Studies were identified on a priori criteria that included: (1) a randomized controlled design comparing an exercise intervention with a control group (i.e., no exercise program prescribed and instructions to maintain current activity levels or no exercise related information); (2) report of depression outcomes; and (3) adults diagnosed with any type of cancer, regardless of stage of diagnosis or type or stage of treatment. Exercise interventions occurring in any setting, with or without supervision, were eligible.

The databases PubMed, PsycINFO, CINAHL Plus, SPORTSdiscus, OregonPDF in Health and Performance, and ProQuest Theses and Dissertations were searched through Nov 18, 2010. We searched all databases using a Boolean search strategy [i.e., (cancer OR neoplas* OR tumor OR chemo* OR radiat* OR malign* OR carciniom*) AND (depress* OR anxiety OR anxious OR worried OR scared OR nervous OR cognitive OR biofeedback OR relaxation OR social support OR mind-body) AND (exercise OR physical activity OR aerobic OR cardiovascular OR resistance OR strength OR muscular OR flexibility OR walking OR program OR interval OR sport OR fitness OR performance OR movement OR stretching OR tai chi OR yoga OR dance OR body OR composition)]. Journals focusing on cancer survivorship (Journal of Clinical Oncology, Breast Cancer Research and Treatment, Journal of Cancer Survivorship, Oncology Nursing Forum), and the reference lists of included studies were also searched for additional reports.

We estimated the intensity of exercise using the compendium of metabolic equivalent units (METs), where 1 MET represents sitting quietly (3.5 ml O₂·kg^-1·min^-1) and <3 METs, 3 to <6 METs, and ≥6 METs represent low, moderate, and vigorous intensity exercise, respectively ^[21]. We calculated the weekly volume of aerobic exercise as the product of minutes of daily exercise and frequency of exercise sessions per week (min·wk⁻¹). We used the Physiotherapy Evidence Database scale (PEDro) to gauge methodological quality of the trials in terms of internal validity and statistical reporting ^[22]. Four independent, trained raters extracted information related to the study with high inter-rater reliability, mean Cohen's κ = 0.90, for categorical variables, and mean intra-class correlation r = 0.94 for continuous variables.

The studies assessed depressive symptoms among cancer survivors as a continuous outcome variable assessed as a component of a comprehensive psychological questionnaire with a depression subscale ^[23] or a questionnaire solely assessing depression levels ^[24]–^[27]. To assess baseline levels of depressive symptoms on a common metric across depression questionnaires, we used a 0–100 scale, where ‘0’ implies absolutely no depressive symptoms, and ‘100’ implies the highest level of symptoms possible on a given scale. We used the standardized mean difference effect size (d) to quantify the difference in depression from baseline to follow-up between the exercise and control groups, correcting for small sample size bias ^[28], ^[29]. The effect size d denotes the difference between the mean depression values of the control and exercise groups, divided by the pooled standard deviation ^[30]; the sign of d values was set to be negative when the exercise group reduced depression more than the control group. The standardized d value can be interpreted as −0.20, −0.50, and −0.80, represent small, medium, and large reductions in depressive symptoms, respectively ^[31]. When trials included more than one exercise group (e.g., aerobic exercise and resistance exercise), we calculated multiple effect sizes. Sensitivity analysis examined the influence of a single study on the overall mean effect size of all trials by iteratively removing a single study and then re-estimating the overall mean effect with 95% confidence intervals ^[32]. We present overall mean effect sizes (d₊) as both, fixed- and random-effects estimates.

