Journal Logo

BASIC SCIENCES: Epidemiology

Patterns of Objectively Measured Physical Activity in the United States


Author Information
Medicine & Science in Sports & Exercise: April 2008 - Volume 40 - Issue 4 - p 630-638
doi: 10.1249/MSS.0b013e3181620ebc
  • Free


Because of the increased concern about the lack of physical activity in the United States, many annual national health surveys such as the Behavioral Risk Factor Surveillance System (BRFSS) and the National Health Interview Survey (NHIS) collect information on physical activity levels. Reports and research from these surveys cover various aspects of the nation's levels of physical activity (or inactivity), such as general physical activity trends (12), trends among certain subpopulations such as Hispanics (1), percent of the U.S. population meeting the Healthy People 2010 leisure-time physical activity goals (6), or associations between physical activity and health care expenditures (26). These reports, however, have in the past relied on self-report, which, when compared with objectively measured physical activity, have been shown to have low correlations in the range of 0.14-0.53 (24). In addition, participants are generally categorized into activity levels on the basis of overall descriptions of activity frequency and intensity, without attempting to categorize the potential differences in the pattern of accumulated physical activity over time.

In the 2003-2004 National Health and Nutrition Examination Survey (NHANES), physical activity as measured by accelerometry was added to the battery of assessments among those participants 6 yr old and older who were ambulatory, providing the first nationally representative sample of objectively measured physical activity in the United States. In addition, the accelerometry data were collected during a 7-d period, allowing for an assessment of the number of minutes of physical activity accumulated by each participant on each day of a 7-d week.

Using latent class analysis (LCA), we assessed whether patterns, or classes, of physical activity exist among adults in this sample during the 7-d period. In this type of analysis, a specified number of classes are requested a priori. Then, LCA finds the requested number of best-fitting, underlying normal distributions for the indicators of these classes (in this case, the daily minutes of physical activity across the 7d of a week). For example, one class may have a very low mean number of minutes of physical activity from Monday to Friday, with a high average number of minutes of physical activity on Saturday and Sunday (i.e., a "weekend warrior"). Another group, possibly active workers, may have a large mean number of minutes of physical activity during the work week but low mean minutes of physical activity on the weekend.

A recent study attempted to assess the effect of the weekend warrior activity pattern on the risk of mortality (14). In this study, the mortality outcomes of the weekend warrior, defined as those who accumulate a large quantity of physical activity (≥ 1000 kcal·wk−1) during a short period of time (1-2 d·wk−1), were compared with those who accumulate a similar amount of activity (≥ 1000 kcal·wk−1) for a longer period of time (3+ d·wk−1), along with those who are insufficiently active (500-999 kcal·wk−1) or sedentary (< 500 kcal·wk−1). Among low-risk men, defined as nonsmokers, BMI < 25 kg·m−2, no history of hypertension, and without hypercholesterolemia, weekend warriors demonstrated the lowest relative risk of mortality, indicating that as long as the accumulated physical activity is sufficient, then the benefits will be accrued. Among high-risk men, defined as having any of the above risk factors, only the regularly active showed improved mortality risks as compared with the most sedentary group.

Much is now known about the overall levels of activity in the United States, but few studies have attempted to define patterns of physical activity. The pattern in which individuals accumulate their physical activity can be highly variable. To meet the recommended guidelines of 30 min·d−1 for 5 d·wk−1 (150 min·wk−1) (23), activity could be undertaken in various ways. For example, one could be physically active all 7 d for a defined duration (~20 min·d−1), whereas another individual may be physically active twice per week at a much longer duration (70 min each). In both cases, the cumulative time spent in activity is the same (140 min·wk−1). Varying the intensity of the activity could also influence activity patterns. Using LCA may help highlight strategies for how inactive people may accumulate additional physical activity, such as by adopting a weekend warrior pattern of activity. In addition, LCA may allow future analysis to assess whether different patterns of physical activity are associated with improved health outcomes.


