Registration and protocol

This systematic review was reported in accordance with the preferred reporting items for systematic review and meta-analysis protocols (PRISMA-P) 2015 statement [22]. This review has been registered with PROSPERO [23] under the record number CRD42017076786 on 31^st of October 2017. The protocol for this review has been published in protocols.io in order to enhance the reproducibility of the results. The protocol can be accessed from: http://dx.doi.org/10.17504/protocols.io.tsvene6.

Literature search strategy and assessment

The study considered any intervention which relates to physical activity and step counts. To find the relevant research articles, an individualized search strategy was built for the PubMed, Web of Science and PsychINFO databases. This literature search was limited to the journal articles which were published from January 2000 to August 2017. The above databases all show an increasing trend in physical activity research from 2000 onwards, making research prior to this date of less relevance. Moreover, the usage of pedometers or accelerometers in physical activity research has also particularly increased since 2000 due to technological advancements. Because of time constraints it was decided to include only articles published in English. It was decided to use individualized key words in the search strategy for the above mentioned databases, in order to extract more relevant articles. It should be noted that for Pubmed and PsycINFO, some of the key words were MeSH terms and this helped to ensure more relevant articles. However, for the Web of Science database, it was decided to use defined key words, since this database does not have a library indexed language. The following list denotes the series of key words used for each database.

PubMed

accelerometer OR pedometer OR step count* OR “Fitness Trackers”[Mesh] AND statistic* OR statistical model* OR quantitative OR “Models, Statistical”[Mesh] OR “Data Mining”[Mesh] OR “Data Interpretation, Statistical”[Mesh] AND “Exercise”[Mesh] OR “physical activity” OR fitness OR Program* OR intervention.

PsycINFO

accelerometer OR pedometer OR step count* OR “Fitness Trackers” AND statistic* OR statistical model* OR quantitative OR “Statistical Analysis”[Index Term] OR “Data Mining”[Index Term] OR “Data Interpretation” AND “Exercise”[Index Term] OR “physical activity” OR fitness OR Program* OR intervention.

Web of Science

accelerometer OR pedometer OR step count* OR “Fitness Trackers” AND statistic* OR statistical model* OR quantitative OR “Data Mining” OR “Data Interpretation” AND “Exercise” OR “physical activity” OR fitness OR Program* OR intervention.

Inclusion and exclusion criteria

To select articles better matched to the objective of this systematic review, five main inclusion criteria were developed. Articles were selected if (i) the study included data taken from a physical activity program or an intervention; (ii) the data collection did not impact the usual natural lifestyle of the participants (e.g. studies where data was collected through walking on a treadmill or from a walking test were excluded); (iii) the study needed to be carried out with at least one group of healthy participants; (iv) the study needed to include a study variable related to “step counts” and this data needed to be collected using a device such as a pedometer or accelerometer; (v) the study needed to carry out a quantitative analysis using “statistical modeling”.

Selection process and data extraction

The initial database search was able to identify a total of 5133 articles which included 1817 articles from Pubmed/Medline, 2066 articles from PsycINFO and 1250 articles from Web of Science. These articles were uploaded to EndNote software and 881 duplicates were removed by the first author. The other two authors were then given online access to view the results. The first two authors then independently screened the titles and abstracts against the inclusion criteria and shortlisted the most relevant articles. The articles which had been accepted by both reviewers were shortlisted and the articles which had questionable eligibility were directed to the third reviewer for resolution. This process shortlisted 157 articles for a full text review, out of which the team agreed on 78 unique articles for inclusion in this review. The following PRISMA flow chart (Fig 1) outlines the above search and review process.

