Predictors of preterm births in North Dakota: a retrospective study of the North Dakota Pregnancy Risk Assessment Monitoring System (PRAMS)

Ethical statement

This study used data from the Pregnancy Risk Assessment Monitoring Survey (PRAMS), which is a population-based survey developed by the Centers for Disease Control and Prevention (CDC) and conducted in partnership with the North Dakota Department of Health and Human Services (ND HHS) (Centers for Disease Control and Prevention, 2023c). The CDC’s protocol has been approved by the CDC’s IRB (Shulman et al., 2018). The protocol of this study was reviewed and approved by the University of Tennessee Institutional Review Board (IRB Number: 21-06599-XM). The need for consent was waived by the IRB because the study utilizes secondary data. Data were provided by the ND HHS through a special request. All identifying pieces of information were removed from the dataset before it was provided to investigators. Privacy of data were maintained by ensuring only study personnel who signed data sharing agreement with ND HHS had access to the data, storing the data in password protected computers only accessible to study investigors and ensuring that all study results are reported in aggregated format only.

Study area, data source, and study population

Portions of this text were previously published as part of a PhD dissertation (Njau, 2022). North Dakota has a population of approximately 773,000 people (United States Census Bureau, 2022). The racial/ethnic distribution of the state is 83% White, 5% American Indian, 3% Black, 2% Asian, 2% are two or more races and 5% are Hispanic or Latino. Women of reproductive age (i.e., 15–44 years) comprise approximately 18% of the population (United States Census Bureau, 2022).

Study data were obtained from the ND PRAMS Program. ND PRAMS utilizes standard CDC PRAMS data collection methodology, which has been described in detail elsewhere (Shulman et al., 2018). A stratified random sample of women identified in the state’s birth certificate dataset is invited to participate in the survey (Shulman et al., 2018). In ND, approximately 14% of women who had live births were sampled between 2017 and 2021. The ND PRAMS response rates for this period were 70.2% (2017), 59.9% (2018), 59.1% (2019), 60.5% (2020), and 59.2 (2021). These response rates exceeded the CDC’s threshold of at least a 50% response rate required for a weighted statewide representative sample (Shulman et al., 2018; Centers for Disease Control and Prevention, 2023c). The CDC accounts for three components in the complex weighting design (sampling weight) calculated for each state and year of data, allowing for generalizability of results to the entire population of the jurisdiction, an adjustment for nonresponse, and an adjustment for noncoverage (Shulman et al., 2018; Centers for Disease Control and Prevention, 2023c).

The PRAMS questionnaire covers an array of topics on maternal behaviors and experiences. It includes questions on smoking, substance abuse, insurance status, prenatal care, other healthcare provider access, and healthcare outcomes, such as preterm births, maternal diabetes, hypertension, and behavioral health. States participating in PRAMS may add questions to the core survey that reflect emerging or state-relevant needs. Due to interest among stakeholders in the state, ND added the adverse childhood experiences (ACEs) module of questions in 2017 (Anda et al., 1999; North Dakota Department of Health and Human Services, 2024).

Mothers with prior preterm births (4.3% of overall dataset) or who had plural births (1.4% of overall dataset) were excluded because preterm birth incidence is known to be excessively high under those conditions (Ekwo, Gosselink & Moawad, 1992; Murray et al., 2018). These criteria resulted in the exclusion of 5.7% of live births from the original dataset.

Descriptive analyses

Statistical analyses were performed in SAS software version 9.4 (SAS Institute Inc., 2013) using procedures for complex survey designs with Taylor Series variance estimation. Sampling strata and sampling weight were specified in all analyses. For descriptive analyses, the ‘proc surveyfreq’ procedure was used to calculate weighted frequencies for each variable and the prevalence of preterm births corresponding to each variable (SAS Institute Inc., 2016a).

Investigation of predictors of preterm births

Preterm birth was defined as a live birth which occurred before 37 weeks of gestation. In accordance with a standard definition of preterm birth that allows for comparison across states and over time, a dichotomous preterm birth variable was created by categorizing gestational period into births prior to 37 weeks (preterm) and births 37 weeks or later (term) (Quinn et al., 2016). While preterm births may be further categorized by severity according to the number of weeks of gestation, severity of prematurity was not the focus of the present study. More recent recommendations are to classify preterm birth using a more robust, multi-faceted taxonomy that considers maternal placental factors as well as fetal conditions and signs indicating the beginning of labor, and which are then postnatally confirmed or refined (Villar et al., 2024). This classification goes beyond the scope of this study, as variables needed for the taxonomy were not available in the dataset provided by ND HHS and therefore postnatal confirmation was not possible.

