Accelerometers are recognized as a valid and objective tool for assessing free-living physical activity (50), although they are limited in their ability to accurately detect stationary activities (e.g., weight training and cycling) and, due to lack of waterproofing, some cannot measure water-based activities. Despite their limitations, accelerometers have been used extensively in field settings to monitor activity patterns. Technological advances have made these devices easier to use in field settings. Now, accelerometers are small, some can record both activity and step counts, most can record and store activity for extended periods (more than 7 d), and waterproof devices are beginning to emerge. In the past two decades, numerous studies have assessed the validity and reliability of accelerometers (27,82). Some validation studies have focused on interpreting the counts by providing cut points that correspond to light-, moderate-, and vigorous-intensity physical activity (27,57), which typically is achieved by providing a MET, where 1 MET is the amount of energy expended at rest. Although numerous validation studies have been published, standards for accelerometer data reduction have not been established. Consensus on and clarification of a number of methodological issues are needed before standardized methods for accelerometer data reduction can be recommended. This paper focuses on five such issues (Table 1) that are important to consider when processing accelerometer data, including identifying the wearing period in a day, identifying minimal wear requirement for a valid day, identifying spurious data, computing outcome variables and aggregating days of data, and extracting bouts of moderate to vigorous physical activity (MVPA). Varying how activity counts are interpreted (i.e., cut point used to determine which counts correspond to light, moderate, and vigorous physical activity) can affect the results significantly; however, this issue is not addressed in this paper. Clarifying these issues and standardizing accelerometer data reduction would greatly improve the ability to compare results across studies.
To determine how researchers have addressed the issues listed in Table 1, we reviewed the methods sections of field accelerometer studies published in 2003 and 2004. The studies were identified primarily through a keyword search (“accelerometer”) in PubMed and by reviewing the reference sections of published articles (N = 198). Laboratory studies and nonphysical activity studies were excluded from this review (N = 134). This review was not meant to be exhaustive but to report the criteria researchers have used in a select sample of accelerometer studies. Of the 64 studies identified in the search, 30 measured activity patterns in children (6–8, 10, 11, 15, 17, 19, 24, 25, 29, 34–36, 39, 49, 51, 52, 55, 56, 59, 63, 66, 69, 73, 74, 76–78, 81) and 34 measured patterns among adults (1–3, 9, 13, 14, 18, 20–22, 30, 32, 37, 38, 41–47, 54, 58, 60–62, 64, 65, 68, 70, 71, 80, 83, 84). Children and adult studies were combined in the synthesis when no group differences were identified. Most of these studies (73.4%) asked participants to wear the accelerometer only during waking hours, only 26.6% asked the participants to wear the accelerometer continuously. In 62 of the 64 studies (96.9%), investigators reported the number of days participants were asked to wear the accelerometer. This period ranged from 1 to 14 d. In total, 65.5% of the children studies asked participants to wear the accelerometer for fewer than 7 d, in contrast to 27.3% of the adult studies. Only eight studies (12.5%) reported how they estimated interruption in wearing time and thus how they determined the wearing period (Table 1, issue 1). These studies used 10 min (10,11,25,59), 15 min (62), 20 min (73,74), and 30 min (19) of continuous zero counts to estimate interruptions allowing the amount of time the accelerometer was not worn to be estimated. These cut points may significantly impact the minimal wear requirement in order for a day to be included in the analysis (issue 2). Twenty-one studies (32.8%) reported the criteria they used to determine whether the accelerometer was worn a significant proportion of waking time. The studies used cut points of 1 (52), 4 (61), 5 (65), 6 (39), 8 (15,21,55), 9 (19), 10 (10,11,20,24,25,59,62,81), 11.20 (49), 12 (17,22), 15.60 (29), and 16.67 (73) h·d−1. Of these, 38.1% of the studies reported using 10 h as their cut point to determine whether a given day was included in the analyses. The study that used 1 h as the cut point examined physical activity patterns in preschool children. Identifying spurious data (issue 3) was mentioned only in one study, where activity counts greater than 9 SD away from the median served to identify spurious data (11). In total, 29 of the studies (45.3%) reported the minimum number of days needed for the analyses (issue 4). Given that seven studies monitored only 1 d, we inferred the number of days used in the analyses as 1 (8,18,29,45,77,78), the other studies used 2 (14,54), 3 (6,10,11,19,21,39,49,52,55,59,63,69,73,81), 4 (17,61,71), 5 (20,22), 6 (32), and 8 (25) d. Three days of monitoring was the criterion used most often (53.6%). Seven studies required that a certain number of weekend days and weekdays be included to compute an outcome variable (14,17,20,25,32,39,49). Finally, only two studies extracted bouts of MVPA (issue 5) (19,84).
To summarize, this select review of the literature revealed that few studies reported how they handled some of the key issues listed in Table 1. Researchers often did not report the decision rules they used to process the accelerometer data. This review also suggested that the criteria for accelerometer data reduction vary considerably. The impact of using different criteria for data reduction, in terms of outcome variables, is unclear. To determine the impact of using various assumptions and criteria on the interpretation of the results, we analyzed a data set using four different data reduction algorithms that have been used in the published literature. The first purpose of this paper was to quantitatively assess the impact of using various decision rules regarding the minimal wear requirement for a valid day and interruption in wearing time on the estimation of a set of commonly used outcome variables (average counts per minute, counts per day, minutes of MVPA per day, and prevalence of meeting physical activity recommendations). The second purpose of this paper was to identify specific issues of greatest importance for accelerometer data reduction.
Data collected as part of the Women on the Move (WOTM) study were reanalyzed using the four selected algorithms. The WOTM study was a 5-yr study aimed at validating physical activity questionnaires. Study eligibility criteria included a) age 40–70 yr, b) self-reported race/ethnicity as African American or Hispanic, c) no health conditions that preclude being active, d) not pregnant, e) not planning to move from the greater metropolitan area of Houston, and f) literate in English or Spanish. Study participants were recruited through the media (print, television, and radio), community presentations, and posted flyers. In total, 656 women expressed an interest in participating in the validation study. Of these, 260 women enrolled in the study with 242 having accelerometer data used in the present analysis. Escobar-Chaves et al. (26) provide an in-depth description of the recruitment and retention protocol.
