Correlates of the number of children ever born to women in rural Ethiopia: Application of truncated generalized Poisson model with exposure

doi:10.21203/rs.3.rs-5014035/v1

Download PDF

Research Article

Correlates of the number of children ever born to women in rural Ethiopia: Application of truncated generalized Poisson model with exposure

https://doi.org/10.21203/rs.3.rs-5014035/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

Rapid population growth and high fertility rates in low-income countries adversely affect the provision of maternal and child healthcare services and are roadblocks to the achievement of sustainable development goals. A high fertility rate has serious health implications for both the mother and the children she bears, and thus, investigating the factors behind this phenomenon is of paramount importance.

Method

The data for this study were extracted from the 2016 Ethiopia Demographic and Health Survey (EDHS 2016). The response variable was the number of children ever born (NCEB) per woman in rural Ethiopia. This responsible variable was zero-truncated since only women who had at least one live birth at the time of the survey were considered. Data from a total of 6256 women were analyzed using a zero-truncated generalized Poisson model, which takes into account any type of dispersion and the truncated nature of the response variable simultaneously.

Results

The mean NCEB for reproductive-age women residing in rural areas of Ethiopia was found to be about 4.5. Factors that are associated with bearing more children include early age at first birth, number of deceased children in the family, low economic status of the household and land ownership. On the other hand, contraceptive use, women’s education and media exposure had a negative impact on the NCEB per woman. The results also revealed significant regional variation, with women in the Somali region registering the highest number of child births.

Conclusion

To curb high maternal fertility, interventions that selectively target regions with high child births (e.g., Somali), uneducated women and poorest women; awareness creation campaigns to combat early initiation of childbearing, particularly teenage pregnancies; unreserved efforts aimed at reducing child mortality; and promoting the use of birth control measures are recommended.

Fertility

Number of children ever born

Truncated Poisson model

Truncated generalized Poisson model

Exposure

The global human population has increased by more than threefold from what it was in the 1950s, that is, it reached eight billion in 2022 from an estimated 2.5 billion people in 1950. According to UN estimates, this figure is expected to rise to 8.5 billion and 9.7 billion by 2030 and 2050, respectively. These projections are made based on the assumption of a decline in fertility for countries with large family sizes and a slight increase in fertility for countries with an average of fewer than two children per woman [1]. Sub-Saharan Africa is expected to become the most populous of the eight geographic regions in the late 2060s, and its population is expected to reach 3.44 billion by 2100 [2].

One of the crucial determinants of population dynamics is fertility. High maternal fertility has an adverse effect on the provision of maternal and child healthcare services. In several countries of the world, the fertility level (the average number of births per woman in her lifetime) has fallen markedly over recent decades. The average global fertility rate, which was about five births per woman in the mid-twentieth century, declined to 2.3 in 2021. This figure is expected to further decline to 2.1 (95% confidence interval: 1.88–2.42) births per woman by 2050. In 2021, sub-Saharan Africa was the region with the highest fertility level, with 4.6 births per woman. With average fertility levels projected to be around three births per woman in 2050, sub-Saharan Africa will account for more than half of the growth of the world’s population between 2022 and 2050 [2].

When we come to Ethiopia, the total population increased from around 67 million to over 126 million from 2000 to 2023. The share of the rural population was 76.84% in 2023. This increase in the total population occurred despite a decline in the population growth rate by 0.44% during the same period. The fertility rate (births per woman) has declined from 6.374 in 2002 to 4.063 in 2022, which is a reduction of roughly 57% within 20 years [3]. Even though this is a remarkable achievement and relatively lower than the sub-Saharan average (4.6 children per woman), it is still far higher than the global average (2.3 children per woman). Such rapid population growth and high fertility are clearly roadblocks to the achievement of sustainable development goals (SDGs).

In the literature, the number of children ever born (NCEB), which refers to the number of children born alive among women of reproductive age, is commonly used as an indicator and measure of women’s lifetime fertility experience. NCEB is count response data and is typically modelled using Poisson regression. However, owing to the restrictive assumption of equi-dispersion, this distribution is inappropriate for data that exhibit overdispersion or underdispersion. The negative binomial distribution is often used to account for the presence of overdispersion. However, there are further issues that constrain the application of these models. These are the possibility of underdispersion and the truncated nature of the data.

In this study, NCEB is a count response variable that was collected by the EDHS considering reproductive women who had given at least one birth. Thus, this variable is truncated (specifically zero-truncated) during the sampling process since the counts begin with one with no possibility for zero counts. One advantage of this truncation is that it excludes women who have not yet delivered a child due to personal problems (e.g., physical infertility). Moreover, the outcome variable may exhibit underdispersion as well. In such situations, estimating standard Poisson or negative binomial regression models is inappropriate [4–9]. The right choice is the truncated generalized Poisson (TGP) model, which can accommodate both over- and under-dispersion as well as take into account the truncated nature of the data [4, 10].