We analyzed the 2003-2004 NHANES, an ongoing, nationally representative health survey with a target population of civilian, noninstitutionalized U.S. citizens. Certain populations were oversampled, including low-income persons, Mexican Americans, African Americans, and those ages 12-19 and 60 yr or older. The survey consists of an interview, from which sociodemographic information is collected, and a physical examination, from which various biological markers of health are ascertained. In addition, the 2003-2004 NHANES collected seven consecutive days of accelerometry measurements among all ambulatory participants 6 yr old and older who agreed to wear the monitor for a week. Written informed consent was obtained from all participants.

Measuring Physical Activity with Accelerometry

Accelerometers are small, electronic devices that record the acceleration of change in bodily motion, either in one plane or in multidimensions. They are particularly useful in measuring physical activity because they eliminate the potential for recall bias and social desirability bias, and they do not depend on literacy. NHANES 2003-2004 used the ActiGraph Model 7164 accelerometer manufactured by ActiGraph (formerly MTI/CSA) to collect information on participants' physical activity. This lightweight, uniaxial monitor is a technically reliable instrument, both within and across monitors (18). NHANES used 1-min epochs to assign a "count" value, a relative measure of the changes in momentum that occurred during these periods, which may then be translated into an estimate of physical activity intensity.

Moderate and Vigorous Physical Activity Cut Points Based on Calibration Studies

The accelerometer cut point used by this study to translate the count value into an estimate of moderate-to-vigorous physical activity (MVPA) was based on a weighted average of published cut points for adults (3,9,15,27) in accordance with the recommendation of NHANES researchers. Each study listed in Table 1 reports a cut point for MVPA, and these were then weighted by their sample sizes to arrive at an N-weighted average cut point of 2020 counts per minute for MVPA. Cut points for vigorous physical activity (VPA) were also reported in these calibration studies, and, using the same N-weighted average as used with the MVPA, the VPA cut point was 5999 counts per minute.

N-weighted mean cut point for moderate to vigorous physical activity (MVPA) and vigorous physical activity (VPA) based on previously published calibration equations.

Accumulating Minutes of MVPA and/or VPA

Physical activity accumulated in a given day was quantified as 1) minutes per day in which the count was higher than the given MVPA cut point, 2) minutes per day in which the count was higher than the given VPA cut point, and 3) minutes per day of MVPA accumulated in bouts of 10 min or more. The latter classification was motivated by the physical activity recommendations published by the Centers for Disease Control and Prevention (CDC) and the American College of Sports Medicine (ACSM) (23), which call for activity to be accumulated in bouts of 10 min or more to achieve health benefits. To allow for brief periods of rest common during activities-for example, to pause for a water break while playing basketball-the criteria used to define a bout required a running average of 70% of the counts to be above the cut point. Once the series of accelerometer minutes fell below 70% of minutes in MVPA, the bout was considered over. Thus, bout minutes of MVPA were a sum of all minutes of MVPA accumulated in these bouts. Those who never achieved a 10-min bout were assigned their longest bout shorter than 10 min in length. Bout minutes of VPA were not assessed, because too few participants achieved the required 10 consecutive minutes of VPA.

Imputation of Missing Daily Minutes of MVPA and/or VPA

Even though accelerometer data were present for most days, within each day there may be extended periods of zero counts, indicating either a nonwearing period or a period with no detectable movement. Periods consisting of 1 h or more of consecutive zeros were treated as missing data. In addition, periods of monitor malfunctioning were also considered missing (e.g., 10 min of identical consecutive nonzero count values). After dividing each day into segments corresponding roughly to nighttime (midnight to 6:00), daytime (6:00 to 17:00), early evening (17:00 to 21:00), and late evening (21:00 to midnight), data were then imputed using the expectation maximization (EM) algorithm, an iterative imputation technique that uses the values of an individual's other, nonmissing data as predictors to estimate the expected value of the total minutes of MVPA/VPA for each missing segment of time (4). Imputation was only performed for those participants providing one or more days of complete data, and remaining participants were excluded. Given this strategy, 88.5% of the population had at least one time segment imputed on at least 1 d, but 49.7% of the population had no time segments imputed for four or more days; overall, this only led to the imputation of 20.6% of the total time across all days.

