A comprehensive, unbiased search strategy was developed to identify all relevant studied published in English language regardless of the age, gender, and country in PubMed from inception (1966) to February 2007. Broad search terms were used so as not to miss any potentially relevant articles during the search procedure. Detailed search strategies can be seen in Figure 1. In addition, we searched references of all retrieved articles and systematic review to ensure that all eligible studies had been identified.

Figure 1

Figure 1: Review question, eligibility criteria and search strategy

The selection criteria for inclusion were:

• Observational or cross-sectional studies evaluating associations between body composition and physical activity levels;

• Measure of habitual physical activity was included;

• Measure of body fat/relative weight was included;

• Measure of activity did not follow an intervention which might alters habitual daily activity; and

• Measures of fat and activity were taken at the same time. We excluded validation studies, review articles, and qualitative studies.

Abstracts were scanned by the two of authors independently to exclude those out of scope. Discrepancies were discussed until consensus was reached. The full text of each remaining paper was then viewed and papers were again excluded with the consensus of both authors. Finally, the reference lists of all remaining papers were scanned. The selection of studies from the reference lists followed the same steps as outlined above.

Data were extracted by the first author using a specially designed electronic data extraction form and create a database using Epi-Info ₉ , and were checked for accuracy by a second author. All disagreements were resolved by consensus and discussion. Information entered into the database included: study citation, year of publication, study settings, methods of body composition assessment, and main findings. The data extraction form had a semi-structured format. Each variable was allotted one field of the data extraction form, and the field expanded as needed to accommodate the text.

Data are presented as figures, numbers and simple frequencies processed with Stata 9.0 software ₁₀ (Stata Corporation). Descriptive statistics and qualitative information are presented when appropriate. A tabular evidence profile was prepared for all the eligible studies. For study location, we recorded the geographic coordinates for each study and used Stata ₁₀ spmap thematic mapping package to plot the position of each study.

The literature searches yielded 2474 titles of potentially relevant articles (Figure 2). After scanning titles and abstracts, 75 potentially relevant articles were identified. Three studies ₁₁,₁₂,₁₃ could not be retrieved. The reference lists of the remaining 72 selected articles were scanned, which resulted in another four publications for inclusion. Eighteen articles were excluded because they were either review articles (n=4), qualitative (n=2), fitness test (n=6) or validation studies (n=6). Fifty eight articles met inclusion criteria and included in this review (Table 1). These papers and their findings are described below.

Figure 2

Figure 2: Literature flow

Figure 3

Table 1: Summary of studies included in the review

AEE: activity-related energy expenditure, BF: body fat, BIA: Bioelectric impedance analysis BMI: Body mass index, CI: confidence interval, DLW: doubly labeled water, DXA: Dual energy X-ray absorptiometry, HRM: heart rate monitor, EE: energy expenditure, FFM: free fat mass, IC: indirect calorimetry, MPA: moderate physical activity level, MRI: magnetic resonance imaging, MVPA: moderate-vigorous physical activity level, OR: odds ratio, PA: physical activity, PAI: physical activity level index, PAL: physical activity level, PAQ: physical activity questionnaire, SF: skin fold, VPA: vigorous physical activity level, WC: waist circumference, WHR: waist hip ratio, §: finding in male, but not female, #: finding in female, but not male, ¢: Null finding.

Our search identified 58 studies ₈,^14-70 evaluating the association between body fatness and level of physical activity. All studies had cross-sectional design. Of the 58 published papers, 74% (n=43) were published after 2001, 19% (n=11) were published between 1996 and 2000; only 7% (n=4) were published before 1995. We found one study that was published in 1968 ₄₇ . Seventeen had sample size of less 100 participants, 16 had 100 - 500 subjects, six had 500-999 subjects, 14 had 1000-4999 participants, and only five had more than 5000 participants. The sample size ranges from 24 ₆₆ to 137, 593 ₃₅ with mean of 3,888 participants. In total 33% (n=19) of samples were preteen (less than 10y), another 33% (n=19) were adolescents (10-19y), with the remainder being between 19 - 70y (28% [n=16]), or a combination ₃₃,₃₆,₅₈,₆₂ (7% [n=4]).

