Introduction

A growing number of epidemiologic studies use device-based measurements to quantify physical activity and sedentary behavior (e.g., physical behaviors), such as with accelerometers or activity trackers [1]. These devices can measure movement up to multiple times per second using raw data (e.g., 10–100 Hz) or summarize data into short windows of time called epochs (e.g., per minute or 15 seconds) [2]. Using epochs, authors classify the data into physical activity intensity levels (e.g., light, moderate, or vigorous) and sedentary behavior, and use monthly, weekly, or periodic sums to provide time in each category. While these approaches are effective in capturing multiple time frames, this grouping does not capture differences in patterns of accumulated physical activity and sedentary behavior over time [2]. The application of latent class analysis (LCA) to accelerometry can help elucidate patterns of physical activity and sedentary behavior over fixed time periods.

LCA is a measurement modeling technique that assumes the presence of underlying, unobserved, mutually exclusive categories which result in different patterns of variables observed in the data [3]. Using LCA, participant data on one or multiple characteristics can be partitioned into classes with other participants that share a similar pattern or behavior [4]. LCA can identify common classes derived through statistical modeling that otherwise might not have been discovered. This distinction has been described as a data-driven approach (i.e., using the data to identify and group people based on different patterns) in contrast to grouping people based on whether they meet certain criteria determined a priori [3]. LCA allows researchers to identify unique groups, which might not have been identified using a priori classifications, and to investigate how these unique groups vary in terms of participant characteristics or health outcomes.

LCA can be used to both identify and classify individuals into different subgroups that share similar patterns of physical activity and sedentary behaviors. One example is the use of LCA to identify the weekend warrior pattern of lower physical activity on weekdays and higher physical activity on weekends [5]. LCA can also be used to classify subgroups of individuals, based on differences in their patterns of physical activity and sedentary behavior data, and explore whether these subgroups differ by various correlates. LCA also incorporates many important aspects of physical activity and sedentary behavior, including frequency and time. The ability to identify behavioral patterns beyond cumulative time in order to maximize the variance explained is a notable advantage over traditional methods [6].

Multivariable analysis of variables used to fit the LCA model, such as accelerometry-assessed physical activity and sedentary behavior, is important in expanding the use of LCA for future research. To our knowledge, no published review discusses methods for examining accelerometry-assessed physical activity and sedentary behavior using LCA among both youths and adults. Therefore, we conducted a scoping review to systematically summarize how LCA has been applied to accelerometry-assessed physical activity and sedentary behavior in order to describe current research practice and identify methodologic gaps to aid future research using these methods.

Methods

We conducted a systematic scoping review according to the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) Extension for Scoping Reviews [7]. The completed PRISMA checklist can be found in table S1 Appendix.

Search strategy

We searched PubMed/Medline, Web of Science, CINHAL, SportDiscus, and Embase through December 31, 2021. The search terms used in these five databases can be found in table S2 Appendix.

The terms were entered based on previous review papers [8–10] as well as substantive experts from our team and searched using the advanced search field in each database. We downloaded results from each database as an RIS file type to import into Sciwheel, a reference management system. After duplicates were removed, search results from Sciwheel were transferred to Covidence for screening of titles/abstracts and full text. Reviewers were given the inclusion and exclusion criteria to aid in the screening and abstraction process. The screening and selection process were independently conducted between two researchers, with discrepancies resolved by consensus. A third reviewer provided a final check of all abstractions.

Results

Descriptive characteristics of the included papers

The literature search captured 2555 records from five databases. Sixty-six full-text papers were assessed for eligibility and 12 papers from 8 unique studies were included in the final analysis (Fig 1). Papers were included from a range of countries including the United States (n = 8) [2, 5, 6, 13–17], Australia (n = 2) [18, 19], the Netherlands (n = 1) [20], and Brazil (n = 1) [21] (Table 1). All included papers used cross-sectional study designs except one which was a prospective cohort study that collected accelerometry at one time point [2]. Among the 12 included papers, 5 used the adult National Health and Nutrition Examination Survey (NHANES) data [2, 5, 6, 14, 15], and 2 used the youth NHANES data [13, 17].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. PRISMA flow diagram of the literature search and study selection process.

https://doi.org/10.1371/journal.pone.0283884.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive characteristics of included studies (N = 12 papers).

https://doi.org/10.1371/journal.pone.0283884.t001

Data collection years for the studies ranged from 2003/2004 [15] to 2018 [21] and the sample size ranged from 217 [21]- 7931 [5] (Table 1). Ten of the 12 papers [2, 5, 6, 13–15, 17–19, 21] ensured that their sample was representative of their target population. For example, the NHANES data sampled the civilian, non-institutionalized US population [2, 5, 6, 13–15, 17]. The percentage of females in the studies ranged from 48.8% [17] to 58.6% [18].

