2.1 Data source and sampling procedure

We used one of the largest and most highly recognized NDHS 2022 datasets available in Nepal to assess the distribution of the number of under-five mortality. This dataset was obtained from the Demographic and Health Survey (DHS) program, ICF, Rockville, Maryland, USA, after submitting a brief description about the purpose of the study based on this data. The NDHS 2022 dataset incorporated up-to-date information on childhood mortality levels and other child, maternal, demographic, and health-related variables exclusively. The dataset encompassed a nationally representative sample of 14,280 households. Within these selected households, all women aged 15–49, or those within the same age group who remained overnight in the household preceding the survey, were eligible for participation in the study. The data were collected from January 5, 2022 and June 22, 2022.

As per the federal structure declared on September 15, 2015 by the Constitute Assembly of Nepal, an updated sampling frame was prepared to conduct the NDHS 2022 based on the frame used by the Central Bureau of Statistics (CBS) (now the office is named as National Statistics Office (NSO)) in the 2011 Nepal Population and Housing Census (NPHC). The NDHS 2022 adopted a two-stage stratified sampling technique to select the respondents. Altogether 14 sampling strata were created by dividing each of the seven provinces into rural and urban areas. In the initial stage, 476 primary sampling units (PSUs) were chosen. Among the selected PSUs, 248 were from urban areas and 228 were from rural areas. These PSUs were selected independently from each sampling stratum within the sample allocation and with probability proportional to PSU size. Before conducting the main survey, the household listing was finalized in all the selected PSUs. This list served as a sampling frame for the selection of households in the second stage. Moreover, in the cases of households listing when selected sub-wards were found excessively large with a projected number of households more than 300, then such sub-wards were segmented and just one segment was picked with the probability proportional to segment size for the survey. In the next step, 30 households were chosen from each cluster to get a total of 14,280 households in the sample. From the selected households, all women between the ages 15 and 49 years, whether they were permanent residents or visitors, were eligible to participate in the survey if they spent the night in the household just before the survey. These women were asked about numerous health-related issues, including their background characteristics, pregnancy history, year and month of each childbirth, and death. The complete procedure of data collection and sampling techniques used in the survey is described in the NDHS 2022 final report [1].

Altogether 14,845 women were interviewed in the survey. To develop a statistical model relating the under-five mortality to its covariates, respondents with missing values were eliminated. Therefore, the final models were fitted using the information from only 8119 respondents.

2.2 Study variables and data analysis

The response variable of this study was the total number of children who departed before reaching the age of 5 for each woman throughout her lifetime, and it is recorded in terms of counts as 0, 1, 2, 3, ……, n. The comprehensive risk factors associated with under-five mortality suggested by the literature and available in the NDHS 2022 dataset are included in the study [29]. These risk factors, their categories, and some summary statistics related to the number of under-five mortality are presented in Table 1. Most of the covariates were reclassified into new categories, while some retained their original categories as reported in the data. The levels of each covariate were defined as per the literature on childhood mortality, relevance to meaningful interpretation from a policy point of view, and considering the explorative analysis of this dataset. Some covariates had missing values. We did not use any imputation methods to replace the missing values. The cases with missing values were excluded from the analysis ensuring that the dataset has sufficient size for the study. The consistency of the data was confirmed by computing frequency tables and cross-tabulations. For data analysis, three different software programs were utilized. Data management tasks, including coding, recoding, and cleaning, were conducted in SPSS 26. Descriptive analysis, such as generating graphs for the outcome variable, calculating means, and determining standard deviations was also performed in SPSS 26. Additionally, stepwise regression analysis, incorporating both forward selection and backward elimination with an entry and removal probability of 0.05, was executed in STATA 18 using the “stepwise” command. Finally, all univariate and multivariate analyses were carried out in R (version 4.3.2) using the “MASS” and “pscl” packages.

