The Atherosclerosis Risk in Communities (ARIC) Study was designed to examine risk factors for cardiovascular disease and its related morbidity and mortality. Participants were selected from four US communities: Washington County, Maryland; northwest suburbs of Minneapolis, Minnesota; Jackson, Mississippi; and Forsyth County, North Carolina. A probability sample of men and women ages 45 to 64 years was recruited from each community, from which 15,792 adults completed the first clinic visit. The first clinic visit occurred between 1987 and 1989 and included a physical examination and an interview by trained personnel regarding medical history, health status, diet, and physical activity. Clinic visits were repeated at approximately three-year intervals. Further details on the ARIC cohort are available elsewhere 2425.

The Baecke physical activity questionnaire was designed to study habitual physical activity and distinguish between different dimensions of physical activity using semi-continuous indices of sport, leisure, and work activity 26. The questionnaire was interviewer administered at the baseline (1986–1989) and third clinic visit (1993–1995) of the ARIC study. The questionnaire included an item on television watching as part of the leisure activity index. This subjective question of television exposure asked: "During your leisure time do you watch television?" and allowed five responses: never, seldom, sometimes, often, and very often. The answers were ranked on an ordinal scale from 1 (low) to 5 (high). From this ordinal scale, television watching was collapsed into three exposure levels: low (never/seldom, n = 2878), medium (sometimes, n = 7293), and high (often/very often, n = 5403). Because television watching is a component question of the leisure index, the remaining three questions on leisure activity were analyzed separately. These items included information on walking and biking and were ordinally scaled from five responses (never, seldom, sometimes, often, and very often). The third question asked about minutes spent walking or biking to and from work or shopping and respondents could answer < 5 minutes/week, 5–<15 minutes/week, 15–<30 minutes/week, 30–<45 minutes/week, or ≥ 45 minutes/week.

Sport and work physical activity were assessed with semi-continuous indices created from individual component questions and have been described in detail elsewhere 262728. Using the reported activities from the sport index, participants were also dichotomised as sufficiently or insufficiently active. Sufficient activity was defined as regular physical activity, or participation in at least one hour per week of activity for 10 or more months of the year 2728. Insufficient activity was identified by less than one hour per week and/or less than 10 months per year.

The reliability and validity of the Baecke questionnaire have been examined in both European and American populations. Test-retest reliability correlations of total physical activity from the three indices range from 0.65 to 0.93 26293031. Validation of the Baecke has been conducted using physical activity diaries, maximum oxygen consumption, accelerometers, and doubly-labeled water 29303233343536373839. In a doubly-labeled water study of three physical activity questionnaires, the Baecke had the highest agreement for total physical activity (Pearson's r = 0.69) 40.

Usual dietary intake was collected at baseline and at the third clinic visit using a semi-quantitative, interviewer-administered food frequency questionnaire (FFQ). The ARIC FFQ contains 66-items and was based on the original Willett 61-item FFQ 41. The ARIC FFQ is organized into seven sections; 1) dairy, 2) fruit, 3) vegetables, 4) meats and fish, 5) sweets and baked goods, 6) miscellaneous, carbohydrates, fried foods, and 7) beverages. Interviewers asked participants; "In the last year how often, on average, did you consume ______?" Nine responses were available for each food item ranging from "almost never" to "more than 6 times per day". Each response was assigned a weight to estimate servings per day and daily intake of nutrients 42.

Using the standard serving sizes from the FFQ, daily servings of each food item were calculated and summed to create food groups. Dietary outcomes were categorized into the following food groups: fruit, vegetables, fruit and vegetables combined, salty snacks, sweets, sweetened drinks 43. Total caloric intake, total fat and percent saturated fat, and estimated nutrient values were calculated at the Channing Laboratory, Harvard Medical School.

The original Willett FFQ has been validated in a number of populations and is a well-recognized dietary tool with validity correlations between 0.35 and 0. 74 414445. The reliability of the ARIC FFQ has been estimated among both white and African Americans participants in the ARIC study 46. Median reliability coefficients were 0.48 and 0.63 for White women and men, respectively, and 0.45 and 0.50 in African American women and men, respectively.

Anthropometric measurements of weight and height were obtained during both clinic visits and were used to calculate body mass index (BMI) (weight in kilograms divided by height in meters squared). Four weight status groups were created using BMI: underweight (<18.5 kg/m²), normal weight (≥ 18.5–<25 kg/m²), overweight (≥ 25–<30 kg/m²), and obese (≥ 30 kg/m²) 47.