Accelerometry and latent class results for the included studies

Focusing on the 8 unique studies, 4 studies included youth (<18 years) participants [13, 16, 18, 21], with age ranging from 6 [13] to 17 [13] years. Four studies included adult participants [5, 14, 19, 20], with age ranging from 18 [5] to 65 [20] years. These eight studies measured physical activity and sedentary behavior using either combined accelerometer and self-report in the LCA (n = 3) [16, 18, 21] or accelerometry only (n = 5) [5, 13, 14, 19, 20] (Table 2). Studies used either the ActiGraph 7164 accelerometer (n = 4) [5, 13, 14, 16] or the ActiGraph GT3X+ (n = 4) [18–21] (S4 Appendix). The accelerometer was worn on the hip (n = 5) [5, 13, 14, 19, 21], waist (n = 1) [18], or was not reported (n = 2) [16, 20].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Description of the accelerometer-related measures (N = 8 studies).

https://doi.org/10.1371/journal.pone.0283884.t002

Physical activity and sedentary behavior patterns were examined on various temporal scales including by day of week [5, 13, 14], weekend vs weekday [16], hour-by-hour [20], weekly average [21], dichotomized above or below the median number of minutes per day for the specific population [18], and derived sex-specific z-scores for physical activity and sedentary behavior variables [19]. Participants were instructed to wear accelerometers during waking hours (n = 6) [5, 13, 14, 16, 18, 20] or 24 hours (n = 2) [19, 21]. The number of accelerometer wear days reported was 7 days (n = 6) [5, 13, 14, 16, 19, 20] or 8 days (n = 2) [18, 21]. The recorded accelerometer epoch length varied from 15 seconds (n = 1) [21] to 60 seconds (n = 4) [5, 13, 14, 19] or was not reported (n = 3) [16, 18, 20]. Table S5 Appendix describes in detail the accelerometer derived variables and identified classes/cluster for all 12 unique studies.

The maximum number of classes considered ranged from up to 5 classes (n = 2) [18, 21] to up to 12 classes (n = 1) [5] (Table 3). However, despite the wide range explored, the number of final LCA classes was either 3 classes (n = 4) [16, 18, 20, 21], 4 classes (n = 1) [13], or 5 classes (n = 3) [5, 14, 19]. Researchers used a variety of criteria to guide the selection of the final of number of classes and model fit for model selection. The most common criteria to select the final best fitting model included the Bayesian Information Criteria (n = 5) [16, 18–21], substantive knowledge (n = 5) [5, 13, 14, 16, 19], sample size of classes (n = 4) [5, 13, 14, 18], and Akaike Information Criteria (n = 4) [16, 18, 19, 21]. Software used for the LCA analysis included Mplus (n = 4) [5, 13, 14, 18], Latent Gold (n = 2) [19, 20], SAS (n = 1) [16], specifically PROC LCA in SAS, and R software (poLCA package) (n = 1) [21].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Description of the latent class analysis (N = 8 studies).

https://doi.org/10.1371/journal.pone.0283884.t003

Three studies adjusted for covariates when developing their latent class models, including sociodemographic [14], health-related [21], and accelerometer measured variables (e.g., total number days of accelerometer) [19] (Table 3). Separate LCA subgroup analysis was performed by sex (n = 3) [13, 16, 19], by school characteristics (in school or out of school) (n = 1) [13], and by age (n = 1) [13]. In terms of specifying the benefits of LCA, 5 studies reported utilizing LCA to identify unique classes [5, 13, 14, 16, 19], 1 study reported LCA as a data reduction tool [18], and 4 studies reported using LCA to combine multiple physical activity variables [18–21].

In this review of 12 studies, several explored associations between LCA-derived variables with outcomes. Two youth studies explored the association between the physical activity and sedentary behavior latent classes to mental health [21] and physical health [21] outcomes, and cardiovascular disease risk factors [17]. Two adult studies explored associations between the LCA derived physical activity and sedentary behavior variables with all-cause mortality [2] and risk factors for metabolic syndrome [15].