The institutional review board for the protection of human subjects at the University of Texas-Houston, Health Science Center, and the U.S. Centers for Disease Control and Prevention reviewed and approved the original WOTM study protocol. This study analyzed data that were collected as part of the 2-wk validation protocol. To verify eligibility criteria, participants initially were screened by telephone followed by an in-person meeting at which participants signed an informed consent and provided demographic information. During the second week of the validation protocol, participants were instructed to wear an accelerometer for a full week during all waking hours, excluding times during which they were exposed to water (e.g., bathing and swimming). Participants wore the accelerometer under their clothes and fitted over the right hip. Participants were instructed to keep a physical activity diary while wearing the accelerometer.
The ActiGraph accelerometer model 7164 WAM, formerly known as the Computer Science and Applications (CSA) and Manufacturing Technology Inc. (MTI) (Actigraph, LLC, Fort Walton Beach, FL), served to collect activity counts. This small uniaxial accelerometer (dimensions: 5.1 × 4.9 × 1.6 cm; wt: 39.8 g) is designed to measure human motion and to suppress high accelerations (i.e., those higher than 2.0 × g, where g is equal to 9.825 m·s−2 the constant of gravitation) to eliminate nonhuman motion. All units recorded activity counts in 30-s epochs. The validity of the ActiGraph has been evaluated in a number of studies. ActiGraph counts from treadmill walking and running (12,27) and under field conditions (12,33) have correlated well with measured oxygen uptake, although ActiGraph counts have been shown to underestimate oxygen uptake for running speeds of 9 km·h−1 or greater (12), and for weight training, stair climbing, household tasks, and yard work (67). The ActiGraph has demonstrated good reliability for most research applications. In a recent comparison of four accelerometers assessing standardized bouts of treadmill walking across multiple trials, the ActiGraph was found to have the least variability across accelerometers and trials and the highest overall reliability (82).
Physical activity diary.
All participants completed a 7-d diary. They were asked to record activities that lasted at least 10 min in duration. The following was recorded in the diary: a) activity type (occupation, sport and exercise, walking, house/yard, inactivity, personal care, free time/entertainment, transportation, or miscellaneous); b) brief description; c) estimated intensity (low, medium, or high); d) position (lying/resting, sitting, standing, or moving around); e) time spent doing the activity (minutes); f) if appropriate, walking pace (moving about, slow pace, medium pace, brisk walking); and g) if appropriate, carrying/lifting a load as well as the estimated weight. For each activity reported in the diary, participants indicated whether they were wearing the accelerometer. Each day, participants were asked to report waking time, sleep time, time they put on the accelerometer, time they took off the accelerometer, and whether the accelerometer was removed during waking hours. Eason et al. (23) provide an in-depth description of this diary. The analyses used the sleep-time information recorded in this 7-d diary to calculate number of waking hours.
Accelerometer data reduction.
We applied four data reduction algorithms to data from the accelerometer and compared the results. The algorithms were developed independently and have been used in the published literature for accelerometer data processing. This paper did not assess the impact of using different intensity cut points (i.e., calibration equations) for the determination of minutes accumulated in varying intensity levels. Therefore, for the purposes of standardization and simplicity, the cut point of ≥1952 counts per minute was used to quantify time spent in MVPA (27). The cut point was halved to accommodate the 30-s epoch used in the WOTM study. In addition, extraction of MVPA bouts was standardized at a minimum of 10-min duration. Processing differed by algorithms, algorithm 1 processed the raw data file in FORTRAN and computed the final outcome variables in SPSS; algorithms 2 and 4 used SAS to read downloaded accelerometer data and compute all outcome variables; and algorithm 3 processed the accelerometer data files in QuickBasic and computed outcome variables in SAS.
Table 2 summarizes the data-processing assumptions used in each algorithm. Algorithms 2, 3, and 4 used 20 min of consecutive zero counts to identify the nonwear time in a given day, and algorithm 1 used 60 min of consecutive zero counts. Algorithm 2 was the only program that initially partitioned the data into 20-min blocks (i.e., scanning the data in 20-min blocks) and used blocks of zeros to correspond to nonwearing time. Algorithms 3 and 4 used a rolling window (i.e., scanning each 30-s epoch and begin a count each time a zero count is encountered) approach to detect 20 min or more of consecutive zeros. The criteria used to identify the minimal wear requirement for a valid day differed between algorithms. Algorithm 1 defined a valid day as having accelerometer data for 60% of the awake time. Computing the number of valid days consisted of first estimating the number of hours the participant was awake in a day, using the diary data. A day was valid if awake time minus total minutes of nonwear time (i.e., 60 min or more of continuous zeros) divided by awake time multiplied by 100 was greater than 60%. For a participant who slept about 8 h·d−1, this corresponded to 16 h of awake time and 60% of this corresponded to a little less than 10 h·d−1 of wear. Algorithms 2, 3, and 4 estimated wearing time from the accelerometer recording, as they were not provided the dairy information. Days on which participant did not wear the accelerometer for at least 12 h·d−1 for algorithm 2 and 10 h·d−1 for algorithm 3 were eliminated from the analyses (wearing time did not need to be consecutive time). For algorithm 4, if the proportion of nonmissing data in a time period was less than 80% of a standard measurement day, defined as the length of time during which 70% of sample participants were wearing the accelerometer, then the data for that day were not considered valid. Algorithm 4 imputed days of data if a participant had at least one valid day of data (51). As shown in Table 2, algorithms 1 and 2 used different cut point to identify spurious data. Algorithm 4 employed a different strategy by searching for accelerometer malfunction. Accelerometer malfunction was identified as having 10 consecutive minutes of constant counts greater than zero. All spurious data were set to missing values. The minimum number of days required to estimate average (per day) values for the outcome variables differed by algorithm. Algorithms 1 and 3 required four valid days of data, whereas algorithm 2 required a minimum of three valid days. Algorithm 4, the only imputation data reduction algorithm, based the outcome variables on 7 d of data for any participant who had at least one valid day of data, as defined above. Computing the proportion of participants who met the ACSM/CDC physical activity recommendation (i.e., accumulating 30 min of MVPA in at least 8- to 10-min bouts on most days of the week (53), defined in this paper as 5 d·wk−1), was based on a minimum of 4 d of data for algorithm 1 and 5 d of data for algorithms 2 and 3, respectively. Participants who were active 70% of the minimum days were assumed to meet the physical activity recommendation. This assumption likely overestimates the number of participants classified as meeting the recommendation. The minimum days required to assess whether participants met the activity recommendation was unavailable for algorithm 4 as was information about interruption allowance. Three algorithms extracted bouts of MVPA that lasted at least 10-min in duration. Algorithms 1 and 2 allowed an interruption (1 or 2 min) anytime in the bout, whereas algorithm 3 extracted bouts with and without an interruption.