2.3 Statistical models

Identical to other disciplines, in some studies related to public health and epidemiology, the response variable of interest consists of count data like the number of under-five mortality a woman has encountered during her lifetime. The statistical modeling of count data is analogous to the regression-type analyses provided that the outcome variable of interest is non-negative integers 0, 1, 2, etc. The count models are useful for estimating the effect of policy interventions either on the average rate or on the likelihood of no event, a single event, or two or more events [25]. The application of OLS regression for count data may not be appropriate due to the violation of normality and constant variance assumptions. The misspecification of the models can result in erroneous estimates of standard errors, leading to ambiguous p-values and confidence intervals [26]. In such a context, count models better analyze the outcome of interest from its covariates. These models enable us to evaluate how the count outcome changes as its covariates increase. In this study, we applied different statistical models, including PR, NBR, ZINBR, and HNBR models to investigate and measure the effect of various factors on the number of under-five mortality a mother experiences during her lifetime as per the nature of the data.

The count data models associated with the generalized linear models (GLMs) are developed on the foundation of Poisson distribution. PR is the most basic model for count data to determine the probability of count by Poisson distribution in which the mean of the distribution is a function of the independent variables. The inimitable feature of this model is that the conditional mean of the response variable is equal to the conditional variance. When conditional variance exceeds the conditional mean, it is a case of overdispersion. When the count data contain overdispersion of the outcome variable, then the NBR model is a better alternative to the PR model [30]. In the PR model if denote the number of under-five mortality for the i^th woman, and denote the mean number of under-five mortality, then in the case when the number of under-five mortality follows Poisson distribution, its probability mass function is given by:

(1)

The mean and variance of the Poisson distribution is . If denotes the number of independent variables, are the covariates included in the model, and are the corresponding coefficients, then the PR model using equation (1) can be expressed as:

(2)

The NBR model also referred to as negative binomial (NB) with two parameters, is constructed using the Poisson-gamma mixture distribution. The NBR model is recommended in the case of overdispersion or modeling the Poisson heterogeneity using a gamma distribution. This model provides the adjusted standard errors of regression coefficients and a more flexible method to predict the count response [31]. The probability distribution function of the NBR model for i^th observation is expressed as:

(3)

The NB distribution is characterized by two parameters, including a mean parameter , which denotes the mean number of under-five deaths per mother, and an overdispersion parameter . The function is the gamma function. When , the negative binomial distribution is the same as Poisson distribution, this means PR model is nested within the NBR model. The mean and variance are and .

Although the NBR model overcomes the issue of overdispersion in the PR model in count data analysis, it cannot adequately model over-dispersed count data caused by an excessive number of zeros. The appropriate model in such cases is to fit the data with a zero-inflated Poisson regression (ZIPR) model or ZINBR model [32]. The zero-inflated models inform about what predicts the occurrence and nonoccurrence of an event and what predicts the frequency of occurrences if the event occurs regarding the count outcome [33]. Moreover, these models account for the modeling of zero counts using both binary and count processes. These models assume two separate data generation processes for count data: the first process generates zeros only and the second data generating process is Poisson or negative binomial [31,34].

Moreover, the ZIPR model fails to fit the count data having excessive zeros with overdispersion. In such a case, the ZINBR is the recommended model. This model also consists of two distinct data generation processes, where is the probability for excess zeros and for the remaining counts that follow a negative binomial distribution [31]. If denote the count of under-five deaths per mother, then the zero-inflated negative binomial (ZINB) distribution is given by:

(4)

Where denote the mean of the negative binomial distribution. The over-dispersion parameter is and it is independent of covariates. The mean of the ZINB distribution is with variance . The ZINB distribution approaches to zero-inflated Poisson (ZIP) distribution when . Likewise, the ZINB distribution converges to the negative binomial distribution as . If and the ZINB distribution reduces to the Poisson distribution [35].

