Differences in risk factors for incident and recurrent preterm birth: a population-based linkage of 3.5 million births from the CIDACS birth cohort

This population-based cohort study used data from the CIDACS Birth Cohort. This cohort was created by linking data from Brazil’s National Live Birth System (SINASC) and the 100 Million Brazilian Cohort baseline from Jan 1, 2001, to Dec 31, 2015. This study adhered to the RECORD (Reporting of studies Conducted using Observational Routinely-collected Data) statement.

The SINASC gathers information on birth notifications across the country, including information about the mother, pregnancy, newborn, and gestational age at birth, which allows the estimation of the prematurity rate for the country. The live birth declaration forms that provide information to SINASC adopt the last menstrual period (LMP) as the standard method for estimating the gestational age in weeks. Results of physical examinations and other methods are alternatively accepted [12]. The 100 Million Brazilian Cohort is primarily built from Cadastro Único (CadUnico) which covers the poorest half of the Brazilian population (families with monthly income equal to or below three minimum wages ~750 USD) [13]. The CIDACS Birth Cohort is composed of 24,695,617 live births. Children included in the cohort were generally born from younger, unmarried, and less educated mothers and are more likely to be born via vaginal delivery than children in the general Brazilian population [14]. In this study, we identified successive pregnancies using the unique maternal identifier and the newborn’s date of birth.

The linkage process used CIDACS-RL (Record Linkage), a novel record linkage tool developed to link large-scale administrative datasets at CIDACS [15]. The linkage was based on the similarity index using common attributes (mother’s name, maternal age at birth, maternal date of birth, and the municipality of residence of the mother at the time of delivery) between the databases. The linkage process is described in detail elsewhere [16].

In this study, we included live births of multiparous women, aged 14–49 years, who entered the CIDACS Birth Cohort as nulliparous women. We excluded (a) all multiple births and live births with congenital anomalies as these conditions are known to be strongly associated with premature birth [2, 6], (b) live births weighing < 500 g and gestational age < 22 weeks [17‐19], (c) those with a birth date prior to the mother’s entry date into the cohort, and (d) live births with missing information on gestational age and for at least one live birth at the first or second pregnancy (Fig. 1).

The study outcome was the occurrence of PTB in the second pregnancy. Incident PTB was defined as a live birth with a gestational age less than 37 weeks preceded by a term delivery. Recurrent PTB was defined as a live birth with a gestational age less than 37 weeks preceded by a previous PTB.

The following covariates were considered in the analyses: (i) sociodemographic characteristics of the mother: residence area (urban; rural), household overcrowding (no: ≤2 inhabitants per room; yes: >2 inhabitants per room), self-declared race/ethnicity (white, black/mixed-race, or indigenous), level of education (<8 years: 8–11 years; ≥12 years of schooling), marital status (married: married or in a stable relationship; unmarried: single, divorced or widowed); (ii) prenatal assistance: the number of visits (none; 1 to 3 visits; ≥ 4 visits); (iii) maternal, delivery, and newborn-related variables: maternal age (14–19; 20–34; 35–49 years), type of delivery (vaginal or cesarean), and sex of the newborn (female or male), and interbirth interval (<12 months, 12 to 24; ≥24 months). Household overcrowding was calculated by dividing the number of individuals living in the house and the number of rooms. The interbirth interval was estimated (in months) by the difference between the second child's birth date and the previous child’s birth date.

Socioeconomic, maternal, and birth characteristics were summarized using frequency distributions. We used longitudinal transition models with multivariate logistic regression to investigate the risk factors for PTB in a second pregnancy, which is different in women whose first pregnancy was delivered at term (incidence) or preterm (recurrence). Thus, a multiplicative interaction between each incidence and recurrence of PTB was included in the model to assess the risk factors’ chances for incidence versus recurrence in the second pregnancy. Transition models, also called dynamic models or Markov models, are relevant when any response in a sequence of repeated measures is modeled conditionally under one or more sets of past or previous measurements [20]. The specific statistical model for each objective is presented in a footnote for Tables 2 and 3. The transition model used here is a concise form proposed by Diggle et al. [21] to express recurrence and incidence in a single expression as predictors of the previous PTB in the first pregnancy as well as the interactions. The interaction test only identifies risk factors that are different between incidence and recurrence (α=0.05). However, the interaction test does not explicitly test whether the risk factor is significant for the incident and recurrent PTB, and thus the odds ratio (OR) and 95% confidence intervals (95% CI) for each risk factor were reported separately according to the first pregnancy PTB status. All models were adjusted for the complete set of risk factors considered to be relevant and plausible in the literature on preterm birth [2, 6, 11, 22‐25]. Individuals with missing observations in any of the variables were excluded from the multiple models.

