It is generally believed that the association between physical activity and health outcomes might be stronger if physical activity measurements were more accurate (16). More accurate assessments of free-living physical activity would help to characterize the relationship between physical activity and disease prevention (the dose/response relationship), assess the efficacy of intervention strategies, and monitor the physical activity patterns of various populations (21). With these goals in mind, researchers are actively searching for valid and reliable measures of physical activity (2). This search has led to the increased availability of a wide variety of objective monitoring technologies. The research application of these technologies has resulted in the accrual of a significant body of literature on objective physical activity assessment, a large portion of which involves accelerometers (up to 90 articles per year in 2003 and 2004) (12). Indeed, accelerometry-based physical activity monitors are one of the most commonly used devices for assessing free-living physical activity (18). Moreover, the use of accelerometers in large-scale surveillance studies is on the rise (e.g., the National Health and Nutrition Examination Survey (7) and the Physical Activity Levels in Children and Youth study) (11).
Although the literature suggests that accelerometer technology and its applications have progressed significantly, information is lacking in key areas. As a result, in December 2004, a conference entitled "Objective Monitoring of Physical Activity: Closing the Gaps in the Science of Accelerometry" was hosted by the University of North Carolina School of Public Health, department of nutrition. The proceedings from the 3-d meeting were assembled in article format and published as a supplement. The final paper, authored by Ward et al. (17) and entitled "Accelerometer Use in Physical Activity: Best Practices and Research Recommendations," summarized the salient points of the meeting. In it, the authors identified the need for studies comparing the validity and interinstrument reliability of different models of accelerometers. Studies of this nature were seen as critical for accelerometer-model selection but were deemed equally important as a means to scrutinize the quality and objectivity of the available reliability evidence (13).
The accelerometer reliability research published to date can be divided into two categories: studies conducted using a mechanical apparatus or those employing some form of subject-mounted setup. The subject-mounted setups can be further subdivided into laboratory-based activity assessments and the more practical, but less controlled situations of free-living activity assessment. Mechanical setups, by virtue of the precise control of the experimental conditions, can determine the variability attributed solely to the accelerometer. As with any method of measurement, it is important to identify and quantify the different sources of variation so that actions can be taken to reduce or control them. This is important because if the measurement error intrinsic to the accelerometer is found to be small, the focus can shift to other sources of variation (e.g., position worn on the body, or variation over time) (6). Moreover, quantifying the inherent variation in accelerometer models allows for better interpretation of results and helps inform accelerometer-purchasing decisions.
Researchers have used various mechanical apparatuses to oscillate accelerometers in various axes to assess reliability. Examples include turntables (6), rotating wheel setups (1), and vibration tables (9). These apparatuses allow researchers to control the magnitude of the acceleration as well as the frequency of the oscillation, two key variables that contribute to the accelerometer's output. However, technical reliability studies to date have assessed only one accelerometer model and could only accommodate a small number of instruments at one time. The purpose of this study was to determine which of the three most commonly used accelerometer models (Actical (Mini Mitter Co., Inc., Bend, OR), Actigraph model 7164 (Actigraph, Fort Walton Beach, FL), or RT3 (Stayhealthy, Inc., Monrovia, CA); see Table 1 for specifications) has the best intra- and interinstrument reliability, using a mechanical laboratory setup. Secondly, this study aimed to determine the individual and combined effects of acceleration and frequency of movement on accelerometer count output. To the authors' knowledge, this study is the first to simultaneously assess the reliability of multiple accelerometers and multiple models in a mechanical setup.
MATERIALS AND METHODS
Hydraulic Shaker Table.
All reliability testing was completed using a hydraulic shaker table (Fig. 1). The shaker table was driven by a hydraulic cylinder (Sheffer, 1-1/18HHSL6ADY) controlled by an electrohydraulic servo valve with cylinder-position feedback. A position transducer (Lucas 5000, DC-E) was used to measure position of the table, and a high-grade control accelerometer (calibrated at 98.1 mV·g−1 ± 3.6%) (B and K model 4371) was attached to the table to measure vertical acceleration. The acceleration signal was transmitted to a charge amplifer (B and K model 2635) and bandpass filtered at 3 kHz. The amplifier input was provided by a function generator, which was programmed to accurately and reliably oscillate the platform at the various testing conditions using a sinusoidal oscillation procedure. The separation of the hydraulic power supply unit from the shaker table helped to minimize the mechanical vibration in the mechanical setup.