The following variables were compared across algorithms: number of participants with valid days, average wearing time, average counts per minute, average counts per day, average minutes of MVPA per day (i.e., ≥1952 counts per minute), average minutes of MVPA bouts per day (MVPA bouts), and proportion of participants meeting the ACSM/CDC moderate-intensity physical activity recommendation. Univariate ANOVA served to determine significant differences across algorithms. For these analyses, the alpha level was set a priori at 0.05. Post hoc Tukey analyses followed all significant ANOVA. A repeated ANOVA was not employed because listwise deletion would have eliminated some of the differences among the algorithms. Finally, χ2 tests of differences were used with categorical data to assess differences among algorithms.
Number of valid days.
The number of participants with at least one valid day was similar for each algorithm (240 for algorithms 1, 2, and 4, and 241 for algorithm 3) as shown in Table 3. Although the frequency distribution for valid days was shifted downward for algorithm 2, there were no significant differences in the frequency of valid days by algorithm (χ2 (df = 12) = 17.19, P = 0.143).
Table 4 contrasts the set of commonly reported outcome variables across the different algorithms. Wearing time ranged from 779 to 960 min·d−1, and it differed significantly across algorithms (F (df = 3) = 97.24, P = 0.000). Post hoc Tukey revealed that all pairwise comparisons were significant (P < 0.05). Average activity counts per minute varied from 285 to 309 counts per minute, and differed significantly across algorithms (F (df = 3) = 2.90, P = 0.034). Post hoc Tukey found that only the activity counts per minute of algorithm 2 were significantly lower than those of algorithm 4 (P < 0.05). Similarly, the average activity counts per day differed significantly across algorithms (F (df = 3) = 3.14, P = 0.025), with the activity counts per day of algorithm 2 being significantly lower than those of algorithm 3 (P < 0.05). The average minutes of MVPA per day ranged from 17 to about 23 min·d−1. The ANOVA results indicated that differences existed among the algorithms (F (df = 3) = 8.99, P = 0.000). Post hoc Tukey found that algorithm 2 yielded significantly fewer minutes of MVPA than all other algorithms (P < 0.05). Finally, the average minutes of MVPA per day accumulated in bouts significantly across algorithms (F (df = 2) = 3.75, P = 0.024), with algorithm 3 having significantly fewer minutes of MVPA bouts than algorithm 2 (P < 0.05). Modifying algorithm 3 to allow for a 2-min interruption anywhere in the bout eliminated this difference between algorithms (P > 0.05).
Table 5 shows the proportion of participants who met the moderate-intensity physical activity recommendation as estimated by each algorithm. The results show that all the algorithms estimated that a very low prevalence of participants would have met the recommendation. Results differed across the algorithms; however, when the percentage of participants who did not meet the recommendation was considered (χ2 (df = 6) = 47.76, P = 0.000). The difference resulted because the percentage of participants classified as not having enough data differed by algorithm, with algorithms 1 and 4 having fewer missing participants than algorithms 2 and 3.
To further explore the impact of using different criteria on the number of participants included in the analyses (sample size) and its impact on outcome variables, a sensitivity analysis was conducted. To eliminate variations, all post hoc exploratory analyses were conducted with algorithm 2. These analyses replicated the conditions used in algorithms 1, 2, and 3. The 20/10 condition corresponds to 20 min of continuous zero counts and 10 h·d−1 of wear (algorithm 3), the 20/12 condition corresponds to 20 min of continuous zero counts and 12 h·d−1 of wear (algorithm 2), and the 60/10 condition corresponds to 60 min of continuous zero counts and 10 h·d−1 of wear (algorithm 1). These three conditions (20/10, 20/12, and 60/10) were run with a minimum of three, four, five, six, and seven valid days of data for a total of 15 conditions.
Figure 1 presents the impact of varying these criteria on sample size. As expected, sample size decreased as the minimum number of valid days required for inclusion in the analyses increased. Varying the length of accelerometer inactivity assumed to be associated with accelerometer removal and the minimum wearing time for determining a valid day both had an impact on sample size. Wearing time appeared to have a greater impact on sample size retained for analysis than did length of accelerometer inactivity (believed to reflect accelerometer removal).
Figure 2 shows the impact of these conditions on average counts per minute. Varying the wearing interruption (20 vs 60 min) had the greatest impact on the counts per minute. Although lengthening the minimum wearing time (10 vs 12 h·d−1) per valid day increased the counts per minute, the observed changes in the outcome variables were minimal compared to the reduction in counts observed when the length of wearing interruption was increased.
The impact of using different criteria on the average number of inactive minutes and minutes of light activity and MVPA per day is shown in Figure 3. The least restrictive 60/10 condition had more minutes of inactivity than the other conditions, indicating that varying the length of zeros associated with accelerometer removal had the greatest impact on minutes of inactivity. For all conditions, there is a slight upward trend with greater number of valid days required, indicating that using more stringent inclusion criteria resulted in an upward trend in minutes of inactivity. A similar pattern was observed for the average minutes of light activity and MVPA. Minutes of light activity and MVPA increased slightly as the valid day requirement criteria became more stringent.
