Approximately 521,000 members joined the program during February 1 – April 30, 2008. Systematic random sampling produced the initial cohort of 26,582 individuals with a baseline weight. This de-identified dataset included all available follow-up data for these members through May 10, 2010.

Members were then excluded from analysis if they reported extreme outlying values for baseline characteristics: age greater than 100 years, weight less than 100 lb (45.4 kg) or greater than 800 lb (362.9 kg) with no follow-up weight entries, or height greater than 10 ft (3.05 m). These extreme outlier values were considered unrealistic, data entry errors, and/or not relevant to the analysis. Exclusion of 469 members with outlier values left an interim cohort of 26,113 members. Although we did not implement a filter for very short heights, the filtering process eliminated members with extreme height values, as the height range was 49-78 inches in the final cohort (n=1258) described below.

Self-reported weights may be inaccurate due to data entry errors, such as omitting, adding, or transposing digits. All 58,574 weight entries were analyzed for potential data entry errors with a series of automated and manual procedures. There were three automated filters (F1, F2, F3) for members with three or more weight entries, and a fourth automated filter (F4) for members with two or more weight entries. The first filter (F1) entailed fitting a second order polynomial regression line on the weight entries and dates; the distance of each point from the expected line (residual) was calculated. Points were flagged for visual evaluation if the residual was more than three times the standard deviation of the residuals for the individual record, and the difference between the actual and expected weight was greater than 10 lb (4.5 kg). The second filter (F2) flagged members that had a change of more than 2% body weight per day (whether over a short or long period of time). The third filter (F3) flagged members who had an absolute change of 50 lb (22.7 kg) or more between any two weight entries. The fourth filter (F4) flagged members with a weight change of 100 lb (45.4 kg) or more over the complete recorded period.

The automated filters identified 301 potentially erroneous weight entries among 248 unique SparkPeople members. Two independent observers (KOH and AWT) manually reviewed these 301 weight entries in the context of the other weight entries for a given individual (the trend) to determine whether the entry was erroneous. The observers demonstrated high interrater reliability, with an overall 95% agreement and Cohen’s Kappa .90. When the observers disagreed about a weight entry, they discussed and reached a consensus determination. After the automated filters and manual review, 73 of 301 weight entries (24%) were deemed erroneous and excluded from analysis. This represents 0.12% of all weights available (73/58,574).

Because the main objective was to evaluate the relationship between website usage and weight loss, 20,518 members with only one weight entry were excluded. The subsequent interim cohort of 5595 members had at least two weight entries. Members with at least two weight entries were younger, with a mean (SD) age 35.3 (11.5) years vs. 37.2 (13.4) years, P<.001, heavier, with a mean (SD) BMI 31.6 (7.7) vs. 30.8 (20.2), P<.001, and more likely to be female (91.3% vs. 85.9%, P<.001) than those with only one weight entry.

Each weight entry and log-in event was date-stamped. Because the data did not include the dates for use of other website features (eg, exercise diary), it was possible that a member used a website feature after his or her last weight entry, thereby complicating the interpretation of associations between website usage and weight change. However, it was possible to determine when a member stopped all website usage activity because each website activity generated a date-stamped log-in. Therefore, the final cohort of n=1258 was defined as those who had at least two weight entries, with the last weight entry on the same day as or after utilization of other website features. In other words, this final cohort consists of members who used the website features between their first and last weight entries. While the study design does not allow definitive analysis of causation, the final cohort at least meets the temporality criteria for causation. This final cohort was analyzed to examine the relationships between website usage and weight change. The flow diagram (Figure 1) depicts how the final cohort for analysis was defined.

Usage of website features included the following variables:

The weight entry span was defined as the number of days between first and last weight entry. Weight change, defined as the last recorded weight minus baseline weight, was stratified by weight entry span.