Download:

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

Fig 1. PRISMA flow diagram.

https://doi.org/10.1371/journal.pone.0206763.g001

Quality assessment

The Downs and Black assessment tool [24] for health care interventions was used to measure the study quality of the included articles. This tool is appropriate for assessing both randomized and non-randomized studies of health care interventions. This checklist consists of 27 items distributed among five sub scales, namely; reporting (10 items), external validity (3 items:), bias (7 items:), confounding (6 items:) and power (1 item:). According to the tool, the maximum score that a study can obtain is 32, with higher scores indicating better quality. In this review the Downs and Black scores have been categorized as follows; high quality (22-25), moderate quality (18-21) and low quality (14-17).

Taxonomy for statistical techniques

There are many different statistical techniques that can be used to address the aims and objectives of a study with the available data. Therefore, the “guide for choosing among statistical techniques” published in [25] has been used to classify the various techniques found in the shortlisted articles.

Risk of bias assessment

A risk of bias assessment was carried out for all the 78 studies using the Downs and Black study quality assessment tool. The quality scores ranged between 14 and 25 with a mean score of 19.7. Four studies were categorized as high quality (22-25), 63 studies as moderate quality (18-21) and 11 studies as low quality (14-17).

Family, community and sociodemographic scope

It was noted that 31% of the family, community and sociodemographic studies were classified as “effect ON step counts”. Out of these studies 55% used general linear models while 41% used generalized linear models, with one using multilevel modeling. The majority of the general linear model and generalized linear model studies had a cross sectional design (75% and 89% respectively), while the study which used multilevel modeling had a cluster randomized design.

The majority (58%) of the general linear model analyses used ANOVA or linear regression. Several socio demographic studies examined the significance of predictor variables such as age, sex, marital status, employment status and race, as well as the interactions of these variables, in terms of physical activity [9, 33, 38–42]. These studies considered participant samples ranging from 8 years to 83 years. It was found that for several studies age and gender were significant predictors of physical activity [33, 38, 40]. Moreover it was also found that the use of new technological devices such as physical activity watches, was more likely to increase peoples’ physical activity [42]. Ordinary least squares regression models were also used to examine the association between income and physical activity [43], and multivariate linear regression models were used in order to test the relationship between parental beliefs and support for the physical activity levels of children, measured using pedometers [34, 44]. However, the effect of climate and weather and neighbourhood environment are also important determinants of physical activity levels. The influence of neighborhood walkability has been tested by estimating the mean differences in steps per day across neighborhoods with different walkability scores, finding no significant relationship [45]. A similar type of analysis has been performed using gender stratified multilevel modeling to analyse the association between neighborhood safety and pedometer measured physical activity, allowing for within neighbourhood clustering [46]. This study used a randomized cluster design with 12 urban public housing communities as the primary sampling units and individuals within these communities as secondary sampling units. The authors have considered day-time safety and night-time safety separately.

All the studies which used generalized linear regression methods have used logistic regression models to investigate significant sociodemographic effects for the achievement of various levels of physical activity [33, 47–51]. For these analyses variables such as city of residence, dog ownership, motor vehicle ownership, education level, occupation type, self rated income level, usual daily occupational or domestic activity level and transport mode to work have been considered as predictors. It was found that the odds of being somewhat active (7500- 9999 steps/day) as opposed to being sedentary to low active (<7500 steps/day) required “good” self-rated health as well as ownership of a dog. Being active to highly active (> 10,000 steps/day) as opposed to sedentary to low active (<7500 steps/day), was related to participant’s city of residence, “good” self-reported health, not owning a vehicle, having a dog and being employed [47]. In another study [49] the data was dichotomised for ≥10,000 steps/day and the logistic regression was run using demographic and individual factors as explanatory variables for both males and females separately. From the analysis it was found that men who were less likely to achieve ≥ 10, 000 steps/day were more likely to be at least 60 years of age and overweight, while men who regularly walked in the workplace, who did vigorous activity at work or were employed in a blue collar occupation were more likely to achieve the above target.

