Predictors of breastfeeding duration among Ethiopian women of childbearing age with infants; application of failure time acceleration and parametric shared instability models | BMC Nutrition

by admin
Predictors of breastfeeding duration among Ethiopian women of childbearing age with infants; application of failure time acceleration and parametric shared instability models | BMC Nutrition

[ad_1]

Descriptive statistics

Baseline characteristics of the covariates are listed in Table 1. Table 1 indicates that approximately 15,400 lactating women of reproductive age with infants were included in this study. Among them, about 27.6% of participants were censored, with the actual duration of breastfeeding unknown, and the rest were events with known duration of breastfeeding, 65.7% lived in rural areas, about 42.6% of women were from poor households, 18.2% of them are from middle-income households and the rest are from high-income households. Table 1 also shows that about 63.6% of the participants are women with female children and 65.4% of the participants are out of work (unemployed). Out of 10,072 unemployed women, 27.7% of them are still breastfeeding and 72.3% of women have stopped breastfeeding (event). Out of a total of 5,328 employed women, 27.5% of women are still breastfeeding and 72.5% of women have stopped breastfeeding.

Among the participants, about 92.1% were non-smokers, 17.9% of women had secondary and higher education, 49.7% of women had primary education, and the rest were uneducated women. The religion category in Table 1 shows that about 41.1% of the participants are Orthodox, 39.4% are Muslim and 17.9% of them are Protestant. As for their regions, about 11.8% of them are from Addis Ababa city, 12.0% from Oromia, 11.7 from Old SNNP, 11.0% from Amhara, about 7.1% of them are from Afar, 7.2% from Benishangul-Gumuz, 7.3% from Dire Dawa, 6.5% from Gambela, 5.8% from Harari and about 10.8% of them are from the Tigray region.

Nonparametric survival analysis

Non-parametric analysis was performed using Kaplan Meir curves plots of survival and risk experience for duration of breastfeeding as shown in Fig. 1 and Fig. 2. The result revealed that the survival graph decreases at an increasing rate at the beginning and decreases further from time to time. This means that most women breastfed heavily soon after giving birth. On the other hand, the hazard diagram grows at an increasing rate at the beginning and increases with increasing time.

Fig. 1

Survival chart for duration of breastfeeding

Fig. 2
figure 2

Risk chart for duration of breastfeeding

Log-rank test

The log-rank test was used at the 5% significance level to validate differences in survival time of each factor. The difference between the probabilities of an event occurring at each point in time was the null hypothesis that was tested.

Cox proportional hazard regression model

After we compared survival experience across groups of covariates, the next important step was to develop a model. An initial step in the model building process was conducted to identify sets of explanatory variables that had the potential to be included in the linear components of a multivariable proportional hazards model. A Cox proportional hazards regression model was then fitted, including the univariate Cox proportional hazards models. From univariate analysis; place of residence, family wealth index, child sex, child age, women’s smoking status, birth interval, place of residence, wealth index, and women’s education level were statistically significant for the variable of interest.

Cox Proportional Hazards Mode Model Diagnostics1

In the present investigation, the two main assumptions of the Cox regression model, log-linearity and proportional hazards, were tested as shown in Table 2. The log-linearity test revealed that the relationship between the log-hazard or the log-cumulative hazard and the covariate was linear. The proportional hazard test in this investigation shows that the ratio of the hazard function for two individuals with different regression covariates does not change over time.

Table 2 Result of the test of the assumption of proportionality for each covariate

The global goodness-of-fit test in Table 2 also shows that the Wald chi-square test statistic is significant, indicating that the proportional hazards assumption is violated. In other words, the plots of the breastfeeding duration covariates are not parallel to each other. Thus, there is a violation of the proportional hazards assumption. This shows that the residuals are not random, there is a systematic pattern and the smoothed plot does not look like a straight line and has some deviation from the horizontal line. Therefore, there is a violation of the proportional hazards assumption.

Since the proportional hazard assumptions were not met, the accelerated failure time (AFT), including univariate and multivariate analysis, model should be run for the current data analysis. Univariate analyzes were fitted for each covariate, considering AFT models for participants’ baseline characteristics. AFT Weibull, exponential, log-logistic and log-normal distribution models were considered for duration of breastfeeding data. To select the best model for the present analysis, the AFT models, namely Weibull, exponential, log-logistic and log-normal distribution were compared using AIC and BIC, considering that the model with the smallest AIC and BIC is the one that fits the data well, as shown in Table 3. Table 3 shows that the Weibull distribution has the smallest AIC and BIC. Therefore, it was selected for univariate and multivariate data analysis in the present investigation.