The shaker table testing conditions were restricted by the displacement amplitude of the shaker plate (approximately 6.5 cm). Within this amplitude range, the possible conditions of acceleration and frequency of oscillation are described by the equation: acceleration (m·s−2) = (amplitude (m) · frequency2 (rad·s−1). The six different conditions chosen were selected to produce a range of physiologically relevant accelerometer counts from light to moderate to hard within the limitations of the shaker plate (Table 2). Compared with treadmill-calibration studies reported in the literature on the Actical (8), these six conditions range in locomotive speed from approximately 2.5 to 4.75 mph. Compared with treadmill-calibration studies using the Actigraph (14), the conditions range in locomotive speed from approximately 2.5 to 6.75 mph. Finally, compared with treadmill calibration studies using the RT3 (10), the conditions range from approximately 1.0 to 3.25 mph. These conditions also were chosen to allow for independent assessments of both acceleration and frequency on accelerometer reliability. This was achieved by selecting three conditions at 0.5 g, allowing only the frequency of oscillation to change and, similarly, by selecting three conditions at 2.5 Hz, allowing only acceleration of the shaker plate to change.
Fifteen accelerometers, five of each of three models (Actical, Actigraph, and RT3), were initialized to collect data using 1-min epochs. The computerized initialization function of the Actical and the Actigraph made time synchronization of these two models easy to attain. In the case of the five RT3, the external start buttons were simultaneously pressed at the exact time (using the initialization PC clock) the Actical and Actigraph were set to begin data collection. The triaxial RT3 was set to vector magnitude mode, thereby combining count data from all three axes. The accelerometers were mounted to the surface of the shaker plate (surface area approximately 1500 cm2) using industrial wax. Care was taken to ensure that the monitors were secured firmly and were positioned vertically along their sensitive axis to maximize and standardize the output of the piezosensor. In the case of the triaxial RT3, it was positioned so the vertical oscillation was along the x-axis.
The hydraulic shaker table was switched on once all 15 accelerometers were in place and the first of the random ordered conditions was set, thereby accelerating all 15 monitors simultaneously in the vertical plane. The shaker table was warmed up to achieve optimal functioning of the hydraulics and the control electronics, thereby ensuring the proper execution and maintenance of each of the six conditions for the 7-min test periods. All conditions began at the turn of a new minute on the PC clock, which was recorded along with the condition end time for data analysis purposes. After approximately 60 min of data collection (12-min warm-up + (six conditions × 7 min per condition) + (6 × 1-min transitions between conditions)), the accelerometers were removed from the shaker plate and downloaded to the initialization PC for further analysis. Data were imported into a customized spreadsheet application using the common epoch-by-epoch time stamp to align the data vertically across models. The recorded condition start and end times were identified, and the middle 5 min of each condition were identified and exported from the spreadsheet application into a statistical package for further analysis (SPSS 13.0).
Experiments 2 and 3.
Based on promising reliability data from experiment 1, the Actical and Actigraph accelerometer models were selected to undergo more robust reliability assessments. In experiment 2, using exactly the same data collection and analysis procedures as described above, 39 Actical accelerometers from a different lot of 40 devices (one was found to be faulty on delivery) were simultaneously accelerated in the vertical plane using the same six conditions already described (Table 2). In experiment 3, again using a similar data collection and analysis procedure as in the first experiment, 50 Actigraph accelerometers from a different lot were simultaneously accelerated in the vertical plane. However, because two devices malfunctioned, all analyses were performed on a sample of 48 Actigraphs.
To determine the variability within a given accelerometer, standard deviation (SD), standard error of the measurement (SEM), and coefficient of variation (CVintra) were calculated from the replicate minutes (minutes 1-5) within each condition. This minute-by-minute variability characterizes the accelerometers' ability to consistently measure the given condition rendered by the shaker table. This is a noteworthy distinction because most intrainstrument reliability analyses focus on within-accelerometer, between-trial variability. As a result, less variability (i.e., technological error) is expected using the present calculation methods because no trial effect is present.
To determine the variability between like-model accelerometers (i.e., between units), SD, SEM, and coefficient of variation (CVinter) were calculated for each of the six testing conditions. In addition, intraclass correlation coefficients (ICC) with a two-way random-effects model for absolute agreement were calculated. To determine the independent effect of acceleration and frequency on count output across models, repeated-measures ANOVA were used. Where significance was found, post hoc analyses were conducted via paired t-tests. In all cases, alpha was set at P < 0.05.