Most studies 32% (n=19) were carried out in the USA ₁₈ (see Figure 3) ₂₄,₂₅,₂₆,₂₇,₃₁,₃₇,₄₄,₄₅,₄₆,₄₇,₄₈,₅₂,₅₄,₅₅,₅₆,₅₉,₆₆,₆₈ . The rest were either from UK ₂₀,₂₂,₃₀,₆₃,₆₅ , Australia ₃₉,₄₉,₆₀,₆₄,₇₀ (both n=5); Sweden ₃₈,₄₁,₆₉ , France ₃₄,₄₃,₆₇ , Portugal ₂₉,₃₂,₆₂ (all n=3); Belgium ₈,₅₀ , Canada ₃₃,₄₀ , Swtizerland ₁₉,₂₃ (all n=2); or Denmark ₅₃ , Germany ₁₇ , Iran ₃₆ , Japan ₂₁ , Mexico ₄₂ , Slovenia ₁₄ , South Africa ₅₁ , Spain ₁₅ (all n=1).

Figure 4

Figure 3: Studies reviewed according to country

Circles indicate site of the study on geographic coordinates

About half (41% [n=24]) of the 58 studies reviewed used objective measures of physical activity. The others relied on self-reported measures, either by the physical activity questionnaire alone (52% [n=30]) or in combination _{25 48},₅₄ with objective measures (n = 3). Duration of pedometer and accelerometer usage differed across studies. For example accelerometer was worn for three days ₄₆ , six days ₃₇ and seven days ₂₀,₂₆,₃₃ ; and while pedometer was worn for three days ₁₆ and seven days ₅₆ .

Most of the studies reviewed rely on body mass index (BMI=weight in kg/square of height in meters), or on a measure that is based on BMI, to reflect weight status or adiposity. Three of the 58 studies reviewed here use more direct and precise measures of adiposity, such as dual-energy X-ray absorptiometry ₆₆ , ₁₈ O dilution ₆₃ , or percent body fat by bioelectrical impedance analysis ₅₈ alone. Of the 58 studies, six (10.34%) ₂₅,₂₇,₃₁,₃₅,₄₉,₅₉ relied on self-reported height and weight for measurement of body fatness. Classifications of 'overweight' and 'obese' also differed across studies. More than half (55% [n=32]) of the studies reviewed reported the classification of body composition used. Half (16 of 32) studies used CDC cut-off point for Age- and sex-specific BMI percentile and another 14 of the 32 studies classified participant as 'overweight if they were above BMI of 25 kg/m2 and 'obesity' if there BMI were greater than 30 kg/m2. Kagamimori et al ₂₁ used the Kaup index (see equation 1) used for evaluating the degree of obesity at three, obese children were those for whom Kaup's index was 18 or more; and Garaulet et al ₁₅ used BMI; Queletet index of 23 or above for overweight and obese subjects.

Kaup index = weight (g) / (height (cm) x height (cm)) x 10 (1)

Most of the studies controlled for one or more confounders (74.14% [n=43]). Gender (63.79% [n=37]), age (32.76% [n=19]), and race/ethnicity (10.34% [n=6]) were the most common confounders adjusted for during analysis. Most of the studies (88% [n=51]) ₈,₁₄,₁₅,₁₆,₁₇,₁₈,₁₉,₂₀,₂₁,₂₂,₂₃,₂₄,₂₅,₂₆,₂₈,₂₉,₃₀,₃₁,₃₂,₃₅,₃₆,₃₇,₃₈,₃₉,₄₀,₄₁,₄₂,₄₃,₄₄,₄₅,₄₆,₄₇,₄₈,₅₀,₅₁,₅₂,₅₃,₅₄,₅₅,₅₆,₅₇,₅₈,₅₉,₆₀,₆₁,₆₃,₆₄,₆₅,₆₇,₆₈,₇₀ reviewed in this study found a statistical significant correlation between body composition and level of physical activity. Only 24% (n=14) studies reported the effect size of the association as negative correlation coefficient (r), with values from -0.5 ₆₃ to -0.16 ₅₃ (see Figure 4 and 5). Two studied assessed physical activity in form of sedentary behaviours. Deheeger M et al ₆₇ found that TV watching was positively correlated with BMI (r=-0.27, p= 0.01) and Trueth MS et al ₅₅ found that fat mass and percentage fat was positively correlated to time spent in sedentary activity (r=0.47, p<0.001) in girls.