In contrast to zero-inflated models, the hurdle models are two-part models. In these models, the first part is a binary response model and the second part is a zero-truncated count model [25]. The hurdle model assumes that all zero observations arise from structural sources only. The positive counts (i.e., non-zero) are due to sampling origin and follow either truncated Poisson or negative binomial distribution [36]. In the study of under-five mortality, it appears well to assume that only those women who did not have childbirth will possess zero under-five mortality during a given period, and the women who had given childbirth will have some positive (non-zero) number of under-five mortality during that period. Here the count zero can be obtained only from those women who did not have childbirth, originating from a structural source. If denote the count of under-five deaths per mother, i = 1, 2, ……, n, where n is the total number of mothers, then the general structure of a hurdle model can be expressed as:

(5)

Where is the probability of mother belonging to the zero under-five mortality, denote the probability mass function for a positive count distribution with parameter and is the distribution assessed at zero count.

Additionally, in the hurdle models, the hurdle negative binomial model is the most commonly used model for non-zero count when the Poisson distribution shows overdispersion [37,38]. The hurdle negative binomial model for the count response variable , (i = 1, 2, ……., n) is written as:

(6)

Where the dispersion parameter , which is considered to be independent with covariates. Moreover, and . The logistic regression model used to model the probability of excess zero is given by:

(7)

Where is the i^th row of covariate matrix Z. are unknown q-dimensional column vectors of parameters. The influence of covariates on count outcomes using NBR can be modeled as:

(8)

Where are the covariates, are the regression coefficients, and p denotes the number of covariates.

Fig 1 illustrates a visual representation of the model selection process, developed using insights from the prior study [39]. The presence of overdispersion in the number of under-five deaths was assessed using the likelihood ratio test (LRT). The LRT evaluates the nested models (i.e., NBR vs. PR, ZINBR vs. ZIPR, and HNBR vs. HPR) and confirms the necessity of the overdispersion parameter. This study employs LRT to compare two nested models, PR versus ZINBR to account for overdispersion. As mentioned earlier, the Poisson model is a special case of the negative binomial model with a dispersion parameter . The statistical hypotheses for the LRT to test overdispersion are defined as: the null hypothesis (H₀): (no overdispersion) against the alternative hypothesis (H₁): (overdispersion). The LRT statistic is expressed as;

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Flowchart for the selection of count regression model to analyze the under-five deaths in Nepal.

https://doi.org/10.1371/journal.pone.0324321.g001

(9)

The LRT statistic follows the Chi-square distribution with a degree of freedom equal to the difference in the number of parameters in the model. The significant result of LRT indicates the presence of overdispersion. In this situation, negative binomial models are more appropriate than the Poisson models as shown in (Fig 1) [25,38]. Further, the Vuong test was used to confirm the suitability between pairs of non-nested models, particularly to compare the zero-inflated count models with their standard analogs. In this study, it is used to compare ZINBR versus NBR. The null hypothesis under the test assumes that the two models NBR and ZINBR are indistinguishable versus the alternative hypothesis, the ZINBR is better than the NBR model. The Vuong test statistic is asymptotically distributed as a standard normal variate [40]. The significant result of the Vuong test suggests that either the ZINBR or HNBR model is suitable for predicting the number of under-five mortality. Finally, the ZINBR and HNBR models were compared using the maximum log-likelihood value and the lowest Akaike information criterion (AIC) and Bayesian information criterion (BIC) values. The overall goodness fit of the most plausible model was assessed using the Pearson Chi-square test statistic [41].

In our study, the ZINBR model emerged as the most appropriate count model for evaluating the factors influencing under-five mortality. Multiple factors may render the ZINBR model more appropriate than the HNBR model for explaining under-five mortality. The zero-inflated models assume that excess zeros (no under-five mortality) result from two distinct processes. One process identifies the presence of excess zeros resulting from individuals having strong protective factors such as adequate nutrition, access to healthcare facilities, and favorable socioeconomic conditions, while the other process models the actual count of under-five mortality among those at risk. The results of this study support this notion, as under-five mortality exhibited disparities across different groups and circumstances. However, in the HNBR model, all zero outcomes are attributed solely to the binary process, assuming that zero under-five mortality cannot occur from the count component of the model. In other words, the ZINBR model accounts for both structural zeros (cases where there is no risk of under-five mortality) and sampling zeros (cases where children are at risk but do not experience under-five mortality). In contrast, the HNBR model assumes that if a child is at risk, the absence of under-five mortality cannot occur by chance. This implies that the ZINBR offers a more suitable framework for managing these excess zeros, whereas the hurdle model assumes all children are at some risk, which may not accurately reflect reality.

