Mediators and moderators of the effects of a year-long exercise intervention on endogenous sex hormones in postmenopausal women

The ALPHA Trial study population and recruitment details have been described previously [6, 8]. Inactive women (<90 min/week recreational activity over the past year or, if between 90 and 120 min/week of physical activity had a VO_2max < 34.5 mL/kg/min), aged 50–74 years, postmenopausal for at least 24 months, body mass index (BMI) 22–40 kg/m², with no major co-morbidities or previous cancer diagnosis besides non-melanotic skin cancer, were recruited from the general population in Calgary and Edmonton, Alberta, Canada. Eligibility criteria also included the following: English-speaking, acceptable heart and lung function as assessed by a baseline fitness test, able to undertake unrestricted physical activity as assessed by their physician [11], normal levels of fasting lipids, glucose, thyroid-stimulating hormone and alanine aminotransferase, breast tissue density above a zero density level, non-diabetic, non-smoker, alcohol intake <14 drinks/week, no medications or exogenous hormones that might influence estrogen metabolism and not currently or planning to undertake a weight loss program or planning any extended absences in the 18 months subsequent to enrollment.

The study participants were randomly allocated either to a year-long, moderate-to-vigorous-intensity, aerobic exercise intervention (n = 160) or to a control group (n = 160) assigned to no exercise. Details of randomization have been described previously [6]. The exercise intervention was individually prescribed program of at least 45 min of any type of aerobic exercise done 5 days per week for 12 months at 70–80% heart rate reserve. At least three sessions per week were supervised by exercise trainers at our fitness centers, and the rest were home-based. The frequency, duration, and intensity were gradually increased during the first 3 months of the intervention from 3 weekly sessions of 15- to 20-min duration at an intensity of 50–60% heart rate reserve to the final prescription achieved in week 12. Women in the control group were asked to maintain their regular inactive lifestyle. No changes to usual dietary intake were to be made in either group. Ethics approval was obtained from the Alberta Cancer Board and Universities of Calgary and Alberta, and all participants provided written informed consent.

Baseline data on demographics, medical and medication history, self-rated health, and past year physical activity [12] were obtained by a self-administered questionnaire. Anthropometric measurements were taken in duplicate; if the two measurements were discrepant (i.e., not identical), a third measurement was taken and the average of the two closest was used in the analyses. Weight and height measurements were taken using a balance beam scale and a stadiometer. Body mass index was calculated as weight/height² (kg/m²). At baseline and 12 months, total body fat and body fat percentage were measured using whole-body dual X-ray absorptiometry scans. In Calgary, scans were done on a Hologic QDR 4500 W^® scanner in whole-body mode and analyzed with software version 11.2.1 (Hologic^®, Inc., Bedford, MA, USA). In Edmonton, scans were done on a Lunar Prodigy^® scanner in either standard mode or thick mode and analyzed with enCORE Software 6.70.01 before November 2004 and 8.60 after November 2004 (Lunar^® General Electric Medical Systems, Madison, WI, USA). Percent body fat was calculated as 100% × {fat mass/(fat mass + lean mass)}. Intra-abdominal fat was measured with a single computed tomography slice at the umbilicus. In Calgary, scans were done using a Marconi PQ5000 VisionMaster/Picker, and in Edmonton, using a Phillips/Marconi MX8000 multislice scanner. The study radiologist used an image analysis software program (Phillips Medical Systems PQ5000) to identify and demarcate the thresholds between the subcutaneous and intra-abdominal area. Physical fitness was assessed at baseline and 12 months using a modified Balke treadmill protocol to estimate maximum oxygen consumption (VO_2max) from submaximal exercise intensities. Oxygen consumption at the age-predicted maximum heart rate was estimated by extrapolating from the two last completed stages using the American College of Sports Medicine metabolic equations for estimating oxygen consumption at the workload of each stage [13].

Blood was collected after a minimum 10-h fast at baseline (60 mL), 6 and 12 months (40 mL), and medications taken in the past 24 h were recorded. Participants were asked not to exercise for 24 h pre-blood draw. All blood samples were collected, processed, and stored within 12 h of collection and then shipped and stored in −86°C freezers until time of assay. Analyses were conducted by the Reproductive Endocrine Research Laboratory at the University of Southern California, a laboratory with well-established protocols and quality control procedures. Serum estradiol and testosterone levels were quantified by radioimmunoassay after organic solvent extraction and Celite column partition chromatography [14‐16]. Chromatographic separation of the steroids was achieved by using different concentrations of toluene in iso-octane and ethyl acetate in iso-octane. Sex hormone–binding globulin (SHBG) and insulin were quantified via chemiluminescent immunometric assay using the Immulite Analyzer (Siemens Healthcare Diagnostics, Deerfield, IL). The SHBG concentration and an assumed albumin concentration of 43 g/L were then utilized in a validated algorithm with total estradiol to calculate free estradiol [17].