In the introduction, we reviewed the decision rules researchers use for accelerometer data reduction. Given the lack of agreement and variability observed in the literature, one of the purposes of this study was to assess quantitatively the impact of using different decision rules in the data reduction process. Analyzing the same data set with four different algorithms provided some insights as to this impact. Overall, it appears that using different decision rules affected several important outcome variables of physical activity. More stringent inclusion criteria significantly influenced wearing time, and average activity counts per minute, activity counts per day, and minutes of MVPA per day. This was seen with algorithm 2, which had the most stringent inclusion rules. Algorithm 2 resulted in a lower wearing time, had the lowest activity counts per minute, counts per day, and had fewer minutes of MVPA per day. The magnitude of the differences observed with algorithm 2 resulted in meaningful differences, for example, the average MVPA for algorithm 2 was 35 min·wk−1 lower than the other algorithms. Algorithm 2 always was implicated in the significant differences among algorithms, but these effects were not consistent: outcome variables were affected differently by different algorithms (e.g., counts per minute under algorithm 2 differed significantly from those obtained with algorithm 4 and the counts per day significantly differed from those obtained with algorithm 3). The algorithm comparisons also showed that MVPA bouts did not differ significantly when all algorithms allowed a 1- or 2-min interruption anywhere in the bout; however, when no interruption was allowed, the number of minutes of MVPA bouts decreased significantly. Given that a small interruption is plausible (i.e., stop at light while brisk walking or jogging or take a break to drink water), it is suggested that a 1- or 2-min interruption should be included when classifying MVPA bouts. The proportion of participants classified as meeting the moderate-intensity physical activity recommendation did not differ across algorithms. The fact that this adult population was quite inactive may have biased these results toward a null finding. Finally, algorithm 4 (the only imputation-based data processing algorithm) yielded findings similar to algorithms 1 and 3, except for the wearing time variable, which differed significantly across all algorithms. At the group level, the results were comparable across all three algorithms. The slight increase in sample size observed with algorithm 4 suggests an increase in power when the accelerometer data are regressed on other extraneous variables. All these effects have an impact on the validity of the data; however, these results are limited clearly in their ability to identify which algorithm approximates the true distribution of the data and hence which decision rules are most appropriate.
This study examined the impact of varying certain decision rules. Varying other decision rules, such as cut points for sedentary, light, moderate, and vigorous activity, may further exacerbate the differences observed. The algorithms used in this study tested a more narrow range of decision rules than was observed in the literature, which suggests that the impact on the outcome variables may be even more pronounced, given the variability observed in the literature. The analyses tested a range the authors felt comfortable in using based in their collective experience. The comparisons made in this paper are not meant to be exhaustive but to highlight the need to develop standards for accelerometer data reduction.
A sensitivity analysis was conducted to elucidate the significant effects observed among the algorithms. As expected, the most stringent criteria negatively affected sample size, and varying the minimum wear requirement appeared to have the greatest effect on sample size. In contrast, the activity (counts per minute) and minutes of inactivity were affected mostly by varying the periods of accelerometer activity included in the data summary. The sensitivity analysis appeared to contradict the comparison of the algorithms, specifically with regard to counts per minute variable for which those of algorithm 1 were higher than those of algorithm 2. Recall that the minimal wear requirement for algorithm 1 was equal to 60% of waking time, based on the diary information. This difference alone can explain the differences observed with condition 60/10, which the minimal wear requirement was fixed at 10 h·d−1. Therefore, the sensitivity analysis clarified that the counts per minute were affected mainly by the amount of accelerometer inactivity that was allowed into the data summary. The more stringent criteria (i.e., 20 min of continuous zeros and 12 h·d−1 of wear time) reduced the size of the denominator in the calculation of average counts per minute and therefore increased the average counts value. In general, having a shorter interval of continuous bouts of zeros, increasing required wearing time, and increasing minimum number of valid days required were consistently associated with an increase in light activity and MVPA. Note the method employed by algorithm 2 to identify 20 min of continuous zeros is less stringent than algorithms 3 and 4. Algorithm 2 scanned the data in 20-min blocks and treated only those blocks as nonwear, which may underestimate nonwear as compared to the other algorithms as they scanned each 30-s epoch and started a counter to identify continuous zeros. Such a systematic pattern in the data suggests that these criteria may affect the inclusion of inactive participants in the summary. Specifically, more stringent criteria may result in the inclusion of fewer inactive participants in the analyses, thereby potentially shifting the distribution of the data. This may affect studies that attempt to characterize activity patterns of a population but might not affect studies that use accelerometer data as a predictor or a correlate. To some extent, this may depend on the outcome variable studied. Similar to the algorithm comparison, the sensitivity analysis focused on a narrow set of decision rules. These differences may be greater, given the variability found in the literature.
This is the first study that examined the impact of using different algorithms for accelerometer data reduction on a subset of outcome variables. However, there are a number of limitations to this study. First, the sample of minority women aged 40–70 yr may have attenuated some of the effects observed because the sample is more homogeneous and less active than certain groups. In addition, children have very different patterns of activity than adults, and this sample was limited in its ability to address issues that may be specific to children. Algorithm 4 imputed data if the participant had at least one valid day of data. This approach has been validated in a sample of adolescent African-American girls (51); however, in a less homogeneous sample 1 d may not be enough to truly capture one’s physical activity pattern. Future studies should determine whether the imputation algorithm should have a minimum number of valid days before proceeding with the imputation. Finally, the analyses examined the impact of the four algorithms on levels of physical activity. The implication of these findings are most relevant for surveillance and validation studies than for interventions studies, because such studies are more concerned with change in behavior. It remains unclear how these algorithms would influence change in MVPA or MVPA bouts. In light of these limitations, this is the first study that demonstrates the need to develop standard for accelerometer data reduction given the impact of using different decision rules on a subset of outcome variables.
Recommendations for Developing Standards for Accelerometer Data Reduction
Our review of studies and analyses clearly shows that, given the lack of overall standards, it is important for investigators to describe explicitly the decision rules they use in their studies, particularly because of the impact that specific decision rules can have on a summary of accelerometer data. To assist in the broader effort to develop overall standards for accelerometer data reduction, the following section proposes specific recommendations with respect to the issues listed in Table 1.
Identification of wearing period.
