Predictors of singleton preterm birth using multinomial regression models accounting for missing data: A birth registry-based cohort study in northern Tanzania

Background Preterm birth is a significant contributor of under-five and newborn deaths globally. Recent estimates indicated that, Tanzania ranks the tenth country with the highest preterm birth rates in the world, and shares 2.2% of the global proportion of all preterm births. Previous studies applied binary regression models to determine predictors of preterm birth by collapsing gestational age at birth to <37 weeks. For targeted interventions, this study aimed to determine predictors of preterm birth using multinomial regression models accounting for missing data. Methods We carried out a secondary analysis of cohort data from the KCMC zonal referral hospital Medical Birth Registry for 44,117 women who gave birth to singletons between 2000-2015. KCMC is located in the Moshi Municipality, Kilimanjaro region, northern Tanzania. Data analysis was performed using Stata version 15.1. Assuming a nonmonotone pattern of missingness, data were imputed using a fully conditional specification (FCS) technique under the missing at random (MAR) assumption. Multinomial regression models with robust standard errors were used to determine predictors of moderately to late ([32,37) weeks of gestation) and very/extreme (<32 weeks of gestation) preterm birth. Results The overall proportion of preterm births among singleton births was 11.7%. The trends of preterm birth were significantly rising between the years 2000-2015 by 22.2% (95%CI 12.2%, 32.1%, p<0.001) for moderately to late preterm and 4.6% (95%CI 2.2%, 7.0%, p = 0.001) for very/extremely preterm birth category. After imputation of missing values, higher odds of moderately to late preterm delivery were among adolescent mothers (OR = 1.23, 95%CI 1.09, 1.39), with primary education level (OR = 1.28, 95%CI 1.18, 1.39), referred for delivery (OR = 1.19, 95%CI 1.09, 1.29), with pre-eclampsia/eclampsia (OR = 1.77, 95%CI 1.54, 2.02), inadequate (<4) antenatal care (ANC) visits (OR = 2.55, 95%CI 2.37, 2.74), PROM (OR = 1.80, 95%CI 1.50, 2.17), abruption placenta (OR = 2.05, 95%CI 1.32, 3.18), placenta previa (OR = 4.35, 95%CI 2.58, 7.33), delivery through CS (OR = 1.16, 95%CI 1.08, 1.25), delivered LBW baby (OR = 8.08, 95%CI 7.46, 8.76), experienced perinatal death (OR = 2.09, 95%CI 1.83, 2.40), and delivered male children (OR = 1.11, 95%CI 1.04, 1.20). Maternal age, education level, abruption placenta, and CS delivery showed no statistically significant association with very/extremely preterm birth. The effect of (<4) ANC visits, placenta previa, LBW, and perinatal death were more pronounced on the very/extremely preterm compared to the moderately to late preterm birth. Notably, extremely higher odds of very/extreme preterm birth were among the LBW babies (OR = 38.34, 95%CI 31.87, 46.11). Conclusions The trends of preterm birth have increased over time in northern Tanzania. Policy decisions should intensify efforts to improve maternal and child care throughout the course of pregnancy and childbirth towards preterm birth prevention. For a positive pregnancy outcome, interventions to increase uptake and quality of ANC services should also be strengthened in Tanzania at all levels of care, where several interventions can easily be delivered to pregnant women, especially those at high-risk of experiencing adverse pregnancy outcomes. We utilized secondary birth registry data from a prospective cohort of women who delivered singletons in the Kilimanjaro Christian Medical Center (KCMC) between the years 2000-2015. A detailed description of the KCMC Medical birth registry is also available elsewhere [38–43]. Briefly, KCMC is one of the four zonal referral hospitals in the country and is located in the Moshi municipality, Kilimanjaro region, northern Tanzania. The centre primarily receives deliveries of women from the nearby communities, but also referral cases from within and outside the region. On average, the hospital has approximately 4000 deliveries per year [41, 42, 44]. The study population in this study was singleton deliveries for women of reproductive age (15-49 years) recorded in the KCMC birth registry between 2000-2015, a total of 55,003 deliveries from 43,084 mothers. We excluded 3,316 multiple deliveries, 49 records missing hospital numbers (i.e. unique identification number used to link mothers and their subsequent births), 791 observations with a mismatch between dates of births of children from the same mother or were of unknown sequence (i.e. whether a singleton or multiple births), and 6,730 deliveries with gestational age <20 weeks and >42 weeks. Data was, therefore, analyzed for 44,117 deliveries born from 35,871 mothers (Fig 1). Data from the KCMC Medical Birth registry, 2000-2015. As we have also described the data collection methods elsewhere [43], birth data at KCMC have been recorded using a standardized questionnaire and is collected by specially trained project midwives. The KCMC Medical birth registry collects prospective data for all mothers and their subsequent deliveries in the hospital’s department of obstetrics and gynecology. Following informed consent, mothers were interviewed within the first 24 hours after birth given a normal delivery or as soon as a mother has recovered from a complicated delivery. The questionnaire used for data collection is available elsewhere [45]. Although the printed questionnaires were in the English language, the Project Midwives performing the interviews were well versed in English, Swahili, and one other tribal language. Furthermore, additional information during data collection were extracted from patient files and antenatal cards for more clarification of prenatal information. Data are then transferred, entered and stored in a computerized data base system at the birth registry located at the reproductive health unit of the hospital. A unique identification number was assigned to each woman