The incidence rate ratio (IRR) in the count part of the ZINBR model is computed based on the association between the expected under-five mortality count and the linear predictors of the model. In the count part of ZINBR model, the outcome variable is assumed to follow a negative binomial distribution characterized by a mean µ and dispersion parameter The expected mean count is . Here X denotes the matrix of predictors for the count part and β denotes the vector of regression coefficients. The IRR quantifies the multiplicative change in the expected count associated with a one-unit increase in the predictor , holding all other predictors constant. For a particular predictor , it’s coefficient determines the IRR by using the relation .

Similarly, the adjusted odds ratio (AOR) in the zero-inflated part of the ZINBR model is derived from the coefficients of the zero-inflation model, which uses a binomial distribution with a logit link function. In this model, the dependent variable is modeled as a binary outcome. The probability of observing the structural zero is determined using a logistic regression, expressed as;

Where is the probability of structural zero, Z denotes the matrix of predictors for the zero-inflation part, and denotes the vector of regression coefficients of the model [31].

For a given predictor , the odds of being a structural zero can be written as

The AOR for a particular predictor can be obtained as

2.4 Ethical approval

This research used the secondary data accessed from the Demographic and Health Survey (DHS) program. The NDHS 2022 data collection procedures were approved by ICF Macro International (Calverton, Maryland, USA) Institutional Review Board and the survey protocol was reviewed by the Nepal Health Research Council (NHRC). Written consent was obtained from the household head to conduct the interviews.

3.1 Descriptive analyses

The analysis of NDHS 2022 data after accounting for the sampling weights revealed that 14,821 women aged 15–49 were included in this study. Among them, most of the mothers 13,541 (91.4%) had no under-five mortality during their lifetime. Of those involved in the survey 1,280 (8.6%) mothers had one up to six under-five mortality in their lifetime. Among the total 1,055 (7.1%) mothers experienced one under-five mortality followed by 167 (1.1%) mothers with two under-five mortality in their lifetime.

Fig 2 illustrates that the average number of under-five mortality a mother had encountered over her life was 0.11, with a 95% confidence interval of (0.10, 0.11), and a variance of 0.154. The variance greater than the mean indicates the possibility of overdispersion. Moreover, the number of under-five mortality demonstrates the right-skewed distribution with excess zero observations. Zero-inflated models or hurdle models might be appropriate count models to better explain such data structure.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Distribution of the number of under-five deaths experienced by a mother during her lifetime.

https://doi.org/10.1371/journal.pone.0324321.g002

Table 1 shows that the average number of under-five mortality was highest in the mountain region (0.18) followed by terai (mean = 0.11), and the lowest was reported in the hill region (mean = 0.09). The Karnali province recorded the highest anticipated number of under-five mortality (mean = 0.18), followed by Sudurpaschim (mean = 0.15), Madhesh province (mean = 0.14), and Lumbini province (mean = 0.11), whereas the lowest was reported in Bagmati and Gandaki province (mean = 0.07). Likewise, concerning the place of residence, rural areas showed a higher mean number of under-five mortality (mean = 0.15) as compared to urban areas (mean = 0.09). The anticipated number of under-five deaths exhibits an inverse correlation with the educational attainment of both mothers and fathers of children. Mothers with no education reported a significantly higher mean number of under-five mortality (mean = 0.27), whereas the lowest was reported for those mothers achieving secondary and higher education (mean = 0.02).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics of under-five mortality during the lifetime of a mother by various circumstances, NDHS 2022.