Identifying when a participant is not wearing the accelerometer is difficult because long continuous bouts of accelerometer inactivity may mean either that the participant removed the accelerometer or was completely inactive. Distinguishing between the two can be difficult as continuous zero reading may occur for the following reasons: a) having removed the accelerometer during a water based-activity (e.g., showering or being physically active); b) sleeping or napping with or without the accelerometer; c) sitting still for long periods; d) having removed the accelerometer for no reason; e) having removed the accelerometer during certain activities (e.g., contact sports) for safety reasons; or f) device malfunction. The algorithms compared in this study used 20 and 60 min of continuous zeros to estimate interruption in wearing time, although our review indicated that a 10, 15, and 30 min of continuous zeros (10,11,19,25,59,62,73,74) were used in previous studies. To our knowledge, no evidence in the literature supports an age-specific cut point. Assuming that adults can remain still for longer periods than can children, it may be more appropriate to use longer bouts of zero counts with adults and older adults. Verifying what participants are doing during these computer-identified interruptions may be needed to guide the selection of an age-appropriate cut point. In the meantime, researchers should report the number of wearing interruptions observed in their data. Participants are not likely to remove the accelerometer multiple times in a day, so examining these patterns may be a first step in identifying whether the selected cut point appears appropriate for the studied population. Researchers also should clearly specify the decision rules that they are using to identify periods of accelerometer inactivity and presumed nonwear, as well as the average wearing time for their study sample.
Identify minimal wear requirement for a valid day.
As noted in the introduction, most field studies ask participants to wear the accelerometer during waking hours. Clearly, the times at which individuals start wearing the accelerometer during the day and when they remove it at night fluctuate considerably. These periods of the day when the accelerometer is removed probably are not associated with a great deal of MVPA (at least not accumulated in 10-min bouts) but likely are associated with sedentary or light activity. If the main interest is to report patterns of sedentary and/or light activity, it is critical that the wearing periods be standardized and that the accelerometer be worn continuously. If participation and compliance are not affected by wearing protocol (waking hours only vs continuous wear), it may be best to report certain outcomes only when participants are wearing the accelerometer continuously.
If the accelerometer is worn during waking hours only, it may be important to develop age-specific criteria for minimal wear because children on average may sleep more than adults. Alternatively, it may be simpler to base the criteria on percent of awake time as this would take into account intraindividual and intraday variability in sleeping patterns. The latter is especially important, as sleeping time may differ between workdays and nonworkdays for adults and between school days and weekend days for children. Although only one of the algorithms used sleep time, it is an approach that the other algorithms have incorporated in the past.
Identifying spurious data.
Our algorithm comparisons and sensitivity analyses were not able to demonstrate the impact of spurious data on the outcome variables. Combining the approaches employed in algorithms 1, 2, and 4 can provide the optimal data cleaning by standardizing the cut point associated with spurious counts. In addition, this approach would look for accelerometer malfunctioning employing the algorithm 4 approach, where counts that are greater than a specified value or counts that are constant for an extended period of time (e.g., 10 min) are excluded. Automated error checking is useful for identifying periods of transient accelerometer malfunctioning and/or participant tampering that may not appear during routine calibration checks. Although it is important to identify spurious counts, it is important also to consider the impact of setting these spurious data point to missing on outcome variables.
Computing outcome variables and aggregating days of data.
Varying the minimum number of valid days included in the analyses had a significant impact on the data. The primary concern in selecting such criteria is the validity of the data. Some studies have suggested a minimum number of days needed to obtain a reliable estimate of the recorded week (16,40,48,51,72,73,75). These studies, with the exception of Gretebeck and Montoye (31), largely have ignored differences in patterns of activities that exist between weekdays and weekend days (5,48,75,79). There is evidence that adult participants are more active on Saturdays than Sundays, with Sunday being the least active day of the week (5,48). Children’s patterns appear to shift over time. Younger children tend to be more active on weekends; the pattern reverses as children age (72,75). Although not addressed in any of the four algorithms, balancing the number of weekdays and weekend days included in the analyses may be important for ensuring an accurate estimate of the pattern of behavior for the recorded week.
In our analyses, algorithms 1 and 2 varied the number of valid days used to estimate the proportion of participants who met current recommendations compared to the criteria used for other outcome variables. Requiring more days for assessing the physical activity recommendations appears reasonable given that MVPA bouts vary by day and are not observed frequently. In the literature, there is evidence that the number of monitoring days needed to get a stable estimate varies by outcomes of interest (48). Therefore, it may be important to vary the criteria by groups of outcome variables. In addition, suggesting which outcome variables are included in all accelerometer studies is important for comparing results across studies. A subset of descriptive variables should also be provided with these outcome variables to facilitate comparison across studies. Researchers are encouraged to include average wearing time and the average number of valid days used to calculate the outcome variables.
Extracting MVPA bouts.
Extracting MVPA bouts is used to determine whether the participant met the current physical activity recommendations. For adults, clear guidelines suggest that 30 min of MVPA can be accumulated in 8- to 10-min bouts (53). This study found that allowing a 1- or 2-min interruption anytime during the bout resulted in higher MVPA means than did allowing no interruption. In a physically active population, allowing an interruption may have a greater impact on the minutes of MVPA. It appears reasonable to allow a 1- or 2-min interruption anytime during the bout.
No cut point has been established for bouts of physical activity in children. Children are encouraged to do 30–60 min of physical activity daily (28). It is unlikely, however, that young children will sustain a high level of physical activity for 30 or 60 min, instead they have short bouts of activity (4). Providing recommendations regarding children appears necessary for analyzing accelerometer data. We have observed in this adult data set that minutes per day accumulated in bouts are much lower than all minutes per day of MVPA. Therefore, counting every burst of MVPA likely will affect the percent of children classified as meeting physical activity recommendation. Given that such criteria have not been established for children, it is important to provide guidelines to better enable comparison of results across studies.
The purpose of this study was to determine the impact of using different accelerometer data reduction algorithms on outcome variables. The analyses in this paper demonstrated that varying the assumptions and criteria used to summarize accelerometer data affected a number of outcome variables. To strengthen the foundation upon which accelerometer studies are building, there is a need to standardize accelerometer data reduction. This paper identified a number of issues and provided some key suggestions that may inform development of standards for accelerometer data reduction. For many of the issues raised in this paper, we unfortunately lack the empirical data to inform our decision process. Future research should focus on providing the empirical knowledge needed to standardize accelerometer data reduction. Therefore, at this time, we may be limited to base most of our recommendations on current practice. The suggestions provided in this paper will serve to begin this process and improve our ability to compare results across studies.
1.Aadahl, M., and T. Jorgensen. Validation of a new self-report instrumentfor measuring physical activity. Med. Sci. Sports Exerc
. 35:1196–1202, 2003.
