Data from a randomized controlled trial of prevention of excessive GWG and postpartum weight retention with women who were 18-35 years of age, normal range to obese class I body mass index (BMI), socially and racially diverse, and relatively healthy (N=1689) and conducted in a midsize city in northeastern United States were used in this study. This trial, conducted from 2011 to 2014, is described in detail elsewhere [18,19], and its ClinicalTrials.gov identifier is NCT01331564. Pregnant women were screened in prenatal clinics, private obstetric practices, practices that provide ultrasounds, and over the phone and online. To meet eligibility criteria, participants had to (1) consent at or before 20 weeks’ gestation, (2) be available for a 24-month intervention, (3) plan to carry their pregnancy to term and keep their baby, (4) read and understand English, and (5) have an email address. Exclusion criteria included the following: BMI less than 18.5 kg/m² (underweight) or greater than or equal to 35.0 kg/m² (obese class II or greater), multiple gestation (eg, twins), having had eating disorders or gastric bypass surgery in the past, having had 3 or more consecutive miscarriages, and the presence of prepregnancy medical conditions that could influence weight loss or gain.

All study participants were sent an email describing the tools on the website. Email, postcard, and telephone reminders were used as prompts to encourage participants to visit the website the first time. A US $5 incentive was also given for the first website visit. The sample for this analysis includes women who were eligible, entered the study during pregnancy as indicated by at least one website log-in or completion of the baseline questionnaire, and had a singleton pregnancy that lasted at least 20 weeks (n=1335). The study protocol was approved by the University of Rochester Research Subjects Review Board and the Cornell University Institutional Review Board.

Fishbein and Yzer’s Integrative Model of Behavioral Prediction [20] was the guiding theoretical framework for the Internet-based intervention to prevent excessive GWG. Fishbein and Yzer’s framework was combined with Fogg’s Behavior Model for Persuasive Design [21] to link weight-associated behaviors and their predictors to intervention features. Michie’s behavior change techniques [7,22] informed the development of the website features for the intervention arm (Figure 1). The website featured blogs, local resources, articles, frequently asked questions (FAQs), and events that were available to women in both the intervention and control arm. Weekly reminders to use the website and to highlight new content were emailed to participants in both arms. In addition, intervention arm participants had access to the weight gain tracker and diet and physical activity goal-setting tools. Intervention participants were emailed weekly with reminders to use the weight gain tracker and diet and physical activity goal-setting tools. Intervention features are described in more detail in the study by Graham et al [23].

Use of the Internet-based features was automatically captured by the website. Utilization of the following 6 intervention features plus log-ins was used to characterize usage in the intervention arm: health-related information (articles and FAQs), blogs, local resources, diet goal-setting tools, physical activity goal-setting tools, and a weight gain tracker (Figure 1). For the control arm, log-ins plus the use of the first 3 features mentioned above were included in the patterns-of-usage measure. For some features, the amount of use in relation to expected use was captured in the patterns-of-usage measure. Consistent use was expected for log-ins and entry of weights into the weight gain tracker. We expected women to track their weight in 30-day intervals, but, to allow for difference in timing of prenatal care, we created 45-day intervals from time of enrollment to delivery. If a woman entered a weight during each of the 45-day intervals of her study participation, she was categorized as a consistent tracker. If during at least half of the intervals a woman entered a weight, she was categorized as an almost consistent tracker. If a woman had entered at least one weight but not during more than half of her intervals, she was categorized as an inconsistent tracker. Finally, if she never entered a weight during pregnancy, she was categorized as a nontracker. We counted website log-ins as feature usage given the amount of content that was visible on the website dashboard after log-in for both intervention and control arm participants (Figures 2 and 3). The same procedure for consistent weight tracking was used to categorize log-ins.

For all other features, consistent use was not expected. Use was expected on an “as needed” basis. Therefore, quantity of use was used for the following features: health-related information, blogs, resources, diet goal-setting tools, and physical activity goal-setting tools. The usage of a feature by a woman was categorized into 3 levels for each of these features: high (≥median among users), low (<median among users), or none (0).

Latent class analysis (LCA) was used to identify usage patterns by women in the intervention arm and the control arm. This analysis was used to group individuals based on their similar usage patterns. All analyses were conducted using a SAS procedure, PROC LCA version 9.2 (SAS Institute). The creation of the intervention usage pattern variables is described in greater detail in the study by Demment et al [17]. The sample used in this analysis excluded women who had never logged in and had not completed a questionnaire, which influences the latent classes that emerge.