Quality assessment

The quality assessment tool, consisting of 16 questions, as applied to the 8 studies is provided in table S3 Appendix. Few studies reported their missing data mechanism (n = 3, question 1) and variables related to missingness (n = 2, question 2) (Table 4). Two studies considered covariates or the inclusion of the covariates into the LCA model (n = 2, question 7). Two studies included information about final replications (n = 2, question 9). None of the studies reported about the number of random start values (n = 0, question 8). Three studies made their supplementary files available (n = 3, question 16).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Quality assessment tool assessing LCA in the included studies in publication date order (N = 8 studies).

https://doi.org/10.1371/journal.pone.0283884.t004

Discussion

This systematic scoping review examined LCA modeling strategies using accelerometer-assessed physical activity and sedentary behavior. We modified and applied an existing quality assessment tool to provide important information on the quality and reporting of these LCA studies. We identified 12 papers from 8 unique studies using LCA methodology to accelerometry or combined accelerometry and self-report.

Our review focused on accelerometry-assessed physical activity and sedentary behavior using LCA, which provides important insight into how valuable LCA can be for this type of data. While studies in this review used different combinations of physical activity and sedentary behavior measures, including differences in cut points, all involved some degree of summarizing more detailed higher dimensional accelerometer-derived data using latent class analysis. This highlights one of the strengths of LCA, to identify physical activity variation across a wide range of movement profiles and summarize this into a smaller subset, in the case of LCA into the latent classes.

LCA used both physical activity and sedentary behavior data to generate classes [16]. Additionally, several studies found that classes differed by sex [13, 16, 19] and by age [13], thereby providing insights into correlates with LCA patterns over a particular time frame. One study [8] explored the use of LCA to identify patterns of physical activity and sedentary behaviors, and while this study did not focus exclusively on accelerometry-based data, they also found differences in the latent classes by sex, age, and school characteristics. DeFaria et al. [21] found that females with mental health disorders were more likely to belong to inactive classes. Metzger et al. [15] reported that metabolic syndrome risk factors were lower among classes with more moderate-to-vigorous physical activity.

Using these two examples, identifying specific classes might provide important insight into interventions addressing these outcomes.

Understanding class membership is important in order to develop interventions that can be tailored to groups. Including individual covariates (e.g., sociodemographic) in LCA models to help predict class membership has been shown to improve model interpretation, identification, and fit [23]. Wurpts et al. [23] suggests that the use of a larger number of high-quality variables and the inclusion of at least one covariate under a strong theoretical basis positively affect LCA model estimation, and that these factors can sometimes compensate for other suboptimal conditions (e.g., a relatively small sample size). In our review, about half of the unique studies reported including covariates in the model.

Almost all studies in this review listed their criteria to guide the selection of the final number of classes and models fit. The most common criteria used to test the relative fit of models were BIC, AIC, BLRT, substantive knowledge including practical interpretation of what each class represented, visual inspection (to ensure classes were sufficiently separated from each other or large enough to be of the researcher’s interest), and entropy. Multiple criteria in the selection of final classes and model fit such as large sample sizes, AIC, BIC, and entropy are recommended to detect minor differences in class samples [8]. Entropy is frequently used in studies to quantify how well participants are classified into latent classes and to quantify the quality of classification and separation between groups in a log based model [24]. Hence, we included entropy in our quality assessment tool to evaluate its adoption in the reviewed studies.

The quality assessment tool identified areas that generally needed further clarification such as variable missingness, description of covariates (if used), number of final replications/iterations used, and total number of fitted models reported. A majority of the studies clearly stated the study objective in relation to preferences and described methods and procedures. LCA models assume data are missing at random and missing data on a covariate or grouping variable can present a complex issue, therefore, confirming this assumption prior to the analysis is important [3]. As the application of LCA in physical activity and sedentary behavior expands, investigators should improve the quality and reporting of LCA studies. This can be achieved by addressing missing data in the LCA and explain how and where the missingness is attributed. Moreover, future studies should describe model selection criteria, the final number of classes which will allocate sufficient sample size in each class, and covariates included in the model. Utilizing an appropriate checklist when writing the methods and presenting the results. For this study we expanded upon an existing checklist for LCA [9] and a checklist for latent trajectory studies is available elsewhere [25].

Conclusions

The aim of this scoping review was to describe methods and procedures used in applying LCA to accelerometry. Across a variety of study populations, LCA identified unique subgroups of physical activity and sedentary behavior, and these unique classes were associated with both health outcomes and individual characteristics. LCA was used to identify unique groupings as a data reduction tool, to combine self-report and accelerometry, and to combine different physical activity intensities and sedentary behavior in one LCA model or separate models. This review provided insight into the reporting of LCA in accelerometry studies and identified areas of improvement for future studies leveraging LCA.

Supporting information

Acknowledgments

The authors thank the University of North Carolina–Chapel Hill librarians for their assistance in the literature search of this review.

Patient and public involvement: Patients and/or the public were not involved in the design, conduct, reporting, or dissemination of this research.