Self-Reported Variables

Age was recorded at the time of the interview, and those over 85 yr of age were assigned a truncated value of 85. For descriptive purposes, age was categorized into decades but was left continuous in the final LCA. Gender was recorded at the time of interview. Race, ethnicity, and country-of-origin questions were recoded into the following categories: 1) Mexican American, 2) other Hispanic, 3) Non-Hispanic (NH) black, 4) NH white, and 5) other race, including multiracial. Education was categorized as less than high school, high school or GED, and more than high school. The poverty income ratio (PIR) was recorded as a ratio of the self-reported family income to the poverty threshold based on family size. The smallest value of 0 indicated no family income, and the highest value was truncated at 5, indicating a family income at least five times the poverty threshold for family size. For descriptive purposes, the poverty ratio was categorized into integer values but was left continuous in the final class analysis.

Statistical Methods

Descriptive statistics.

One thousand two hundred thirty-nine participants from our base study population were excluded because of missing covariate information or because their accelerometer data were missing or invalid (no days of adherent wear time). To determine whether the analysis sample differed from the subgroup excluded because of missing/invalid data, we used chi-square tests to compare categorical variables, and we used t-tests to compare the mean of continuous variables. Using SAS (Cary, NC) survey procedures, sample weighted means, standard deviations, standard errors, and 25th, 50th, and 75th percentiles were computed for overall MVPA, bout minutes of MVPA, and VPA for each day of the week. To account for item nonresponse, the sampling weights were reweighted within strata of age (20-34, 35-49, 50-64, and 65+), gender, and race (Mexican, NH black, and white/other) (21).

Latent class analysis.

Employing LCA, we used each participant's 7 d of total MVPA/VPA to determine whether natural groupings, or classes, of people exist who tend to accumulate their minutes of physical activity in a similar pattern during the 7 d. Classes can be thought of as groups of people who share similar means for the various indicators of class-in this case, the 7 d of accumulated MVPA/VPA. The most general probability density function used to define the LCA model is as follows:

On the left side of this equation, the probability distribution function defines a distribution for y, the vector of minutes ofMVPA/VPA across the 7 d, conditioned on μ(θ), the vector of mean MVPA across the 7 d, and Σ(θ), the covariance matrix for the multivariate normal distribution of the 7 d. On the right side, for each underlying class (indexed by G), the probability function for the vector y is weighted by probability of being in the each of these specific classes. μg(θ) is the vector of predicted means for the gth group, and Σg(θ) is the covariance matrix for the gth group. The overall probability distribution for y is, thus, a probability-weighted sum of each class's probability density for the given values of y (7).

If the distribution f is assumed to be a multivariate normal distribution with G components, then the probability density function of an individual in the Gth class is

where yi is the vector of minutes of MVPA/VPA across the seven times for the ith subject, μg(θ) is the vector of predicted means across the seven times t for the ith subject, Σg(θ) is the covariance matrix for the multivariate normal distribution of the 7 d of MVPA/VPA, and ∣Σg(θ)∣ is the determinant of the covariance matrix (7).

Thus, the likelihood function used to maximize this model for all participants across all class possibilities is

where N indexes the subjects and G indexes the classes, and, again, the pi weights the probability function for the gth class. This likelihood function is identical to the multivariate normal distribution with the addition of the probability-weighted class memberships (2).

The probability of being a member of a particular class is assigned to an individual according to Bayesian posterior probabilities, using a prior probability proportional to the size of the particular class relative to the entire population. Thus, the probability of an individual being a member of class g is

where Pg is the prior probability of being in class g, conditioned on the covariates. The numerator in this case is the prior probability that subject i belongs to class g, multiplied by the probability density for the observed 7 d of MVPA for yi, given the predicted means and covariance of the 7 d of MVPA in class g. The denominator is the sum of the probability densities for all possible class memberships given the individual's set of indicator values yi, weighted by each class's specific prior probability (11). Individuals were assigned to the class with their highest posterior probability of class membership, referred to as modal allocation.

Selecting the number of classes.

