Fertility Decision and Its Associated Factors in Sub-Saharan Africa: A Multilevel Multinomial Logistic Regression Analysis | BMC Women’s Health

Study design and population

Data from the most recent Demographic and Health Surveys (DHS) of 35 SSA countries, conducted between 2008 and 2020, were used. These DHSs used a two-stage stratified sampling technique. A total of 284,744 (weighted) married women who had complete information about fertility decisions were used for this study (Fig. 1, Table 1). Details of the DHS methodology are reported elsewhere [21].

Fig. 1

Schematic presentation of sample extraction

Table 1 A detailed description of the study sample used for the study

Study variables

Result variable

The outcome variable was fertility decisions. Women were asked if they wanted to have a child in the future with the following options: wanting more children, not wanting more children, undecided fertility desire, declared sterilized and declared infertile. Women who reported infertile, sterilized, and with missing information were excluded because their responses were unclear about their fertility decisions. Finally, the fertility decision was calculated on the three responses (those who have a desire to have more children, who have no desire to have more children and an undecided fertility desire). Then, no desire was coded 0, desire for more children was coded 1, and undecided fertility desire was coded 2.

Independent variables

The independent variables were grouped into individual-level and community-level factors. Individual-level factors included respondent’s age, respondent’s education level, husband’s education level, parity, media exposure, current contraceptive use, ideal number of children , decision-making autonomy, number of living children, employment of women, gender. household head and wealth status. Community-level factors were place of residence and region of Africa (particularly SSA).

Operational definition

Media exposure

Constructed from three variables (frequency of listening to radio, television and reading newspapers/magazines). For this study, it was recoded as yes (if women were exposed to at least one medium) and no (otherwise).

Decision-making autonomy

It was constructed from three variables (decision on the respondent’s health care, decision on a major household purchase, and decision on visits to family or relatives) and recoded as “respondents alone” if a woman is the decision maker and “otherwise” if another person (husband, parents or friends) is the decision maker [16].

Data management and statistical analysis

STATA version 16 was used for data management (extraction, recoding and cleaning) and statistical analyzes (descriptive and analytical analysis). All analyzes were weighted to make the data representative, to take into account the non-response rate and to obtain a better statistical estimate. [22]. The proportion of fertility decisions with their 95% CI was estimated and to assess the factors associated with fertility decisions, random-effects and fixed-effects analyzes were used.

Fixed effects analysis

We used multinomial multinomial logistic regression analysis. During the analysis, we fitted four models; null model, model 1, model 2 and model 3. The null model was fitted only with the outcome variable. Models 1, 2 and 3 were fitted using individual variables, community variables and individual and community variables respectively. To select variables eligible for multilevel analysis, a bivariate multilevel multinomial regression was fitted first and variables with a p-value less than 0.20 in the bivariate analysis were considered eligible. Then, in the multivariate analysis, the adjusted relative risk (aRRR) with its 95% confidence interval (CI) was reported. Finally, variables with a p-value

Random effects analysis

It was conducted to assess the cluster-level variability of fertility decisions. Intraclass correlation coefficient (ICC) and proportional variation of variance (PCV) were calculated. Log-likelihood and deviance were used to check model fit, and a model with the highest log-likelihood and lowest deviance was considered the best-fitting model.

Comments are closed.