https://doi.org/10.1371/journal.pone.0324321.t001

Considering the time to get the water source, the household with the water facility outside the premises had a higher number of under-five deaths (mean = 0.16), whereas the lower expected number (mean = 0.1) was observed in the household having water facility within the premises. Similarly, the higher anticipated count of under-five mortality (mean = 0.19) was recorded in the households without toilet facility than in the households with any type of toilet facility (mean = 0.1). It is also found that households with no access to electricity (mean = 0.17) and having an unfinished type of house (mean = 0.15) experience a higher average number of under-five deaths. The expected number of under-five deaths was found identical among the respondents belonging to the Hindu faith and adhering to other religions. Regarding the mother’s religious affiliation, Table 1 revealed that the average under-five mortality was elevated among the mothers from Dalit, Muslim, and other Terai castes, whereas it was lowest among Brahmin/Chhetri and Janajati mothers. Furthermore, the households having 3 or more children aged 5 or below had high under-five deaths (mean = 0.13), in contrast, the households with 1–2 or no children aged 5 or below experienced a lower mean number of under-five mortality (mean = 0.11). Mothers who had access to mass media experienced a lower expected number of under-five mortality (mean = 0.09) meanwhile, a greater average number of under-five mortality was observed among those who had no access to mass media (mean = 0.16). Similarly, households that utilize wood for cooking exhibited a higher mean number of under-five deaths (mean = 0.14) compared to those using LPG for cooking (mean = 0.06). Table 1 also shows an inverse correlation between households’ wealth index and the expected number of under-five mortality. The households categorized under the poorest wealth index encountered the highest under-five mortality (mean = 0.19), whereas the lowest mortality was observed in those households belonging to the richest wealth index (mean = 0.04). On the other hand, mothers with 3 or more CEB were found to have a higher expected number of under-five deaths (mean = 0.33) than those mothers with 2 or fewer CEB (mean = 0.02).

Mothers who smoke cigarettes experienced a higher mean number of under-five mortality (mean = 0.3) than those who did not smoke. Likewise, mothers who traveled for more than 30 minutes from their house to the nearest healthcare facility exhibited the highest average number of under-five mortality (mean = 0.17) compared to those who traveled 30 minutes or less. The average number of under-five deaths was highest among those mothers where the primary occupation of both husband and wife was agriculture (mean = 0.18) and (mean = 0.456) respectively. It was also found that the mean number of under-five mortality is positively correlated with birth order number. The expected number of under-five mortality was highest (mean = 0.53) for those mothers with birth order 4 or more as compared to those with 1^st-3^rd birth order. The expected number of under-five mortality was also found to differ according to types of birth whether it was single or multiple. The average number of under-five mortality was highest in multiple births (mean = 0.29), while for single births the mortality was (mean = 0.15). Furthermore, the highest mean number of under-five mortality was reported for a preceding birth interval less than 24 months (mean = 0.29) and the lowest was reported for a preceding birth interval 24 months or above (mean = 0.18) suggesting that the preceding birth interval was inversely associated with the under-five mortality. The mean count of under-five mortality was reported lowest for those mothers who received 2 or more TT injections before childbirth (mean = 0.07) whereas the highest mortality (mean = 0.15) was observed for those mothers who received less than 2 TT injections. Mothers who received delivery assistance from traditional birth attendants exhibited the highest expected number of under-five mortality (mean = 0.18) compared to those assisted by health professionals. It was also found that the average number of under-five mortality was greater among individuals who underwent home deliveries (mean = 0.22) than those mothers who gave birth at healthcare centers. Similarly, mothers who visited 4 or more times for ANC detected the lowest mean number of under-five mortality (mean = 0.09) as compared to those who had less than 4 ANC visits. Additionally, the mothers who had a normal delivery (mean = 0.12) and did not receive postnatal care within the first two days after childbirth had the highest expected number of under-five mortalities. Furthermore, there is a positive correlation between the age of the mother at childbirth and the mean number of under-five mortality. The highest average number of under-five mortality was observed among mothers aged 35 and above, followed by those in the age group of 30–35. There was no noticeable difference in the expected number of under-five mortality between children who received basic antigen vaccines and those who did not. Nevertheless, the lowest mean number of under-five mortality was recorded among children who were fully vaccinated compared to those who were not fully vaccinated.

