Beyond the Cox model: a log-logistic accelerated failure time frailty model for analysing first birth timing in Zimbabwe Authors: Hilary Takunda Takawira Published: 2026 ABSTRACT AbstractBackgroundThe Cox proportional hazards (PH) model is the most common approach for analyzing time-to-event data in public health. However, its reliance on the proportional hazards’ assumption is a key limitation, often violated in demographic research. This study illustrates a comprehensive model selection framework, comparing the performance of various parametric accelerated failure time (AFT) and frailty models as robust alternatives when the PH assumption fails, using data on the timing of first birth as a case study.MethodsWe conducted a secondary analysis of time to first birth among 9,955 women from the 2015 Zimbabwe Demographic and Health Survey. After confirming violations of the PH assumption using Schoenfeld residuals, we systematically compared the fit of several parametric models’ exponential, Weibull, log-normal, and log-logistic with and without gamma-shared frailty terms. Model selection was based on the Akaike (AIC) and Bayesian (BIC) Information Criteria.ResultsThe log-logistic AFT model with gamma frailty showed improved relative fit for this dataset, accommodating a non-monotonic hazard and accounting for unobserved regional heterogeneity (θ = 0.671, p<0.001). The model yielded interpretable time ratio (TR) estimates, where higher education (TR=1.11) and later cohabitation (TR=1.25) were associated with delayed first birth, while marriage (TR=0.98) and contraceptive use (TR=0.97) were associated with earlier timing. In contrast, standard parametric models and the Cox model were suboptimal for this dataset.ConclusionThis study provides an empirical demonstration of how AFT models with frailty terms can be applied in settings where the proportional hazards assumption is violated. The log-logistic frailty model showed improved relative fit for this dataset and offers an interpretable framework for analysing time-to-event outcomes with clustered data.Keywords: Model selection, Survival analysis, Accelerated failure time, Frailty model, Time to first birth, Parametric models Introduction The Cox proportional hazards (PH) model has long been the cornerstone of survival analysis in demographic and public health research, prized for its semi-parametric nature that requires no assumption about the baseline hazard shape [1, 2]. Its widespread use in modelling time-to-event outcomes, such as the timing of first birth, is evident across sub-Saharan Africa (SSA). A woman’s first birth timing is a critical demographic indicator, intricately linked to fertility patterns, maternal and child health outcomes, and long-term socio-economic development [3, 4]. Early childbearing remains a global public health concern, with the World Health Organization identifying adolescent pregnancy as a priority area for intervention (WHO, 2011; WHO, 2014). Understanding factors associated with age at first birth is therefore essential for effective public health programming [5–7]. However, the Cox model’s validity is critically dependent on the proportional hazard’s assumption, which is frequently violated in complex demographic data where the effect of covariates (e.g., education, age at marriage) on the hazard of first birth may change over time [8]. Furthermore, data from national surveys like the Demographic and Health Surveys (DHS) have a hierarchical structure, with women nested within clusters and regions, creating unobserved heterogeneity that can lead to correlated failure times and biased estimates if ignored [7, 9]. Early marriage remains a critical pathway to early childbearing, with studies across multiple contexts documenting its persistent influence [7, 10, 11].The standard Cox model is ill-equipped to handle this clustering, a significant limitation given that factors influencing first birth operate at multiple levels, from the individual (e.g., education, age at first sex) to the household and community (e.g., wealth, region) [5, 12]. In response to these limitations, methodological advancements have highlighted the utility of parametric accelerated failure time (AFT) models and shared frailty models. AFT models provide a direct and often more intuitive interpretation by modelling the log of survival time, yielding time ratios (TRs) that quantify the acceleration or deceleration of the event time associated with a covariate [13, 14]. When coupled with a frailty term (a random effect that accounts for unobserved cluster-level heterogeneity), these models produce more efficient and reliable estimates, allowing for the identification of contextual “hotspots” of risk beyond individual-level characteristics [4]. Recent studies in SSA have begun to demonstrate the superiority of these approaches. For instance [8], employed an inverse Weibull gamma shared frailty model in Ethiopia to address PH violations and quantify regional clustering. The Ethiopian Demographic and Health Survey has been extensively used to examine reproductive health outcomes [15]. Similarly [14] and [7], applied log-normal and Bayesian frailty models, respectively, highlighting the value of flexible parametric distributions and frailty structures for accurately modeling time to first birth. Collectively, this emerging body of work underscores that moving beyond the Cox paradigm can yield significant methodological and substantive insights. Despite this progress, a systematic comparison of parametric AFT models with frailty corrections, specifically to identify the optimal baseline distribution for a given dataset, remains underexplored in the context of first birth. This study aims to fill this gap by evaluating whether a log-logistic AFT model with gamma frailty provides improved fit for this dataset. The log-logistic distribution is particularly well suited for demographic processes like first birth, as it accommodates a non-monotonic hazard function that can increase to a peak during late adolescence and early adulthood before decreasing. Therefore, this study uses data on time to first birth among Zimbabwean women as a case study to demonstrate a comprehensive model selection process. We first confirm the inadequacy of the Cox PH model due to assumption violations. We then systematically compare the performance of parametric AFT models (exponential, Weibull, log-normal, log-logistic) with and without a gamma shared frailty term. We hypothesize that the log-logistic AFT model with gamma frailty may provide improved fit for