One of the most difficult tasks of LCA is determining the proper number of classes that adequately describe the population without overspecifying the number of class groupings, thereby losing the interpretative value of the classes. Several criteria were used to select the appropriate number of classes. We first used the bootstrap likelihood ratio test (BLRT), which compares the fit of k classes to k − 1 classes, because it outperformed the Lo-Mendell-Rubin likelihood ratio (16) in controlling both type I and type II error (22). Second, we considered a measure referred to as "entropy" that is the average highest predicted probability of class membership (20). This measure ranges from 0 to 1, with lower entropy indicating little confidence that individuals belong in the class with their highest assigned probability, whereas an entropy of 1 would indicate certainty that individuals belong in their assigned class. Third, if one or more class sizes were too small to be of any public health relevance, the number of classes was reduced. Finally, substantive knowledge was used to establish the appropriate number of classes (19). There should be a correspondence between the established classes and some practical interpretation of what the classes indicate. As Muthen (the author of the MPLUS statistical software) concludes, "Substantive theory, auxiliary information, and practical usefulness will continue to have to guide the statistical analysis" (19).

Specifying Variance Estimates

In a completely unrestricted model, LCA will estimate a separate variance-covariance matrix for each class. If the number of latent classes or the number of variables is large, so are the parameters to be estimated. Thus, a typical strategy is to impose restrictions, such as constraining covariances to zero or constraining classes to have the same variance-covariance matrix (i.e., Σg(θ) = Σ(θ)). Constrained models allow for more parsimonious and stable results (11).

Because the mean minutes of physical activity for the lowest activity class had significantly lower variances than in the more active classes, a model that allowed this lowest activity class to have variances that differed from all of the other activity classes was selected. For all but the lowest activity class, we allowed weekend and weekday variances to differ, but we constrained them to be equal across classes. In this way, we tried to create a parsimonious, stable model that still captured some of the complexity of the substantive issues of the analysis.

The LCA was performed using MPLUS. The modeling was conducted by requesting a range from 1 to 6 classes a priori as the number of group memberships to predict. Beyond six classes, the sample size of the more active classes became very small, and the activity patterns during the 7 d became highly unstable. Whereas MPLUS allows for complex survey sampling in conjunction with LCA modeling, the software does not currently account for survey sampling when computing the BLRT statistic. As such, this analysis, designed to establish an appropriate number of activity classes, was performed without sample weights and cluster sampling.

Structural Equation Modeling Perspective

Figure 1 provides a visual representation of the LCA model using the graphic presentation typical of structural equation modeling. In this perspective, the latent classes (represented by circles) are derived from the patterns of physical activity across the 7 d of the week (squares indicate observed variables). Simultaneously, the sociodemographic characteristics are used to help predict who falls into each of the derived activity classes by providing more refined prior probabilities based on the distribution of the sociodemographic characteristics within each class. The Bayesian posterior class membership probabilities are then based on these prior probabilities. Because the reporting of the direct effects of the sociodemographics on class membership is complicated and beyond the scope of the current analyses, these results will be presented in future work.

Structural equation model for the prediction of the latent classes based on the minutes of physical activity across the 7 d of the week, with sociodemographic characteristics predicting class membership.

This research was approved by the public health institutional review board of the University of North Carolina at Chapel Hill.


A total of 10,122 participants completed the 2003-2004 NHANES physical examination. We excluded those under age 20, to more clearly focus this research on an adult population. Of the remaining 5041 participants, 4252 provided data for the accelerometer portion of the survey. Finally, an additional 450 participants were excluded because of missing covariate data (e.g., education, household income) or because they had no days of valid accelerometer data with which to impute the other, missing days, leaving a final population of 3802 participants. Because imputation was not performed on the assessment of bout minutes of MVPA, only the 3462 participants who provided at least three valid days of data were included in the analysis of bout minutes.

The sociodemographic characteristics of the final sample and those excluded because of missing data are presented in Table 2. Statistically significant (P < 0.05) differences were found between the categorical distributions of age, education, the poverty index, and race/ethnicity. Whereas the chi-square tests for the categorized age and poverty index ratio were significant, the t-tests for the continuous age (P value = 0.82) and poverty index ratio (P value = 0.21) were not. For age, this seems to be attributable to the missing data being overrepresented by the younger and older participants in the various age categories, leading to a similar mean age (50.8 yr old for those in the final sample vs 51.0 yr old for those not included in the final sample). Similarly, for the poverty index, the poorest and richest were least likely to provide complete data, again leading to a similar mean (2.6 for those in the final sample vs 2.5 for those not included in the final sample).