Table 3 Comparison of AFT models using AIC, BIC criteria for breastfeeding duration data

In all univariate analyzes of Weibull AFT models, women’s age, smoking status, place of residence, household wealth index, child’s gender, women’s education level, children’s age, and women’s birth interval were significantly associated with duration of breastfeeding at the 5% significance level. A summary of the univariate analysis is provided in Table 4. Variables that were significant at the 5% significance level in the univariate model were included in the multivariate analysis of the data. The multivariate analysis of the data within this study is shown in Table 5. In Table 5, the main effect of the covariates, namely place of residence, women’s age, birth interval, women’s education level, smoking status were taken into account. and child age as potential predictors of the variable of interest. Multivariate analysis of data with Weibull AFT models and corresponding AIC and BIC values ​​were performed as indicated in Table 5.

Table 4 Univariate data analysis for Weibull ATF model
Table 5 The result of the multivariate analysis of the data is the final AFT Weibull model

The result of the Weibull AFT model in Table 5 shows that education level, women’s age, birth interval, place of residence, child’s gender, smoking status and wealth status are statistically significant variables for the variable of interest.

Analysis and model comparisons for a parametric shared frailty model

To test the effect of regions on the variable of interest, a multivariate survival analysis including the Gamma shared frailty term was performed. This was done using the covariates; place of residence, women’s education level, religion, employment status, child’s age, child’s gender, wealth index, women’s education level, smoking status and birth interval. In this study, the AIC and BIC criteria were considered to compare different candidate parametric shared frailty models, considering the model with the smallest AIC and BIC as the best model.

Parametric output distributions, namely gamma frailty distribution, log-normal and inverse Gaussian distributions were fitted and compared by considering women’s regions as frailty terms. The effect of the random component (frailty) was significant for Weibull gamma shared frailty due to the smallest AIC and BIC values [20]. The final shared Weibull gamma frailty model is shown in Table 6.

Table 6 Parametric estimates of the Weibull-Gamma Frailty Model

Table 6 shows that residential area has a significant effect on breastfeeding duration. Therefore, the expected duration of breastfeeding for urban women was reduced by 4% compared to rural women, holding other covariates constant (Φ = 0.96; 95% CI; (0.94,0.97); p-value = 0.001). This indicates that women in rural areas had a longer duration of breastfeeding than women in cities.

Level of education has a significant impact on variation in breastfeeding duration. Therefore, comparing uneducated women with medium and higher level, the expected duration of breastfeeding for uneducated women increased by 3% compared with medium and higher level, holding other covariates constant (Φ = 1.03; 95% CI ;(1.00) ,1.06); p-value = 0.039). Similarly, the expected duration of breastfeeding for women with primary education was increased by 13% compared with secondary education and above, holding other covariates constant (Φ = 1.13; 95% CI; (1.11,1.15 ); p-value < 0.001). This result indicates that more educated women had a short duration of breastfeeding compared to uneducated or less educated women.

The age of the child also plays an important role in the change in the duration of breastfeeding. Therefore, as the child’s age increases by one month, the expected duration of breastfeeding decreases by 1%, holding other covariates constant (Φ = 0.99; 95% CI; (0.76, 0.99); p-value < 0.001). Therefore, the increase in the age of the child leads to a decrease in the duration of breastfeeding.

Smoking status of mothers/women also plays a significant role in variation in breastfeeding duration. Comparing female smokers to female nonsmokers, the expected duration of breastfeeding for female nonsmokers increased by 60% compared with female smokers, holding other covariates constant. (F = 1.60; 95% CI; (1.57, 1.63); p-value < 0.001).

Birth interval between successive births also has a significant effect on variation in breastfeeding duration. Therefore, comparing a woman whose birth interval is between 2-3 years with that of less than 2 years, the expected duration of breastfeeding for a woman whose birth interval is 2-3 years increases by 2% compared to woman whose birth interval, < 2 years, holding other conditions constant (Φ = 1.02; 95% CI; (1.09, 1.25); p-value < 0,027). По подобен начин, очакваната продължителност на кърмене за жена, чийто интервал на раждане > 4 years, was increased by 28% compared to a woman whose birth interval < 2 years (Φ = 1.28; 95% CI; (1.06, 1.43); p-value < 0.01).

Women’s age is also another significant variable in changing breastfeeding duration. Therefore, the expected duration of breastfeeding for a woman aged 40–44 years increased by 4% compared to a woman aged 15–19 years, holding other covariates constant (Φ = 1.041; 95% CI (1.01 , 1.22); p-value = 0.019) and expected breastfeeding duration for a woman aged 35–39 years increased by 3% compared to a woman aged 15–19 years (Φ = 1.030; 95% CI; (1.01, 1.55 ); p-value = 0.005). Therefore, the more women’s age leads to longer duration of breastfeeding for the present study.

[ad_2]

Source link

You may also like