3.2 Inferential analysis

3.2.2 Model selection to predict the number of under-five mortality.

The presence of overdispersion abolishes the possibility of Poisson models, so the suitable alternative model is the NBR model. Nevertheless, the existence of excess zero as shown in (Fig 2) directs towards the likelihood of the zero-inflated models. The Vuong test was used to decide whether a zero-inflated model is more suitable than a standard count model. The calculated result of Vuong z-statistic = −2.814, p-value = 0.00244 at 0.05 level of significance indicates in favor of the alternate hypothesis that the zero-inflated model is more appropriate than the standard count model for this dataset. Finally, the ZINBR and HNBR models were fitted to assess the number of under-five deaths using these seven covariates. After fitting each model, the final model was selected based on the corresponding values of log-likelihood, AIC, and BIC. The lowest AIC and BIC values alongside the maximum value of log-likelihood as illustrated in Table 2 suggest that the ZINBR model is the better model than HNBR to estimate the number of under-five mortality encountered by a mother throughout her lifetime based on the NDHS 2022, however, these results showing no substantial differences. Moreover, the overall goodness of fit was evaluated by computing Pearson Chi-square statistic, which provided a value of 7800.16 with a p-value of 0.9909 at 8098 degrees of freedom. This result offers evidence in favor of the null hypothesis, indicating that the ZINBR model adequately fits the data.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Comparison of different count models to predict the number of under-five mortality.

https://doi.org/10.1371/journal.pone.0324321.t002

Fig 3 illustrates the curves for the difference between observed and predictive probabilities derived by fitting three count models related to the number of under-five mortality. Despite the visual similarities among the fitted curves, examination of the log-likelihood, AIC, and BIC values in Table 2 leads to the conclusion that the ZINBR model appears as the most suitable choice among the considered count models.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Plots of difference between observed and predicted probability of the number of under-five deaths for different models.

https://doi.org/10.1371/journal.pone.0324321.g003

3.2.3 Factors linked with the number of under-five mortality using the ZINBR model.

The fitted ZINBR model consists of two parts. First is the count model (negative binomial with log link), having a similar interpretation as negative binomial regression. The second part is the zero-inflation model (binomial with logit link). The estimates of this model are interpreted similarly to the standard logistic regression model. Table 3 shows the IRR, standard error, 95% confidence interval, z-values, and corresponding p-values for the various factors linked to the occurrence of under-five mortality. In the count model, the covariates including ecological region, mother smokes cigarettes, total CEB, source of drinking water, preceding birth interval, and birth order number were identified as statistically significant factors allied with the number of under-five mortality a mother has experienced throughout her lifetime, however, the place of residence of mothers whether they stayed in rural or urban areas and some levels of the covariates ecological region and source of drinking water remained insignificant at the 0.05 level of significance.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Factors linked to the number of under-five mortality using the ZINBR model.

https://doi.org/10.1371/journal.pone.0324321.t003

As shown in Table 3, mothers from the mountain region as compared to the hill region has 1.176 (IRR: 1.1756, 95% CI: 1.0095–1.369) times higher incidence of experiencing under-five mortality, whereas those living in the terai region incidence rate for the number of under-five mortality increased by 1.079 (IRR: 1.0793, 95% CI: 0.9129–1.2761) times while holding the other covariates constant although it was not statistically significant. Likewise, the smoking habit of mothers has a substantial impact on the incidence rate of under-five mortality. The mothers who smoke cigarettes have a 1.207 (IRR: 1.2067, 95% CI: 1.0392–1.4011) times higher incidence of under-five mortality than those who did not smoke cigarettes. The fitted model also revealed that the incidence rate for the number of under-five mortality is positively and significantly related to the total CEB. The incidence of under-five deaths among mothers with a total CEB 3 or more is expected to have 1.59 (IRR: 1.588, 95% CI: 1.3116–1.9226) times greater than those having a total CEB 2 or less, while holding the other covariates constant. The source of drinking water has also emerged as a significant variable related to the under-five mortality. Those who drink water from tube-wells experienced approximately a 23% lower incidence of under-five deaths (IRR: 0.7719, 95% CI: 0.6191–0.9623) compared to those who drink water from other sources (like bottled water, Tanker truck, cart with small tank, etc.). Preceding birth interval is another covariate that shows a negative and significant correlation with the number of under-five deaths encountered by a mother during her lifetime. As compared to the preceding birth interval of 24 months or more, while holding the other variable constant in the model, the incidence of under-five deaths was 1.206 (IRR: 1.2061, 95% CI: 1.0788–1.3485) times higher among those mothers having a preceding birth interval less than 24 months. On the other hand, the incidence of experiencing under-five deaths was 3.568 (IRR: 3.5681, 95% CI: 3.1512–4.0401) times higher among those children with birth order 4^th or above than those with 1^st- 3^rd birth order.