this dataset compared to standard parametric and Cox models., offering not only robust statistical estimates but also more interpretable results in the form of time ratios, all while accounting for the unobserved regional heterogeneity that is central to understanding reproductive timing in a country like Zimbabwe. This study uses time to first birth among Zimbabwean women as an empirical case study to demonstrate a systematic framework for comparing survival models when the proportional hazards assumption is violated. The aim is not to identify a universally superior model, but rather to illustrate how flexible parametric AFT models with frailty structures may provide improved fit and interpretability under specific empirical conditions. Accordingly, all model comparisons and conclusions are interpreted as dataset-specific rather than generalizable methodological claims. Therefore, this study evaluates the performance of alternative survival models for this dataset, with a focus on assessing whether flexible parametric AFT models with frailty structures provide improved fit under conditions where the proportional hazards assumption is violated. Study design and data source To demonstrate the proposed modeling framework, we conducted a secondary analysis of the 2015 Zimbabwe Demographic and Health Survey (ZDHS). The DHS employs a two-stage stratified cluster sampling design: first, enumeration areas (EAs) are selected with probability proportional to size, followed by the systematic selection of households within each EA. This complex survey design induces clustering and stratification, creating an ideal use-case for evaluating frailty models that account for such data structures. The 2015 ZDHS was selected for this methodological demonstration for several key reasons. First, as a publicly available and well-documented dataset, it ensures the full reproducibility of our analysis, a cornerstone of robust methodological research. Second, its substantial sample size (N = 9,955 women) provides the necessary statistical power for fitting and comparing multiple advanced survival models, including those with frailty terms, without convergence issues. Finally, using a single, established dataset al.lows for a clear and unambiguous comparison of model performance, isolating the methodological contribution from confounding factors like temporal trends. This establishes a benchmark for future methodological applications using subsequent DHS rounds. Study Population and inclusion Criteria The analysis included all women aged 15 to 49 years from the 2015 ZDHS with complete data on the variables of interest. This cohort provides a complete picture of the reproductive lifespan, capturing the period of highest risk for first birth and ensuring the results are generalizable to all reproductive-age women in Zimbabwe. Outcome variable The outcome was time-to-first birth, measured as the woman’s age in years at her first live birth. For women who had not given birth by the time of the survey, the data were right censored at their current age. This handling of censoring is a fundamental strength of survival analysis that we leverage in our model comparison. Independent variables and rationale for categorization Independent variables were selected a priori based on the socio-ecological model and prior literature. To ensure model parsimony, statistical stability, and enhanced interpretability critical for a clear methodological demonstration variables were strategically categorized. Age at first sex (v525) and age at first cohabitation (v511) were categorized into < 15, 15–17, and ≥ 18 years. This captures critical developmental and legal thresholds and aligns with established risk groupings in adolescent health literature (Mangiaterra et al., 2008; WHO, 2011). Wealth index (v190) was kept as five categories (Table 1). Variables with skewed distributions were dichotomized to yield more robust and interpretable coefficient estimates. This included: Media exposure: A composite variable (derived from v157–v159) dichotomized as Exposed (weekly access to newspaper, radio, or TV) versus Not exposed. Spousal age gap: Calculated and coded as < 5 years versus ≥ 5 years to isolate the effect of a pronounced age-power differential. Other covariates included educational attainment, place of residence (urban/rural), geographical region, current working status, religion, and knowledge of the ovulatory cycle. A full description of variables is provided in Table 1. Table 1Description of study variables, their sources from the DHS, measurement types, and categorization used in the analysisVariable CategoryVariableDHS VariableTypeLevels / CodingDependent VariablesEventDerived from v201Binary0 = No children ever born 1 = Has at least one childAge first birthv212ContinuousAge in yearsSocio-economic FactorsWealth indexv190Categoricalpoorest, poorer, middle, richer, richesteducation levelv106Categoricalno education, primary, secondary, higherCurrently workingv714Binaryno, yesGeographic FactorsResidence typev025Binaryurban, ruralregionv024CategoricalManicaland, Mashonaland Central, Mashonaland East, Mashonaland West, Matabeleland North, Matabeleland South, Midlands, Masvingo, Harare, BulawayoMarital/Sexual FactorsMarital statusDerived from v501BinaryUnmarried, Married/cohabitingAge first cohabitationDerived from v511Categorical< 15 years, 15–17 years, ≥ 18 yearsAge at first sexDerived from v525Categorical< 15 years, 15–17 years, ≥ 18 yearsSpousal age gapDerived from v730 & v012Binary< 5 years, ≥ 5 yearsMedia/Knowledge FactorsMedia exposureDerived from v157–v159BinaryNo media, Exposed to mediaOvulation knowledgev217Categoricalduring, after, middle, before, at any time, other, don’t knowReproductive Health FactorsContraceptive useDerived from v313BinaryNo, yesReligionv130Categoricaltraditional, roman catholic, protestant, Pentecostal, apostolic, other_christian, muslim, none, other Description of study variables, their sources from the DHS, measurement types, and categorization used in the analysis Statistical analysis Analyses were conducted in Stata version 18.5. Sampling weights, primary sampling units, and stratification variables were incorporated in the descriptive analyses, Kaplan–Meier survival estimates, log-rank tests, and Cox proportional hazards models to account for the complex survey design of the Zimbabwe Demographic and Health Survey and to produce nationally representative descriptive estimates. Descriptive and preliminary survival analysis The study population was characterized using weighted descriptive