Empirical studies that have explored the relationships between NCEB and various demographic, social, and economic characteristics in Ethiopia and elsewhere have utilized the standard Poisson and negative binomial regression models [11, 12], multilevel versions of these models [13, 14], zero-truncated Poisson (ZTP) and zero-truncated negative binomial (ZTNB) regression models [15] and generalized Poisson (GP) regression [7, 16]. None of these approaches take into account underdispersion and truncation of the outcome variable at the same time. Thus, the main objective of this study is to identify and analyze the correlates of NCEB in Ethiopia using a zero-truncated generalized Poisson (ZTGP) model that accounts for these two issues simultaneously. The analysis is specifically for women in rural areas of the country since more than three quarters of the Ethiopian population resides in rural areas where the level of fertility is a serious issue.

The other point is the exposure time. Models that do not include exposure time implicitly assume that each subject was “at risk” of an event occurring for the same amount of time. However, this assumption may not be plausible. For example, we might predict more births just because a woman is exposed to risk (that is, pregnancy) for a longer period of time. Thus, failing to control for exposure time could lead to misleading results. Including an exposure variable, such as cohabitation duration (number of years at risk), allows the counts of children to be comparable across subjects that are observed for different durations of time. An illustration of count data models incorporating an exposure variable for insurance claims data is discussed in Faroughi et al. [17].

This paper is organized as follows: Section one presents the background and rationale of the study. The second section discusses the source of the data, the response and explanatory variables and the specifications of various count data models. The results and discussion of the study findings are presented in Section three. Section four presents conclusions and recommendations.

2.1 Sources of data and variables

The data utilized in this study were obtained from the 2016 Ethiopian Demographic and Health Survey (EDHS 2016). The study population comprises all reproductive-age women (15–49 years) who reside in rural areas of the country and had given at least one birth at the time of the survey.

The response variable is the number of children ever born (NCEB) (alive) per woman in rural areas of Ethiopia, that is, it is the sum of the number of children surviving and dead per woman at the time of the survey. The explanatory variables included in the study are women’s current age, age at first sex, age at first birth and years of education, husband/spouse years of education, region, religion, wealth status, number of years of contraceptive use, number of deceased sons/daughters, land ownership status and frequency of listening to the radio. Moreover, the number of years since first cohabitation (that is, cohabitation duration), which adjusts for the amount of opportunity an event (birth) has, is used as an exposure variable. Ideally, the mean NCEB is expected to be proportional to the size of this exposure.

2.2 Models for count data

2.2.1 Poisson regression model

Poisson regression is often used to model count data. Suppose that are independent random variables. Let denote the value of an event count outcome variable for the i^th subject (in our case, the number of children ever born for the i^th mother) with mean ${\theta _i}$. The probability mass function (PMF) of a Poisson distributed response variable ${Y_i}$ is given by:

The parameters ${\theta _i}$ are related to the covariates ${X_i}=({x_{i1}},\,\,{x_{i2}},\,\,...,\,\,{x_{ip}})^{\prime}\,\,,\,\,\,\,i=1,\,\,2,\,\,.\,.\,.,\,n$, through the log-link function $\log ({\theta _i})\,\,=\,\,{X^{\prime}_i}\beta$, where $\beta =({\beta _1},\,\,{\beta _2},\,\,...,\,\,{\beta _p})^{\prime}$ is a vector of unknown regression coefficients. In terms of the inverse of the log-link function, we can express the expected response as ${\theta _i}=\exp ({X^{\prime}_i}\beta )$. If an exposure variable is incorporated, the model can be expressed as: $\log ({\theta _i})\,\,=\,\,{X^{\prime}_i}\beta +\log ({E_i})$, where ${E_i}$ is the exposure for the i^th individual [5, 18–19].

2.2.2 Generalized Poisson model

One of the drawbacks of the Poisson model is the assumption of equality between the conditional variance and the conditional mean (equi-dispersion). When this assumption is violated, the standard errors of the parameter estimates are biased, resulting in misleading inferences [20]. An alternative specification, the negative binomial distribution, is still not permissible in cases where the response exhibits underdispersion [19].

Consul and Jain [21] and Consul and Famoye [8] proposed the generalized Poisson (GP) distribution that can accommodate any type of dispersion. The PMF of a response variable ${Y_i}$ following the GP distribution is given by:

where ${\theta _i}>0$ and $\delta$ is the dispersion parameter such that $\hbox{max} ( - 1\,,\, - {\theta _i}/4)<\delta <1$. We may introduce covariates into the regression model through the relationship: $\log \left( {{{{\theta _i}} \mathord{\left/ {\vphantom {{{\theta _i}} {(1 - \delta )}}} \right. \kern-0pt} {(1 - \delta )}}} \right)={X^{\prime}_i}\beta$ [8].