The second part of the ZINBR model in Table 3 shows the AOR, standard error, confidence interval for AOR, z-value, and p-value for the factor change in the odds of presence in the zero-count group (no under-five mortality) in comparison to non-zero count group (with at least one under-five mortality) in the zero-inflation model (binomial with logit link). Table 3 exhibits that the baseline odds of belonging in the zero-count group was 1.954. The preceding birth interval has a significant impact on the odds of being in the zero-count group, and it was the only significant covariate in the zero-inflation model. The odds of always being in the zero-count group for those children who had a preceding birth interval less than 24 months was 87% (AOR: 0.1295, 95% CI: 0.0259–0.6474) lower than those children with a birth interval 24 months or more. The variables ecological region, place of residence, mother smokes cigarettes, total CEB, source of drinking water, and birth order number were not statistically significant in this model. The broad confidence intervals observed for the AOR concerning the variables total CEB and birth order could potentially stem from convergence problems in the data.

5.1 Strengths and limitations of the study

One strength of this study is its employment of the most comprehensive nationally representative NDHS 2022 data to identify and measure the effects of various factors related to under-five mortality in Nepal. The detailed analysis using descriptive and statistical models addresses the research gap in determining the factors associated with under-five mortality, especially in the context of Nepal. The conclusions drawn from this study, which utilizes a countrywide dataset, could benefit policymakers and concerned authorities in devising effective healthcare strategies aimed at the most at-risk groups to reduce childhood mortality. However, the assessment of under-five mortality was based on the available variables in the data and their suitability for analysis. Some variables, such as the duration of breastfeeding and gestational age of the child, were excluded due to computational complexity and a significant amount of missing values. The clinically significant causes of mortality under the age of five were not included in the study. Moreover, the NDHS 2022 utilized a women’s questionnaire to gather data from all eligible women aged 15–49. They were asked about various topics, including child mortality history, ANC, childbirth, PNC, working status, and their husbands’ background characteristics. Such information was obtained through the respondents’ self-reports, which could lead to potential recall errors, particularly over extended periods. For instance, older mothers may struggle to recall specific details about past events accurately due to memory lapses. They might not remember the exact birth date of their child, the age at which the child died, or whether the death occurred during the neonatal, infant, or under-five periods. This type of recall bias can introduce inaccuracies, including the underestimation or overestimation of under-five mortality rates, misreporting the timing of a child’s death, leading to biased age-specific mortality rates, and recalling recent or traumatic events more accurately than older or less significant ones. Our study did not incorporate several qualitative factors that could impact under-five mortality, such as awareness of childcare practices and hygiene, cultural beliefs and practices, and healthcare-seeking behaviors. Future research may focus on qualitative factors, such as exploring the importance of maternal participation in child healthcare decision-making, reforming detrimental social norms, building community trust in health systems, consciousness on childcare & hygiene practices, and examining environmental influences on child health. Similarly, it would be valuable to design a longitudinal study that assesses the effectiveness and sustainability of interventions at the household or community level aimed at reducing under-five mortality over time. Such studies may evaluate different interventions, including immunization programs, nutrition initiatives, sanitation and hygiene improvements, disease prevention and treatment efforts, strengthening health systems, and the effectiveness of awareness programs focused on maternal and child health.