statistics, presented as frequencies and percentages for categorical variables. The distribution of time-to-first birth was summarized using medians and interquartile ranges (IQR) and explored graphically. To conduct a preliminary, unadjusted assessment, we generated survey-adjusted Kaplan-Meier survival curves for key categorical predictors and used the log-rank test to evaluate differences across categories. Violation of proportional hazards The validity of the standard Cox proportional hazards (PH) model was formally assessed using Schoenfeld residuals. The global test indicated a significant violation of the PH assumption (χ²= 1599.40; df = 38; p < 0.001), which rendered the Cox model inadequate for our data. This finding provided the necessary justification for proceeding with alternative parametric survival modeling approaches. Framework for accelerated failure time models In response to the PH violation, we implemented a systematic model comparison framework based on parametric accelerated failure time (AFT) models. We estimated a suite of eight models: four standard parametric distributions (exponential, Weibull, log-normal, and log-logistic) and four parallel models that incorporated a gamma-shared frailty term. The frailty was specified at the regional level (v024) to account for unobserved geographical heterogeneity, a critical adjustment for the DHS’s clustered sampling design. Selection of the final model Model fit was assessed using the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), with lower values indicating a superior balance of fit and parsimony. The log-logistic AFT model with gamma-shared frailty demonstrated the lowest AIC and BIC values and was therefore selected for further analysis in this dataset. Model specification The log-logistic distribution accommodates non-monotone hazards, making it suitable for reproductive timing data. where Ti is the survival time, Xi are covariates, β are coefficients, ϵi is a standard logistic error term, and σ is the scale parameter. Time ratios (TR = e^β) indicate relative acceleration of the event: TR > 1 implies longer time to first birth; TR < 1 implies shorter time. To account for clustering by region, we included a gamma shared frailty term zj allowing survival times within regions to be correlated. The frailty-adjusted hazard is: Where h0(t) is the baseline log-logistic hazard and zj ∼ Gamma (θ, θ). This approach accommodates non-monotone hazards and unobserved heterogeneity across regions. To account for the multi-stage clustered design of the DHS, we employed a shared frailty model. We specified the region (v024) as the frailty level, rather than the primary sampling cluster (v021), for three key reasons. Regions are administratively meaningful policy units in Zimbabwe; identifying region-level frailties directly informs strategic resource allocation and targeted interventions. We also hypothesized that unobserved contextual factors such as regional disparities in health infrastructure, economic development, and cultural norms create a shared risk environment that operates at this broader geographic level. In addition, using the region provides more stable statistical estimates of contextual effects by aggregating a larger number of individuals and events, thus mitigating the risk of spurious estimates that can occur with smaller, individual clusters. The log-logistic AFT model with gamma-shared frailty demonstrated the lowest AIC and BIC values and was therefore selected as the final and best-fitting model for this analysis. Model interpretation Covariate effects are reported as Time Ratios (TR = e^β) with 95% confidence intervals. A TR > 1 indicates a factor is associated with a delay in the time to first birth (longer survival time), while a TR < 1 indicates an acceleration (shorter time to event). The variance of the frailty term (θ) quantifies the magnitude of unobserved regional-level heterogeneity. Statistical significance was determined at α = 0.05. Survey design and unobserved heterogeneity accounting The 2015 Zimbabwe Demographic and Health Survey (ZDHS) employed a multi-stage stratified cluster sampling design. Survey weights (v005), primary sampling units (v021), and stratification variables (v022 and residence) were incorporated in the descriptive analyses and Cox models to account for the complex survey design and to obtain appropriate descriptive estimates and standard errors. For the parametric accelerated failure time models with frailty, a model-based approach was adopted. Survey weights were not applied in these models; instead, clustering and unobserved heterogeneity were addressed by including a gamma shared frailty term at the regional level (v024). The frailty term was used to capture intra-regional correlation and unmeasured contextual differences across regions, including variation in social norms, health service access, economic conditions, and reproductive health environments. The regional level was selected because regions are meaningful administrative and public health units in Zimbabwe and provide a more stable contextual grouping than smaller primary sampling units. Therefore, estimates from the frailty models should be interpreted as model-based associations within the analytic sample after accounting for unobserved regional heterogeneity, rather than as design-weighted nationally representative estimates. Ethical considerations The 2015 Zimbabwe Demographic and Health Survey (ZDHS) received ethical approval from the ICF International Institutional Review Board and the Medical Research Council of Zimbabwe. Written informed consent was obtained from all participants prior to data collection. For respondents aged 15–17 years, consent procedures were conducted in accordance with DHS ethical guidelines, including parental or guardian consent where required. This study is a secondary analysis of anonymized, publicly available DHS data. No additional ethical approval was required. The author did not have access to any identifiable information, and all analyses were conducted on fully anonymized datasets. Results Time to first birth among study participants The median observed analysis time was 19 years, with values ranging from 15 to 49 years. For women who had experienced a first birth, analysis time corresponded to age at first birth; for women who had not yet given birth, analysis time was censored at current age. Table 2 summarizes the socio-demographic characteristics of reproductive-age women in Zimbabwe (DHS 2015) by childbirth status. Among 9,955 women, most had ever given