The summary of accelerometer count data across all six conditions and models (Table 3) suggests that the Actical accelerometer had better intrainstrument reliability (mean CVintra = 0.4%), followed by the Actigraph (4.1%) and the RT3 (46.4%), respectively. However, the Actigraph accelerometer had better interinstrument reliability (mean CVinter = 4.9%), followed by the Actical (15.5%) and the RT3 (42.9%), respectively (Table 3). The same hierarchy of interinstrument reliability was found with the calculation of the average-measure intraclass correlation coefficients (R = 0.995, 0.985, and 0.910 for the Actigraph, Actical, and RT3, respectively).
Holding the frequency of oscillation of the shaker plate constant at 2.5 Hz allowed for independent assessment of the impact of varying acceleration conditions on the magnitude of the count output. As expected, increasing the magnitude of acceleration increased the count output in all accelerometer models (Fig. 2). However, holding acceleration constant at 4.9 m·s−2 and increasing movement frequency produced seemingly counterintuitive results: no consistent relationship was found between models with increasing frequency of oscillation of the shaker plate (Fig. 3). In fact, the Actical count output increased with increasing frequency, whereas the Actigraph counts decreased and the RT3 counts both decreased and increased.
Experiments 2 and 3.
Further testing of a larger sample of Actical accelerometers (N = 39) showed that the intrainstrument reliability remained relatively stable in experiment 2 (CVintra = 0.5%) compared with experiment 1 (0.4%) (Table 4). However, the interinstrument reliability improved markedly from an average CVinter of 15.5% in experiment 1 to 5.4% in experiment 2 (Table 4). Further testing of Actigraph accelerometers in experiment 3 (N = 48) also produced differing results compared with experiment 1. The second set of analyses on the Actigraphs produced better intrainstrument reliability (3.2% compared with 4.1% in experiment 1) (Table 5). However, the interinstrument reliability of the Actigraph decreased from a CVinter of 4.9% in experiment 1 to 8.6% in experiment 3 (Table 5).
Presenting the relative variability data of the two accelerometer models in graphic rather than tabular form highlights the intensity effect. The variability of the Actical accelerometer is negatively related to the intensity of the shaker plate testing condition (Fig. 4, top). As the acceleration of the condition increases, the interinstrument variability decreases; thus, it is the acceleration, rather than the frequency, that affects the variability of the Actical output. However, a comparable graph (Fig. 4, bottom) depicting the Actigraph data shows no relationship between acceleration and relative interinstrument variability; rather, it suggests that frequency is more closely related (again negatively) to accelerometer variability. This apparent heteroscedasticity went undetected by the ICC values, which increased for the Actical from 0.985 in experiment 1 to a perfect value of 1.00 in experiment 2. Likewise, the already high ICC values of the Actigraph increased from 0.995 to 0.999 from experiment 1 to experiment 3.
In experiments 2 and 3, when the frequency was held constant at 2.5 Hz and the acceleration increased from 4.9 to 9.81 to 12.26 m·s−2, both the Actical and Actigraph count output responded by increasing in magnitude, the same as in experiment 1. Meanwhile, when the acceleration of the shaker table was held constant at 4.9 m·s−2, the Actical count output did not follow the same pattern of increasing magnitude, as the frequency of oscillation increased from 1.5 to 2.0 to 2.5 Hz (Fig. 5). However, the Actigraph count output showed its characteristic decrease in count output as the frequency increased (Fig. 5).
Actigraph count output increased as frequency decreased at a given acceleration, resulting in a graded count output across the intensity spectrum (Fig. 6). However, because the Actical accelerometers showed very little count variation across the three testing frequencies at 4.9 m·s−2 (i.e., conditions 1-3), a gap appears in the middle of the count output intensity spectrum (Fig. 6). In addition to the differences in the distribution of count outputs across the intensity spectrum, the order of the conditions also differs between accelerometer models.
