Methods
Source Population
The study was based on the Nurses' Health Study II, an ongoing prospective cohort that was established in 1989 by enrolling 116,671 female nurses aged 24-44 at baseline. Questionnaires are administered biennially to update lifestyle characteristics and health related outcomes. The 1989 baseline questionnaire collected information on medical, diagnostic, and prescription history, date of birth, occupational status, family history of diabetes, gravidity, height and weight, and a range of other characteristics. Time varying characteristics are updated every two years.
After 2001, most participants had passed the usual reproductive age (all were aged ≥35, 70% were aged ≥40, and 20% aged ≥45) and information on the diagnosis of gestational diabetes was not collected in the questionnaire. We included in the analysis individual singleton live births up to the end of 2001 if the participant did not report gestational diabetes in a previous pregnancy, a diagnosis of type 2 diabetes or cancer, or a cardiovascular disease event (myocardial infarction, stroke, coronary artery bypass graft procedure, or angina) before an otherwise eligible pregnancy (see Appendix 1). Pregnancies after gestational diabetes were not included because women with gestational diabetes in a previous pregnancy were likely to change their diet and lifestyle during the next pregnancy to prevent a recurrence. We conducted sensitivity analyses with all pregnancies regardless of previous gestational diabetes status. Follow-up for each questionnaire cycle is >90% through 2001.
(Enlarge Image)
Appendix 1.
Flow-chart of study participants [posted as supplied by authors].
^Exclusion criteria are not mutually exclusive and individual reasons may not total the number of excluded participants.
Ascertainment of Lifestyle Factors
Twice a year we obtained information on participants' smoking status and weight. Self reported weight was highly correlated with weight measured by a technician among a random subset of participants in the Boston area cohort (r=0.97).
Beginning in 1991 and every four years thereafter, participants were asked to complete a semiquantitative food frequency questionnaire in addition to the main questionnaire. The food frequency questionnaire is designed to measure average dietary intakes over the past year and has been extensively validated. To characterize overall dietary quality, we created a summary diet score based on the Alternate Healthy Eating Index-2010, which is based on the US Department of Agriculture Healthy Eating Index, designed to measure adherence to US dietary guidelines. We computed scores for adherence to the index before pregnancy from each food frequency questionnaire cycle occurring before a given pregnancy. We included 10 of 11 the components of the index in our diet score: higher intakes of vegetables, fruit, nuts, whole grains, polyunsaturated fatty acids, and long chain omega 3 fatty acids and lower intakes of red and processed meats, sugar sweetened beverages, trans fats, and sodium. Alcohol, a component of the Alternate Healthy Eating Index-2010, was not included in this analysis because of the concern of its potential adverse impact on fetal outcomes. Participants' scores for each of the individual components ranged from 0 to 10 based on level of intake, with 10 representing optimal dietary behavior. The overall diet score ranged from 2.5 (worst) to 77.5 (best). Pregnancies were excluded if the participant did not return a prepregnancy food frequency questionnaire, left more than 70 food frequency questionnaire items blank, or reported unrealistic total daily energy intakes (<2090 kJ (<500 kcal) or 14,630 kJ (>3500 kcal)).
Physical activity was assessed on the main questionnaire in 1989, 1991, and 1997. Participants were asked to report the average amount of time they spent on each of the following moderate to vigorous recreational activities each week: bicycling, calisthenics or use of a rowing machine, lap swimming, squash or racquetball, jogging, running, tennis, or other aerobic exercise. Women were also asked to report the amount of time they spent each week walking (that is, walking or hiking outdoors, including walking to work) and to categorize their usual walking pace outdoors as easy or casual (slower than 2.0 mph), normal (2.0-2.9 mph), brisk (3.0-3.9 mph), or very brisk or striding (faster than 4.0 mph). Usual total physical activity before pregnancy was calculated as the sum of the minutes spent each week in each activity, including brisk or very brisk walking. Validity and reproducibility of the questionnaire were evaluated and reported previously. Briefly, the correlation between physical activity as reported in one week recalls and that reported on the questionnaires was r=0.79. The correlation between moderate to vigorous activity reported in diaries and that reported on the questionnaires was r=0.62.