Potential confounders of the relationship between website usage and weight change included baseline BMI, age, gender, and weight entry span. Registration source (referred by friend, search engine ad, other) was examined as a potential confounder because it is possible that members who are introduced to the online program by a personal friend may differ in website usage and weight change than those who find the site only after an Internet search. Likewise, geographic location (zip code in midwest, northeast, south, or west region of the United States or other country) was examined as a potential confounder because of possible geographical variation in Internet use and dietary and physical activity factors related to body weight. These potential confounders were examined for association with weight change by using univariable regression analyses. Variables with a significant association (P<.05) with weight change were adjusted for in subsequent analyses.

Regression analysis was conducted with each website usage variable as predictor and net weight change as the outcome with adjustment for the identified confounders (univariable analyses). Multiple regression analysis with backward stepwise selection was conducted to identify the best subset of predictive covariates for net weight change. All covariates were included in the full model before model selection. Variables were included in the final model if P<.20.

The distribution of website usage was highly skewed, such that many members did not use a given feature before their last weight entry. Therefore, website usage variables were coded as binary variables for the initial analysis (never using a website feature vs. using it at least once). Binary coding of website usage has been used in a prior study of an online weight loss program [21]. Weight change was defined as the last recorded weight minus baseline weight. Because all members in the sample had two or more log-ins and weigh-ins, and most members had at least one SparkPoint, those variables were not included in this analysis.

To further analyze associations between website usage and weight change, one option would be to assume a linear regression model, but the linearity assumption may be not be valid. For example, the change in weight associated with an increase from 3 to 4 forum posts may be different from the weight change associated with an increase from 150 to 151 forum posts. We considered categorizing the website usage variables into tertiles, but the data distributions were too skewed, so that observations in the top tertile may be equivalent to some observations in the middle tertile.

Weekly use of online weight loss program features has been associated with greater perceived social support [20] and weight loss [3,4,6,7,14]. Therefore, we categorized log-in days, weight entry days, food entry days, exercise entry days, and forum posts as greater than or less than 4 events per 30 days, corresponding to approximately once per week. The median value of SparkPoints was 20.8 per 30 days, so SparkPoints was categorized as above or below 20 per 30 days. We initially categorized exercise minutes as above or below 600 minutes per 30 days, corresponding approximately to 150 minutes per week, but there were not enough members who met this threshold to allow meaningful analysis. Therefore, we categorized exercise minutes as above or below 120 minutes per 30 days, corresponding to 4 exercise entries of 30 minutes length per 30 days. Zero use was also included as a category for all variables except log-in days and weight entry days because, by definition, all members in the analysis had at least one log-in day and weight entry day.

Weight change was defined as (last recorded weight minus baseline weight) divided by 30 day unit of weight entry span. For example, if a member’s last weight was 100 kg and baseline weight was 108 kg, and weight entry span was 60 days, then the weight change was -8 / 2 = -4 kg per 30 days.

Matlab version R2008b (MathWorks, Natick, MA) was used to identify potential outlying values for baseline characteristics and self-reported weights. The R statistical package (R Development Core Team, Version 2.14.1) was used for descriptive and regression analyses.

The study was approved by the Committee for the Protection of Human Subjects at the University of Texas Health Science Center at Houston.

In the analysis with binary website usage variables, having at least one forum post was the only website variable significantly associated with greater weight loss. Members with at least one forum post reported an additional weight loss of 1.55 kg (95% CI 0.55 kg to 2.55 kg) relative to those with no forum posts (Table 4).

In the analyses with categorical website usage variables, weekly log-ins, weight entries, and food entries (at least 4 in 30 days) were associated with greater reported weight loss than less frequent use (Table 5). More exercise entry days per 30 days was also positively associated with weight loss. SparkPoints was associated with relative weight gain compared to having zero SparkPoints (although only 74 of 1258 members had zero SparkPoints). On multiple regression analysis including baseline BMI and all website usage variables, weight entry days was the only variable significantly associated with weight change. Members with at least 4 weight entry days per 30 days (weekly) reported an additional 5.09 kg weight loss per 30 days (95% CI 3.29 kg to 6.88 kg) relative to those with fewer weight entry days.

Average weight loss, based on self-report, was modest in this free online weight loss program, but active users had better outcomes. Making weekly weight entries (at least 4 weight entry days per 30 days) was associated with an additional 5 kg weight loss. After controlling for weight entry days, the other website usage variables were not associated with weight loss.