We used Stata 11.1 (StataCorp, College Station, TX) with macros developed for meta-analysis ^[33] to perform all statistical analyses. Begg's test ^[34] (z = −1.67, p = 0.10), Egger's test ^[35] (t = −0.12, p = 0.90), and the trim-and-fill method ^[36] identified no asymmetries in the effect size distribution suggestive of publication bias. Potential heterogeneity or between-study variance was calculated as Q and I² (and 95% CI) ^[37], ^[38]. The Q statistic follows an approximate χ² distribution with k-1 degrees of freedom, where k is the number of studies included in the meta-analysis ^[38]. The Q statistic can be standardized to I² with values ranging from 0% (homogeneity) to 100% (heterogeneity). To explain variance of depressive symptom reduction—the relation between study-level characteristics and the magnitude of the depression reduction effect size (d₊)—a modified, weighted least squares regression was used with weights equal to the inverse variance of each exercise intervention effect size (viz., fixed-effects meta-regression). The underlying assumptions of meta-regression are similar to that of ordinary least-squares regression, including independence of errors, homoscedasticity of variance, and normally distributed variables ^[28], ^[33], ^[39], ^[40]. Statistically significant bivariate regression analyses were integrated into a multiple-moderator fixed effects regression to determine which variables could explain unique between study variance. To reduce multicollinearity in multiple meta-regression models, all continuous variables were zero centered based on their means; categorical variables were contrast coded (−1/+1). Beta-values (β) appear in standardized form in order to quantify the amount of variability in ds associated with each moderator of interest. All meta-regression model estimated effect sizes are depicted using the moving constant technique, entering multiple predictor variables simultaneously ^[41]. Two-sided statistical significance was p<0.05.

Qualifying for inclusion in the meta-analysis were 37 relevant randomized controlled exercise interventions ^[17]–^[20], ^[42]–^[74] (N = 2,929) with a total of 40 comparisons (k = 40) of exercise versus control conditions (Figure 1). Thirty-four studies provided one effect size, and three provided two effect sizes ^[20], ^[44], ^[52]. Exercise interventions were published in 2006±4.2 (range: 1994–2010) with most studies (70%) conducted in North America. The mean PEDro score of the exercise interventions was 7.0±1.0 suggesting relatively high methodological quality ^[22]. A minority of studies (20%) reported that they implemented at least one theory of behavior change (Table 1). Studies assessed depression using the Center for Epidemiologic Studies-Depression questionnaire (40%) ^[25], Profile of Mood States (23%) ^[23], the Beck Depression Inventory (18%) ^[24], Hospital Anxiety and Depression Scale (12%) ^[26], or Symptom Assessment Scale (7%) ^[27].

Cancer survivors participating in the exercise trials averaged 51.3±6.5 yr (range: 39–70). The majority of cancer survivors participating in the exercise interventions were white, non-Hispanic (n = 2,255; 77%), and women (n = 2,548; 87%) with a time since cancer diagnosis of 25.3±19.6 months (range: 2.8–73.0). Exercise interventions were more common during curative therapy with 29 of the 40 exercise interventions (73%) occurring during treatment (i.e., chemotherapy or radiation treatment). Trials most often examined breast cancer survivors (k = 24) ^[17], ^[20], ^[48]–^[69]. Two trials each focused on prostate cancer ^[19], ^[70], leukemia ^[72], ^[73], and lymphoma ^[18], ^[74] survivors, and only one trial examined colorectal cancer survivors ^[71]. The remaining 6 trials examined survivors with diverse types of cancer diagnoses ^[42]–^[47]. At baseline, the standardized metric of depressive symptoms was 34.2±26.9 and ranged from 3.49 to 81.5.

The mean length of the exercise interventions was 13.2±11.7 wk with an average of 3.0±2.5 sessions per week lasting 49.1±27.1 min·session⁻¹. Average weekly volume of all exercise was 129.4±64.9 min·wk⁻¹. Exercise modalities included walking (k = 16; 40%), stationary cycling (k = 5; 13%), weight machines (k = 2; 5%), resistance bands (k = 3; 8%), and yoga (k = 8; 20%). In addition, flexibility exercises were prescribed in 50% of the exercise interventions. The absolute intensity of exercise was 3.9±1.3 METs indicating they were of low (i.e., <3 METs) to moderate (i.e., ≥3 to <6 METs) intensity. A majority of exercise interventions (60%) was supervised. Table S1 summarizes methodological characteristics of the included trials.

Exercise provided a small overall reduction in depressive symptoms compared to standard care among all types of cancer [d₊ = −0.13 (95% CI: −0.26, −0.01)]. Subgroup analysis by cancer type revealed significant reductions in depressive symptoms among breast cancer survivors [d₊ = −0.17 (95% CI: −0.32, −0.02)], but no significant difference in depressive symptoms among prostate, leukemia, lymphoma, and colorectal cancer survivors (Table 2). Figure 2 depicts the fixed-effects mean reduction in depressive symptoms, stratified by type of cancer. Collectively, the 40 effect sizes lacked homogeneity [I² = 55% (95% CI: 35–68), p<0.001], as did the analysis when restricted to breast cancer survivors [I² = 59% (95% CI: 37–73), p<0.001; Table 2].