GWG data were obtained through an audit of the participants’ prenatal, labor and delivery, and 6-week postpartum medical records. Total GWG was calculated as the difference between the first weight and the last weight in pregnancy. Overall, 12.80% (171/1335) of the sample did not have sufficient weight information in the prenatal chart to yield a valid measured GWG. Sufficient weight information required having a measured weight at both <14 weeks’ and ≥37 weeks’ gestation. Missing data were handled using multiple imputation to address issues of bias [24], which may result from analyzing only complete cases using SAS PROC MI and MIANALYZE procedures. A total of 60 imputed datasets were created for the primary analysis and also for the models presented below. The weekly rate of GWG was calculated as the difference between the last pregnancy weight and the weight measured closest to 20 weeks (±3 weeks), or the imputed average weight at 18-21 weeks if missing, divided by the number of weeks between these weights. Next, the binary outcomes of excessive total GWG and excessive rate of GWG after 20 weeks were calculated using the Institute of Medicine guidelines for each BMI group as determined at randomization. For each of the 3 BMI groups, the cutoff values for excessive total GWG and excessive weekly rate of GWG are as follows: normal range BMI, greater than 16 kilograms and greater than 0.23 kilograms per week; overweight BMI, greater than 11.5 kilograms and greater than 0.15 kilograms per week; and obese class I BMI, greater than 9 kilograms and greater than 0.12 pounds per week.

Among those participants with measured and available GWG data, chi-square analysis was used to examine the relationship between usage patterns and (1) demographics and (2) the binary outcomes of excessive total GWG and excessive rate of GWG. For total GWG, analysis of variance was used to examine the relationship between usage patterns and GWG. Since the created usage patterns differed by arm, these analyses were performed separately within intervention arm participants and within control arm participants.

Using a modified Poisson regression approach [25], the relative risk of excessive total or weekly GWG was estimated for different usage patterns within strata. Similarly, least squares multiple regression models assessed the mean difference in total GWG (kilograms). Usage was parameterized with two approaches. First, we examined by strata whether or not various patterns of usage were associated with relative risk of excessive GWG. Second, we examined by strata how the combined usage patterns that include most frequent usage of the active ingredients of the intervention, that is, weight gain tracking and behavioral goal setting and self-monitoring (Figure 1), were associated with relative risk of GWG. All models were adjusted for BMI, age, race, and parity, as well as pregnancy timing variables including gestational age at delivery, the weeks between the first and last pregnancy weight, and the weeks between the last pregnancy weight and delivery. Significance level was set at P ≤.05.

We applied a novel approach, LCA, to understand usage of website features included in a GWG eHealth intervention. We examined patterns of usage for both the intervention participants and the control arm participants. This approach is a unique methodological contribution to process evaluations for self-directed Internet-based interventions where the most appropriate measures of engagement are not yet well defined. Usage patterns for both intervention and control arm participants varied by demographic characteristics. Higher-income, older, white, and married women in both arms were more likely to be higher users of the website. In the control arm, where the content of the website was primarily informational, women who were having their first baby were greater users.

In the control arm, GWG outcomes did not differ by usage pattern. While this was expected because behavior change and weight management tools were not included on the control website, documenting that lack of effect by amount of use helps in interpreting the relationship between amount of use of features in the intervention arm and weight outcomes. The concern is that amount of use is associated with a personality type that will have better outcomes, no matter what the content of the intervention is. The control group results indicate that this concern is likely not relevant to this study.

In the intervention arm, GWG outcomes did differ by usage pattern. Among participants with lower income and normal range BMI (stratum 1), almost consistent or inconsistent trackers had a higher relative risk of excessive GWG and inconsistent trackers gained more weight during pregnancy compared with the nonuser usage pattern. Among participants with normal range BMI and higher income (stratum 2), the consistent trackers had a lower relative risk of excessive weekly GWG rate compared with nonusers. Among overweight and obese higher-income participants (stratum 4), the consistent trackers gained less weight during pregnancy than nonusers.

In order to better understand the patterns of usage and GWG outcomes, we examined 2 of the usage groups together that included most frequent usage of the active ingredients of the intervention, weight gain tracking and behavioral goal setting and self-monitoring. We compared super users and consistent trackers with the 3 usage groups that used less of the hypothesized active ingredients in the pregnancy intervention (medium users, almost consistent or inconsistent trackers, and nonusers). In the higher-income stratum (strata 2 and strata 4), higher users of the active intervention ingredients were associated with reduced risk of excessive GWG total and in the normal range BMI women (stratum 2) for weekly rate. Across BMI categories, total GWG (kilogram) was significantly lower in the super and consistent users compared with the medium users, almost consistent or inconsistent trackers, and nonusers.

The strengths of this study are as follows: the intervention’s measures of usage are objectively measured by the website and as such there is no study staff reporting bias for the intervention use variables; a large randomized effectiveness trial with an economically and racially diverse sample; a theory-based and formative research-informed Internet-based intervention.

A limitation of this research is that consistent weight tracking and logging in to the website was low and was particularly low among lower-income participants, with only 25% of low-income participants in either the consistent tracker or super user usage patterns. This affects the ability to detect statistically significant differences in the GWG outcomes between groups defined by usage.

This study used a novel, data-driven approach to process evaluations that may be particularly helpful for self-directed Internet-based interventions on any topic. This approach may further the understanding of how self-directed Internet-based intervention tools are used and whether there are benefits associated with different patterns of use. The implications for this particular self-directed Internet-based intervention to prevent excessive GWG vary by socioeconomic status of the women. For higher-income women there was a reduction in GWG, but not necessarily a significant reduction in rate or excessive GWG for overweight or obese higher-income women. For lower-income women there were no detectable effects of usage on GWG. Future self-directed Internet-based interventions should consider best approaches for consistently engaging lower-income women when the success of interventions is anticipated to depend on consistent use.