unemployed women and households with at least one bank account holder with a bandwidth of 75 and 212 nearest neighbours, respectively. A regional scale of influence defined as bandwidths between 418 and 1254 was only observed in female-headed households, indicating moderate spatial variation.

The level of influence of the determinants varied from a low of 5.5% for the proportion of women with no education to a high of over 95% for the proportion of children born in health facilities and the proportion of married women. The percentages for other variables varied between 25% and 62% of the clusters, as shown in Table 3, and the corresponding spatial distribution of significant coefficients is mapped in Figures 4A-Q.

Figure 4. Spatial distribution of significant coefficients across clusters.

Panels A-G show positive associations (shades of blue), H-M negative association (shades of red) and N-Q show mixed associations and non-significant coefficients (p-value > 0.05) in grey. Most variables (A-M) exhibited global influence (bandwidths >1254 neighbors). Regional influence (bandwidths 418–1254 neighbors) only for female-headed households (panel P). Local influence (bandwith < 418) across unemployed women and bank account holders (panles N and Q).

3.5. Spatial distribution of MGWR coefficients

The spatial distribution of MGWR coefficients was categorized based on the direction of association with birth registration, which included positive (Figures 4A-G), negative (Figures 4H-M), and mixed associations (Figures 4N-Q). The coefficients of all determinants were mapped regardless of the scale of influence (i.e., global, regional, or local).

Positive association

Figures 4A-C show that indicators of healthcare utilization (facility-based deliveries and complete immunization of children) and married women had a near-universal positive influence, with significant effects in at least 80% of the clusters (Table 3). The strength of association were strongest in the central regions of the country, and extended to northeastern and western regions for facility-based deliveries.

The influence of older women (35+) was observed across 62.3% of the clusters with minimal spatial variation in strength across regions (Table 3, Figure 4D). Households affiliated with the Kikuyu ethnic group and catholic faith, showed significant influence across 25.2% and 48.5% of clusters respectively (Table 3). The influence of Kikuyu households was geographically concentrated in central Kenya, while Catholic households exhibited significant associations across all regions except the northeastern parts (Figure 4E & F). MGWR revealed that an increase in average travel time to health facilities was associated with a rise in birth registration across 43.6% of clusters (Table 3, Figure 4G). This unexpected trend had its strongest association in the southwestern regions of the country, including all the coastal area counties.

Negative association

Higher proportions of younger children in large family sizes (birth order 4+), the poorest households, and longer average travel time to CROs were significantly associated with low birth registration across 59.8, 42.7%, and 46.0% of the clusters (Table 3, Figures 4H, J & K). Among these, the negative impact of poverty on birth registration was strongest in the northern regions of the country. The effect of uneducated women was limited showing significance in only 5.5% of the clusters (Table 3, Figure 4L). In contrast, urbanicity showed a more substantial negative influence, across 53.3% of the clusters (Table 3), with strongest effects concentrated in the central and western regions (Figure 4M).

Mixed association

The unemployment rate of women (mothers) demonstrated a distinctly mixed association with birth registration across 54.8% of the clusters (Table 3, Figure 4N). Specifically, a strong negative association was observed in the northeastern, southern, and certain western regions of the country, while a significant positive association was noted in parts of western Kenya and some coastal counties (Kilifi, Tana River, and Lamu).

Similarly, the proportion of female-headed households and households with bank account holders showed mixed associations in 56.0 % and 52.8% of the clusters, respectively (Table 3, Figures 4P and Q). However, significant negative associations were limited to a small number of clusters primarily in the northeastern regions for female-headed households and in the central and southern regions of the country for households with bank account holders.

4. Discussion

Kenya has yet to achieve complete birth registration coverage, with some counties displaying alarmingly low coverage below 20% (22,43). This incompleteness undermines the government’s ability to accurately assess population demographics, limiting evidence-based policy-making and effective planning, including public health. Many global studies on determinants of birth registration assume constant determinants across regions. However, local factors and attitudes can vary significantly influencing spatial heterogeneity in birth registration coverage, as evidenced by our findings in Kenya (Moran’s I 0.3; p-value <0.001). Using the best fitting local model (MGWR), we highlight the importance of spatial variations in the relationship between birth registration rates and their determinants.