Three moderators explained unique variance relating to the efficacy of exercise to reduce depressive symptoms when entered in a multiple meta-analysis regression model. Weekly volume of aerobic exercise reduced depression in dose-response fashion (β = −0.24, p = 0.03), a pattern that was only evident in higher quality trials. Exercise reduced depressive symptoms most when exercise sessions were supervised (β = −0.26, p = 0.01) and cancer survivors were between 47–62 yr [(β = 0.27, p = 0.01); Table 3]. These three moderators together explained 35% of the variance in depression reduction resulting from exercise; yet, variability beyond that expected by sampling error alone remained unexplained.

In our bivariate analyses (Table S2), three other features related to the magnitude of exercise-induced reduction of depressive symptoms: explicit use of theory in intervention development, the percentage of non-Hispanic Whites in the sample, and time since cancer diagnosis. Interventions were more successful in reducing depressive symptoms when they used psychological theory, sampled greater percentages of non-Hispanic Whites, and were more proximal to the date of cancer diagnosis. Still, none of these moderators explained significant variability in combined moderator analysis, suggesting that their influence is explained by the variables in the combined moderator model (viz., supervision, volume of exercise, methodological quality, and age; see Table 3). Standardized baseline depressive symptom scores were not associated with depressive symptom improvements resulting from exercise (p = 0.71).

The major limitation of this meta-analysis is that depressive symptoms were a secondary outcome in almost all exercise interventions. As such, cancer survivors in the exercise trials that we meta-analyzed were not recruited based on depression levels; moreover, they may have exhibited few depressive symptoms at baseline. Nonetheless, our analysis suggests exercise is efficacious to improve depressive symptoms among cancer survivors. Our analysis may underestimate the efficacy of exercise to reduce depressive symptoms among cancer survivors with higher levels of depression or those with a diagnosis of depression. Furthermore, our analysis and interpretation of our findings have focused on the statistical associations, not the clinical implications of exercise and the experience of depressive symptoms.

Despite our intention to include all types of cancer of any race, 26 of the 40 effect sizes (65%) targeted white, non-Hispanic, breast cancer survivors exclusively, which has been a limitation of previous meta-analyses examining a variety of health-related outcomes among cancer survivors ^[14]–^[16]. Additionally, the number of exercise interventions among breast cancer survivors limits the generalizability of our findings to other types of cancer, even though there was no significant difference in reduction of depressive symptoms attributable to type of cancer. These limitations should provide an impetus for researchers to continue investigating the effects of exercise among other ethnic groups and underrepresented cancer types.

Despite an overall rating of high methodological quality (7.0±1.0 of 11), we noted some consistent methodological weaknesses throughout the literature, such as inclusion of small sample sizes, inconsistent criteria with respect to study entry eligibility and baseline depressive symptoms levels, poor reporting of adverse events, and failure to follow intent-to-treat analytic strategies. Indeed, the dose-response pattern of exercise to the reduction of depression symptoms materialized most clearly only in the studies with the highest methodological quality.

In closing, we confirmed that exercise provides a small overall reduction in depressive symptoms among cancer survivors. Depressive symptom reduction occurred in dose-response fashion with weekly volume of aerobic exercise, with high-quality trials documenting large changes for cancer survivors accruing larger weekly volumes of aerobic exercise. Larger reductions in depressive symptoms also occurred with supervised exercise, and among cancer survivors 47–62 yr. Cancer survivors should strive to avoid physical inactivity; discuss the safety and feasibility of exercising with their medical care provider to optimize depressive symptoms management; and eventually aim to achieve larger weekly volumes of aerobic exercise if possible, complimented with resistance training twice-weekly, and flexibility activity on days of non-exercise ^[2]. Furthermore, clinicians now have evidence to advocate the potential benefit of aerobic exercise as a modality to manage depressive symptoms among cancer survivors.