For a GP-distributed random variable ${Y_i}$, the mean and variance, conditioned on the predictor variables, are given by:

$E({Y_i}|{X_i})=\frac{{{\theta _i}}}{{1 - \delta }}$ ,$Var({Y_i}|{X_i})=\frac{{{\theta _i}}}{{{{(1 - \delta )}^3}}}=\frac{1}{{{{(1 - \delta )}^2}}}E({Y_i}|{X_i})=\phi E({Y_i}|{X_i})$

The term $\phi =1/{(1 - \delta )^2}$ determines the status of dispersion of the outcome variable. The assumption of equi-dispersion holds when $\delta =0$. Under this situation, the GP distribution reduces to the standard Poisson distribution with parameter ${\theta _i}$. On the other hand, we have overdispersion and underdispersion when $\delta >0$ and $\delta <0$, respectively.

2.2.3 Truncated count data regression models

When we use Poisson or negative binomial regressions to model a count outcome variable, we are implicitly assuming the existence of zero counts owing to the distributional properties underlying both models. However, we may encounter situations in which zero counts never surface. As stated in the previous section, data on the response variable in this study (i.e., the number of children ever born) were collected by the EDHS considering reproductive women who had given at least one birth. This simply means that the outcome variable can never assume zero values. In such instances, count regression models that take into account this truncation need to be considered since modelling zero-truncated count data using the standard (non-truncated) distributions may result in biased estimates of the parameter vectors, and consequently, erroneous inferences [4].

Zero-truncated Poisson regression model

Suppose that ${Y_i}$ is a random variable that takes on positive integer values (that is, the number of events of interest in which at least one event must occur). Then, ${Y_i}$ follows the zero-truncated Poisson (ZTP) distribution with mean ${\theta _i}$ if its PMF is given by [22]:

Zero-truncated generalized Poisson distribution

In addition to the presence of only positive integer counts in the response variable, the assumption of equi-dispersion may also be violated. In such situations, the zero-truncated generalized Poisson (ZTGP) distribution is more appropriate. The ZTGP distribution with parameters ${\theta _i}$ and $\delta$, denoted by $ZTGP\,(\delta \,,\,\,{\theta _i})$, can be expressed as [10, 23]:

$$f({y_i};{\theta _i},\,\delta )=\frac{1}{{{y_i}![\exp ({\theta _i}/(1+\delta {\theta _i})) - 1]}}\,\,{\left( {\frac{{{\theta _i}}}{{1+\delta {\theta _i}}}} \right)^{{y_i}}}{\left( {1+\delta {y_i}} \right)^{{y_i} - 1}}\exp \left( { - \frac{{\delta {\theta _i}{y_i}}}{{1+\delta {\theta _i}}}} \right)\,\,\,\,,\,\,{y_i}=\,1,\,2,\,3,\,\,...$$

When the null hypothesis ${H_0}:\delta =0$ holds, the assumption of equi-dispersion is valid, and the ZTGP distribution reduces to ZTP.

2.3 Comparison of models

To choose the best-fit model among candidate models, the Akaike information criterion (AIC) and Bayesian information criterion (BIC) can be utilized. The model with the smallest information criteria is considered to be the one that better fits the sample data.

3.1 Descriptive statistics

The sample average number of children ever born (NCEB) among married reproductive-age women who reside in rural areas of Ethiopia and had at least one live birth at the time of the survey was found to be 4.4966 (95% CI: 4.4316–4.5617). Table 1 pertains to a summary of the socio-demographic, maternal and household characteristics of the respondents and their relationship with NCEB. As one would expect, the mean NCEB increased with the current age of women. About two-fifths of women had their first birth before reaching their 18th birthday (i.e., childbearing at an early age) and had an average of about five children. We also notice that the mean NCEB steadily decreased as the age at first birth increased. In terms of level of education, over 70% and 55% of women and husbands/partners, respectively, had no education, whereas those with higher education accounted for less than five percent of both groups. We can observe a decreasing pattern in the mean NCEB as the level of education increases for both groups. About three-quarters of the women had never used contraceptives, and registered a higher mean NCEB (4.68) as compared to those who used the same. Regarding the number of sons/daughters who have deceased, about two-thirds of women had never lost a child. The mean NCEB was considerably lower for these women (3.66) as compared to those who had lost at least one child.

In general terms, the mean NCEB exhibited a downward trend with increasing wealth index. The other covariate considered was land ownership status. About three-fifths of women owned land alone or jointly with their husband/partner. The mean NCEB for these women (4.82) was considerably higher than that for those who did not own land (4.00). Moreover, the mean NCEB was slightly higher for women who had never listened to the radio. Nearly half of the sample women were Muslim. The mean NCEB was found to be the smallest for followers of the Protestant Church, followed by those who followed the Orthodox (Coptic) Church. We also observe variation in the mean NCEB across regions, with women from Gambella (3.67) and Somali (5.14) registering the lowest and highest mean NCEB, respectively.