at first admission and used to trace her medical records at later admissions. Access to data analyzed in this study followed ethical approval granted on June 26, 2019. The response variable was preterm birth, defined as any birth before 37 completed weeks of gestation and further categorized based on gestational age as <28 weeks (extremely preterm), [28, 32) weeks (very preterm), [32, 37) weeks (moderate to late preterm), and ≥37 weeks (term) for a full-term pregnancy [2]. Gestational age was estimated from the date of last menstrual period of the mother and recorded in completed weeks [4]. Independent variables included maternal background characteristics, particularly age categories (15-19, 20-24, 25-34, 35-39 and 40+) in years, area of residence (rural vs urban), education level (none, primary, secondary and higher), marital status (single, married and widow/ divorced), occupation (unemployed, employed and others), parity (primipara vs multipara (para 2-6)), referral status (referred for delivery or not), number of antenatal care visits (<4 and ≥4 visits), and body mass index (underweight [<18.5 Kg/m2], normal weight [18.5–24.9 Kg/m2], overweight [25–29.9 Kg/m2], and obese [≥30 Kg/m2]). Maternal health before and during pregnancy included, alcohol consumption during pregnancy, maternal anemia, malaria, systemic infections/sepsis and pre-eclampsia/eclampsia (all categorized as binary, yes/no). Maternal HIV status was categorized as positive or negative. Complications during pregnancy and delivery included premature rapture of the membranes (PROM), postpartum hemorrhage (PPH), placenta previa and placenta abruption also categorized as binary, yes/no, with “yes” indicating the occurrence of these outcomes. Newborn characteristics included sex (male vs female), perinatal status (dead if experienced stillbirth/early neonatal death vs alive) [43], and low birth weight (LBW) defined as an absolute infant birth weight of <2500g regardless of gestational age at birth [46, 47]. Data were analyzed using STATA version 15.1 (StataCorp LLC, College Station, Texas, USA). The primary unity of analysis was singleton deliveries for women recorded in the KCMC Medical Birth Registry between the years 2000 and 2015. We summarized numeric variables using means and standard deviations, and categorical variables using frequencies and percentages. The Chi-square test was used to compare the proportion of preterm birth by participants characteristics. We used multinomial logistic regression models to determine the predictors of preterm birth as opposed to previous studies [4, 6, 7, 12, 18, 28–30, 32, 33, 48] that performed a binary regression analysis. The multinomial/polytomous regression model is an extension of the logistic model for binary responses to accommodate multinomial responses which does not have any restrictions on the ordinality of the response [27]. Let Yi denote a nominal response variable for the ith subject, and Yi = c (the response variable occuring in category c), while Pr(Yi) defines the probability that Yi = c. The multinomial logit model can be written as A nominal model to allow for any possible set of c − 1 response categories is written as where the multinomial logit ηic=Xic′βc. In this model, all of the effects βc vary across categories (c = 1, 2, …, C) and makes comparisons to a reference category compared to the ordinal regression model that uses cumulative comparisons of the categories [49]. We used robust standard errors adjusted for clusters to account for nested observations/ deliveries within mothers. We would like to indicate here that we performed preliminary analysis using the binary and ordinal logistic regression models. There were a couple of variables that did not satisfy the proportional odds (PO) assumption, hence the ordinal logistic regression model could not be used. The close alternative model that relaxes the PO assumption are the generalized ordered logistic regression models. However, we encountered a non-convergence problem, especially with four preterm birth categories and appropriate interpretation of results. For instance, the order of gestational age categories is <28 weeks (extremely preterm), [28, 32) weeks (very preterm), [32, 37) weeks (moderate to late preterm), and 37+ weeks (term/normal). Assuming the variable is coded as 0 to 3 (with 0 being term birth), the first panel of coefficients will be interpreted as; 0 vs. 1+2+3, then 0+1 vs 2+3 etc [50]. This will imply modeling the probability of delivering at a normal gestational age (category 0) compared to preterm (categories 1-3), probability of delivering term and very preterm vs other preterm categories, etc. Similar interpretations will apply even if preterm birth is coded from extremely preterm (0) to term (3). Such interpretation could be somehow misleading given the nature of this outcome and may not be appealing to clinicians or public health practitioners. Nevertheless, the choice of regression models often depends on the research question one would like to address. In this study, the choice of multinomial regression model was relevant to determine preterm birth predictors across different preterm birth categories, other than performing a binary or an ordinal regression analysis. As previously indicated, data analysis in this study considered missing values in the covariates. A description of how missing data were imputed is also reported in [43]. Data were imputed using a multiple imputation technique, which is a commonly used method to deal with missing data, which accounts for the uncertainty associated with missing data [34, 37, 51]. We assumed the missing data were missing at random (MAR) where the probability of data being missing does not depend on the unobserved data, conditional on the observed data [34–37]; hence the variables in the dataset were used to predict missingness [43]. We also assumed a nonmonotone pattern of missingness in which some subject values were observed again after a missing value occurs [35, 43, 51]. Under a