Figure 5

Figure 4: Effect size of the reviewed studies

Figure 6

Figure 5: Effect size and p-value displayed as function of sample size

We performed a systematic review of literature on the association between body composition and physical activity level, using 58 studies primarily from the last 40 years incorporating over 225, 500 subjects.

In our review of these studies, we found, like others ₇₁,₇₂,₇₃,₇₄ , that most cross-sectional studies ₁₄ reported that overweight subjects had lower physical activity levels than their non-overweight peers. Null findings are not uncommon ₂₇,₃₃,₃₄,₄₉,₆₂,₆₆,₆₉ . Some of the studies, however, illustrate the variety of findings: body mass index was negatively associated with physical activity in females ₄₀,₄₃,₅₅ , but not males, whereas fat mass was negatively associated with physical activity in males ₈,₁₅,₂₉,₃₁,₃₂,₃₈,₄₆,₅₈,₆₀,₆₅ , but not females.

The studies reviewed reported relatively moderate to weak (r=-0.5 to r=-0.16), but statistical significant associations between body fatness and physical activity levels. Physical activity explained only 2.5 to 25 percent shared variation in body composition in the subjects. Interpreting the size of a correlation coefficient has always caused difficulty ₇₅ . Several authors have attempted to provide a practical guideline; the most commonly quoted are those advocated by Cohen ₇₆ . However, as Rosenthal and colleagues state in their excellent discussion of this issue ₇₇ , “mechanical labelling … [correlation coefficient] … automatically as 'small', 'medium' and 'large' can lead to a later difficulties,, even 'small' effects can turn out to be practically important”.

Sample size is also crucial, for many reasons. As overweight develops due to small caloric imbalances over time, it is likely that researchers are trying to identify relatively small effects. Must et al ₇₄ comments that “a large sample size is also needed to examine possible interactions, such as that between physical activity and initial body fat status, or between physical activity and gender”. Contrast to this belief, we found that the largest reported effect sizes ([r=-0.50, n=77, p=0.001] ₆₃ , [r=-0.43, n=47, p=0.001] ₇₀ , [r=-0.42, n=34, p<0.05] ₂₂ , [r=-0.43, n=80, p=0.006] ₅₈ ) (Figure 5) had sample size of less than 100. However, the sample size could explain why most of the studies were underpowered and reported results of borderline correlation coefficients. Also this could explain the null findings, especially in studies with the smallest samples ₆₆ . Without an adequate number of subjects, stratification limits the power of the analysis to detect significant differences.

The gender of the subjects in the sample also appears to be important. Twelve studies reported gender dimorphism. For example, Rush et al ₅₈ found that for boys physical activity level were negatively correlated with %body fat (r=-0.43, p=0.006), while for girls no significant correlate was found (r=-0.02, p=0.90). Suter et al ₄₀ found that the sum of 10 skin fold thickness showed a significant negative association with PA in girls (r=-0.38, p<0.01), but not in boys (r=-0.23, ns). The gender difference of the relation between physical activity and body composition is consistent with findings from Westerterp et al ₇₂ . The differences between the sexes are probably based on a difference in the ability to compensate for an increase in energy expenditure. Males tend to compensate less for an increase in energy expenditure by a change in energy intake when body fat stores permit ₇₈ . Thus, exercise is not an effective modality to reduce body fat in females unless accompanied by restriction of energy intake ₇₉ .

Other possible explanations for this gender dimorphism could be related to gender differences in body fat distribution. Males tend to have more abdominal fat than females, and abdominal fat has been shown to be more responsive to exercise interventions ₈₀ . The implications of the results are that female probably does not lose much fat when they adopt a higher level of physical activity ₇₂ . In males on the other hand, as expected, a higher level of physical activity is associated with a lower percentage body fat ₇₂ .