2.Adriaens, M. P., P. F. Schoffelen, and K. R. Westerterp. Intra-individual variation of basal metabolic rate and the influence of daily habitual physical activity before testing. Br. J. Nutr
. 90:419–423, 2003.
3.Ayabe, M., P. H. Brubaker, D. Dobrosielski, et al. The physical activity patterns of cardiac rehabilitation program participants. J. Cardiopulmol. Rehabil
. 24:80–86, 2004.
4.Bailey, R. C., J. Olson, S. L. Pepper, J. Porszasz, T. J. Barstow, and D. M. Cooper. The level and tempo of children’s physical activities: an observational study. Med. Sci. Sports Exerc
. 27:1033–1041, 1995.
5.Baranowski, T., M. Smith, W. O. Thompson, J. Baranowski, D. Hebert, and C. de Moor. Intraindividual variability and reliability in a 7-day exercise record. Med. Sci. Sports Exerc
. 31:1619–1622, 1999.
6.Beech, B. M., R. C. Klesges, S. K. Kumanyika, et al. Child- and parent-targeted interventions: the Memphis GEMS pilot study. Ethn. Dis
. 13:S40–53, 2003.
7.Beech, B. M., S. K. Kumanyika, T. Baranowski, et al. Parental cultural perspectives in relation to weight-related behaviors and concerns of African-American girls. Obes Res
. 12(suppl 1):7S–19S, 2004.
8.Beghin, L., L. Michaud, R. Hankard, et al. Total energy expenditure and physical activity in children treated with home parenteral nutrition. Pediatr. Res
. 53:684–690, 2003.
9.Beunen, G. P., J. Lefevre, R. M. Philippaerts, et al. Adolescent correlates of adult physical activity: a 26-year follow-up. Med. Sci. Sports Exerc
. 36:1930–1936, 2004.
10.Brage, S., N. Wedderkopp, U. Ekelund, et al. Features of the metabolic syndrome are associated with objectively measured physical activity and fitness in Danish children: the European Youth Heart Study (EYHS). Diabetes Care
11.Brage, S., N. Wedderkopp, U. Ekelund, et al. Objectively measured physical activity correlates with indices of insulin resistance in Danish children. The European Youth Heart Study (EYHS). Int. J. Obes. Relat. Metab. Disord
. 28:1503–1508, 2004.
12.Brage, S., N. Wedderkopp, P. W. Franks, L. B. Andersen, and K. Froberg. Reexamination of validity and reliability of the CSA monitor in walking and running. Med. Sci. Sports Exerc
. 35:1447–1454, 2003.
13.Buchheit, M., C. Simon, A. U. Viola, S. Doutreleau, F. Piquard, and G. Brandenberger. Heart rate variability in sportive elderly: relationship with daily physical activity. Med. Sci. Sports Exerc
. 36:601–605, 2004.
14.Buchowski, M. S., S. Acra, K. M. Majchrzak, M. Sun, and K. Y. Chen. Patterns of physical activity in free-living adults in the Southern United States. Eur. J. Clin. Nutr
. 58:828–837, 2004.
15.Burdette, H. L., R. C. Whitaker, and S. R. Daniels. Parental report of outdoor playtime as a measure of physical activity in preschool-aged children. Arch. Pediatr. Adolesc. Med
. 158:353–357, 2004.
16.Coleman, K. J., and L. H. Epstein. Application of generalizability theory to measurement of activity in males who are not regularly active: a preliminary report. Res. Q. Exerc. Sport
17.Cooper, A. R., A. S. Page, L. J. Foster, and D. Qahwaji. Commuting to school: are children who walk more physically active? Am. J. Prev. Med
. 25:273–276, 2003.
18.Coronado, M., J. P. Janssens, B. de Muralt, P. Terrier, Y. Schutz, and J. W. Fitting. Walking activity measured by accelerometry during respiratory rehabilitation. J Cardiopulm. Rehabil
. 23:357–364, 2003.
19.Cradock, A. L., J. L. Wiecha, K. E. Peterson, A. M. Sobol, G. A. Colditz, and S. L. Gortmaker. Youth recall and TriTrac accelerometer estimates of physical activity levels. Med. Sci. Sports Exerc
. 36:525–532, 2004.
20.Craig, C. L., A. L. Marshall, M. Sjostrom, et al. International physical activity questionnaire: 12-country reliability and validity. Med. Sci. Sports Exerc
. 35:1381–1395, 2003.
21.Dubbert, P. M., M. W. Vander Weg, K. A. Kirchner, and B. Shaw. Evaluation of the 7-day physical activity recall in urban and rural men. Med. Sci. Sports Exerc
. 36:1646–1654, 2004.
22.DuVall, C., M. K. Dinger, E. L. Taylor, and D. Bemben. Minimal-contact physical activity interventions in women: a pilot study. Am. J. Health Behav
. 28:280–286, 2004.
23.Eason, K. E., L. C. Masse, S. H. Kelder, and S. R. Tortolero. Diary days needed to estimate activity among older African-American and Hispanic women. Med. Sci. Sports Exerc
. 34:1308–1315, 2002.
24.Ekelund, U., J. Aman, and K. Westerterp. Is the ArteACC index a valid indicator of free-living physical activity in adolescents? Obes. Res
. 11:793–801, 2003.
25.Ekelund, U., A. Yngve, S. Brage, K. Westerterp, and M. Sjostrom. Body movement and physical activity energy expenditure in children and adolescents: how to adjust for differences in body size and age. Am. J. Clin. Nutr
. 79:851–856, 2004.
26.Escobar-Chaves, S. L., S. R. Tortolero, L. C. Masse, K. B. Watson, and J. E. Fulton. Recruiting and retaining minority women: findings from the Women on the Move study. Ethn. Dis
. 12:242–251, 2002.
27.Freedson, P. S., E. Melanson, and J. Sirard. Calibration of the Computer Science and Applications, Inc. accelerometer. Med. Sci. Sports Exerc
. 30:777–781, 1998.
28.Fulton, J. E., M. Garg, D. A. Galuska, K. T. Rattay, and C. J. Caspersen. Public health and clinical recommendations for physical activity and physical fitness: special focus on overweight youth. Sports Med
. 34:581–599, 2004.
29.Going, S., J. Thompson, S. Cano, et al. The effects of the Pathways Obesity Prevention Program on physical activity in American Indian children. Prev. Med
. 37:S62–S69, 2003.