Logistic regression models were also used to investigate whether the parent’s achievement of a daily step count goal, and non-excessive screen time on weekdays and weekends, were associated with children’s achievement of their daily step count goals [34, 50]. In Sigmundova’s study [50], the odds ratios were derived for meeting the recommended level of child step counts on weekdays and weekend days, separately for mother-child and father-child pairs, while controlling for the children’s and parent’s age and overweight/obese condition. It was found that on weekdays, the children are six times more likely to achieve their recommended level of step counts, where the daily kindergarten routine accounted for more than the median step count. Similarly, it was found that children are 10 times more likely to achieve the recommended level of step counts on weekends when their mothers achieve 10,000 steps per day or more.

Interventions

Thirty six percent of the “effects ON step counts” studies have analysed effectiveness of an intervention using various statistical methods. Most of the studies (72%) out of 25 have used general linear models. It was noted that more than 50% of the studies using general linear models were analysing data for randomized control trials, followed by quasi-experimental, cross sectional and longitudinal designs. All the studies which used multilevel modeling involved randomized control trial designs.

It has been confirmed that well-structured pedometer walking programs with goals and proper physical activity consultations are very effective in promoting and changing the walking behavior in various communities [18, 52–54]. In one study [52] participants were randomly assigned to four groups, where group 1 and 2 had initial goals of 5000 steps per day and group 3 and 4 had initial goals of 10000 steps per day. Half-way through of the intervention groups number 2 and 4 were given new goals of 10000 steps per day and 5000 steps per day respectively. It was found that less walking behavior was observed when less ambitious goals were assigned regardless of whether the goals were met. However, a reduction in step goals also resulted in reduced walking behavior. It was found that pedometer-based physical activity interventions with counseling were effective for community-based as well as for work based individuals [13, 15, 16, 18]. These interventions promoted physical activity as well as reduced sitting time [16]. Similarly, it was found that creating awareness of the importance of physical activity within a pedometer based intervention for adolescents was also helpful in promoting physical activity [53]. The effectiveness of a competition based employer-sponsored physical activity program has been analysed using a one-way repeated measures analysis of variance with Duncan’s post hoc tests to examine in which program weeks significant changes occurred [55]. It was found that step counts tended to increase significantly in the middle of the 12 week program [55]. Apart from ANOVA, several studies have used analysis of covariance (ANCOVA) [4, 13, 13] and multivariate analysis of covariance (MANCOVA) analyses [4]. However, it should be noted that these models do not accomadate missing data [56]. If there is any missing data for an individual, then the entire data vector of that individual needs to be removed from the analysis [56]. Moreover, generally it is observed that the effectiveness and success of a physical activity program depends on the strategies and components included in the program. The long term effects of these strategies, helping participants to maintain levels of physical activity and health outcomes even after the program ends, are also very important. However, one study which examined a 4-year follow up effect for a 12 month pedometer based physical activity intervention found no support for this hypothesis, when using a repeated measures analysis of variance [17].

Apart from ANOVA, linear regression models have been used to examine the effect of environment on changes in pedometer recorded step counts during an intervention [20, 21]. An inverse association was found between density of private gyms and pedometer measured steps per day, such that those who live near higher density gym areas had a lower increase in step counts from baseline to six months [20] when controlling for potential confounders including age, race, education and sex. In the second study [21], the influence of the neighborhood environment on step counts during an intervention was assessed using hierarchical multiple linear regression models, while controlling for demographic variables. In the 12-month intervention one group was provided with a map of local areas to promote walking whereas the other group was only provided with pedometers to count their steps. Principal component analysis was used to reduce sixty-nine environment variables to 8 principal component factors, explaining 80.7% of the variability. The demographic variables were added in block 1 and 8 environmental factors including dangerousness of roads in block 2. Different environmental factors were found to be significant at different times (baseline, 3 months, 9 months and 12 months), however, none of the demographic variables (age, gender, annual household income and SIMD rank) were significant at any time [21]. This suggests that neighborhood environment is also important when promoting physical activity through an intervention.