Sociodemographic characteristics of the overall study population, analytic sample, and the population excluded because of missing data, among those 20 yr old and older.

The weighted median minutes of MVPA/VPA and bout minutes of MVPA are presented in Table 3 by day of the week. The median number of minutes of overall MVPA ranged from 17.0 on Monday and Tuesday to 12.0 on Sunday, whereas the median minutes for overall VPA were zero for all days but Tuesday and Thursday. Bout minutes of MVPA had a median of 2.0 for all days. The median values were substantially lower than the mean for all types of physical activity and for all days of the week, suggesting nonnormally distributed data, but the means and standard deviations are presented for comparison with other published data.

Sample weighted mean, standard deviation, standard error, 25th-, 50th-, and 75th-percentile for minutes of moderate to vigorous physical activity (MVPA), bout minutes of MVPA, and vigorous physical activity (VPA), by day of the week.

The analyses of the class memberships for the overall minutes of MVPA produced a BLRT test statistic that detected a statistically significant improvement in fitness at the < 0.0001 level for all numbers of classes, ranging from 1 to 6. This indicates that separating the population into six classes was justified on the basis on these criteria. Whereas entropy decreased as the number of classes increased, it was still high (0.94) for five and six classes. Because the six-class analysis resulted in very small active classes, we settled on a more parsimonious, five-class model for presentation.

Figure 2 shows the plot of the mean minutes of total accumulated MVPA for five classes. The largest sample-weighted percentage of the population fell into the lowest two classes-roughly 79% of the total population. These two lowest classes encompassed activity averaging less than 25 min·d−1 of MVPA. The highest activity class, with a mean of 134 min of MVPA per day, only constituted 0.9% of the population. This class also demonstrated a high level of activity Monday through Friday, with less activity on the weekend, with a particularly pronounced decrease on Sunday. All classes demonstrated this dip on weekends compared with weekends, to varying degrees.

FIGURE 2-Five:
latent classes: MVPA. Class proportions have been weighted to account for the complex sample design.

When participants were required to accumulate their minutes in bouts, the BLRT test justified as many as six classes at the < 0.0001 level; however, the two most active classes were very sparsely populated (0.4% and 0.5% of the population). Given this, we settled on five classes, which also allowed for a direct comparison with the overall minutes of MVPA. Figure 3 shows the class means for the five-class results for the bout minutes of MVPA. The two most active classes represented 0.6% and 4.0% of the sample-weighted population, just slightly less than the two most active overall MVPA classes, which had 0.9% and 4.5% of the population, respectively. Roughly 93.5% of the entire population is now classified into the lowest two groups. In addition, for the bout analysis, the second-least-active group now has a mean of 10.3 bout minutes of MVPA per day, whereas the second-least-active group in the overall MVPA analysis had a mean of 21.0 min across all 7 d. Thus, when only bout minutes are counted, a much larger percentage of the population was classified into the lowest groups, and these groups are less active.

FIGURE 3-Five:
latent classes: bout minutes of MVPA. Class proportions have been weighted to account for the complex sample design.

Even though the number of minutes accumulated in bouts tends to shift the classes into lower levels of mean activity level, the general patterns are very similar to the overall MVPA analysis, with one exception. A class emerged with moderate levels of physical activity Monday through Friday, but with a much higher level of activity on the weekend, particularly on Sunday. This class, representing 1.8% of the population, will be referred to as the "weekend warrior" class. The six-class analysis of the overall minutes of MVPA (data not shown) also demonstrated this weekend warrior class, although with a smaller percentage of the population (0.9%).

The VPA analysis did not produce stable results (very large and very small class sizes), because of the very low number of participants accumulating any VPA. For example, both the three- and four-class VPA analyses produced a most active class with only two and one participants, respectively-a trivial class assignment from both a substantive and analytical perspective.