In experiment 1, the Actical was found to have the best intrainstrument reliability, whereas the Actigraph had the best interinstrument reliability, with the RT3 generally performing poorly. The exceptionally poor reliability of the RT3 accelerometers may be explained by the fact that the RT3 has a much wider frequency range than both the Actical and Actigraph (upper cutoff frequency of 10 Hz compared with 3.0 and 2.5 Hz, respectively). It has been suggested that an overly wide bandwidth filter could allow physiologically unrelated vibrations (i.e., noise) to be included in the signal (3). Although the separation of the hydraulic power supply unit and the shaker table in our experimental setup helped minimize vibration in the mechanical setup, it could not ensure it. Further complicating the issue of reliability is the fact that the lower cutoff frequency of the RT3 is 2.0 Hz, which is greater than one, and equal to two, of the six testing conditions used in the present study. Finally, the fact that the RT3 accelerometer is triaxial may have increased its ability to detect vibrations in the mediolateral and anteroposterior axes, something neither of the other monitors were capable of doing. These explanations suggest that a large portion of the variability in the RT3 may be attributable to hardware and setup issues. Nevertheless, the following attributes of the RT3 can be considered limitations: its large size, its external display, the presence of an external button, the accessible battery compartment, and the fact that it is not waterproof.
The excellent intra- and much weaker interinstrument reliability of the Actical in experiment 1 were surprising and disconcerting. This level of interinstrument variability raises quality-assurance concerns. Most accelerometer companies perform some form of calibration procedure as part of a quality-assurance check before filling an order (19). Experiment 2 was conducted to assess the extent of the quality-assurance concerns across a larger number of Actical accelerometers.
Although the Actigraph performed well, with both the intra- and interinstrument variability falling below 5%, the existence of a discrepant trend in count output across frequencies between the Actical and the Actigraph suggested a validity concern (discussed in the next section). Experiments 2 and 3 were performed to further assess this concern.
Experiments 2 and 3.
The results of experiments 2 and 3 clearly indicate that under the mechanical testing conditions of these experiments, the Actical (CVintra = 0.5%, CVinter = 5.4%) is more reliable than the Actigraph (CVintra = 3.2%, CVinter = 8.6%). This suggests better interinstrument calibration of the Actical by the manufacturer compared with the lot of devices from experiment 1. To the authors' knowledge, this is the first technical reliability data published on the Actical. However, there are three such technical reliability studies available for comparison on the Actigraph accelerometer (1,5,6).
In the study by Brage et al. (1), six Actigraph accelerometers were exposed to a host of acceleration and frequency conditions via a dual rotating-wheel setup. The mean intrainstrument variability of the six Actigraphs was slightly higher (within-instrument, between-trial CVintra = 4.4%) but comparable with that of the 48 Actigraphs in the present study. Likewise, over similar acceleration and frequency conditions, the range of interinstrument reliabilities reported (CVinter = 5-12%) matched quite well with those in the present study. When presented with large interinstrument variability, Brage et al. (1) suggest that either a multipoint, unit-specific calibration be used before and after each measurement period, or that some form of post measurement adjustment be employed (i.e., covariate) during statistical analyses.
In a preliminary study by Fairweather et al. (5), four Actigraphs were oscillated at 2.0 Hz using a mechanical shaker system. The only reliability data reported were for interinstrument reliability (CVinter ≈ 3.0%), which was much better than that found in the present study. The difference in these results can be explained first by the small number of accelerometers tested (i.e., homogeneous sample) and second by the fact that only one testing condition was used. In a study by Metcalf et al. (6), Actigraphs were rotated at medium and fast speeds via a turntable setup. Intrainstrument reliability (within instrument, between trials; N = 7) ranged from 0.8 to 1.4%, whereas interinstrument reliability (N = 23) was found to be 3.3% at both fast and medium speeds. At first glance, the intrainstrument reliabilities look much better than those in the present study. However, if only similar frequency conditions from the present study are compared (i.e., fast speed of 120 rpm = 2.0 Hz and medium speed of 72 rpm = 1.2 Hz), the reliability results align much better with the present study (aligned CVintra = 0.17 and 0.56%, respectively). However, the interinstrument reliability did not align as well (aligned CVinter = 7.1 and 7.3%, respectively), likely because of the larger, more heterogeneous sample of accelerometers in the present study.
To date, only one study has compared the interinstrument reliability of different accelerometer models (20). This study assessed four accelerometer models using a more applied approach, employing standardized bouts of treadmill and outdoor running activity. The results of the generalizability study concluded that, overall, the Actigraph (N = 10) was the most reliable accelerometer (CVinter = 8.9%) compared with the Tritrac (the predecessor version of the RT3; N = 9; CVinter = 9.4%), Biotrainer (IM Systems, Baltimore, MD; N = 9; CVinter = 10%), and the Actical (N = 7; CVinter = 20.0%). Although not a technical reliability study, the interinstrument reliability of the Actigraph was nearly equal to that of the present study. However, the high degree of variability in the Actical was much greater than in the present study, especially in experiment 2. It is possible that the Actical monitors were acquired before the manufacturer was aware of potential issues in their calibration quality assurance.
