Methods
Study Design and Setting
This was a retrospective cohort study investigating the association between interpregnancy interval and the incidence of preterm birth, small for gestational age birth, and low birth weight among the second and third births of mothers in Perth, Western Australia, in the period 1980-2010.
Data Source
We sourced birth data from the Midwives’ Notification System, a population-wide database of all births in Western Australia (of at least 20 weeks’ gestation, more than 400 g birth weight, or both), based on statutory collection since 1980. Statistical linkage of births to the same mother was provided by Data Linkage Western Australia, located in the Western Australian Department of Health. We selected all mothers who had their first three births as liveborn singletons within the period 1 January 1980 to 31 December 2010 while resident in Perth at the time of each birth (Figure). We excluded births of 45 weeks’ gestation or longer, as national reference data were unavailable for calculating small for gestational age. We also excluded births to mothers younger than 14 years.
(Enlarge Image)
Figure 1.
Selection of records used in study
Of the 84 151 mothers whose third births were live singletons while mothers were resident in Perth in the study period, 40 443 had all their first three births as live singletons in the same period. Reasons for this reduction include mothers having at least one birth outside the study period or outside Western Australia (n=38 652), mothers being resident outside Perth but elsewhere in Western Australia for at least one birth (n=3900), mothers not having liveborn singletons for one or more pregnancies (n=631), and records missing data (for example, gestational age or birth weight) for one or more pregnancies (n=525). Interpregnancy interval was erroneously negative for two mothers, reducing the sample to 40 441. We were unable to determine whether these two records had erroneous birth dates or parities; we assume a low level of such errors in the database and a negligible effect on the precision and bias of results.
Variables
Outcome variables were preterm birth (gestational age less than 37 completed weeks), small for gestational age birth (less than the 10th centile of Australian national birth weight centiles by sex and gestational age in weeks), and low birth weight (less than 2500 g). For consistency with previous studies, we modelled interpregnancy interval as a categorical variable, classed as: 0-5 months, 6-11 months, 12-17 months, 18-23 months (as the reference category), 24-59 months, 60-119 months, and 120 months or longer. We calculated interpregnancy interval as the time between one birth and the estimated start of the pregnancy of the subsequent birth (birth date minus estimated gestational age). We also adjusted for possible confounders (see below). The proportion of records for which pregnancy dating, and hence gestational age, was estimated by ultrasonography (versus last menstrual period) increased to more than 70% throughout the study period. We expect the primary effects of this variation in method to be that some records have more precise estimates of prematurity and small for gestational age than others; we assume a very small effect on assignment of interpregnancy intervals given the scale of interpregnancy interval categories relative to uncertainties around pregnancy dating.
Statistical Modelling
We used a maternally matched design to model the odds of preterm birth, small for gestational age birth, and low birth weight as a function of interpregnancy interval. Whereas previous studies have observed an association between interpregnancy interval and adverse birth outcomes as a comparison among women, here we used conditional logistic regression to measure this association within individual women. Conditional logistic regression is commonly used for matched case-control studies and longitudinal studies. In the context of our study, in which we matched birth outcomes by mother, this conditional approach accounts for each woman’s overall risk of adverse birth outcomes among all of her children included in the analysis. This design thereby removes the effects of measured or unmeasured maternal factors that are either fixed (such as genetic predisposition) or strongly correlated over time (such as long term health). Essentially, this enables inferences that are based purely on within mother effects. In comparison, the traditional approach of unconditional logistic regression is based on differences between women. In the absence of confounding by persistent maternal factors, the two models will report the same effects of interpregnancy interval. However, where unmeasured persistent confounders exist, the unconditional model will give biased estimates of the effects of interpregnancy interval.
The conditional model required data on three births per mother, with the first and second births defining the start of the interpregnancy intervals of the second and third births. We explicitly controlled for factors that vary between births as possible within mother confounders of an effect of interpregnancy interval: maternal age (categorical variable: 14-19, 20-24, 25-29, 30-34, 35-39, and ≥40 years), parity, and birth year. We also controlled for socioeconomic status as a factor that potentially varies (using the area level Index of Relative Socio-Economic Disadvantage from the 1996 Australian Bureau of Statistics’ Census of Population and Housing, categorised as national fifths). The conditional model was fitted in R 2.15.1, using the clogit() function in R’s Survival package.
For comparison with the unmatched design of previous studies, we also generated results by using an unconditional logistic regression model for each type of birth outcome, with interpregnancy interval as the predictor variable of interest. We also adjusted for possible confounders of an effect of interpregnancy interval: maternal age (in five year categories as for the conditional model), parity, birth year, socioeconomic status (in nationally defined fifths as for the conditional model), ethnicity (white versus non-white), and the outcome of the previous birth. Confidence intervals for the unconditional model were based on robust standard error estimates that take within mother clustering into account. The unconditional model was fitted in R 2.15.1, using lrm() and robcov() functions in R’s rms package.
Throughout this paper, we use the term “matched” to refer to the within mother design based on conditional logistic regression and “unmatched” to refer to the between mother design based on unconditional logistic regression.