30.Goris, A. H., M. A. Vermeeren, E. F. Wouters, A. M. Schols, and K. R. Westerterp. Energy balance in depleted ambulatory patients with chronic obstructive pulmonary disease: the effect of physical activity and oral nutritional supplementation. Br. J. Nutr
. 89:725–731, 2003.
31.Gretebeck, R. J., and H. J. Montoye. Variability of some objective measures of physical activity. Med. Sci. Sports Exerc
. 24:1167–1172, 1992.
32.Hayden-Wade, H. A., K. J. Coleman, J. F. Sallis, and C. Armstrong. Validation of the telephone and in-person interview versions of the 7-day PAR. Med. Sci. Sports Exerc
. 35:801–809, 2003.
33.Hendelman, D., K. Miller, C. Baggett, E. Debold, and P. Freedson. Validity of accelerometry for the assessment of moderate intensity physical activity in the field. Med. Sci. Sports Exerc
. 32:S442–449, 2000.
34.Hoos, M. B., H. Kuipers, W. J. Gerver, and K. R. Westerterp. Physical activity pattern of children assessed by triaxial accelerometry. Eur. J. Clin. Nutr
. 58:1425–1428, 2004.
35.Hoos, M. B., G. Plasqui, W. J. Gerver, and K. R. Westerterp. Physical activity level measured by doubly labeled water and accelerometry in children. Eur. J. Appl. Physiol
. 89:624–626, 2003.
36.Hoos, M. B., K. R. Westerterp, H. Kuipers, L. Schuwirth, and W. J. Gerver. Physical activity as measured by accelerometry in children receiving growth hormone. Acta Paediatr
. 93:1307–1311, 2004.
37.Inoue, Y., T. Kimura, M. H. Noro, et al. Is laparoscopic colorectal surgery less invasive than classical open surgery? Quantitation of physical activity using an accelerometer to assess postoperative convalescence. Surg. Endosc
. 17:1269–1273, 2003.
38.Inoue, Y., T. Kimura, S. Fujita, et al. A new parameter for assessing postoperative recovery of physical activity using an accelerometer. Surg. Today
. 33:645–650, 2003.
39.Jackson, D. M., J. J. Reilly, L. A. Kelly, C. Montgomery, S. Grant, and J. Y. Paton. Objectively measured physical activity in a representative sample of 3- to 4-year-old children. Obes. Res
. 11:420–425, 2003.
40.Janz, K. F., J. Witt, and L. T. Mahoney. The stability of children’s physical activity as measured by accelerometry and self-report. Med. Sci. Sports Exerc
. 27:1326–1332, 1995.
41.Karolkiewicz, J., L. Szczesniak, E. Deskur-Smielecka, A. Nowak, R. Stemplewski, and R. Szeklicki. Oxidative stress and antioxidant defense system in healthy, elderly men: relationship to physical activity. Aging Male
42.Kirk, A., N. Mutrie, P. MacIntyre, and M. Fisher. Effects of a 12-month physical activity counselling intervention on glycaemic control and on the status of cardiovascular risk factors in people with type 2 diabetes. Diabetologia
43.Kirk, A., N. Mutrie, P. MacIntyre, and M. Fisher. Increasing physical activity in people with type 2 diabetes. Diabetes Care
44.Kirk, A. F., N. Mutrie, P. D. Macintyre, and M. B. Fisher. Promoting and maintaining physical activity in people with type 2 diabetes. Am. J. Prev. Med
. 27:289–296, 2004.
45.Le Masurier, G. C., C. L. Sidman, and C. B. Corbin. Accumulating 10,000 steps: does this meet current physical activity guidelines? Res. Q. Exerc. Sport
46.Lof, M., U. Hannestad, and E. Forsum. Comparison of commonly used procedures, including the doubly-labelled water technique, in the estimation of total energy expenditure of women with special reference to the significance of body fatness. Br. J. Nutr
. 90:961–968, 2003.
47.Mâsse, L. C., J. E. Fulton, K. L. Watson, M. T. Mahar, M. C. Meyers, and W. W. Wong. Influence of body composition on physical activity validation studies using doubly labeled water. J. Appl. Physiol
. 96:1357–1364, 2004.
48.Matthews, C. E., B. E. Ainsworth, R. W. Thompson, and D. R. Bassett. Jr. Sources of variance in daily physical activity levels as measured by an accelerometer. Med. Sci. Sports Exerc
. 34:1376–1381, 2002.
49.McMurray, R. G., K. B. Ring, M. S. Treuth, et al. Comparison of two approaches to structured physical activity surveys for adolescents. Med. Sci. Sports Exerc
. 36:2135–2143, 2004.
50.Montoye, H. J., H. C. G. Kemper, W. H. M. Saris, and R. A. Washburn. Measuring Physical Activity and Energy Expenditure
. Champaign, IL: Human Kinetics, 1996, p. 72.
51.Murray, D. M., D. J. Catellier, P. J. Hannan, et al. School-level intraclass correlation for physical activity in adolescent girls. Med. Sci. Sports Exerc
. 36:876–882, 2004.
52.Pate, R. R., K. A. Pfeiffer, S. G. Trost, P. Ziegler, and M. Dowda. Physical activity among children attending preschools. Pediatrics
53.Pate, R. R., M. Pratt, S. N. Blair, 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
54.Powers, G. C., T. Zentner, F. Nelson, and N. Bergstrom. Validation of the mobility subscale of the Braden Scale for predicting pressure sore risk. Nurs. Res
. 53:340–346, 2004.
55.Prochaska, J. J., and J. F. Sallis. A randomized controlled trial of single versus multiple health behavior change: promoting physical activity and nutrition among adolescents. Health Psychol
. 23:314–318, 2004.
56.Proctor, M. H., L. L. Moore, D. Gao, et al. Television viewing and change in body fat from preschool to early adolescence: The Framingham Children’s Study. Int. J. Obes. Relat. Metab. Disord
. 27:827–833, 2003.
57.Puyau, M. R., A. L. Adolph, F. A. Vohra, and N. F. Butte. Validation and calibration of physical activity monitors in children. Obes. Res
. 10:150–157, 2002.