At baseline, participants reported cigarette smoking (current, former, never) and years of education (<high school, high school, some college, or higher). Also at baseline, participants were asked to rank their general health as excellent, good, fair, or poor. Because one of the study centers was entirely African American (Jackson, MS) and others predominately white, a race-center variable to control for effects of race and center was created by combining each participant's race with their respective study center (e.g., Forsyth County African American; Forsyth County white).

In order to examine the associations between television, physical activity, diet, and body weight status participants were excluded who did not answer the television question at baseline (n = 32) or were missing information on age, race, gender, body mass index, general health status, or smoking (n = 83). Additionally, in order to control for the effect of race and center, individuals other than white or African American and all non-white participants from Minneapolis and Washington were excluded (n = 103). Excluded from the prospective analysis were those individuals who died (n = 722) or did not return to the third clinic visit (n = 2,125), as well as those missing information on both physical activity and diet outcomes at the third clinic visit (n = 49). Final sample sizes were 15,574 for the cross-sectional analysis and 12,678 for the prospective analysis, the latter referred to as "the cohort".

The associations of baseline television watching with baseline and six-year physical activity, diet, and body weight status were estimated from multivariable linear and logistic models. Analysis of variance (ANOVA) was used to estimate the mean of each outcome variable at each of the three levels of television exposure (low, medium, high). For the linear models, analyses were also performed using log transformations of all non-normally distributed outcomes. The results were unchanged using log-transformed variables and therefore, for ease of interpretation, they were left in their original state. For logistic models, outcomes were dichotomized as: having a BMI >= 25, being below the median for positive outcomes (e.g., physical activity), or above the median for negative outcomes (e.g., total energy intake). Low television exposure was defined as the referent and the odds of each outcome were estimated for medium and high television exposure.

To assess the temporal association of baseline television watching with physical activity and diet over the six years of follow-up, the baseline value of each outcome was included in all prospective models for a further "baseline-adjusted" model. In a longitudinal analysis, adding the baseline outcome variable to the model will adjust for the cross-sectional associations at baseline between television and each outcome of interest.

All statistical models included the following covariates measured at baseline: age, race-center, gender, BMI, education, smoking, and general health, with categorization indicated in Table 1. An a priori conceptual model based on a literature review and previous ARIC analyses helped direct model building with regard to confounder/covariate cut-points, television categories, and outcome categorizations. Post-hoc analyses were performed to assess if different outcome cut-points (median, vs. mean dichotomization; tertile vs. quartile categorization) resulted in different associations or conclusions. Analyses were also performed to assess any effect of outliers (defined as ≥ 3 standard deviations from the mean) and missing data. The statistical models, measures of effect, and 95% confidence intervals were computed using SAS V9.1 (Cary, NC; 2002).

The analysis for this study was approved by the Public Health-Nursing IRB at the University of North Carolina at Chapel Hill, and participants provided written, informed consent at each ARIC center.

At baseline, watching television "often" or "very often" (high exposure) and "sometimes" (medium exposure) was associated with a 40% and 20% greater odds of being insufficiently active, respectively, compared to watching television "never" or "seldom" (low exposure) (Table 3). High television exposure was also associated with a 50% greater odds of being below the median of the sport index, but it was not associated with the work index. Participants with low television exposure had almost two-fold greater odds (1.96, 95% CI 1.78, 2.17) of walking during leisure time compared to those with medium television exposure. High television exposure was also associated with less walking and biking during leisure and for transportation. Increasing television exposure exhibited a graded relationship with unhealthy dietary choices (Table 3). High television exposure was associated with an approximate 20 to 30% greater odds of being above the median for servings of salty snacks, sweets, and sweetened drinks, total fat, and percent of calories from fat and saturated fat, and below the median for fruit and vegetable servings.

In a linear model, differences were examined between the adjusted means of each outcome at the three levels of exposure. These models indicated that individuals with higher television exposure participated in less sport physical activity than people who were exposed to medium or low amounts of television. There was no apparent relationship between television exposure and physical activity from the work index (Figure 1). Examining diet, we found that participants who reported high television exposure ate more servings per day of salty snacks (0.07, 95% CI 0.04, 0.09), sweets (0.16, 95% CI 0.10, 0.22), sweetened drinks (0.11, 95% CI 0.07, 0.15), percent of calories from fat (0.78, 95% CI 1.09, 0.47), and percent of calories from saturated fat (0.27, 95% CI 0.47, 0.13) than people who reported low television exposure (Figures 2, 3, 4). Those with high exposure also consumed almost one-half serving fewer fruits and vegetables per day (-0.41, 95% CI -0.52, -0.30) (Figure 2).