This modeling represents the first time that objectively measured physical activity data have been analyzed using LCA. Whereas a large portion of the population presented little activity, a weekend warrior class did emerge, as did a highly active class with less activity on the weekend. We are unsure whether these activity patterns were driven by specific types of activity, such as work-related activities.

In the analysis of the overall minutes of MVPA, the statistical significance of the BLRT statistic indicated that six classes were justified on the basis of the data. However, class size for the more active groups displayed a small sample size, and, as such, the reduced model with only five classes provided a more parsimonious model.

An inactive class emerged in overall MVPA that represented nearly 34% of the entire population. This class averaged 5.3 min of MVPA per day. The second-least-active class, with a mean of 21.0 min of MVPA per day, also represents a class of which many, if not most, would not have accumulated 30 min of MPVA on most days of the week. Together, these two groups represent a very large proportion of the population with PA levels significantly below the recommended levels (23). Determining the sociodemographic and behavioral characteristics of these groups, to target appropriate physical activity interventions, could lead to significant improvements in activity levels in the United States and, thus, in the overall health of the nation.

The analysis of the bout minutes of MVPA produced patterns similar to those found in the analysis of the overall minutes of MVPA, with the important exception of the weekend warrior class. From Monday through Friday, this class had significantly less activity than the class above it (e.g., the class representing 4.0% of the population), but because of the significant increase in physical activity of the weekend warriors on Saturday and Sunday, and the decrease in the other group during this period, their means were relatively similar. The weekend warriors accumulated a daily mean of 31.5 min of MVPA, whereas the more active group was only slightly more active, averaging 37.0 min. The health outcomes of these two groups would be interesting to compare, because they represent two populations that both achieved the recommended minimum total accumulation of MVPA, accumulated their minutes in bouts, and have similar mean minutes of MVPA, although they accumulated their MVPA minutes in very different manners.

A recent article using self-report data from NHANES and the BRFSS reports a prevalence of the weekend warrior pattern in the U.S. population-defined as accumulating ≥ 150 min of MVPA on 1 or 2 d in a week-of approximately 3% and 1%, respectively (13). Our results, although based on objectively measured accelerometer data, found that a similar proportion of the population (1.8%) could be classified as weekend warriors according to bout minutes of MVPA.

Another important difference between bout minutes of MVPA and the analysis of overall MVPA minutes is that the more active groups in the bout-minutes analysis had significantly fewer participants than in the overall MVPA analysis. In fact, the two most active classes in the six-class analysis of bout minutes of MVPA (data not shown) only contained 0.4% and 0.5% of the population. The utility of class assignments with such small populations is an important consideration in mixture modeling, especially if the analyses include associations between class assignment and any outcomes. If the six class levels of bout minutes of MVPA were used to simultaneously model health outcomes, the two most active classes would likely not be large enough to generate proper associations with the health outcomes, particularly if the outcomes were rare.

This challenge was made most explicit in the analysis of the VPA. Only 1.4% of all days had 10 min or more of VPA, and in 91.1% of all days, participants accumulated less than 1min of VPA (data not shown). Because of the highly skewed data, the model assumptions generally failed to produce results. When classes were successfully modeled, their sizes were too small to serve any useful analytic purposes.

These low levels of VPA are troubling when considered in light of the Healthy People 2010 goals of increasing to 30% "the proportion of the adults who engage in vigorous physical activity that promotes the development and maintenance of cardiorespiratory fitness three or more days per week for 20 or more minutes per occasion" (25). Not only does this analysis reflect a much lower percentage of the adult population achieving this goal than desired, it also reflects a much lower percentage than the 27.4% of adults reported to have met this goal in 2005 by the BRFSS (5). In fact, only 23 participants registered 20 min of VPA on at least 3 d·wk−1, representing only 0.9% of the sample-weighted total population. Because the BRFSS assessment was based on self-report, this discrepancy could reflect a large amount of overreporting present in the BRFSS levels. It could also reflect that the cut point for VPA was too high, thereby missing many minutes of activity in our population that should have been classified as VPA.