birth. The proportion was highest among those with no education (92.5%) and decreased with increasing education and wealth levels. Rural women (76.4%) were more likely to have given birth than urban women (68.5%). By religion, the highest proportion was among women with no religious affiliation (84.9%) and Apostolic sect members (78.5%), while lower proportions were observed among Pentecostal (68.6%) and Protestant (66.3%) women. Table 2Socio-demographic characteristics among reproductive age women in Zimbabwe DHS, 2015VariableTotal N (%)CensoredN (%)Ever had birth.N (%)χ²(p-value)Education levelp < 0.001 No education106 (1.1)8 (7.6)98 (92.5) Primary2,385 (24.0)386 (16.2)1,999 (83.8) Secondary6,637 (66.7)2,069 (31.2)4,568 (68.8) Higher827 (8.3)239 (28.9)588 (71.1)Wealth indexp < 0.001 Poorest1,499 (15.1)279 (18.6)1,220 (81.4) Poorer1,452 (14.6)325 (22.4)1,127 (77.6) Middle1,549 (15.6)408 (26.3)1,141 (73.7) Richer2,558 (25.7)641 (25.1)1,917 (74.9) Richest2,897 (29.1)1,049 (36.2)1,848 (63.8)Residence typep < 0.001 Urban4,521 (45.4)1,422 (31.5)3,099 (68.5) Rural5,434 (54.6)1,280 (23.6)4,154 (76.4)Religionp < 0.001 Apostolic sect3,829 (38.5)823 (21.5)3,006 (78.5) Pentecostal2,679 (26.9)840 (31.4)1,839 (68.6) Protestant1,618 (16.3)545 (33.7)1,073 (66.3) Roman Catholic670 (6.7)216 (32.2)454 (67.8) None471 (4.7)71 (15.1)400 (84.9) Other Christian589 (5.9)183 (31.1)406 (68.9) Others (Traditional/Muslim/Other)99 (1.0)24 (24.2)75 (75.8)Regionp < 0.001 Manicaland1,019 (10.2)259 (25.4)760 (74.6) Mashonaland Central993 (10.0)221 (22.3)772 (77.7) Mashonaland East910 (9.1)239 (26.3)671 (73.7) Mashonaland West1,054 (10.6)253 (24.0)801 (76.0) Matabeleland North849 (8.5)208 (24.5)641 (75.5) Matabeleland South829 (8.3)218 (26.3)611 (73.7) Midlands1,062 (10.7)271 (25.5)791 (74.5) Masvingo1,046 (10.5)293 (28.0)753 (72.0) Harare1,235 (12.4)378 (30.6)857 (69.4) Bulawayo958 (9.6)362 (37.8)596 (62.2) Socio-demographic characteristics among reproductive age women in Zimbabwe DHS, 2015 Table 3 presents behavioral and reproductive characteristics of reproductive-age women by childbirth status. About 8.6% of women who initiated sex before 15 years of age were censored, whereas the majority who initiated at 15–17 years (90.7%) or ≥ 18 years (88.0%) had given birth. Women who began cohabitation before age 18 were more likely to have given birth compared to those who started later (65.4%). Those exposed to media (71.5%) had a slightly lower proportion of childbirth experience compared to those without media exposure (77.6%). Nearly all married women (93.8%) and contraceptive users (97.1%) had ever given birth, compared with 40.9% of unmarried and 49.7% of non-users. Table 3Behavioral and reproductive characteristics by birth experience statusVariableTotalN (%)CensoredN (%)Ever had birth.N (%)χ² (p-value)Age at first sexp < 0.001 < 15 years488 (4.9)42 (8.6)446 (91.4) 15–17 years3,123 (31.4)291 (9.3)2,832 (90.7) ≥ 18 years4,491 (44.7)542 (12.1)3,949 (87.9)Age at first cohabitationp < 0.001 < 15 years377 (3.8)8 (2.1)369 (97.9) 15–17 years2,141 (21.5)122 (5.7)2,019 (94.3) ≥ 18 years7,437 (74.7)2,572 (34.6)4,865 (65.4)Media exposurep < 0.001 No media exposure2,230 (22.4)500 (22.4)1,730 (77.6) Exposed to media7,725 (77.6)2,202 (28.5)5,523 (71.5)Marital statusp < 0.001 Unmarried3,940 (39.6)2,330 (59.1)1,610 (40.9) Married6,015 (60.4)372 (6.2)5,643 (93.8)Contraceptive usep < 0.001 No5,086 (51.1)2,559 (50.3)2,527 (49.7) Yes4,869 (48.9)143 (2.9)4,726 (97.1)Spousal age gapp < 0.001 < 5 years2,571 (25.8)148 (5.8)2,423 (94.2) ≥ 5 years7,384 (74.2)2,554 (34.6)4,830 (65.4)Knowledge of ovulationp < 0.001 After period ended4,282 (43.0)711 (16.6)3,571 (83.4) Middle of cycle1,527 (15.3)471 (30.8)1,056 (69.2) Before period1,023 (10.3)292 (28.5)731 (71.5) At any time1,446 (14.5)417 (28.8)1,029 (71.2) Don’t know1,557 (15.6)759 (48.8)798 (51.2) Behavioral and reproductive characteristics by birth experience status A wider spousal age gap (≥ 5 years) was associated with a lower proportion of women who had given birth (65.4%) compared to smaller gaps (< 5 years: 94.2%). Knowledge of ovulation timing also varied women who correctly identified ovulation as occurring after menstruation were more likely to have given birth (83.4%), while those uncertain about timing had the lowest proportion (51.2%). Kaplan–Meier estimates of time to first birth by sociodemographic and reproductive characteristics The median age at first birth among reproductive-age women in Zimbabwe was 19 years (IQR: 17–21). Median age at first birth varied by socio-demographic and behavioral characteristics. Women with higher education had the oldest median age at first birth (23 years, IQR: 20–25), while those with no education had the youngest (17.5 years, IQR: 16–20). Women in the richest wealth quintile had a later median age at first birth (21 years, IQR: 19–23) compared to those in the poorest quintile (18 years, IQR: 17–20). Urban residents (20 years, IQR: 18–22) experienced first birth slightly later than rural residents (19 years, IQR: 17–21). Age at first sex and age at first cohabitation were associated with age at first birth; women initiating sex or cohabitation before 15 years had the earliest births. Women who used contraceptives or had a spousal age gap of less than 5 years also had slightly older median ages at first birth (Table 4). Variation in median age at first birth was observed across religious groups and knowledge of ovulation. Regionally, women in Masvingo, Harare, and Bulawayo had the oldest median ages at first birth (20 years), whereas women in Mashonaland West had the youngest (18 years). Table 4Median age at first birth (years) by socio-demographic, behavioral, and reproductive characteristics, Zimbabwe DHS 2015VariableCategoryMedian (IQR), yearsDFLog-rankp-valueMedia exposureNo19 (17–21)1185.03< 0.001Yes19 (18–22)Age at first sex< 1515 (14–16)23268.11< 0.00115–1717 (17–18)≥ 1821 (19–23)Age at first cohabitation< 1515 (14–16)25832.83< 0.00115–1717 (16–18)≥ 1820 (19–22)Contraceptive useNo19 (17–21)1740.98< 0.001Yes19 (18–21)Spousal age gap< 5 years20 (18–22)15.430.020≥ 5 years19 (17–21)Education levelNo education17.5 (16–20)31084.41< 0.001Primary18 (16–20)Secondary19 (18–21)Higher23 (20–25)Wealth indexPoorest18 (17–20)4952.38< 0.001Poorer18 (17–20)Middle19 (17–21)Richer19 (18–22)Richest21 (19–23)Residence typeUrban20 (18–22)1597.89< 0.001Rural19 (17–21)Marital statusUnmarried19 (17–21)1689.59< 0.001Married19 (17–21)ReligionTraditional19 (17–21)8469.57< 0.001Roman Catholic20 (18–22)Protestant20 (18–22)Pentecostal20 (18–22)Apostolic