Potential predictors were identified using a conceptual causal model based on existing knowledge (Fig. 1). The predictors consisted of demographic characteristics (i.e., maternal race, maternal age, marital status, residence), behavioral factors (i.e., drinking, smoking cigarettes, and using e-cigarettes three months prior to pregnancy, using marijuana one month prior to pregnancy), maternal characteristics (i.e., adverse childhood experiences (ACE) score, intendedness of the pregnancy, experience of physical abuse during pregnancy), and healthcare access (i.e., whether prenatal care visits were paid by Medicaid, adequacy of prenatal care). The predictors also included pregestational health conditions (i.e., pregestational diabetes, pregestational hypertension, and pregestational body mass index (BMI)) and gestational health conditions (i.e., gestational diabetes, gestational hypertension, depression during pregnancy). While fertility treatments are considered a risk factor for preterm birth (Salmeri et al., 2024; Sanders et al., 2022), the unweighted count of births to women in our original dataset (i.e., prior to removing multiple births and births to women with a prior preterm birth) and in our subset of data from 2017 to 2021 who identified as using assisted reproductive technology or using fertility enhancing drugs were each fewer than 10. Due to the small numbers, fertility treatment was not assessed as a potential predictor in the analyses.

Since the outcome (preterm births) is dichotomous, logistic regression was used to assess associations between preterm births and several potential predictors listed above and to compute adjusted odds ratios (Sperandei, 2014). Binary logistic regression models were fitted using the ‘proc surveyreg’ procedure of SAS (SAS Institute Inc., 2016b). Model building followed a two-step process (Fig. 2). First, bivariate associations of each potential predictor and preterm birth were assessed by fitting univariable logistic regression models of preterm births for each predictor. Second, predictors were retained if the p-value of a likelihood ratio test comparing the univariable model to the null model was <0.2 (Dohoo, Martin & Stryhn, 2012) and then assessed in a full binary multivariable logistic regression model. A final reduced, parsimonious multivariable model was produced using a manual backward elimination process in which covariates with the highest p-values were removed, one at a time, until all remaining variables were significantly associated with the outcome at a significance threshold p < 0.05. At each elimination step, the coefficients of remaining predictors before and after the removal of each covariate were compared to identify confounders. Non-significant variables were retained as confounders if their removal resulted in >20% change in the coefficient of another variable in the model (Dohoo, Martin & Stryhn, 2012). Age and race were retained in the multivariable model regardless of their statistical significance because they were expected to be important confounders in the conceptual model (Fig. 1). ACE score was not retained as a confounder because it did not have a significant association with preterm birth. The models were refitted using the R package ‘survey’ version 4.2.1 to calculate Akaike Information Criterion (AIC) with consideration of complex survey design (Lumley & Scott, 2015; Lumley, 2004).

Descriptive analysis

The weighted survey responses represented 47,195 singleton births among ND women who never had a previous preterm birth from 2017 to 2021. Stratified population frequencies and preterm birth prevalence proportions are presented in Table 1. The study’s target population reflected greater racial diversity than the state’s population overall; 75% of the study population was White, 8% was American Indian, and 17% was other races (Table 1). The average maternal age was 29 years; 13% of the study population was aged 35 years or older, 83% was aged 20–34 years, and 4% was younger than 20 years. Most of the population was married (68%). The population was nearly evenly distributed between counties classified as urban (51%) and rural (49%). The average maternal ACE score was 1.8; 41% of the population had an ACE score of 0, 20% had a score of 1, 11% had a score of 2, 7% had a score of 3, and 21% had an ACE score of 4 or more. Two in five live births (39%) resulted from unintentional pregnancies, 3% of women experienced physical abuse during pregnancy, and 17% experienced depression during pregnancy. A minority of the population smoked three months prior to pregnancy (19%), used electronic cigarettes three months prior to pregnancy (5%), or used marijuana in the month prior to pregnancy (7%), but the majority did consume alcohol 3 months prior to pregnancy (69%). Approximately one in four births (24%) was to a woman who had prenatal care paid for by Medicaid and 1 in five births (22%) was to a woman who received less than adequate prenatal care. More than half of the population was either overweight (28%) or obese (28%) prior to pregnancy, 4% had diabetes prior to pregnancy, and 5% had high blood pressure prior to pregnancy. Approximately 1 in 10 ND women with a live singleton birth and no prior history of preterm birth developed diabetes during pregnancy (9%) and more than 1 in 10 developed hypertension during pregnancy (12%).