58.Rafamantanantsoa, H. H., N. Ebine, M. Yoshioka, et al. The role of exercise physical activity in varying the total energy expenditure in healthy Japanese men 30 to 69 years of age. J. Nutr. Sci. Vitaminol. (Tokyo)
59.Riddoch, C. J., L. Bo Andersen, N. Wedderkopp, et al. Physical activity levels and patterns of 9- and 15-yr.-old European children. Med. Sci. Sports Exerc
. 36:86–92, 2004.
60.Saelens, B. E., J. F. Sallis, J. B. Black, and D. Chen. Neighborhood-based differences in physical activity: an environment scale evaluation. Am. J. Public Health
. 93:1552–1558, 2003.
61.Samuel-Hodge, C. D., L. M. Fernandez, C. F. Henriquez-Roldan, L. F. Johnston, and T. C. Keyserling. A comparison of self-reported energy intake with total energy expenditure estimated by accelerometer and basal metabolic rate in African-American women with type 2 diabetes. Diabetes Care
62.Schmidt, M. D., P. S. Freedson, and L. Chasan-Taber. Estimating physical activity using the CSA accelerometer and a physical activity log. Med. Sci. Sports Exerc
. 35:1605–1611, 2003.
63.Sherwood, N. E., W. C. Taylor, M.Treuth, et al. Measurement characteristics of activity-related psychosocial measures in 8- to 10-year-old African-American girls in the Girls Health Enrichment Multisite Study (GEMS). Prev. Med
. 38 (suppl):S60–S68, 2004.
64.Stach-Lempinen, B., C. H. Nygard, P. Laippala, R. Metsanoja, and E. Kujansuu. Is physical activity influenced by urinary incontinence? BJOG
65.Steele, B. G., B. Belza, J. Hunziker, et al. Monitoring daily activity during pulmonary rehabilitation using a triaxial accelerometer. J. Cardiopulm. Rehabil
. 23:139–142, 2003.
66.Story, M., N. E. Sherwood, J. H. Himes, et al. An after-school obesity prevention program for African-American girls: the Minnesota GEMS pilot study. Ethn. Dis
. 13:S54–S64, 2003.
67.Strath, S. J., Bassett, Jr., D. R. A. M. Swartz, and D. L. Thompson. Simultaneous heart rate-motion sensor technique to estimate energy expenditure. Med. Sci. Sports Exerc
. 33:2118–2123, 2001.
68.Talbot, L. A., J. M. Gaines, T. N. Huynh, and E. J. Metter. A home-based pedometer-driven walking program to increase physical activity in older adults with osteoarthritis of the knee: a preliminary study. J. Am. Geriatr. Soc
. 51:387–392, 2003.
69.Taylor, W. C., T. Baranowski, L. M.Klesges, et al. Psychometric properties of optimism and pessimism: results from the Girls’ Health Enrichment Multisite Studies. Prev. Med
. 38 (Suppl):S69–S77, 2004.
70.Timperio, A., J. Salmon, and D. Crawford. Validity and reliability of a physical activity recall instrument among overweight and non-overweight men and women. J. Sci. Med. Sport
. 6:477–491, 2003.
71.Timperio, A., J. Salmon, M. Rosenberg, and F. C. Bull. Do logbooks influence recall of physical activity in validation studies? Med. Sci. Sports Exerc
. 36:1181–1186, 2004.
72.Treuth, M. S., N. F. Butte, A. L. Adolph, and M. R. Puyau. A longitudinal study of fitness and activity in girls predisposed to obesity. Med. Sci. Sports Exerc
. 36:198–204, 2004.
73.Treuth, M. S., N. E. Sherwood, T. Baranowski, et al. Physical activity self-report and accelerometry measures from the Girls health Enrichment Multi-site Studies. Prev. Med
. 38(suppl):S43–S49, 2004.
74.Treuth, M. S., N. E. Sherwood, N. F. Butte, et al. Validity and reliability of activity measures in African-American girls for GEMS. Med. Sci. Sports Exerc
. 35:532–539, 2003.
75.Trost, S. G., R. R. Pate, P. S. Freedson, J. F. Sallis, and W. C. Taylor. Using objective physical activity measures with youth: how many days of monitoring are needed? Med. Sci. Sports Exerc
. 32:426–431, 2000.
76.Trost, S. G., J. R. Sirard, M. Dowda, K. A. Pfeiffer, and R. R. Pate. Physical activity in overweight and nonoverweight preschool children. Int. J. Obes. Relat. Metab. Disord
. 27:834–839, 2003.
77.Tudor-Locke, C., B. E. Ainsworth, L. S. Adair, and B. M. Popkin. Objective physical activity of Filipino youth stratified for commuting mode to school. Med. Sci. Sports Exerc
. 35:465–471, 2003.
78.Tudor-Locke, C., B. E. Ainsworth, L. S. Adair, and B. M. Popkin. Physical activity in Filipino youth: the Cebu Longitudinal Health and Nutrition Survey. Int. J. Obes. Relat. Metab. Disord
. 27:181–190, 2003.
79.Tudor-Locke, C., S. A. Ham, C. A. Macera, et al. Descriptive epidemiology of pedometer-determined physical activity. Med. Sci. Sports Exerc
. 36:1567–1573, 2004.
80.Vogels, N., G. Egger, G. Plasqui, and K. R. Westerterp. Estimating changes in daily physical activity levels over time: implication for health interventions from a novel approach. Int. J. Sports Med
. 25:607–610, 2004.
81.Wedderkopp, N., C. Leboeuf-Yde, L. Bo Andersen, K. Froberg, and H. Steen Hansen Back pain in children: no association with objectively measured level of physical activity. Spine
discussion 2024, 2003.
82.Welk, G. J., J. A. Schaben, and Morrow, Jr. J. R. Reliability of accelerometry-based activity monitors: a generalizability study. Med. Sci. Sports Exerc
. 36:1637–1645, 2004.
83.Westerterp, K. R., E. P. Meijer, A. H. Goris, and A. D. Kester. Alcohol energy intake and habitual physical activity in older adults. Br. J. Nutr
. 91:149–152, 2004.
84.Whitt, M., S. Kumanyika, and S. Bellamy. Amount and bouts of physical activity in a sample of African-American women. Med. Sci. Sports Exerc
. 35:1887–1893, 2003.