The discrepant trend in count output across frequencies between the Actical and the Actigraph (Fig. 3) was confirmed in experiments 2 and 3 (Fig. 5). This result is indeed intriguing, because it suggests that there is a validity issue at play. How can two accelerometer models designed to measure the same thing produce very different trends when presented with the same testing conditions? Which one, if any, is correct? It is important to understand that accelerometers are accelerometer-based physical activity monitors, not instruments that merely record acceleration. As such, these instruments must consider both the frequency and acceleration of movement to validly assess physical activity.
The six testing conditions imposed on the Actical resulted in a bimodal distribution of the six mean count outputs (Fig. 6). The gap in Actical output occurs because conditions 1-3 remain virtually unchanged despite changes in frequency and, hence, work performed. These results call into question the validity of the Actical. Conversely, these same six conditions imposed on the Actigraph resulted in a distributed count output across the intensity spectrum (Fig. 6). With the exception of a reversed order in conditions 5 and 6, the graded output of the Actigraph matches the theorized intensity spectrum based on the quotient of work and body mass (Table 2). The fact that there is general agreement between the Actigraph count output and mass-specific work may provide evidence of instrument validity. However, data from the present study do not allow definitive conclusions regarding the validity of either accelerometer model.
The Actigraph user's manual presents the accelerometer frequency rage as 0.25-2.5 Hz, which may be misleading because it implies that movements inside this range are measured full scale, whereas those outside this range are not registered at all. However, in the original Actigraph design study, Tryon and Williams (15) describe the filter as a weighting function with optimal weight at 0.75 Hz that decreases as frequency increases or decreases. However, unlike Tryon and Williams, Brage et al. (1) explicitly state that Actigraph output is only proportional to acceleration if frequency is held constant, thus suggesting that some form of frequency-dependent filter is present. The authors go on to develop a frequency-based correction factor that can be applied to raw Actigraph counts to restore linearity. Applying this correction factor to the Actigraph data in experiment 3 results in increased mean count output across all six conditions (because of the reweighting) (Fig. 7). Likewise, conditions 5 and 6 become properly ordered (i.e., aligning to work per kilogram) as a result of the frequency-correction equation. The corrected data are consistent with the notion that at least a portion of the decline in Actigraph output with increasing frequency may be a result of bandwidth-filtering procedures.
Unfortunately, no design-specifications research comparable with that of the Actigraph (15) has been published on the Actical; therefore, the filtering specifics of this accelerometer model are unknown and are in need of future research. Likewise, because the Actigraph 7164 has been phased out and replaced by the GT1M model, further studies are required to examine the comparability and technical reliability of this new Actigraph model.
That reliability sets the limit on validity is a fundamental tenet of science and, as such, justifies the need for quality reliability research. Researchers employing accelerometers to assess physical activity would do well to start treating their accelerometers with the same care as those using metabolic carts. This means the initiation of proper calibration checks with each and every use and, obviously, substituting the calibration gas with some form of mechanical apparatus that reliably oscillates the accelerometer across a host of intensity conditions. And, of course, an a priori variability limit must be set (e.g., mean difference ≤ 5%). If such a calibration check was implemented with the data from experiments 2 and 3, seven (18%) of the Actical and 16 (33%) of the Actigraph accelerometers would be rejected as too variable for use (Fig. 8). That the accelerometer units along the x-axes are ordered according to serial number is also of interest; in this manner, visual checks for batch/lot effects can easily be made. For example, looking at the data from the Actical, one can easily determine that units 19-23 are clustering (i.e., come from the same homogeneous batch). Further, one can readily see that units 1-18 are subject to more variability than units 24-39. These data clearly illustrate that batch effects can greatly influence reliability and, therefore, deserve consideration in reliability study designs.
The popularity of accelerometry as an objective measure of physical activity stems from its ability to provide direct, objective, and detailed physical activity information (4). However, the quality of information from accelerometers is only as good as the devices themselves. Therefore, it is important that researchers and manufacturers work together to ensure the reliability and, ultimately, the validity of these measurement devices. Finally, journal editors and peer reviewers will have to do their part by demanding that proper reliability procedures be followed and reported for successful publication.