Our hypothesized mediators were body fat change (for sex hormones and SHBG) and circulating insulin change (for SHBG). These hypotheses were supported by our earlier review of the scientific literature on proposed biologic mechanisms relating physical activity to postmenopausal breast cancer risk [4]. Moreover, the proposed mediators, by definition [9], were associated with our exercise intervention; relative to controls, the exercise group in the ALPHA Trial experienced significantly greater decreases in total body fat, intra-abdominal fat area [8], and circulating insulin levels [7] over 12 months.

Potential moderators were baseline characteristics for ALPHA Trial participants that were selected for their hypothetical relations either to exercise adherence during the trial (e.g., baseline fitness level, age, years postmenopause, self-rated health, previous recreational activity, baseline BMI) or to changes in sex hormone levels (e.g., circulating levels of sex hormones at baseline, previous HRT use, baseline BMI). We hypothesized that women who exercised more might experience greater changes in sex hormone levels and that women with higher sex hormone levels or lower SHBG levels at baseline might experience greater changes as a result of the exercise intervention.

The analyses included participants with complete data and were based on assigned groups at randomization regardless of adherence. Hormone levels were log-transformed to achieve an approximate symmetric distribution. Intervention effects were evaluated with general linear models considering the repeated measures of sex hormone concentrations at 6 and 12 months. The model took the following form: Y _ij  = β₀ + β₁ X _i  + γZ _ij  + ε _ij , where, for the ith subject, Y _ij is a log-transformed outcome measure at jth time (j = 1 for 6 months, j = 2 for 12 months), X _i is an indicator variable for the exercise intervention arm (0 for control, 1 for exercise intervention), Z _ij is a vector of adjustment variables, and (ε _i1, ε _i2) is a correlated error vector following a bivariate Gaussian distribution with mean 0. Of the parameters β₀, β₁, and γ (associated with the intercept, exercise intervention, and adjustment variables, respectively), β₁ is the target parameter, representing the adjusted difference in the mean log outcome between the exercise intervention group and the control group over the two follow-up time points, adjusting for Z. We refer to the exponential of β₁ as the treatment effect ratio (TER), as it is a ratio of the adjusted geometric means of the outcome for the exercise intervention group over the control group. In this paper, we also refer to the TER as the ‘intervention effect’.

Our assessment of mediation was based on findings from the general linear model with both the intervention assignment and changes in the hypothesized mediators predicting changes in sex hormone or SHBG levels. Mediation by adiposity and insulin levels was investigated in two ways. Primarily, mediation was investigated by examining the change in the TER pre- versus postadjustment for changes in the hypothesized mediators at 6 and 12 months from baseline. If the adjustment attenuated TER, then we considered this finding to be suggestive of mediation [10]. Secondarily, the main effect of each potential mediator on sex hormone or SHBG changes was also evaluated in these models, adjusting for the intervention assignment. Kraemer et al. [9] explain that a mediator must measure a change occurring during treatment (e.g., changes in body fat and insulin levels occurred during the 12-month intervention period [7, 8]), must correlate with treatment assignment (e.g., we observed significantly greater changes in body fat and insulin levels in the exercise group than in controls [7, 8]), and must have either a main or interactive effect on the outcome. This secondary analysis is informative because it is these conditions under which our main analysis would suggest mediation.

Moderation was investigated by determining the statistical significance of the interaction term between the intervention assignment and each hypothesized moderator (described above), adjusting for the baseline outcome (e.g., baseline estradiol level), the intervention group assignment, time, and the hypothesized moderator (e.g., baseline age) [9]. All baseline characteristics were treated as continuous variables in these analyses with the exception of HRT use, which was dichotomized. To help explain our findings from these analyses, intervention effects were also estimated within subgroups of these characteristics. We reported intervention effects within all subgroups regardless of the statistical significance of the interaction term in general linear models. Subgroups were created by dichotomizing each variable at the median with the exception of self-rated health, which was measured by questionnaire using the SF-36 scale [18] (range, 0–100): ‘low’ was defined as a score <82 and ≥82 was considered ‘high’. All statistical tests were two-sided with a significance level set at 0.05. Statistical analyses were performed using SAS software (version 9.1; SAS Institute, Inc., Cary NC).