sect19 (17–21)Other Christian19 (17–21)Muslim19 (18–20)None18 (16–20)Other19 (17–20)Knowledge of ovulationDuring period19 (17.5–22)6182.83< 0.001After period19 (17–21)Middle of cycle20 (18–22)Before period19 (18–22)At any time19 (17–20)Other18 (15.5–18)Don’t know19 (17–21)RegionManicaland19 (17–21)9322.42< 0.001Mashonaland Central19 (17–21)Mashonaland East19 (17–21)Mashonaland West18 (17–21)Matabeleland North19 (17–20)Matabeleland South19 (17–21)Midlands19 (17–21)Masvingo20 (18–22)Harare20 (18–22)Bulawayo20 (18–22.5) Median age at first birth (years) by socio-demographic, behavioral, and reproductive characteristics, Zimbabwe DHS 2015 Women’s timing of first birth varies by education, media exposure, and place of residence. Women with higher education tend to have their first birth later than those with lower education (Fig. 1a), with median age at first birth increasing from no education/primary to secondary and highest education levels. Media exposure is associated with modestly delayed first births, with survival curves diverging during the late teens to early twenties (Fig. 1b). Residence also influences timing: rural women experience earlier first births, whereas urban women show a slower decline in survival probability, indicating later childbearing (Fig. 1c). Kaplan–Meier curves showing time to first birth among reproductive-age women in Zimbabwe: (a) by education level, (b) by media exposure, (c) by residence type, and (d) by marital status The Kaplan-Meier survival curve illustrates significant differences in time to first birth by marital status (Fig. 1d). Married women exhibit a steep decline in survival probability, indicating a rapid transition to first birth following marriage. In contrast, unmarried women maintain higher survival probabilities across reproductive ages, reflecting delayed childbearing. Proportional-hazards assumption The proportional hazards assumption of the Cox model was assessed using Schoenfeld residuals. The findings suggested that majority of the covariates were highly significant violators of the Proportional Hazards Assumption (p < 0.05). Age at first sex, age at first cohabitation, contraceptive use, education level, wealth index, residence type, region, current age, religion, and ovulation knowledge violated the assumption of proportionality (p < 0.05, Table 5). There were also statistically significant deviations of media exposure and marital status of the proportional hazard’s assumption. Table 5Test of proportional-hazards assumption for covariates in the multivariable Cox model of time to first birthVariableχ²dfp-valueAge at first sex1321.522< 0.001Age at first cohabitation733.892< 0.001Marital status12.2910.00046Media exposure5.1610.02317Contraceptive use89.231< 0.001Education level315.283< 0.001Wealth index92.564< 0.001Residence type44.611< 0.001Region61.569< 0.001Current age20.121< 0.001Working status3.5710.05893Religion77.887< 0.001Ovulation knowledge20.2050.00115Global test1599.4038< 0.001 Test of proportional-hazards assumption for covariates in the multivariable Cox model of time to first birth Only working status did not violate statistically significantly (p = 0.0589), which means that its impact on the hazard did not change significantly over the years. The proportionality test was significant (χ²= 1599.40, df = 38, p < 0.001), which suggests that the model, in general, failed the proportional hazards test. Hence alternative models were considered. Comparison of the survival models Among the standard parametric survival models, the log-logistic model showed the showed the best relative fit among the candidate models for this dataset (AIC = -7114.16; BIC = -6857.63), followed by the lognormal model (AIC = -6253.59; BIC = -5997.06). The Weibull model showed poorer fit (AIC = -2929.62; BIC = -2673.09), while the exponential model showed the poorest fit (AIC = 10,782.25; BIC = 11,032.21), suggesting that the assumption of a constant hazard over time was not appropriate for modelling time to first birth. To account for unobserved heterogeneity, frailty models were estimated. Among these, the log-logistic frailty model showed the showed the best relative fit (AIC = -8683.97; BIC = -8563.40), outperforming the Weibull frailty model (AIC = -2698.93; BIC = -2578.36). Across all models considered, including both standard and frailty specifications, the log-logistic frailty model had the lowest AIC and BIC values, indicating better relative fit for this dataset. However, this finding should be interpreted as specific to the current data and modelling context rather than indicative of a universally superior model (Table 6). Table 6Comparison of parametric survival modelsModelLog-likelihood (Null)Log-likelihood (Model)dfAICBICWeibull169.3861503.81239-2929.624-2673.093Exponential-5417.698-5353.1273810782.25011032.210Lognormal1377.2853165.79539-6253.590-5997.059Log-logistic1512.9433596.07839-7114.156-6857.625Weibull Frailty48.6951367.46218-2698.925-2578.360Log-logistic Frailty1800.2914359.98418 -8683.968 -8563 .404 Lower AIC and BIC values indicate better model fit. BIC uses N equal to the number of observationsBold text is highlingting the model with AIC and BIC Comparison of parametric survival models Lower AIC and BIC values indicate better model fit. BIC uses N equal to the number of observations Bold text is highlingting the model with AIC and BIC Log-logistic AFT models with gamma frailty Results from the multivariable log-logistic accelerated failure time (AFT) model with gamma frailty are presented (Table 7). Exponentiated coefficients were expressed as time ratios (TRs). Values greater than 1 indicate a longer time to first birth, and values less than 1 indicate a shorter time. Table 7Bivariate and multivariable log-logistic AFT models with gamma frailty for time to first birth, Zimbabwe DHS 2015 (N = 9,955)PredictorBivariateMultivariableCategoryTR (95% CI)p-valueaTR (95% CI)p-valueAge at first sex< 15 (Ref)1.001.0015–170.99 (0.98–1.01)0.0860.98 (0.97–0.99)< 0.001*≥ 181.17 (1.16–1.19)< 0.0011.08 (1.07–1.09)< 0.001*Media exposureNot exposed (Ref)1.001.00Exposed1.04 (1.04–1.05)< 0.0011.00 (0.99–1.01)0.420Age at first cohabitation< 15 (Ref)1.001.0015–171.15 (1.13–1.16)< 0.0011.15 (1.13–1.16)< 0.001*≥ 181.35 (1.33–1.36)< 0.0011.25 (1.23–1.27)< 0.001*Marital statusNot married (Ref)1.001.00Married0.94 (0.93–0.95)< 0.0010.98 (0.98–0.99)< 0.001*Contraceptive useNo (Ref)1.001.00Yes0.94 (0.93–0.95)< 0.0010.97 (0.96–0.97)< 0.001*Spousal age gap< 5 years (Ref)1.001.00≥ 5 years0.98 (0.97–0.99)< 0.0010.98 (0.97–0.99)< 0.001*Education levelNone (Ref)1.001.00Primary1.01 (0.99–1.04)0.3071.01 (0.99–1.04)0.307Secondary1.03 (1.01–1.06)0.0121.03 (1.01–1.06)0.012Higher1.11 (1.08–1.14)< 0.0011.11 (1.08–1.14)< 0.001*Wealth indexPoorest (Ref)1.001.00Poorer0.99 (0.98–1.01)0.4560.99 (0.98–1.01)0.456Middle0.99 (0.98–1.01)0.6060.99 (0.98–1.01)0.606Richer1.00 (0.99–1.02)0.6781.00 (0.99–1.02)0.678Richest1.03 (1.01–1.04)< 0.0011.03 (1.01–1.04)< 0.001*Residence typeUrban (Ref)1.001.00Rural1.01 (0.99–1.01)0.2601.01 (0.99–1.01)0.260Employment statusNot working (Ref)1.001.00Working0.99 (0.98–0.99)0.0020.99 (0.98–0.99)0.002*ReligionTraditional (Ref)1.001.00Roman Catholic1.02 (0.98–1.05)0.3671.02 (0.98–1.05)0.367Protestant1.03 (0.99–1.07)0.1021.03 (0.99–1.07)0.102Pentecostal1.03 (0.99–1.06)0.1611.03 (0.99–1.06)0.161Apostolic sect1.02 (0.98–1.05)0.3991.02 (0.98–1.05)0.399Other Christian1.00 (0.97–1.04)0.8371.00 (0.97–1.04)0.837Muslim1.00 (0.94–1.06)0.9141.00 (0.94–1.06)0.914None1.00 (0.97–1.04)0.9171.00 (0.97–1.04)0.917Knowledge of ovulationDon’t know (Ref)1.001.00After period1.01 (0.98–1.04)0.6571.01 (0.98–1.04)0.657Middle of cycle1.01 (0.98–1.05)0.4021.01 (0.98–1.05)0.402Before period1.01 (0.98–1.04)0.4331.01 (0.98–1.04)0.433At any time1.01 (0.98–1.04)0.6411.01 (0.98–1.04)0.641Other0.97 (0.90–1.04)0.3970.97 (0.90–1.04)0.397Covariates measured at survey time; associations should be interpreted as correlational rather than causal* represents significant predictors at 5% signficance level Bivariate and multivariable log-logistic AFT models with gamma frailty for time to first birth, Zimbabwe DHS 2015 (N = 9,955) Covariates measured at survey time; associations should be interpreted as correlational rather than causal * represents significant predictors at 5% signficance level Age at first sexual intercourse was significantly associated with time to first birth. Women who initiated sexual activity at ages 15–17 years had a 2.0% shorter time to first birth compared with those who started sexual activity before age of 15 (TR = 0.980, 95% CI: 0.970–0.990, p < 0.001), while those who started at ≥ 18 years had an 8.0% longer time (TR = 1.080, 95% CI: 1.070–1.090, p < 0.001). Age at first cohabitation was associated with time to first birth. Women who began cohabitation at ages 15–17 years had a 15.0% longer time to first birth compared with those who started before age 15 (TR = 1.150, 95% CI: 1.130–1.160, p < 0.001), while those who started at ≥ 18 years had a 25.0% longer time (TR = 1.250, 95% CI: 1.230–1.270, p < 0.001). In the multivariable model, timing of first birth was associated with age at first sex, age at first cohabitation, marital status, contraceptive use, spousal age gap, education, wealth, and employment. Women who initiated sexual activity at 15–17 years had a slightly shorter time to first birth (aTR = 0.98, 95% CI: 0.97–0.99) compared with those < 15, whereas initiation at ≥ 18 years was associated with an 8% longer interval (aTR = 1.08, 95% CI: 1.07–1.09). Age at first cohabitation showed a strong effect: cohabiting at 15–17 years increased the time to first birth by 15% (aTR = 1.15, 95% CI: 1.13–1.16), and cohabiting at ≥ 18 years increased it by 25% (aTR = 1.25, 95% CI: 1.23–1.27) compared with those cohabiting before age 15. Marriage was associated with a modest 2% shorter interval to first birth (aTR = 0.98, 95% CI: 0.98–0.99), indicating an accelerated transition once unions are formed. Contraceptive use reduced time to first birth by 3% (aTR = 0.97, 95% CI: 0.96–0.97), and a spousal age gap ≥ 5 years shortened it by 2% (aTR = 0.98, 95% CI: 0.97–0.99). Educational attainment delayed first birth, with secondary education extending the interval by 3% (aTR = 1.03, 95% CI: 1.01–1.06) and higher education by 11% (aTR = 1.11, 95% CI: 1.08–1.14) relative to women with no education. Wealth was associated with birth timing only at the richest level, with a modest 3% longer interval (aTR = 1.03, 95% CI: 1.01–1.04). Employment slightly accelerated first birth by 1% (aTR = 0.99, 95% CI: 0.98–0.99). In the model-based log-logistic AFT frailty analysis, later age at first cohabitation, later sexual debut, higher education, and higher wealth were associated with delayed first birth, while marriage, contraceptive use, wider spousal age gap, and employment were associated with shorter time to first birth. Significant frailty variance (θ = 0.671, p < 0.001) indicated unobserved regional heterogeneity, and model fit was adequate (LR: χ² (29) = 4935.67, p < 0.001). These findings should be interpreted as model-based associations after accounting for unobserved regional heterogeneity, rather than as design-weighted nationally representative estimates. Discussion This study confirms that the transition to first motherhood in Zimbabwe occurs at an early age, with a national median of 19 years. This finding is consistent with the broader demographic landscape of Sub-Saharan Africa, where early childbearing remains a persistent phenomenon. Similar median ages have been reported in Ethiopia (20 years, IQR: 16–24) using an inverse-Weibull gamma shared frailty model [8] and across multiple high-fertility countries [4]. However, our analysis reveals significant stratification; while women with no education had a median age at first birth of 17.5 years, those with higher education delayed this transition to 23 years. This stark disparity underscores that national medians can mask profound inequalities driven by socioeconomic and educational factors, a pattern consistently observed in the region [3, 14]. Our application of the log-logistic Accelerated Failure Time (AFT) model with gamma frailty showed the best relative fit for this dataset, offering direct interpretations of how covariates accelerate or decelerate the time to first birth. The results delineate clear pathways shaping reproductive timing. The strong delaying effect of later age at first sex (Time Ratio, TR = 1.08 for ≥ 18 years) and, most notably, later age at first cohabitation (TR = 1.25 for ≥ 18 years) highlights the pivotal role of delayed union formation. This aligns with extensive research identifying early marriage as a primary driver of adolescent childbearing across sub-Saharan Africa [16, 17]. The 25% longer time to first birth for women initiating cohabitation after 18, compared to those starting before 15, This study illustrates the powerful normative pressure for childbearing within unions. The protective effect of higher education (TR = 1.11) reinforces a cornerstone of demographic theory: schooling is a critical mechanism for delaying motherhood [5, 18]. Education fosters