Outcome Assessment
The outcome of interest was incident gestational diabetes. Gestational diabetes diagnosed by a physician was ascertained by self report on each biennial questionnaire through 2001. Investigators ceased the update of occurrence of gestational diabetes in 2001 as most women had passed reproductive age. In a validation study among a subgroup of the Nurses' Health Study II cohort, 94% of self reported gestational diabetes events were confirmed by medical records. Among the confirmed diagnoses of gestational diabetes, physicians were most likely to use the National Diabetes Data Group criteria. From a supplemental questionnaire sent to a random sample of parous women who did not report gestational diabetes (n=114), 83% reported undergoing a 50 g glucose screening test during pregnancy and 100% reported frequent prenatal urine glucose screening, suggesting a high level of gestational diabetes surveillance among both cases and non-cases in this cohort.
Definition of the Low Risk Group
We considered four low risk lifestyle factors before pregnancy based on the strength of the evidence supporting their role in the development of gestational diabetes: a healthy body weight, adherence to a healthy dietary pattern, regular exercise, and abstinence from cigarette smoking. Not being overweight or obese was classified as a BMI <25. Participants in the upper two fifths for their scores before pregnancy on the Alternate Healthy Eating Index-2010 were considered to be in the lower risk category for adherence to a healthy diet. Low risk usual exercise was defined as an average of at least one half hour a day of vigorous or moderate physical activity for five days a week (150 minutes/week), including brisk walking, in keeping with published US guidelines. Women who reported being current non-smokers in the questionnaire preceding the pregnancy were counted as low risk for smoking status. For each of the four factors, we created a binary variable, with participants receiving 1 if they met the criteria for low risk and 0 otherwise.
Statistical Analysis
Cumulative average amounts of dietary food and levels physical activity before pregnancy were calculated at each time period to reduce variation within participants and reflect long term habitual levels. If a woman was pregnant while she was filling out a given questionnaire, data from the previous questionnaire were carried forward for exposure variables to better represent usual lifestyle before pregnancy, when available. We used multivariable log binomial models (that is, binomial regression models with a log link function) with generalized estimating equations, specifying an exchangeable correlation structure, to estimate the relative risks and 95% confidence intervals for associations between the four major potentially modifiable factors (smoking status, score on Alternate Healthy Eating Index-2010 (fifths), physical activity (fifths), and BMI) and risk of gestational diabetes. We used generalized estimating equations to account for potential correlations between repeated observations contributed by a single participant (that is, more than one pregnancy). Covariates in the multivariable models included the above variables, age (continuous), parity (nulliparity or 1, 2, 3, or ≥4 previous pregnancies), race/ethnicity (white, African-American, Hispanic, Asian, other, missing), family history of diabetes (yes, no), history of infertility (yes, no), year of pregnancy, total energy intake (fifths) before pregnancy, and alcohol intake before pregnancy (none, 1-5, 6-15, ≥16 g/day).
We calculated the population attributable risk percentages and 95% confidence intervals to estimate the proportion of gestational diabetes in this cohort that hypothetically would not have occurred if all pregnancies were in the low risk group by comparing pregnancies in the low risk category (for each factor individually or in combination) with all the others. To allow valid calculation of the population attributable risk percentage, we estimated the relative risks with age explicitly in the multivariate model based on log binomial generalized estimating equation models, as mentioned above. For simplicity and increased statistical efficiency, we used a single binary categorical variable to calculate the population attributable risk percentage pertaining to the impact of more than one low risk factor. In addition, we derived similar percentages and 95% confidence intervals based on the distributions of the four low risk factors according to the recent prevalence among women of reproductive ages (ages 18-45) in the nationally representative US National Health and Nutrition Examination Survey (2007-10), which is a program of studies designed to assess the health and nutritional status of adults and children in the US.
We also carried out a sensitivity analysis to assess the robustness of our findings, restricting our analysis to first births from nulliparous women to reduce possible confounding by experiences in previous pregnancies. We evaluated effect modification by established risk factors for gestational diabetes including age (<35 v ≥35), BMI before pregnancy (<25 v ≥25), family history of diabetes (yes, no), and race (white vnon-white) by conducting stratified analyses. P values for heterogeneity were derived from the cross product interaction term coefficient (exposure variable × binary effect modifier variable) added to the main effects multivariable model. All statistical analyses were performed with SAS software (version 9.1; SAS Institute). P<0.05 was considered significant.