The authors acknowledge Doug Bitner for his assistance with the mechanical setup and Dr. Kong Chen for his kind review of the manuscript. This study was funded by the Canadian Population Health Initiative of the Canadian Institute for Health Information.
This study was not funded in any way by any of the accelerometer manufacturers. The results of the present study do not constitute endorsement by the authors or ACSM of the products described in this paper.
1. Brage, S., N. Brage, N. Wedderkopp, and K. Froberg. Reliability and validity
of the Computer Science and Applications accelerometer in a mechanical setting. Meas. Phys. Ed. Exerc. Sci.
2. Caspersen, C. J. Physical activity epidemiology: concepts, methods, and applications to exercise science. Exerc. Sport Sci. Rev.
3. Chen, K. Y., and D. R. Bassett, Jr. The technology of accelerometry-based activity monitors: current and future. Med. Sci. Sports Exerc.
4. Esliger, D. W., J. L. Copeland, J. D. Barnes, and M. S. Tremblay. Standardizing and optimizing the use accelerometer data for free-living physical activity monitoring. J. Phys. Act. Health
5. Fairweather, S. C., J. R. Reilly, S. Grant, A. Whittaker, and J.Y. Paton. Using the Computer Science and Applications (CSA) activity monitor in preschool children. Pediatr. Exerc. Sci.
6. Metcalf, B. S., J. S. Curnow, C. Evans, L. D. Voss, and T. J. Wilkin. Technical reliability of the CSA activity monitor: the Early Bird Study. Med. Sci. Sports Exerc.
7. National Center for Health Statistics. National Health and Nutrition Examination Survey Laboratory Protocol.
Available at: http://www.cdc.gov/nchs/data/nhanes/nhanes_03_04/LAB.pdf
. Accessed February 13, 2006.
8. Puyau, M. R., A. L. Adolph, F. A. Vohra, I. Zakeri, and N. F. Butte. Prediction of activity energy expenditure using accelerometers in children. Med. Sci. Sports Exerc.
9. Powell, S. M., D. I. Jones, and A. V. Rowlands. Technical variability of the RT3 accelerometer. Med. Sci. Sports Exerc.
10. Rowlands, A. V., P. W. Thomas, R. G. Eston, and R. Topping. Validation of the RT3 triaxial accelerometer for the assessment of physical activity. Med. Sci. Sports Exerc.
11. Thompson, A. M., P. D. Campagna, L. A. Rehman, R. J. Murphy, R. L. Rasmussen, and G. W. Ness. Physical activity and body mass index in grade 3, 7, and 11 Nova Scotia students. Med. Sci. Sports Exerc.
12. Troiano, R. P. A timely meeting: objective measurement of physical activity. Med. Sci. Sports Exerc.
13. Trost, S. G., K. L. McIver, and R. R. Pate. Conducting accelerometer-based activity assessments in field-based research. Med. Sci. Sports Exerc.
14. Trost, S. G., D. S. Ward, S. M. Moorehead, P. D. Watson, W. Riner, and J. R. Burke. Validity
of the computer science and applications (CSA) activity monitor in children. Med. Sci. Sports Exerc.
15. Tryon, W. W., and R. Williams. Fully proportional actigraphy: a new instrument. Behav. Res. Methods Instrum. Comput.
16. U.S. Department of Health and Human Services. Physical Activity and Health: A Report of the Surgeon General
, 1st ed. McLean, VA: International Medical Publishing Inc., pp. 144-145, 1996.
17. Ward, D. S., K. R. Evenson, A. Vaughn, A. B. Rodgers, and R. P. Troiano. Accelerometer use in physical activity: best practices and research recommendations. Med. Sci. Sports Exerc.
18. Welk, G. (Ed.). Physical Activity Assessments for Health-Related Research
. Champaign, IL: Human Kinetics, pp.125-141, 2002.
19. Welk, G. J. Principles of design and analyses for the calibration of accelerometry-based activity monitors. Med. Sci. Sports Exerc.
20. Welk, G. J., J. A. Schaben, and J. R. Morrow, Jr. Reliability of accelerometry-based activity monitors: a generalizability study. Med. Sci. Sports Exerc.
21. Wood, T. M. Issues and future directions in assessing physical activity: an introduction to the conference proceedings. Res. Q. Exerc. Sport