Spatially varying relationships (Local influence) was observed between birth registration and women’s unemployment and bank account ownership in both directions of influence (positive and negative). Factors such as social protection programmes in Turkana County, located in the northwestern part of the country, necessitate birth registration to access cash transfers and open bank accounts, as these require birth certificates to prevent fraud and may explain the positive influence (69). Similarly, a birth certificate or national ID is necessary to open a bank account. Adults engaging with these systems may enhance household awareness of the importance of birth registration, paralleling findings from Nigeria (25). However, this association varies, with many northern, northeastern, and some western counties showing no significant link or negative influence, likely due to limited financial services access in arid regions like Garissa, Wajir, and Tana River (70). Conversely, urban areas like Nairobi and agricultural regions like Kirinyaga (central Kenya), with better financial service access, show stronger positive associations.

The association between households headed by females and birth registration varied regionally, reflecting broader spatial bandwidths that capture broader contextual influences. Overall female-headed households were associated with higher birth registration. While poverty can impede birth registration in these households, when financial barriers are removed, they often demonstrate increased maternal autonomy—specifically, a woman’s capacity to manage household economic resources (71). Women in such roles, are more likely to prioritize their children’s needs, including birth registration, over other demands (72,73). The rise in registered births among single mothers in Kenya between 2018 to 2022 in Kenya (74) may reflect increased maternal autonomy and sensitization campaigns. The strong association in certain counties in western Kenya may relate to high male migration rates to urban areas in search of employment (75), leaving women as temporary heads of households. However, in northern counties like Mandera, cultural practices and religious beliefs, such as women’s remarriage, may reduce female-headed households, contributing to insignificant association in those areas.

The MGWR analysis showed that all the determinants that had a consistent positive association with birth registration maintained a uniform relationship across the study area (global influence), aligning with the global model. However, the level of influence varied with health care utilization indicators (facility-based deliveries and full immunization of children), and marriage as the most significant predictor of birth registration, covering at least 80% of the clusters. This observation highlights the role of matrimony in influencing birth registration through financial and social stability provided by partners and the transfer of information from partners who are knowledgeable about the birth registration process (21). Furthermore, incorporating interventions to enhance maternal and child healthcare programmes alongside birth registration could effectively improve national coverage. This is in line with WHO-led global initiatives that advocate for integrating health services with the CRVS to boost birth registration coverage (12).

The influence of religious practices and cultural and historical norms on birth registration was geographically specific, varying according to ethnic and religious affiliations. Households belonging to the Kikuyu ethnic group (the largest in the country) demonstrated significant positive associations concentrated in the central part of Kenya, where the community is indigenous. Furthermore, this region holds historical importance for birth registration, as Nyeri, the headquarters of the former central province, was among the first districts (now counties) alongside Nairobi to implement birth registration after Kenya gained independence in 1963 (22,43). This early adoption may have fostered a strong awareness of birth registration among the Kikuyu community. The positive correlation observed with the Catholic faith in other areas may be attributed to the practice of issuing baptismal certificates for young children, which can facilitate birth registration and identification documents. These religious documents are accepted as alternatives for late birth registration (beyond six months after birth) (76,77).

Notably, the local model showed that increased average travel time to health facilities was associated with higher birth registration rates. This observation, although counterintuitive, had been overlooked in the global model. A possible explanation for this pattern could be the predominance of lower-level facilities in Kenya (dispensaries, health centres), many of which may lack delivery services or are not recognized by the government for birth registration services (78). Additionally, individuals may have preferred hospitals for delivery and be willing to incur one-off travel costs to reach these facilities, regardless of distance.

All variables associated with low birth registration rates also had global influences. However, the level of influence of uneducated women was low (significant in only 5.5% of the clusters), indicating that this could be a less critical barrier to birth registration. Instead, focus should be directed towards other barriers, such as addressing socioeconomic challenges that disadvantage younger children in large family sizes (birth order 4+), poverty, particularly in the northern parts of the country, and improving coverage to CROs. The local model also highlights the challenges urban areas, particularly Nairobi, central, and western Kenya, face in birth registration due to urbanisation, which may dilute the perceived benefits of better service access in these urban areas (53,54,79). About 60% of Kenya’s urban households live in informal settlements with limited access to services and high rates of poverty that lead to prioritization of immediate needs over administrative tasks (80). The focus has been on rural areas, where targeted outreach programs emphasize registration, whereas urban settings may lack such initiatives. Further, overburdened civil registration offices in dense urban centers can deter registration due to delays (44).