nonmonotone pattern of missingness, it is recommended to use chained equations, which goes with several names such as the Markov chain Monte Carlo (MCMC), and the fully conditional specification (FCS), to impute missing values [37, 51–55]. Furthermore, the FCS method allows imputation of all types of variables simultaneously, namely some continuous and other categorical. For the illustration of FCS algorithm, we let Y denote the fully observed outcome in this study i.e., preterm birth, X denote the partially observed covariates X = X1, …, Xp, and W denote the fully observed covariates W = W1, …, Wq. Let Xo and Xm denote the vectors of observed and missing values of X for n subjects. For each partially observed covariate Xj, we posit an imputation model f(Xj|X−j, W, Y, θj) with parameter θj where X−j = (X1, …, Xj−1, Xj+1, …, Xp) [56]. This according to [56] is typically a generalized linear model chosen according to the type of Xj (e.g. continuous, binary, multinomial, and ordinal). Furthermore, a noninformative prior distribution f(θj) for θj is specified. We further let xjo and xjm denote the vectors of observed and missing values in Xj for the n subjects and y and w denote the vector and matrix of fully observed values of Y and W across n subjects. Let xm(t) denote imputations of the missing values xjm at iteration t and xj(t)=(xjo,xjm(t)) denote vectors of observed and imputed values at iteration t. Let x-j(t)=(x1(t),…,xj-1(t),xj+1(t-1),…,xp(t-1)). The tth iteration of the algorithm consists of drawing from the following distributions (up to constants of proportionality) [56]; The FCS starts by calculating the posterior distribution p(θ|xo) of θ given the observed data. This is followed by drawing a value of θ* from p(θ|x0) given (xo,x-j(t),w,y), which is the product of the prior f(θj) and the likelihood corresponding to fitting the imputation model for Xj to subjects for whom Xj is observed, using the observed and most recently imputed values of X−j [56]. Missing values in Xj are then imputed from the imputation model using the parameter value drawn in the preceding step [56]. Finally, a value x* is drawn from the conditional posterior distribution of xm given θ = θ*. The process is then repeated depending on the desired number of imputations [36, 53, 55, 56]. Within each imputation, there is an iterative estimation process until the distribution of the parameters governing the imputations have converged in the sense of becoming stable, although more cycles may be required depending on certain conditions such as the amount of missing observations in the data [55, 56]. Rubin’s rule is then used to provide the final inference for θ^ by averaging the estimates across M imputations given by [56]; while the estimate of the variance of θ^M is given by; which is a combination of within and between imputation variances. Detailed descriptions on implementation of the FCS/MICE algorithm in STATA is well-presented elsewhere [54, 57]. Maternal age and education level were imputed as ordinal variables, while maternal occupation, marital status, and BMI (because normal weight (18.5–24.9 Kg/m2) was a reference category) as multinomial variable [43]. The rest of the variables were binary, and so imputed using the binomial distribution. Preterm birth (the outcome in this study), parity, pre-eclampsia/eclampsia, anemia, malaria, systemic infections/sepsis, PROM, PPH, abruption placenta, placenta previa, and year of birth did not contain any missing values, hence used as auxiliary variables in the imputation model. The imputation model generated 20 imputed datasets after 500 iterations (imputation cycles). A random seed of 5000 was specified for replication of imputation results each time a multiple imputation analysis is performed [51]. We developed a multivariable analysis model by including all covariates in the multinomial logit analysis model [54]), with standard errors adjusted for clusters (i.e., deliveries nested within mothers). We then performed stepwise regression, in which variables with p < 0.1 or p < 10% were retained in the model. The next steps entailed performing a series of adjusted analysis to test the effect of retaining and dropping variables in the multivariable model. Variables in the final model were evaluated at p-value<0.05 level of statistical significance. We used AIC to compare model performance and non-nested models [58], and Likelihood ratio test to compare nested models. After the imputation of missing values, we estimated parameter estimates adjusting for the variability between imputations [54, 57]. Before the analysis of imputed data, we firstly performed complete case analysis using multivariable multinomial regression model. The final model from this analysis was then compared to those from the multiply imputed dataset. We followed the recommendations suggested by Sterne et al., [34] for reporting and analysis of missing data. As described in [43], this study was approved by the Kilimanjaro Christian Medical University College Research Ethics and Review Committee (KCMU-CRERC) with approval number 2424. For practical reasons, since the interview was administered just after the woman had given birth, consent was given orally. The midwife-nurse gave every woman oral information about the birth registry, the data needed to be collected from them, and the use of the data for research purposes. Women were also informed about the intention to gather new knowledge, which will, in turn, benefit mothers and children in the future. Participation was voluntary and had no implications on the care women would receive. Following consent, mothers were free to refuse to reply to single questions. For privacy and confidentiality, unique identification numbers were used to both identity and then link mothers with child records. There was no any person-identifiable information in any electronic database, and instead, unique identification numbers were used. Necessary measures were taken by midwives to ensure privacy during the interview process.