Similar relationships were observed between increasing television watching and body weight status. At baseline, participants with medium and high television exposure had a 22 and 43 percent greater odds of being overweight or obese than individuals with low exposure (Table 4). However the absolute magnitude of effect was less dramatic when comparing adjusted means in a linear model, and the only significant comparison was between low and high television exposure (Figure 5). Further adjustment for sport physical activity and total kilocalories had no appreciable affect on these results.

Compared with low exposure, high and medium exposure to television remained a predictor of physical activity, diet, and body weight status over the six years of follow-up (Table 4, 5). Although the six-year associations were attenuated compared to the baseline associations, a consistent relationship was still observed between higher television exposure and more unhealthy physical activity and dietary outcomes. Adjusting for the baseline association of exposure and outcome in the models resulted in further attenuation of these associations, although a significant effect was observed for most physical activity and dietary outcomes.

The associations between television watching and body weight status were almost the same for both medium and high television exposures in the prospective model (Table 4). However, both exposure levels were still associated with greater odds of being overweight or obese at six-years follow-up relative to low exposure. Additionally, when the baseline term for body weight status was added, the prospective association attenuated to the null.

Post-hoc exploratory analyses were conducted to examine the effect of different cut-points and model choices. Regardless of the model choice (linear versus logistic model), consistent associations were observed between television and each outcome using different exposure and outcome categorizations (continuous, tertile, quartile, different dichotomization). Even when outliers were removed, the conclusions were the same (data not shown). Lastly, we also restricted the sample to only participants with complete data however the conclusions were again the same (data not shown).

This study is the first large cohort analysis to examine the associations of reported television exposure with physical activity, diet, and body weight status. However, it is important to recognize that the question measuring television exposure was subjective and semi-quantitative, and could result in misclassification. Additionally, although the Baecke questionnaire has shown validity and reliability in other populations 293940 the television question used in this analysis has not been validated as a single item, to our knowledge. Although better measurement tools have been developed and great strides have been made in physical activity epidemiology, 52 no tool has been designed to adequately measure sedentary behaviors 53. The television question in the ARIC study is also not an absolute measure of time, and is a subjective measure of exposure. To minimize this limitation, ancillary data was obtained (M. Hulens, personal communication, 2004 54) which compared the Baecke television question with a second, concurrent report of the continuous number of hours of television exposure in a Belgian population. Agreement between continuous hours of television and the single item Baecke was acceptable (chi square 92.3, p < 0.0001).

Food frequency questionnaires like the ARIC-FFQ collect information on a limited number of food items and do not assess total energy intake; therefore, these dietary outcomes contain measurement error. Yet until better methods are developed, FFQs are a practical method for gathering information on dietary behaviors in large cohort studies 44. Although validation of the ARIC FFQ is not available, the Willet FFQ, on which the ARIC questionnaire is based, has been validated and used in other well-recognized cohort studies 415556. Reliability of this tool has been shown to be between 0.45 – 0.63 across visits (3 years) in this study population 46. Future studies could consider exploring measures of diet quality, in addition to the components that we assessed.

Another limitation of this study is the loss of participants between visits. These individuals (n = 2,847) were less healthy, reported higher television exposure, and had more negative diet and physical activity patterns than those who remained in the study. Because these individuals fall into both the high exposure and more unhealthy outcome categories at baseline, it is unlikely that the results would be different had these participants remained in the study.

This study is observational and relies on recall of both exposure and outcomes. Therefore, in a study of this nature it is impossible to untangle any effect of previous television, activity, and dietary behaviors from our observations. Additionally, baseline data collection began in the mid-eighties and societal television, physical activity, and diet behaviors will not be the same as present day. It is important to recognize that the generalizability of this study to current sedentary behaviors and other populations is limited.

Additionally, physical activity and diet are arguably mediators on the pathway between television exposure and overweight or obesity. For example, physical activity has been previously found to be an effect modifier of the association between television and obesity 19. A fundamental principle of adjusting for a confounder is that the covariate is not an effect modifier or mediator of the relationship between exposure and outcome. Therefore, in our example adjusting for a mediator (e.g., physical activity) like a confounder may not result in the best estimate of the association 57. Lastly, although an attempt was made to establish temporality, true cause and effect cannot be ascertained from this study design.