To our knowledge, this is the first analysis of a naturalistic cohort of members of a free online weight loss program available to the general public. Prior studies evaluated naturalistic cohorts of members who paid for monthly subscriptions to commercial online weight loss programs in Australia [12], the UK [13], and Sweden [14]. Because free programs may have greater potential to reach people in need of weight loss assistance [15,16], evaluating this free program is an important advancement in our understanding of weight loss resources. Another strength of this study was the use of multiple regression to identify variables that are independently associated with weight loss.

Our results are consistent with prior studies documenting a positive relationship between engagement in online programs and weight control. Most of these involved research volunteers and strict eligibility criteria [3-11], while others were naturalistic studies of commercial online programs [12-14]. Overall, favorable weight outcomes were associated with higher frequencies of log-ins [3-7,9,10,12,14], weight entries [6,10,11,14], self-monitoring entries for diet and/or exercise [6-8,10-13], and use of social support tools [6,8,12,13].

Although the current study discovered that several website usage variables were associated with weight loss, the multiple regression analysis indicates that weight entry days is the most important. It is possible that weight self-monitoring leads to frequent modification of diet and exercise behavior in response to weights. An alternative explanation is that members made weight entries only when they were losing (rather than gaining) weight. However, a post-hoc analysis found that 71% percent of members posted at least one weight reflecting a higher weight than a previous entry. In other words, they did not avoid documenting weight regain. The evidence in favor of weight self-monitoring improving weight control is strengthened by the consistency of the association in studies of online [6,10,11,14] and traditional [22-28] programs. When the analysis controlled for weight entry days, the other website usage variables were not related to weight change. It may be that the main benefit of these other website features is to maintain interest in the online program and maximize the opportunity for weight self-monitoring.

When website usage variables were coded as binary variables, making at least one forum post was associated with greater weight loss, although this analysis did not adjust for weight entry days. Because the content of the messages was not available for this study, we could not determine the nature of interactions on the forums. Furthermore, it was not possible to assess how often members read messages on a forum without posting (“lurking”). However, prior studies found that SparkPeople members receive high-quality advice [17] and social support [18] on the forums and that use of forums is associated with greater perceptions of social support [19,20]. Forum use may facilitate weight loss in part by boosting adherence to weight self-monitoring. Although the optimal source, frequency, and venue for social support has not been determined, the link between social support and weight loss has been reported in several studies of online [6,8,12,13] and traditional [29-34] weight loss programs.

The study had several limitations. First, attrition was high, which is a common problem plaguing commercial online weight loss programs [12-14]. The majority of members in the initial sample made only one weight entry, so that the main analytic sample represented a small portion of all who registered for the program. Compared to those with only one weight entry, members with multiple weight entries were younger, heavier, and more likely to be female—consistent with the profile of individuals who typically enroll in weight loss programs. Even within the analytic sample, 21% of members quit entering their weight after 30 days. Since it is very easy to join a free online program, it is possible that many people joined without a strong sense of commitment, thereby inflating the initial sample and the subsequent attrition rate. Sending personalized follow-up messages to inactive members may improve adherence to weight self-monitoring [7], although this approach would be difficult to implement without additional resources.

Another limitation of the study is that the available records did not include ethnicity, race, or clinical characteristics. Furthermore, members were predominantly female, as is typical of weight loss programs. Taken together, these limitations indicate that results cannot be extrapolated to the general population, regular Internet users, or those who initially who register for this online program. The study is also limited by reliance on self-reported weights. However, weight reported by members of another online weight loss program was highly correlated with and similar to objective weights [35]. The lack of date stamps on most website usage variables (except log-ins and weight entries) limited the depth of analysis on the intensity and patterns of website usage. Lastly, because this was an observational study, the causal link between usage of website features and weight loss cannot be definitively ascertained without experimental studies.

The public health impact of an intervention is determined by efficacy and dissemination [36]. Because this online program is free, scalable, and widely disseminated, the potential public health impact is significant. The current study suggests that increasing the use of weight self-monitoring among current and future members may boost the impact of this program even more.