As preventive health behaviors can change between the first and second pregnancies, we assessed the chances of the incident and recurrent PTB according to changes in the number of prenatal visits. We created a four-category variable considering the performance of ≥ 4 prenatal visits in both pregnancies: yes-yes (4 or more appointments in the first and second pregnancy, reference category), yes-no (4 or more appointments in the first and fewer than 4 appointments in the second pregnancy), no-yes (fewer than 4 appointments in the first and 4 or more appointments in the second pregnancy), or no-no (fewer than 4 appointments in the first and second pregnancy). The analyses were adjusted by the same variables mentioned above, except for the number of prenatal visits analyzed as the primary exposure variable.

All analyses were performed using Stata version 15.1 (Stata Corporation, College Station, USA).

Due to changes in gestational age information in the SINASC registry [26], we conducted complementary analyses with live births born between 2012 and 2015 to verify the consistency of the results of our initial analyses (Additional file 1).

We applied a cross-validation technique using multivariate logistic regression to assess the reliability of our model using. We used a 70% training sample and validated the model with a 30% sample. Receiver operating characteristic (ROC) curves were generated and sensitivity and specificity were calculated for the optimal cut-off based on the ROC curve. The results were compared with those obtained in the general model (Additional file 1).

This is the first study to assess the factors associated with the incidence and recurrence of PTB in a poor population in a middle-income country with great social and health inequalities. The linkage with the national live birth information system allowed us to track women individually through successive pregnancies. Also, the large number of cases allowed us to simultaneously investigate the factors associated with the incidence and recurrence of PTB, and to carry out our analysis considering the changes in the number of prenatal consultations between pregnancies.

However, this study has some limitations. The use of secondary data, which was not designed primarily for research purposes, may be subject to some limitations related to missing, underestimation, and potential misclassification. The proportion of preterm births recorded in SINASC may be subject to underreporting related to errors in the gestational age [41]. Until 2010, the gestational age at birth was collected over wide intervals of gestational weeks. As of 2011, SINASC started to collect gestational age as a continuous variable; however, the estimation method changed to be mainly based on the LMP [26]. The LMP is the method recommended by WHO, due to its wide accessibility and low cost [42]. Nonetheless, this method can be sometimes not much accurate due to circumstances such as individual variations in the length of the menstrual cycle, especially memory biases [43]. However, these are probably non-differential errors and are unlikely to introduce bias in the measure of association, although the absolute measures of risk may be underestimated. In addition, the proportion of missing data in the cohort could be a limitation for the generalization of our findings. Residual confounding is also possible because important variables for determining PTB, such as maternal comorbidities (e.g., obesity, diabetes, and hypertension), risk behaviors (e.g., smoking, alcohol, or drug use during pregnancy), and access and quality of health services were not available. Also, we were not able to classify the preterm birth subtypes (spontaneous or with medical indication) because of the lack of information in our dataset. The lack of genetic information is another limiting factor; it is known that genetic variants are associated with the duration of pregnancy and the risk of PTB [44]. These limitations may be affecting the differences observed between incident vs. recurrent PTB in this study. In addition, the association between changes in the number of prenatal care visits and PTB according to previous PTB history may present confounding bias. Our findings must be interpreted with caution because the absence of an association between a risk factor and PTB recurrence should not be a reason to dismiss the factor as a potential focus for preventive action. Finally, this study was conducted among the poorest population of a middle-income country with a history of great social and health inequalities which might limit the generalizability of these findings.