Ways of unraveling how and why physical activity influences mental health through statistical mediation analyses

Participation in physical activity has been linked to favorable mental health outcomes, including reduced symptoms of depression (Rethorst, Wipfli, & Landers, 2009), anxiety (Taylor, 2000) and chronic fatigue (Puetz, O’Connor, & Dishman, 2006); lower risk of age-related cognitive decline (Colcombe & Kramer, 2003); and enhanced psychological wellbeing (Bize, Johnson, & Plotnikoff, 2007), sleep (Youngstedt & Kline, 2006), and self-esteem (Spence, McGannon, & Poon, 2005).

Although various biophysical and psychosocial factors have been identified as potential mechanisms responsible for the observed associations, whether and how physical activity causes such favorable mental health outcomes remains fundamentally unknown. Clarifying mechanisms of physical activity influence on aspects of mental health is crucial for demonstrating causality (Hill, 1965) and designing optimal physical activity intervention programs that can confer psychological benefits .

The identification of ‘real’ mechanisms of influence, represented by mediating variables or mediators, is not a simple matter. It requires consistent evidence of mediational processes across different settings, data collection methods, and study designs (MacKinnon, 2008). A research program aiming at unraveling the causal effects of physical activity on particular aspects of mental health would have to resort to both qualitative and quantitative methods.

While the former may provide valuable detailed information on how participation in physical activity affects psychosocial mediators and mental health, and clarify unexpected influences and idiosyncratic responses (Faulkner & Biddle, 2004; Maxwell, 2004), the latter quantifies mediating processes and assesses their statistical reliability. Quantitative methods encompass mediational metaanalyses (Shadish, 1996) and statistical analyses of studies that directly test mediation hypotheses (MacKinnon, 2008).

The strength of evidence for causal mediating processes depends on the study design and assumptions onwhich the analyses are based, with crosssectional studies providing the weakest and randomized controlled trials the strongest evidence (Shadish, Cook, & Campbell, 2002). Although there are experimental designs that permit the evaluation of mediational processes by simply estimating intervention effects of the independent variable on the mediator and of the mediator on the dependent variable (e.g., double randomization or experimental-causal-chain design; Spencer, Zanna, & Fong, 2005), most study designs rely on statistical mediation analyses.

These analyses are the focus of this paper as they pertain to the identification of mechanisms explaining eventual causal relationships between physical activity and aspects of mental health. Central issues related to the conceptualization of, and modeling approaches to, statistical mediation analyses are first discussed. This is followed by practical recommendations for mediation analyses in the field of physical activity and mental health. Finally, promising trends in mediation analyses are outlined.

Statistical mediation analysis: general issues

Statistical mediation analysis refers to statistical procedures aimed at testing the hypothesis that an independent variable (X) affects an outcome (Y) through one or more mediators (M). There is general consensus among proponents of different approaches to mediation analyses that, to be a mediator, a variable needs to be related to the independent variable (e.g., Baron & Kenny, 1986; James, Mulaik, & Brett, 2006; Kraemer, Wilson, Fairburn, & Agras, 2002; MacKinnon, 2008).

Statistically, this is operationalized as the regression coefficient(s). The second primary criterion for mediation is a significant association between M and Y, after adjustment for X (Baron & Kenny,1986; MacKinnon, 2008), which is represented by the regression coefficient b. Although this criterion of mediation is commonly accepted, some authors maintain that it may not be always valid.

Specifically, if the effect of M on Y depends on the values of the X (i.e., there is a significant interaction effect), it is theoretically possible to observe an overall nil association between M and Y (Cerin & MacKinnon, 2008; Judd & Kenny, 1981; Kraemer et al., 2002). However, this would not invalidate the hypothesized mediating process as, for at least some values of X, M would be independently associated with Y, i.e., the regression coefficient b would be statistically significant (assuming the study is powered to detect such effects).

James et al. (2006) raise another reason for not using the second criterion of mediation. Namely, if the relationships between X, M and Y are deterministic (i.e., there is perfect correlation among them), the regression coefficient b will be undefined. In such cases, and in cases of stochastic complete mediation (i.e., M explains the imperfect association of X with Y in its entirety), these authors maintain that the second criterion of mediation should be replaced by the estimation of the relationship between M and Y without adjusting for X.