this delay by expanding life aspirations beyond traditional roles, increasing economic opportunities, and enhancing access to health information [3, 12]. Our finding that contraceptive use was associated with a slightly shorter time to first birth (TR = 0.97) may seem counterintuitive but is logically consistent with the sequencing of life events in a high-fertility context. It suggests that in Zimbabwe, contraceptive use often commences within marriage to manage the timing of a first birth rather than to delay its onset indefinitely, a pattern observed in other settings where marriage and childbearing are closely linked [1, 9, 19, 20]. Similarly, a larger spousal age gap was associated with earlier childbearing, potentially reflecting more traditional and pronatalist marital norms where younger wives have less autonomy over reproductive [1, 21]. A critical finding of this study is the significant regional heterogeneity, confirmed by both the stratified medians (ranging from 18 years in Mashonaland West to 20 years in Bulawayo and Harare) and the highly significant gamma frailty term (θ = 0.671, p < 0.001). This indicates that powerful, unobserved contextual factors at the provincial level exert a substantial influence on reproductive timing. This finding resonates strongly with spatial analyses in other developing countries. For instance, Matabeleland North and South were identified as hotspots for early sexual debut in Zimbabwe [22], while [5] documented a high spatial clustering of adolescent pregnancy in northern Nigeria. Similar geographic disparities in reproductive health outcomes have been documented in Angola [23], and Bangladesh [21]. The significant frailty variance underscores that individual-level models are insufficient; community-level norms, economic structures, and variations in health service accessibility are fundamental drivers of reproductive behavior ([7, 8, 21]). The findings from this study are comparable with a larger range of evidence in varying geographical regions. Ethiopian studies have regularly reported the same trends based on Demographic and Health Survey data (Central Statistical Agency [CSA] [15]. It has been found that in South Asia comparable relationships exist among early marriage, educational level, and first birth time, as has been determined in Bangladesh [21] and Nepal [19]. The cross-cultural similarity of these relationships is also supported by research conducted in Angola [23], Nigeria [4], and Uganda [24]. It has been repeatedly noted by the agencies of the World Health Organization and the United Nations that multifaceted and context-specific interventions are necessary to deal with adolescent pregnancy and early childbearing, which is fully justified by the regional heterogeneity observed in the current research. The implication of these findings on policy is that a standard, country-wide intervention approach is ineffective. We recommend a two-track, geographically specific strategy, which is also suggested by the global health organizations and research on other African contexts [18]. In the areas where the median age at first birth has been lower, like Mashonaland West, the interventions should be highly targeted at the underlying factors: preventing early marriages by conducting community discussions and the legalization of the issue and linking girls to secondary schooling. On the other hand, in some provinces such as Harare and Bulawayo where the shift is already behind schedule then the emphasis can be made on strengthening these positive changes through the implementation of comprehensive sexuality education (CSE) and youth-friendly health services to enable women to plan and space their births. To bridge the gap between policy and practice in Zimbabwe, several actionable strategies are recommended. First, the Ministry of Primary and Secondary Education should strengthen the in-school CSE curriculum by embedding practical life-skills modules focused on goal setting, financial literacy, and critical analysis of gender norms. Second, there is an urgent need to bolster youth-friendly, community-based health services, which includes training community health workers to provide confidential counselling and a broad range of contraceptive methods, including Long-Acting Reversible Contraceptives (LARCs), to all adolescents. Finally, sustained interventions must engage families and community leaders to shift the underlying economic and social drivers of early marriage, consistently communicating the long-term benefits of educating girls and delaying first births. This study, utilizing data from the 2015 Zimbabwe Demographic and Health Survey (ZDHS), provides a critical pre-COVID-19 baseline for understanding the timing of first birth among Zimbabwean women. The choice of this dataset is strategic, as it captures the country’s reproductive health landscape before the potential disruptions caused by the global pandemic, which may have altered access to education, health services, and economic stability, thereby influencing fertility trajectories ([25]. Consequently, our findings represent the established patterns and factors associated with early childbearing in Zimbabwe, offering a vital benchmark against which future, post-pandemic trends can be measured. The key strength of this study lies in its use of a nationally representative sample and its robust methodological approach. The log-logistic AFT model with gamma frailty was specifically suited to manage the non-monotonic hazard of first birth and to account for unobserved regional-level clustering, providing more reliable and interpretable estimates. However, the study is not without limitations. Its cross-sectional design precludes causal inference, and the data on the timing of life events are subject to recall bias. Furthermore, while the model included key covariates, unmeasured confounders such as women’s autonomy, detailed partner characteristics, and specific cultural attitudes influence the results and represent an important avenue for future qualitative and mixed-methods research. Although the findings suggest that the log-logistic accelerated failure time (AFT) frailty model fits the Zimbabwe DHS data best, the same conclusion should be understood in the context of the empirical data used. The paper does not intend to determine the universal methodological superiority of log-logistic AFT frailty model in all the survival data environments. Alternatively, the model is suggested as an interpretable and flexible alternative to the proportional hazards when the proportional