Table 1

Descriptive statistics of individual and household characteristics and the mean NCEB
Variable	Categories	Freq.	Percent	Mean NCEB
Current age	15–19	253	4.04	1.14
	20–24	1,101	17.6	1.92
	25–29	1,455	23.26	3.37
	30–34	1,216	19.44	4.84
	35–39	1,056	16.88	6.14
	40–44	677	10.82	6.93
	45–49	498	7.96	7.57
Age at first birth	Less than 18	2,604	41.62	5.09
	18–25	3,414	54.57	4.11
	Over 25	238	3.8	3.54
Women's education	No education	4,510	72.09	5.02
	Primary	1,471	23.51	3.32
	Secondary	223	3.56	2.17
	Higher	52	0.83	2.13
Husband’s/partner’s education	No education	3,542	56.62	4.92
	Primary	2,061	32.94	4.18
	Secondary	429	6.86	3.18
	Higher	224	3.58	3.16
Number of years contraceptives used	Not ever used	4,661	74.5	4.68
	1–3 years	1,169	18.69	3.80
	4 or more years	426	6.81	4.43
No. of children died	None	4,205	67.22	3.66
	One	1,242	19.85	5.41
	Two	491	7.85	6.77
	Three or more	318	5.08	8.54
Wealth index	Poorest	2,431	38.86	4.67
	Poorer	1,260	20.14	4.41
	Middle	1,125	17.98	4.34
	Richer	1,050	16.78	4.46
	Richest	390	6.23	4.23
Land ownership status	Does not own	2,473	39.53	4.00
Land ownership status	Owns alone or jointly	3,783	60.47	4.82
Freq. listening radio	Not at all	4,937	78.92	4.56
Freq. listening radio	At least once a week	1,319	21.08	4.27
Religion	Muslim	2,984	47.7	4.67
	Orthodox	1,853	29.62	4.40
	Protestant	1,277	20.41	4.22
	Other	142	2.27	4.56
Region	Afar	616	9.85	4.54
	Amhara	838	13.4	4.26
	Benishangul-Gumuz	590	9.43	4.37
	Dire Dawa	201	3.21	4.78
	Gambella	426	6.81	3.67
	Harari	265	4.24	4.33
	Oromia	1,039	16.61	4.59
	SNNPR	973	15.55	4.55
	Somali	679	10.85	5.14
	Tigray	629	10.05	4.52

In this study, cohabitation duration was used as an exposure variable. Table 2 presents a summary of this exposure variable categorized into seven groups. We can observe that the mean NCEB steadily increases as duration of cohabitation increases, making it an ideal measure of exposure (the number of times the event (birth) could have occurred).

Table 2

Summary statistics of cohabitation duration (exposure)
Cohabitation duration	Frequency	Percent	Mean NCEB
0–4	772	12.34	1.25
5–9	1,283	20.51	2.53
10–14	1,312	20.97	4.14
15–19	1,106	17.68	5.44
20–24	863	13.79	6.54
25–29	589	9.41	7.26
30+	331	5.29	7.73

3.2 Standard Poisson regression analysis

We first fit the standard Poisson regression model to assess the amount of dispersion. The results pertaining to the dispersion statistics are presented in Table 3. We can see that the dispersion statistic ($({1 \mathord{\left/ {\vphantom {1 {df}}} \right. \kern-0pt} {df}})Pearson=0.4178786$) is considerably smaller than one (that is, the conditional variance is about 58% smaller than the conditional mean). Thus, conditional on the predictor variables, the responsible variable is underdispersed. Under such situations, the standard errors of the coefficient estimates from the Poisson model are biased, leading to erroneous inferences. This underdispersion also rules out the negative binomial distribution, which assumes equi-dispersion or overdispersion.

Table 3

Dispersion statistics for the fitted standard Poisson model
Generalized linear models	No. of obs = 6,256
Optimization : ML	Residual df = 6,226
	Scale parameter = 1
Deviance = 2691.641490	(1/df) Deviance = 0.4323228
Pearson = 2601.711977	(1/df) Pearson = 0.4178786

3.3 Model comparisons

We used the AIC and BIC to compare the goodness-of-fit of the Poisson distribution as well as its generalizations that are suitable for both types of dispersions and/or take into account the truncated nature of our response variable. We can see from Table 4 that the fitted zero-truncated generalized Poisson model with exposure has the smallest AIC and BIC, and hence, is the preferred model. In contrast, the non-truncated standard Poisson model (with exposure), which assumes equi-dispersion and does not account for truncation, fared the least.

