Operational definition

The sum of the unmet need for limiting and unmet need for spacing is the total unmet need for family planning [19].

includes pregnant women whose pregnancy was mistimed, postpartum ammenohoric whose last birth was mistimed and fecund women who are not using any method of family planning, and say they want to wait two or more years for their next birth, undecided about the timing of next birth or undecided whether to have another children [19].

women who are pregnant whose pregnancy was mistimed, postpartum amenorrheic women whose last birth was mistimed and fecund women who are not using any method of family planning and say they don’t want any more child [19].

Other variable definitioons used in this study are displayed in table below (Table 1).

Data quality was assured for all EDHSs, and was available in respective EDHS reports [13–17]. Missing variables for >5% of cases were excluded from the model and cases missing for outcome variable were excluded. The data management and analysis were performed by STATA version 15. Data was extracted from IR (individual recode) file. The wealth index for the 2000 survey was constructed by using the statistical method of principal component analysis. Wealth status was then created from assets by placing households on a continuous measure of relative wealth after which households were grouped into five wealth quintiles namely poorest, poorer, middle, richer, and richest [19]. Items used to construct the wealth index were household owns a radio, television, electricity, kerosene lamp, bed/table, and electric mitad. And the type of flooring of the house, toilet facility, drinking water facility, number of members per room in the household, cattle, sheep, goats, house, and land were also used to construct the variable wealth index. Variable v005 is divided by one million (1,000,000) to generate population sample weight (wgt) to correct for over and under sampling and applied in all descriptive statistics.

Variables were recorded for the analysis. Exposure to media is categorized in to, have exposure to media and don’t have exposure to media, which is generated from, reading about FP from newspaper, hearding about FP from radio and watching from television. The variable region is recoded into agrarian (Tigray, Amhara, Oromia, and SNNPR), urban (Addis Ababa, Dire dawa, and Harrari), and emerging/ pastoralist (Afar, Benshangul gumuz, Gambella, and Somali) for multilevel analysis. The data was declared as survey data with “svyset” command.

In this study, two-level mixed-effects logistic regression analyses were employed using STATA software version 15. Since the EDHS data was hierarchical, i.e., women were nested in household and household were nested in cluster multi-level analysis is recommended. Because of the sampling approach used in the all EDHSs, women from the same cluster may be more similar to each other than women from the rest of the country. To account for this clustering, two-stage multivariable multilevel logistic regression analysis was used to estimate the effects of individual- and community-level determinants on unmet need and to estimate the between-cluster variability in the odds of unmet need. The datasets of EDHS 2000–2016 were merged together to assess the determinants of unmet needs. A multilevel logistic regression model in a combination of both fixed effect (a measure of association) and random effect (a measure of variation) was performed. Clusters were treated as random effects. The random effects were measured by the intra-class correlation coefficient (ICC), median odds ratio (MOR), and proportional change in variance (PCV). The ICC shows the variation in unmet need for family planning for married reproductive women due to community characteristics.

ICC=Va/(Va+π23) Where, V_a is area (cluster) level variance and (π²/3) ≈ 3.29 refers to the standard logistic distribution, that is, the assumed level-1 variance component [20].

The MOR was calculated by using the formula:

The proportional change in variance was calculated as—PCV = (V_A-V_B)/V_A.

Where, V_A is the variance of the initial model and V_B is the variance in the subsequent models [20], [21].

Four models were fitted. Model Ⅰ, with no determinants (random intercept) to estimate random variation in the intercept and ICC. Model II only included individual-level variables, model III only included community-level variables to estimate the community-level characteristics, and finally, model IV included both individual-level and community-level variables adjusted for both. The information criteria’s Akaike Information Criteria (AIC) and Schwarz’s Bayesian Information Criteria (BIC) were used to compare the models to choose the best fitted. The best-fit model was the model with the lowest AIC and BIC [22, 23]. A multilevel bivariable logistic regression model was employed and variables with p-value less than 0.25 were candidate variables for multilevel multivariable logistic regression model. The fixed effect size of individual and community-level factors using AOR at 95% CI was used to measure the association between outcome and determinant variables. A p-value of 0.05 was used to declare the significance. Multi-collinearity was checked by using variance inflation factor (VIF) and for each variable is less than 2.1 included in the model.