Although we chose LCA as our modeling strategy, other possibilities were also considered for the method of modeling classes. Latent class growth analysis (LCGA) and growth mixture modeling (GMM) have recently found wide application in the social sciences as an effective method for modeling growth trajectories with longitudinal data (8,10). It was determined, however, that the 7 d of PA, although contiguous in time, did not constitute a longitudinal analysis analogous to, for example, the development of childhood obesity according to age in adolescence. This decision was based, in part, on the desire to impose no restrictions on the shape of the classes across 7 d. In other words, we did not want to force the physical activity classes to follow a "growth" pattern. With this conceptualization, physical activity is more analogous to multiple continuous measurements of academic ability such as math, verbal, spatial, etc., and, as such, it should be modeled with LCA.

One question that remains from the LCA is whether any relevant undetected patterns exist that may have been missed by this analysis. For example, although many people work Monday through Friday, others do not; therefore, their "weekend" may fall on days other than Saturday and Sunday. Other missed groups may include active workers who work weekends but do not work some days during the normal work week, or those who participate in leisure-time activity every other day. Future analysis may choose to order days differently, such as by least to most activity, instead of by day of the week, leading to a class analysis that would still capture overall activity but would better assess the regularity of activity. Depending on the purpose of the analysis, this or other strategies may provide preferred results.

Several limitations of this work are worth noting. The analysis population had a sociodemographic distribution that differed from those who were excluded because of nonresponse or inappropriate data, which, in turn, could have affected the overall patterns found. Another weakness was that, because of the small number of participants registering sufficient levels of VPA, this part of the analysis was unsuccessful. This would have been an interesting group to assess, and, possibly, with different analytical techniques or a larger study population, such an analysis will be possible in the future. Accelerometers do not capture all types of physical activity, particularly static activities such as raking leaves or riding a bike (17). Therefore, although the current analyses do not rely on self-report, they may still not reflect the true activity levels of the U.S. population. Similarly, the cut points for what constitutes MVPA and VPA are also sensitive to the types of activities being done. Changing these cut points would affect the amount of physical activity that participants were credited for, which, in turn, could have affected the outcomes of the latent class analysis. An analytical weakness worth noting was that the BLRT test statistic was not available with the simultaneous use of the NHANES sample weights, and, therefore, the sample weights were not used for this analysis. Adding the sample weights could have affected the final decision for the number of classes as well as the pattern that these classes assumed.

These data represent objectively measured physical activity from a large, nationally representative sample of the U.S. population. Physical activity was assessed in several ways, to analyze the class membership patterns for different patterns of physical activity. Our results indicate that a very large portion of the U.S. population may be classified into patterns of physical activity that represent low levels of MVPA throughout the week. The levels of VPA were most surprising, indicating that fewer than 1% of the population engaged in VPA for at least 20 min on three or more days per week. In addition, a weekend warrior class emerged for approximately 2% of the population. The LCA analysis provides a novel approach for assessing patterns of objectively measured physical activity in epidemiologic studies.

This research was supported by the National Institutes of Health Grant NIDDK 56350 award to the Department of Nutrition's Clinical Nutrition Research Center.