Table 4

Comparison of count data models
Model	AIC	BIC
Standard Poisson with exposure	22712.46	22914.7
Zero-truncated Poisson with exposure	21460.33	21662.57
Generalized Poisson with exposure	20748.58	20957.56
Zero-truncated generalized Poisson without exposure	20211.55	20420.53
Zero-truncated generalized Poisson with exposure	20199.97	20408.95

3.4 Analysis with the zero-truncated generalized Poisson model

We fit the zero-truncated generalized Poisson (ZTGP) model to identify and analyze correlates of the number of children ever born (NCEB). The estimated dispersion parameter was $\hat {\delta }= - 0.438$ (95% CI: -0.462, -0.414). Thus, we reject the null hypothesis ${H_0}:\delta =0$ and conclude that there is underdispersion since the 95% confidence limits are both negative. This finding further corroborates the conclusion we reached from the fitted standard Poisson model earlier.

Table 5 pertains to the results of the fitted ZTGP model, where the average marginal effects are reported instead of coefficient estimates. Marginal effects are popular means by which the effects of regressors in nonlinear models can be made more intuitively meaningful. In the Poisson model and its generalizations, for instance, they are more informative since they provide effects on the counts scale (not rates). We can see from the table that, with the exception of husband’s/partner’s education, all the other explanatory variables are statistically significant at the 5% level. Age at first sex was dropped from the pool of explanatory variables because of high multicollinearity with age at first birth (r = 0.667, p < 0.001).

Compared with women in the age bracket of 15–19 years, those in higher age categories have significantly higher NCEB. With respect to age at first birth, a woman with one more age at first birth is predicted to have 0.06 fewer children on average, all other things being equal. The other significant predictor was women’s education. As the number of years of education of women increases by one, the NCEB decreases by 0.056. On average, each additional year of contraceptive use lowers the NCEB by 0.104. Moreover, the study revealed a positive association between the number of children who have died in the family and NCEB per woman.

Religious affiliation was also found to be significantly associated with NCEB. Women who follow the Orthodox (Coptic) Church are predicted to have fewer children than Muslims, Protestants and followers of ‘other’ religions, keeping the other covariates constant. The results also revealed significant regional variation. Compared with those in the Afar region, women residing in the Amhara, Benishangul-Gumuz, Gambella and Tigray regions have significantly lower NCEB. On the other hand, women residing in the Somali region have a significantly higher NCEB than those in the Afar region. Regarding economic status, poorest women have a significantly higher NCEB than those in higher categories of wealth index. The results indicated that land ownership was a push factor for higher number of births. Women who own land (alone or jointly with their partner) are predicted to have 0.118 more children as compared to those who do not own the same. Moreover, holding the other covariates fixed, women who listened to the radio (at least once a week) are predicted to have 0.105 fewer children than those who did not, on average.

Table 5

Results of the fitted zero-truncated generalized Poisson (ZTGP) model
Variable	dy/dx	Std. Err.	z	P > z	[95% Conf. Interval]
Women's current age (Ref. = 15–19)
20–24	1.170	0.087	13.4	0.000	0.999	1.341
25–29	2.026	0.084	24.26	0.000	1.862	2.189
30–34	2.255	0.086	26.34	0.000	2.087	2.422
35–39	2.176	0.087	25.06	0.000	2.006	2.346
40–44	1.768	0.092	19.27	0.000	1.588	1.948
45–49	1.432	0.094	15.28	0.000	1.248	1.615
Age at first birth	-0.060	0.007	-8.84	0.000	-0.073	-0.047
No. of deceased children	0.341	0.018	19.27	0.000	0.307	0.376
No. of years contraceptives used	-0.104	0.011	-9.88	0.000	-0.125	-0.084
Husband's education	0.005	0.006	0.76	0.445	-0.007	0.016
Women's education	-0.056	0.010	-5.71	0.000	-0.075	-0.036
Religion (Ref. = Orthodox)
Muslim	0.144	0.066	2.190	0.028	0.015	0.273
Protestant	0.206	0.076	2.710	0.007	0.057	0.355
Other	0.482	0.147	3.280	0.001	0.194	0.770
Region (Ref. = Afar)
Amhara	-0.809	0.098	-8.28	0.000	-1.000	-0.617
Oromia	0.157	0.091	1.74	0.082	-0.020	0.335
Somali	0.844	0.091	9.30	0.000	0.666	1.021
Benishangul-Gumuz	-0.270	0.100	-2.69	0.007	-0.467	-0.073
SNNPR	-0.069	0.103	-0.67	0.501	-0.270	0.132
Gambella	-0.739	0.108	-6.83	0.000	-0.951	-0.527
Harari	0.210	0.119	1.76	0.079	-0.024	0.443
Tigray	-0.363	0.107	-3.40	0.001	-0.572	-0.154
Dire Dawa	0.128	0.116	1.10	0.273	-0.100	0.356
Wealth index (Ref. = Poorest)
Poorer	-0.207	0.056	-3.73	0.000	-0.317	-0.098
Middle	-0.294	0.059	-4.95	0.000	-0.410	-0.177
Richer	-0.180	0.064	-2.83	0.005	-0.305	-0.055
Richest	-0.440	0.084	-5.23	0.000	-0.605	-0.275
Land ownership status (Ref. = Does not own)
Owns alone or jointly	0.118	0.040	2.94	0.003	0.040	0.197
Freq. of listening to radio (Ref. = Not at all)
At least once a week	-0.105	0.047	-2.24	0.025	-0.196	-0.013