hazards assumption is not met and when the functions of hazards are non-monotonic. This analysis can be further developed in the future with simulation studies to provide a systematic assessment of model performance despite the conditions. These simulations would be able to study the extent of censoring, sample size, frailty variance parameters, and hazard function shape (e.g. monotonic versus non-monotonic), and various levels of proportional hazards violation. Bias, root mean squared error (RMSE), coverage probability, convergence rates, and model selection frequency are performance metrics that would give an extra insight into the strength of competing survival models. Such simulations would enable more generalization of the methodological framework, not only in relation to the empirical context under consideration in the present case. There is no simulation-based checking on the proposed model in the study. Even though we demonstrate that the log-logistic AFT gamma frailty model fits the 2015 Zimbabwe DHS data on the timing of first birth best, we do not claim that the model fits the data best under every data-generating condition. The omission of the simulation analysis is limiting to our ability to assess the model performance under varying degrees of censoring, sample size, frailty variance, or PH violation degrees. We, therefore, offer our conclusion as context-specific and not general. Future research ought to complement the findings of our study by the controlled simulations as well as the replications of our findings with other DHS datasets in other countries or different periods of time to verify the claim that the log-logistic AFT frailty model is invariably the correct specification in the identical kind of usage on another set of demographics. Conclusion This study provides an empirical demonstration of a systematic survival model comparison framework for time-to-event data where the proportional hazards assumption is violated. Through a systematic comparison of survival models, it establishes that the log-logistic Accelerated Failure Time (AFT) model with gamma frailty provides a provided the showed the best relative fit for this dataset to the conventional Cox model. The global violation of the proportional hazard’s assumption invalidates the Cox model, while the log-logistic frailty model’s lower AIC and BIC confirm that it showed the best relative fit for this dataset. Interpreted through Time Ratios (TRs), the findings reveal that higher education and delayed cohabitation postpone first births, highlighting the role of female empowerment and protection from early marriage. Conversely, faster transitions to first birth among married women and contraceptive users emphasize the contextual nature of reproductive behavior, where contraceptive use primarily supports family formation. A major contribution of this analysis is the detection of significant unobserved regional heterogeneity, confirming that reproductive timing is associated with not only by individual factors but also by community-level contexts. Regional differences in median age at first birth from 18 years in Mashonaland West to 20 years in Harare and Bulawayo underscore the need for geographically targeted policies. Methodologically, this study provides a reproducible framework for survival analysis when Cox assumptions fail, highlighting the practical advantages of the log-logistic AFT frailty model for datasets exhibiting non-monotonic hazards and unobserved heterogeneity. Substantively, it offers actionable insights for Zimbabwe and similar settings: delaying first births, improving maternal, and child health require sustained investment in girls’ education, strict enforcement of child marriage laws, and regionally tailored reproductive health interventions. Acknowledgements Not applicable. Clinical trial number Not applicable. Abbreviations Accelerated Failure Time Akaike Information Criterion Adjusted Time Ratio Bayesian Information Criterion Confidence Interval Comprehensive Sexuality Education Demographic and Health Survey Interquartile Range Long-Acting Reversible Contraceptives Sub-Saharan Africa Time Ratio Zimbabwe Demographic and Health Survey Authors' contributions HTT conceptualised the idea, processed data into the software and delivered the results for this manuscript. HTT drafted the initial manuscript as all author made substantial revisions. HTT commented on the final draft of the manuscript. The author read and approved the final manuscript. Funding Not applicable. Data availability This study utilized publicly available data from the Zimbabwe Demographic and Health Survey (ZDHS) 2015, obtained from the Demographic and Health Surveys (DHS) Program. The datasets are accessible to registered users upon reasonable request at the DHS Program website (https://dhsprogram.com/data/). Researchers can access the same data used in this analysis by selecting the Zimbabwe DHS 2015 dataset and submitting a data request that specifies the intended use. The author did not have any special access privileges that others would not have. Declarations Ethics approval and consent to participate The DHS women’s sample includes respondents aged 15–49 years. For participants aged 15–17 years, the original DHS survey obtained consent according to DHS ethical procedures. This secondary analysis used anonymized public-use data and required no additional ethical approval was required. Consent for publication The current research is a secondary data analysis of anonymous publicly available data from the Demographic Health Surveys Program. The primary data collection by the Demographic Health Surveys Program has obtained informed consent from all the participants, including their consent to the data being used for future research and publication after anonymization. The current manuscript does not contain any identifiable information, individual images, or data that require additional consent for publication. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement This study utilized publicly available data from the Zimbabwe Demographic and Health Survey (ZDHS) 2015, obtained from the Demographic and Health Surveys (DHS) Program. The datasets are accessible to registered users upon reasonable request at the DHS Program website (https://dhsprogram.com/data/). Researchers can access the same data used in this analysis by selecting the Zimbabwe DHS 2015 dataset and submitting a data request that specifies the intended use. The author did not have any special access privileges that others would not have.