The choice between applying the former or latter criterion of mediation should be determined a priori based on theoretical considerations, i.e., whether we expect M to completely or partially explain the effect of X on Y. In the field of physical activity and mental health deterministic processes are practically nonexistent and complete mediation processes, as examined in single studies, highly improbable. Therefore, it is recommended that researchers use the regression coefficient b, with adjustment for X, as a criterion of statistical mediation. In doing so, it is also recommended that they routinely examine possible interaction effects between X and M on Y (Cerin & MacKinnon, 2008; MacKinnon, 2008).

Mediation analyses in the field of physical activity and mental health: recommendations

The aim of this section is to provide recommendations on approaches to mediation analyses suitable for specific study designs implemented in the field of physical activity and mental health. This is followed by a brief description of studies specifically aimed at the evaluation of mediating processes.

A considerable number of cross-sectional epidemiological as well as cross-sectional small-scale studies have examined associations of physical activity with aspects of mental health (see Biddle & Mutrie, 2008 for a recent synthesis). However, only a few of these studies have attempted to shed light on the underlying mechanisms (e.g., Dishman et al., 2006; Van de Vliet et al., 2002), which appears to be largely due to a failure to measure plausible mediators.

It is recommended that mediation analyses using data from crosssectional studies be conducted using the methods and modelbuilding strategies described above. In doing so, it is important to acknowledge that such studies cannot prove causality. The observed relationships can be spurious due to violations of the assumption of no omitted influences.

Moreover, reciprocal and reverse causal effects are a possibility. Hence, cross-sectional studies should pay particular attention to minimizing sample bias and confounding, and consider alternative theoretically-plausible models of influence including, where appropriate, reciprocal effects.

Sample bias and confounding can be respectively addressed by recruiting large representative samples of the population and measuring and adjusting for known confounders. Large samples also allow the simultaneous estimation of structural as well as measurement models of the independent variables, mediators and outcomes, which helps reduce bias in regression estimates due to measurement error (Kline, 1998).

Finally, it is noteworthy that extrapolation of findings from crosssectional mediation studies to longitudinal mediation processes should be treated with great caution. In this respect, Maxwell and Cole (2007) have shown that cross-sectional estimates of mechanisms of influence represent substantially biased estimates of longitudinal mediation processes.

Moreover, the mechanisms responsible for inter-individual differences in physical activity and mental health may substantially differ from those underlying within-individual changes in mental health status. For example, selfefficacy and affect were found to be significant mediators of the cross-sectional relationships between physical activity and satisfaction with life in a sample of older adults (Elavsky et al., 2005). However, only affect accounted for the observed relationship between changes in physical activity and changes in life satisfaction.

Limitations of current approaches to mediation analyses and future developments

A perusal of the current literature in the field of physical activity and mental health reveals a shortage of mediation studies, which explains why not much is known about the underlying mechanisms of influence (Biddle & Mutrie, 2008). There is a wealth of experimental and statistical methods waiting to be used by those who wish to unravel how and why physical activity affects various aspects of mental health in diverse populations and circumstances.

Clearly, it is imperative that future studies put increased effort on identifying mediating processes responsible for the observed associations. When it comes to mediation analyses, physical activity researchers are not the only ones facing a big challenge. Methodologists and statisticians have themselves to work out solutions to a wide range of challenging problems. For example, they need to clarify and evaluate approaches to assessing causal inference in absence of randomization.

Longitudinal mediation models require further development as it is still unclear which models produce correct inferences and under what conditions. These advances need to be coupled by substantive efforts to clarify the temporal course and appropriate timing of mediation relations. The development of models able to combine mediation and moderation processes, and quantify non-linear relations among variables with different distributions, is particularly important to the field of physical activity and mental health.

In this regard, applications of generalized additive models (Hastie & Tibshirani, 1990) and mixture models (identifying groups of persons with similar trajectories of change and mediation processes explaining these trajectories; Muthén & Muthén, 1998e2007) are promising venues for future research.


Author: Ester Cerin

back to top