3.5 Discussion

This study considered various count data models to investigate the predictors of the number of children ever born (NCEB) in Ethiopia. This response variable was zero-truncated since data only on women who had given at least one birth at the time of the survey were utilized. Cohabitation duration was used as an exposure variable that allows the counts of children to be comparable across subjects who were observed for different durations of time. The zero-truncated generalized Poisson model, which accounts for both types of dispersions as well as the truncated nature of the response variable, was found to be the best fit model on the basis of model selection criteria (AIC and BIC).

The results of our study revealed that as the age of a woman at first birth increases, the number of children ever born (NCEB) decreases. This could be attributed to the fact that early initiation of childbearing lengthens women’s reproductive lifespan, and consequently, upsurges the level of fertility. This is particularly the case in countries where birth control measures (such as contraception) are not widely used. Moreover, teenage pregnancy is often associated with pregnancy-related complications, exasperating the problem of child (and maternal) mortality, which indirectly affects fertility levels. Several studies reported similar findings regarding the positive effect of first birth at an earlier age on fertility [12, 13, 16, 24, 25]. Moreover, the study found that teenage women (15–19 years) have significantly fewer births as compared to those in higher age categories. This could be attributed to the cumulative number of births during the reproductive life span of women. This finding is supported by studies in Ethiopia, Bangladesh and Ghana [11, 14, 15, 26].

One of the covariates rarely investigated in fertility studies is land ownership. This study revealed that women who own land (alone or jointly with their partner) have a significantly higher number of children than their counterparts who possess no land. Studies in Kenya and Nepal also found a positive relationship between land ownership and fertility [27, 28]. The authors argued that, in poor, rural, agrarian settings, high fertility trends might have arisen from the demand for farm labour. According to a study in Ethiopia by Ali et al. [29], the land tenure regime where household size was used as a criterion for land distribution might have contributed to higher fertility levels in rural areas. In support of this assertion, they reported a significant reduction in the lifetime fertility of women in the post-reform period in which the land tenure regime de-linked land access from household size. Note that the EDHS data in the current study do not allow for disaggregated analysis for the pre- and post-reform periods. Our result is also inconsistent with a study in Nepal which revealed that secure land ownership for women is associated with a reduction in the number of children, possibly through women’s empowerment effect [30].

Our results revealed that the NCEB increases with the number of children who have died. A number of studies have also reported this positive association between child mortality and the number of children ever born [25, 31, 32]. In a high child mortality environment, this could be explained by the behavioral tendency of parents to bear more children than the desired number in anticipation that some will not survive (fertility response to expected mortality or hoarding effect). Such an association could also be the result of a response to an actual child death, that is, couples may try to offset the loss of a child (replacement hypothesis).

We found a negative relationship between NCEB and contraceptive use. Each additional year of contraceptive use lowers the number of births per woman by 0.104, on average. Our result is consistent with studies conducted in Ethiopia, which reported significantly lower NCEB for mothers who used modern contraception methods than those who did not [12, 13]. A study in Ghana also reported that a decline in one child was associated with a 15% increase in the use of contraception [26]. This could be attributed to the vital role of contraception in controlling the incidence of unintended pregnancy. In contrast, studies in Malawi [31] and Nepal [33] reported a positive association between the use of contraceptives and NCEB. Women’s adoption of contraceptives after reaching or exceeding the desired number of children was cited as a possible explanation for this finding.

The other significant factor was women’s education. Holding all other covariates constant, the NCEB decreases as the number of years of education of women increases. Various studies on fertility have shown a consistent negative correlation between NCEB and increasing education [11–15, 24, 26]. Compared with uneducated women, educated women have a better understanding of reproductive health and family planning and are more likely to use contraception and avoid early childbearing. Moreover, education empowers women to make their own decisions regarding sexual and reproductive health issues. The current study revealed no significant relationship between husbands’/partners’ education and NCEB. This finding is not in line with those of Rahman et al. [24], Kiser and Hossain [15] and Nibaruta et al. [25].