1. Berrigan D, Dodd K, Troiano RP, Reeve BB, Ballard-Barbash R. Physical activity and acculturation among adult Hispanics in the United States. Res Q Exerc Sport. 2006;77:147-57.
2. Bollen KA, Curran PJ. Latent Curve Models: A Structural Equation Perspective. Hoboken (NJ): Wiley-Interscience; 2006. p. 188-228.
3. Brage S, Wedderkopp N, Franks PW, Andersen LB, Froberg K. Reexamination of validity and reliability of the CSA monitor in walking and running. Med Sci Sports Exerc. 2003;35(8):1447-54.
4. Catellier DJ, Hannan PJ, Murray DM, et al. Imputation of missing data when measuring physical activity by accelerometry. Med Sci Sports Exerc. 2005;37(11 suppl):S555-62.
5. Centers for Disease Control and Prevention. Behavioral Risk Factor Surveillance System Survey data. Atlanta (GA): U.S. Department of Health and Human Services, Centers for Disease Control and Prevention; 2005. Available at:
6. Chowdhury PP, Balluz L, Murphy W, et al. Surveillance of certain health behaviors among states and selected local areas-United States, 2005. MMWR Surveill Summ. 2007;56:1-160.
7. Dolan CV, Jansen BRJ, van der Maas HLJ. Constrained and unconstrained multivariate normal finite mixture modeling of Piagetian data. Multivariate Behav Res. 2004;39:69-98.
8. Croudace TJ, Jarvelin MR, Wadsworth ME, Jones PB. Developmental typology of trajectories to nighttime bladder control: epidemiologic application of longitudinal latent class analysis. Am J Epidemiol. 2003;157:834-42.
9. Freedson PS, Melanson E, Sirard J. Calibration of the Computer Science and Applications, Inc. accelerometer. Med Sci Sports Exerc. 1998;30(5):777-81.
10. Greenbaum PE, Del Boca FK, Darkes J, Wang CP, Goldman MS. Variation in the drinking trajectories of freshmen college students. J Consult Clin Psychol. 2005;73:229-38.
11. Hagenaars JA, McCutcheon AL. Applied Latent Class Analysis. Cambridge (MA); New York (NY): Cambridge University Press; 2002.
12. Hughes E, McCracken M, Roberts H, et al. Surveillance for certain health behaviors among states and selected local areas-behavioral risk factor surveillance system, United States, 2004. MMWR Surveill Summ. 2006;55:1-124.
13. Kruger J, Ham SA, Kohl HW 3rd. Characteristics of a "weekend warrior": results from two national surveys. Med Sci Sports Exerc. 2007;39(5):796-800.
14. Lee IM, Sesso HD, Oguma Y, Paffenbarger RS Jr. The "weekend warrior" and risk of mortality. Am J Epidemiol. 2004;160:636-41.
15. Leenders NY, Nelson TE, Sherman WM. Ability of different physical activity monitors to detect movement during treadmill walking. Int J Sports Med. 2003;24:43-50.
16. Lo Y, Mendell NR, Rubin DB. Testing the number of components in a normal mixture. Biometrika 2001;88:767-78.
17. Matthew CE. Calibration of accelerometer output for adults. Med Sci Sports Exerc. 2005;37(11 suppl):S512-22.
18. Metcalf BS, Curnow JS, Evans C, Voss LD, Wilkin TJ. Technical reliability of the CSA activity monitor: the EarlyBird study. Med Sci Sports Exerc. 2002;34(9):1533-7.
19. Muthen B. Statistical and substantive checking in growth mixture modeling: comment on Bauer and Curran (2003). Psychol Methods. 2003;8:369-77.
20. Muthen BO. MPlus Technical Appendices. Los Angeles (CA): Muthen & Muthen; 1998-2004.
21. National Center for Health Statistics. Accounting for Item Nonresponse Bias. Hyattsville (MD): National Center for Health Statistics (U.S.); 1994.
22. Nylund KL, Asparouhov T, Muthén BO. Deciding on the number of classes in latent class analysis and growth mixture modeling: a Monte Carlo simulation study. Struct Equation Model. 2007;14(4):535-69.
23. Pate RR, Pratt M, Blair SN, et al. Physical activity and public health. A recommendation from the Centers for Disease Control and Prevention and the American College of Sports Medicine. JAMA. 1995;273:402-7.
24. Sallis JF, Saelens BE. Assessment of physical activity by self-report: status, limitations, and future directions. Res Q Exerc Sport. 2000;71:S1-14.
25. U.S. Department of Health and Human Services. Healthy People 2010. Understanding and Improving Health and Objectives for Improving Health. 2 vols. Washington, DC: U.S. Government Printing Office; November 2000. p. 22-11. Available at:
26. Wang G, Pratt M, Macera CA, Zheng ZJ, Heath G. Physical activity, cardiovascular disease, and medical expenditures in U.S. adults. Ann Behav Med. 2004;28:88-94.
27. Yngve A, Nilsson A, Sjostrom M, Ekelund U. Effect of monitor placement and of activity setting on the MTI accelerometer output. Med Sci Sports Exerc. 2003;35(2):320-6.


©2008The American College of Sports Medicine