The overall prevalence of preterm births was 6.4% (95% confidence interval (CI) [5.4–7.4]), but there were clear differences in preterm birth prevalence when stratified by some potential predictors (Table 1). Notably, prevalence was substantially higher among women who were American Indian (10.6%), those aged 35 years or older (9.0%), those who smoked prior to pregnancy (9.3%), those with pregestational diabetes (18.1%) or pregestational hypertension (13.0%), and those who developed hypertension during pregnancy (17.5%), compared to the overall study population.

Predictor investigation of preterm births

Variables that had unadjusted associations with preterm births included maternal race, maternal age, intentionality of pregnancy, smoking three months before pregnancy, pregestational BMI, pregestational diabetes, gestational diabetes, pregestational hypertension, gestational hypertension, maternal marital status, prenatal care paid by Medicaid, and adequacy of prenatal care (Table 2). The results of the final multivariable model show that the significant predictors of preterm births in this study population were maternal race, maternal age, pregestational diabetes, and gestational hypertension (Table 3). Pregestational hypertension (confounder of pregestational diabetes) and pregestational BMI (confounder of race and age) were retained as confounders in the final model although they were not significantly associated with the outcome. The odds of preterm births were higher among women who were American Indian (adjusted odds ratio (AOR) = 1.7, 95% CI [1.3–2.3]) compared to those who were White and among women who were aged 35 years or older (AOR = 1.6, 95% CI [1.0–2.5]) compared to those aged 20–34. Additionally, the odds of preterm births were higher among women who had pregestational diabetes (AOR = 4.3, 95% CI [2.0–9.3]) and who developed gestational hypertension (AOR = 4.5, 95% CI [3.1–6.7]) than those who did not have pregestational diabetes or develop gestational hypertension, respectively. The AIC of the final multivariable model was 0.66 (95% CI [0.64–0.68]).

Strengths and limitations

This is one of the first studies that has investigated predictors of preterm births in ND. It is also one of a few studies within a growing body of work exploring the association between childhood adversity and poor maternal outcomes. With response rates ranging from 59.1% to 70.2% for the five years of data included in this study, the study’s response rates exceeded the threshold established by the CDC and hence the findings are generalizable to women in ND. The sampling strategy for ND PRAMS also ensured a representative sample of American Indian women. Furthermore, the sampling strategy ensured ample sample size for investigation of urban/rural status as a predictor of preterm births. In limiting our study population specifically to women who did not have a previous preterm birth and who had a singleton live birth, our results provide greater focus regarding the common risk factors for preterm births among women without the strongest risk factors of plural births and prior preterm birth. Another key strength of the study was the utilization of clinically reported data for the outcome of interest (gestational age at delivery) which minimizes recall bias.

A limitation of the study was the use of self-reported data. Some PRAMS elements, including the ACEs module, are self-reported during the postpartum period, hence some self-reporting and/or recall bias is to be expected. One study comparing ACEs information collected prospectively (i.e., during childhood) and ACEs reported retrospectively (i.e., by the same individual as an adult) found that the association between the two calculated scores had modest agreement (Reuben et al., 2016). The prospective ACE score was more strongly associated with poorer objective health outcomes (i.e., collected through tests) among adults and the retrospective ACE score was related to poorer subjective outcomes (i.e., self-reported). The authors emphasize that the two scores should be considered to have assessed complementary sources of information, rather than concluding that the retrospective score is inaccurate. In a study of 2003 PRAMS data among 12 states, validity and reliability for three measures (i.e., participation in the Women, Infants, and Children (WIC) program; Medicaid payment for delivery; and breastfeeding initiation) were high and provide evidence for confidence in self-reported PRAMS information (Ahluwalia, Helms & Morrow, 2013).

Additionally, the birth certificate information and the PRAMS dataset do not encompass all known predictors. For example, clinical-level questions on known mechanical risk factors, such as amniocentesis procedures and placental complications, are not included in the PRAMS survey or birth certificate. Thus, this analysis was limited to the available population-level variables. The number of women that used fertility treatments in this study was not large enough for appropriate assessment of this variable for potential association with preterm births. Additionally, information needed for a more robust classification of preterm birth (Villar et al., 2024) was not available in the dataset used for this study, and the dichotomous variable based solely on gestational age at birth can limit the ability to draw clinical insights. However, the dichotomous definition used in this study is consistent with the definition used in extant literature and allows for comparison across jurisdictions and over time. The above limitations notwithstanding, the findings of this study are important in guiding future studies and public health efforts to address preterm births.