This study has offered important insights into the geographical context for evaluating the relationship between birth registration and its determinants. Using local regression models, we show that the association between birth registration and its determinants is heterogenous varying across different spatial scales (global, regional, and local) which would have been obscured by global models. However, several limitations should be considered when interpreting these findings.

Limitations

The OLS global model violated key assumptions, potentially introducing bias (81) while MGWR model offered a more flexible framework for detecting spatially varying relationships and reducing spatial dependency in the residuals. However, causal inferences remain limited without complementary qualitative data. Local spatial analyses were conducted at DHS cluster level using weighted individual-level data for representativeness. While this approach supports robust spatial analysis, there are limitations related to ecological inference as individual-level interpretations cannot be directly inferred from aggregated data. The analysis did not apply formal correction for multiple local tests (e.g., False Discovery Rate control), as such adjustments are not implemented in GWR/MGWR frameworks. This may increase the likelihood of Type I error in local significance testing.

Other additional limitations should be noted. First, birth registration status was self-reported without document verification which may lead to misclassification particularly if respondents confuse religious documents or birth notifications slips as birth registration documents. However, for health facilities births are typically registered after issuance of birth notification slips, unless delays in registration occur due to systematic issues but we had no means to assess the extent of this gap. Second, the geographical displacement of DHS clusters to protect confidentiality, may affect the spatial precision of our findings. Third, the analysis included all health facilities in Kenya in modelling travel time because we couldn’t distinguish those providing delivery or birth registration services from those that are not. This may underestimate the average travel time in areas where only a subset of health facilities offers services. Individuals may also bypass nearby facilities in favour of those farther away. As such the direction and magnitude of bias are context-dependent. Future analyses could classify facilities by service level and type (public vs. private) to more effectively assess associations between facility accessibility, measured by travel time, and birth registration rates. Fourth while we used Kikuyu and catholic as proxies for ethnicity and religion among several groups, we acknowledge that cultural practices may vary. However, we do not expect substantially different results from minority groups. Lastly, while subnational factors such as administrative capacity or outreach programmes were not included due to data constraints, MGWR model diagnostic indicated no significant positive residual spatial autocorrelation, suggesting that major spatial patterns were adequately captured and the risk of spatial confounding is minimal.

5. Conclusion

This study shows that one-size-fits-all approach may be ineffective for enhancing CRVS systems and for improving birth registration coverage across diverse regions of Kenya. Each location in Kenya exhibited unique characteristics influencing behaviours and attitudes towards birth registration, necessitating context-specific strategies to address the long-standing gaps. Future qualitative studies may complement this study by providing additional insights into the observed patterns.

Insights from this study suggest that the CRS should emphasize on the need to integrate birth registration with MCH services, including establishing a one-stop shop for registration at health facilities to streamline processes and improve accessibility. The rolling out of additional CROs should also be prioritized in regions with poor access and using mobile registration centers as a temporary solution. Reducing or waiving registration fees for low-income families, particularly those with larger family sizes, is recommended to address socioeconomic barriers. The CRS should utilize informal knowledge networks such local community health workers, churches, village elders, or external humanitarian programmes to raise awareness about birth registration within the community. The CRS should not ignore emerging urban challenges to birth registration, such as urban poverty and congestion. Investing in infrastructure and digital solutions could effectively manage the demand in urban areas. Finally, emphasis should be placed on scaling these interventions in areas where gaps are most pronounced.

List of Abbreviations

AICc

Corrected Akaike Information Criterion

CROs

Civil Registration Offices

CRS

Civil Registration Services

CRVS

Civil Registration and Vital Statistics Systems

GHSL

Global Human Settlement Raster Layer (GHSL)

GWR

Geographically Weighted Regression

IQR

Interquartile Range

KDHS

Kenya Demographic and Health Survey

MCH

Maternal and Child Health

MGWR

Multiscale Geographically Weighted Regression

OLS

Ordinary Least Squares

SGWR

Similarity Geographically Weighted Regression

UN

United Nations

VIF

Variance Inflation Factor

WHO

World Health Organization

Availability of data and material

The datasets used in this study were extracted with authorized access provided by the Demographic and Health Survey (https://dhsprogram.com/data/). The data can be availed by the authors upon formal request.