Multilevel models were used to identify significant predictors for differences in step count changes for intervention and control groups [7, 30, 57, 58]. A feasibility study on classroom-based physical activity was analysed, adjusting for within cluster correlation at three different levels, namely student, classroom and school, using random effects for school, classroom and subject and assuming fixed effects for grade, gender, outdoor recess and physical education [7]. The effectiveness of goal manipulation along with social comparison feedback within an intervention have also been tested using multilevel models [57]. The results of this study agree with Anson’s study [52], confirming that there is an effect on physical activity behavior from social comparison feedback.

In two studies [30, 59] it was predicted that the percent of days on which step counts were logged would be higher during intervention periods than control periods, while controlling for age, physical activity level (life style index) and gender. The authors also used a multilevel model and a generalized estimating equation to predict the program’s participation [30]. These models have more flexibility in managing missing data as well as any unbalanced or nested structure in the data [56].

Personal and psychological behavior

This Personal and Psychological Behavior category includes 14% of the “effect ON step counts” studies. All the studies which used structural equation modeling (SEM) fall in this category, with the rest using general linear models, generalized linear models and multi level models. In this category, most of the studies (80%) are cross sectional while only two studies that used structural equation models, have longitudinal designs. The four studies which used structural equation models relate to school-based activity programs [37, 60–62].

Structural equation modeling (SEM) has become a popular data-analytic approach among psychologists [63]. These models accommodate path analysis for measured variables [64] and latent constructs, reflected in their manifest indicators using measurement model [65]. The direct and indirect effects of variables such as self efficacy, social support, parental influence, environment, intrinsic and extrinsic goals and autonomous and controlled motivation on pedometer based physical activity have been tested [37, 60, 61]. In order to test the relationship between physical activity goals and autonomous controlled motivation and also to test the mediation effect of autonomous and controlled motivation for the relationship between goal content and physical activity, the sample was categorized on puberty status (physically immature vs physically mature), physical goal content (intrinsic goals vs extrinsic goals) and motivation for being physically active (autonomous motivation vs controlled motivation). It was found that there is no significant direct path from goal content to physical activity level. However, participant’s autonomous motivation acted as a mediator between goal content and physical activity level [37]. Similarly, SEM has been used for a three-wave prospective design which collected data at three different time stamps. This model was used to examine the positive effect of autonomous motivation towards exercise on predicted health-related quality of life, physical self-concept and the number of steps taken [62]. The model building process was comprised of two stages, a confirmatory factor analysis to test the measurement model followed by a path analysis [62]. Using path analysis, self efficacy was found to have direct and indirect effects on physical activity behavior, with parental influence affecting physical activity behavior directly for both parents and children [60]. However, SEM is also currently used for other important types of analysis such as longitudinal analysis, meta analysis, biometrical genetics, multivariate survival and spatial analysis, suggesting that more use of SEM methods is likely in the future [66].

A multivariate analysis of variance was used to estimate the main and interaction effects for gender, pubertal and weight status on goal content for physical activity behavior [37]. When considering general linear models, hierarchical linear regression was preferred by the researchers for testing the effects of planned behavior and other related psychological factors on physical activity [67, 68]. It was found that planned behavior predicts the intention to walk but not actual pedometer based step counts [68].

Finally, logistic regression was used to examine the effects of intention and actual television viewing on the physical activity of youths [35, 69]. Both these studies found that more hours watching television was associated with reduced step counts.

Weather effects

Physical activity behavior is affected by the weather and seasonal patterns. The four studies which were categorized under this heading included a longitudinal study design, with two studies utilizing analysis of variance and the other two studies using a linear mixed modeling technique to analyse the data. In order to test the overall effect of seasons on step counts along with sociodemographic variables, ANOVA was used [70, 71]. It was found that, both normal and overweight participants living in the UK reduced their step counts in winter, with normal weight individuals more influenced by this seasonal change [70]. Furthermore, a significant day of the week effect (especially high for Sunday) existed in both summer and winter periods of the year [70, 71]. These studies considered daily weather variables for temperature, rainfall, relative humidity, wind speed, sea level pressure, snow on ground etc. along with demographic variables [72, 73].