Regarding religious affiliation, Muslims, Protestants and followers of ‘other’ religions are predicted to have a significantly higher number of children than women who follow the Orthodox (Coptic) Church. This finding is in line with studies conducted in Ethiopia [12, 14] and Nigeria [16]. The study also found significant regional variation in the NCEB. Similar findings have been reported in a number of studies in Ethiopia [12–14].

Our analysis revealed that poorest women have a significantly higher NCEB as compared to those in higher categories of wealth index. This inverse relationship between fertility and economic status has been documented by studies in Ethiopia [11, 12, 14], Bangladesh [15, 24] and Ghana [26]. One possible explanation is that women from economically healthy households are more likely to have access to education and, consequently, have the knowledge as well as the resources to implement family planning methods. The study also revealed that women who listened to the radio are predicted to have fewer children than those without any media exposure. This could be attributed to the role of mass media in disseminating vital information on maternal health and family planning, which could reach even those women with little or no schooling. Various studies have reported similar findings [11, 24, 26].

High fertility rate is associated with increased maternal and child mortality risk, particularly for low-income countries. Thus, exploring the factors that contribute to high fertility is crucial. This study identified and analyzed the correlates of the number of children ever born in rural Ethiopia using the zero-truncated generalized Poisson model. This model is ideally suited for a responsible variable that is truncated as well as exhibit any type of dispersion, as is the case in this study. The results revealed that early age at first birth, women’s land ownership, number of children who have died in the family and low economic status of the household were positively associated with NCEB. Muslims, Protestants and followers of ‘other’ religions were also found to have significantly higher child births than followers of the Coptic Church. On the other hand, contraceptive use, women’s education and media exposure were found to be inversely related to the NCEB.

On the basis of these findings, we recommend that any kind of intervention should selectively target vulnerable groups (e.g., the uneducated and the poorest of the poor). Awareness creation campaigns through mass media aimed at combating the early initiation of childbearing, particularly teenage pregnancy, should be strengthened. Moreover, efforts aimed at reducing child mortality may indirectly serve as a vital tool to curb high maternal fertility. It is also critical to promote the use of birth control measures and contraception.

AIC Akaike information criterion

BIC Bayesian information criterion

EDHS Ethiopia Demographic and Health Survey

GP Generalized Poisson

NCEB Number of children ever born

PMF Probability mass function

SDG Sustainable development goals

TGP Truncated generalized Poisson

UN United Nations

ZTGP Zero-truncated generalized Poisson

ZTNB Zero-truncated negative binomial

ZTP Zero-truncated Poisson

Acknowledgements

I would like to acknowledge the DHS Program for providing the data used in the current study.

Author contributions

EG designed the research concept and exclusively carried out data processing, data analysis and manuscript drafting.

Funding

No funding was received for the current study.

Data availability

The dataset used in the current study can be accessed freely upon registration with the DHS program (http://www.DHSprogram.com).

Ethical approval

The authorization (approval) letter for the use of the dataset was obtained from the DHS Program.

Consent for publication

Not applicable.

Competing interests

The author declares that there is no competing interest in this study.