In this review we found that studies that used preteen (<10y) and adolescents (10 - 19y) consistently reports high effect sizes. Nielsen et al ₅₃ was the only one that used adult subjects and reported effect size (r=-0.16, n=783, p<0.05); which is the weakest correlation of all the reviewed studies.

It is important to note that most primary studies have used proxy measures of body fatness such as subcutaneous fat thickness (i.e., skin folds) or height-to-weight ratios (i.e., BMI) to estimate fat mass. While these measures have been validated previously ₈₁ , this is an important limitation to conclusions drawn from this review. Only eight studies used dual-energy X-ray absorptiometry (DXA), which is considered a 'gold standard' technique for assessing fat mass in children. The use of objective measure of body fatness resulted in variety of findings. In this review, we found as expected that objective measures were able to detect higher effect sizes compared to self-reported activity measures. For example, Davies et al ₆₃ was able to explain 25% variation between the two factors with the use of doubly labelled water. Another important finding is that combination of methods tends to produce a better result.

By searching several PubMed database and the studies' reference lists, we located four other reviews ₇₁,₇₂,₇₃,₇₄ on the association between body composition and physical activity. These reviews differed from ours in terms of focus of the review, search methods, and choice of synthesis of the studies.

Three of these reviews ₇₁,₇₂,₇₃ used meta-analysis, while one used narrative syntheses like our review ₇₄ . Three of these reviews ₇₁,₇₃,₇₄ focused specially on youth; while Westerterp et al ₇₂ re-analyzed existing data including 290 healthy subjects, aged 18-49y, 146 females and 144 males, from 22 different studies. Must et al ₇₄ was the most directly comparable review, being most recent and with a specific focus on relationship between adiposity and physical activity. Must et al ₇₄ in their review reviewed 20 published prospective observational studies of the relationship of PA and sedentary behaviour with the development of overweight and adiposity, with an emphasis on methodological issues. A 2000 meta-analysis ₇₃ encompassed 50 studies, 42 were published as journal articles, four as chapters in books, three as abstracts and one was unpublished. Thirty-two studies assessed activity by questionnaire, eight by heart rate, ten by motion counters and six by observational. A recent 2004 meta-analysis ₇₁ , reviewed the empirical evidence of association between television (TV) viewing, video/computer game use and body fatness, and physical activity using 52 independent samples.

Nevertheless, our review has yielded broadly similar results to previous work. Westerterp et al ₇₂ found that in males, there is a significant inverse cross-sectional relationship between activity energy expenditure and percent body fat, whereas no such relationship was apparent in females. Marshall et al ₇₁ found a statistically significant relationship between TV viewing and body fatness among children and youth although it is likely to be too small to be of substantial clinical evidence. Also, Rowland et al ₇₃ reported that there is a small to moderate relationship between body fat and activity in children.

As with any other review, this study suffers from noteworthy limitations. There are appreciable amounts of missing data in the association analysis. Undoubtedly, these missing effect sizes weakened our discussion on the actual strength of the association. Self-reported activity measures and self-reported height and weight may not be accurate and is subject to subject recall bias, as it requires participants to remember an event or series of events that occur in the past. In addition, obese subjects may tend to over report their activity level. The search strategies did not locate 'grey literature' (e.g. unpublished studies, local reports, PhD and Masters abstracts). It was, however, reasoned that problem with including grey literature (poor study quality owing to lack of peer review ₈₂ and the time and costs involved in identifying and retrieving grey literature ₈₃ outweighed the possible advantage of preventing our results from the influence of publication bias. A thorough search of conference proceedings might also be useful ₈₄ . Although recent research has shown that conference abstracts might not accurately summarize subsequently full reports ₈₅ , the fact that about half the research reported at conference does not get published in full might makes such abstracts a useful source of studies ₈₆ .

This study was originally submitted to Karolinska Institutet, Unit for Preventive Nutrition as partial requirement for the award of Master in Applied Public Health Nutrition. Dr Eric Poorvleit supervised the thesis. Swedish Institutet sponsored AO programme through Guest Scholarship for Advance Studies.