The other studies found using multilevel models that the effect of temperature and humidity were both significant positive determinants of step counts while rainfall, snow depth and wind speed had a significant negative relationship with daily step counts [72, 73]. A significant interaction effect for maximum wind speed and BMI was also found, with maximum wind speed having more of an effect on individuals with lower (BMI = 20) and higher BMI’s (BMI = 35) [73]. Clearly, all these results suggested that the effect of weather and climatic conditions need to be considered when promoting physical activity interventions.

Health

24% of all the papers were categorized as health related studies. This category can be further divided into “effect ON step counts” and “effect OF step counts”.

When considering the “effect ON step counts” one longitudinal study used multilevel modeling [74] to analyse the data. This study predicted the change in physical activity of weight conscious college women. For this analysis body satisfaction and reported eating behavior measures were considered as predictors in two models, one to examine the retrospective and the other the prospective relationship between physical activity and body/eating behavior. It was found that weight conscious college women tend to show a increase in physical activity level after a negative weight experience or negative eating behavior.

All the other health related studies were categorized under “effect OF step counts”. Only general linear models (52%) and generalized linear models (48%) were used in these studies. Most of the studies were cross sectional in design with two longitudinal and one non randomized controlled trial. Linear regression and ANOVA general linear models were used. Logistic regression was the most common (80%) generalized linear model approach, followed by exponential regression and log-binomial regression.

Logistic regression has been extensively used to predict the presence or absence of health related outcomes such as adiposity measures [28, 75], coronary heart disease [36, 76–78], bone and muscle health [27, 29], based on physical activity levels. In all these studies physical activity was measured by step counts and used as a predictor along with other relevant independent variables. In ANOVA analyses, step counts were divided into tertiles [36, 79–81], whereas in linear regression, correlations with step counts were considered [19, 27–29, 82]. Seniors who walked more, maintaining the recommended level of physical activity, have less risk of sarcopenia [27, 29]. Also more steps per day lowered the odds of having Metabolic Syndrome (MetS) [36, 77, 78]. Log binomial is another modeling method that can be utilized depending on the outcome variable and objectives of the study. In these models, a log function is used as the link function instead of a logit function. Log binomial regression is able to produce an unbiased estimate for the adjusted relative risk [83]. These models can be used to derive the risk ratios instead of odds ratios when the outcome event is rare. One study used the log binomial regression to investigate the association between physical activity (steps/day) and depression [84]. According to this study, women with moderate ambulatory levels (>7500 steps/day) had a 50% lower prevalence of depression compared with sedentary women (<5000 steps/day). Hierarchical linear regression was also used to evaluate the relationship between cognition and physical activity [85]. In this study cognitive performance was measured by several scores including verbal episodic memory and visual episodic memory etc.

Other- Miscellaneous

Only eight articles (10%) were categorized in this Other-Miscellaneous category. Most of these studies had a cross sectional design, with some longitudinal and randomized control trials. All these articles used ANOVA and linear regression.

Based on physical activity recommendations [81, 86], linear regression was used to relate step counts to 30 minutes of moderate to vigorous physical activity level per day.

Similarly maximum oxygen usage during high intensity exercise was predicted using a linear regression model based on step counts and other related body composition measures [87]. Other studies examined the influence of monitoring interval and alternative starting days for step count interventions in order to reliably assess and predict the participants’ physical activity [31, 88]. ANOVA, inter class correlation and linear regression methods were used to model the collected data. One study reported that a minimum of four days need to be monitored starting from Sunday [31]. Another study [88] also claimed that four days were required for a pedometer based intervention, with different monitoring periods for different step count collecting devices.