United Nations. Global Issues: Population. https://www.un.org/en/global-issues/population (2023). Accessed 25 June 2024.
United Nations Department of Economic and Social Affairs, Population Division. World Population Prospects. 2022: Summary of Results. https://www.un.org/development/desa/pd/sites/www.un.org.development.desa.pd/files/wpp2022_summary_of_results.pdf (2022). Accessed 25 June 2024.
World Bank Group. World Development Indicators. https://data.worldbank.org/country/ethiopia (2024). Accessed 25 June 2024.
Hardin JW, Hilbe JM. Regression models for count data from truncated distributions. Stata J. 2015;1:226–46.
Harris T, Yang Z, Hardin JW. Modeling underdispersed count data with generalized Poisson regression. Stata J. 2012;12(4):736–47.
Famoye F, Wulu JT, Singh KP. On the generalized Poisson regression model with an application to accident data. J Data Sci. 2004;2:287–95.
Wang W, Famoye F. Modeling household fertility decisions with generalized Poisson regression. J Popul Econ. 1997;10:273–83.
Consul PC, Famoye F. Generalized Poisson regression model. Commun Stat Theor Methods. 1992;21:89–109.
Consul PC. Generalized Poisson distributions: properties and applications. Marcel Dekker; 1989.
Consul PC, Famoye F. The truncated generalized Poisson distribution and its estimation. Commun Stat - Theory Methods. 1989;18(10):3635–48.
Cherie N, Getacher L, Belay A, Gultie T, Mekuria A, Sileshi S, Degu G. Modeling on number of children ever born and its determinants among married women of reproductive age in Ethiopia: A Poisson regression analysis. Heliyon. 2023. 10.1016/j.heliyon.2023.e13948.
Gebre MN. Number of children ever-born and its associated factors among currently married Ethiopian women: evidence from the 2019 EMDHS using negative binomial regression. BMC Womens Health. 2024. doi.org/10.1186/s12905-024-02883-w.
Hussen NM. Application of two level count regression modeling on the determinants of fertility among married women in Ethiopia. BMC Womens Health. 2022. doi.org/10.1186/s12905-022-02060-x.
Melese ZY, Zeleke LB. Factors affecting children ever born among reproductive aged women in Ethiopia: Data from EDHS 2016. World J Public Health. 2020;5(3):66–75.
Kiser H, Hossain MA. Estimation of number of ever born children using zero truncated count model: evidence from Bangladesh Demographic and Health Survey. Health Inf Sci Syst. 2018. 10.1007/s13755-018-0064-y.
Ibeji JU, Zewotir T, North D, Amusa L. Modelling fertility levels in Nigeria using generalized Poisson regression-based approach. Sci Afr. 2020. doi.org/10.1016/j.sciaf.2020.e00494.
Faroughi P, Li S, Ren J. The applications of generalized Poisson regression models to insurance claim data. Risks. 2023. doi.org/10.3390/risks11120213.
Agresti A. An introduction to categorical data analysis. 2nd ed. Wiley; 2007.
Cameron AC, Trivedi PV. Regression analysis of count data. 1st ed. Cambridge University Press; 1998.
Cameron AC, Trivedi PV. Econometric models based on count data: Comparisons and applications of some estimators and tests. J Appl Econom. 1986;1:29–53.
Consul PC, Jain GC. A Generalization of the Poisson distribution. Technometrics. 1973;15:791–9.
Hilbe JM. Zero-truncated Poisson and negative binomial regression. Stata Tech Bull. 1999;47:37–40.
Hardin JW, Hilbe JM. Generalized linear models and extensions. 2nd ed. Stata; 2007.
Rahman A, Hossain Z, Rahman ML, et al. Determinants of children ever born among ever-married women in Bangladesh: evidence from the Demographic and Health Survey 2017–2018. BMJ Open. 2022. 10.1136/bmjopen-2021-055223.
Nibaruta JC, et al. Determinants of fertility differentials in Burundi: evidence from the 2016-17 Burundi demographic and health survey. Pan Afr Med J. 2021. 10.11604/pamj.2021.38.316.27649.
Boateng D, Oppong FB, Senkyire EK, et al. Socioeconomic factors associated with the number of children ever born by married Ghanaian females: a cross-sectional analysis. BMJ Open. 2023. 10.1136/bmjopen-2022-067348.
Chege V, Susuman AS. Landholding and fertility relationship in Kenya: A multivariate analysis. J Asian Afr Stud. 2016;51(1):43–59.
Bhandari P, Ghimire D. Rural agricultural change and fertility transition in Nepal. Rural Sociol. 2013;78(2):229–52.
Ali DA, Deininger K, Kemper N. (2022). Pronatal property rights over land and fertility outcomes: Evidence from a natural experiment in Ethiopia. The Journal of Development Studies. 2022; 58(5):951–967.
Chakrabarti A. Female land ownership and fertility in Nepal. J Dev Stud. 2018;54(9):1698–715.
Palamuleni ME. Determinants of high marital fertility in Malawi: Evidence from 2010 and 2015-16 Malawi demographic and health surveys. Open Public Health J. 2023. 10.2174/18749445-v16-e230419-2022-150.
Angko W, Arthur E, Yussif HM. Fertility among women in Ghana: Do child mortality and education matter? Sci Afr. 2022. 10.1016/j.sciaf.2022.e01142.
Adhikari R. Demographic, socio-economic, and cultural factors affecting fertility differentials in Nepal. BMC Pregnancy Childbirth. 2010. doi.org/10.1186/1471-2393-10-19.

No competing interests reported.

Download PDF

Reviewers agreed at journal
16 Oct, 2024
Reviewers agreed at journal
06 Oct, 2024
Reviewers invited by journal
25 Sep, 2024
Editor invited by journal
02 Sep, 2024
Editor assigned by journal
02 Sep, 2024
Submission checks completed at journal
02 Sep, 2024
First submitted to journal
01 Sep, 2024

You are reading this latest preprint version

Correlates of the number of children ever born to women in rural Ethiopia: Application of truncated generalized Poisson model with exposure

Status:

Version 1

Abstract

Background

Method

Results

Conclusion

1. Introduction

2. Materials and Methods

2.1 Sources of data and variables

2.2 Models for count data

2.2.1 Poisson regression model

2.2.2 Generalized Poisson model

2.2.3 Truncated count data regression models

2.3 Comparison of models

3. Results and Discussion

3.1 Descriptive statistics

3.2 Standard Poisson regression analysis

3.3 Model comparisons

3.4 Analysis with the zero-truncated generalized Poisson model

3.5 Discussion

4. Conclusion

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1