Subjects were recruited into the Look AHEAD Trial as previously described (9,11). Participation was limited to people with type 2 diabetes from 16 clinical centers in the U.S., those 45–76 years old, and those with a BMI ≥25 kg/m² (or ≥27 kg/m² if taking insulin). Subjects were randomized to either ILI, including a weight loss diet and increased physical activity, or DSE. The goal of ILI was a sustained 7% loss of weight during the first year, with weight maintenance thereafter. A substudy was conducted at five clinical sites to measure longitudinal changes in bone density and body composition by DXA scanning as previously described (9). All randomized subjects at these five sites (1,479 of 5,145 total) were approached to participate in the DXA substudy with the exception of those who weighed over the weight limit for the scanner (300 lb). A total of 1,373 subjects consented to and completed baseline assessments of body composition, including total body BMD (including the head), BMD of the lumbar spine (L1–L4), total hip, and the femoral neck of the hip, using a Hologic fan beam densitometer. Of these, 1,274 subjects had a scan 1 year after randomization (Y1) and 1,158 subjects after 4 years (Y4); 1,123 of the 1,158 Y4 subjects undergoing DXA scans also had a Y1 scan. A total of 1,309 subjects within the current database had a DXA scan at baseline and at least one follow-up assessment. The common reasons for missing data were loss to follow-up and weight >300 lb. A complete accounting of the subjects is detailed in the Supplementary Data.

At baseline, Y1 and Y4 clinic visits, weight, height, prescription medication use, demographic characteristics, smoking history, alcohol use, calcium intake in the diet (but not supplement use), and A1C were obtained as described (9). Fitness, expressed in METs, was quantified from a submaximal graded exercise test (12). The MET level presented is that at 80% of the subject’s maximal heart rate. Graded exercise test measures were done at the same visit as the DXA measures or within 1 week of each other if not on the same day. Physical activity was quantified in a subgroup of Look AHEAD participants in kilocalories per week at baseline, 1-year, and 4-year follow-up from a self-reported questionnaire developed by Paffenbarger et al. (13) for assessment of leisure-time physical activity. Of 1,309 participants in the data set for this article, 564 have leisure-time physical activity data (43%). Calcium intake from food was quantified from a food frequency questionnaire in a subgroup of 50% of the subjects seen at each clinic site and accounted for 54% of the current cohort (14). Data related to bone fractures, which was considered a final outcome measure for the larger trial, were not accessible for this article.

Mean and SD were computed for continuous baseline measures, and frequencies and percentages were computed for categorical baseline measures. Baseline data were analyzed to assess the effectiveness of the randomization and differences by sex. All analyses were conducted by sex.

BMD outcome measures were expressed as absolute differences in areal density (g/cm²), an accepted measure of bone mass per a two-dimensional space, which is the actual quantity obtained from a DXA scan, and as percent change from baseline [(Y4 BMD – baseline BMD) / (baseline BMD) × 100]. Additionally, mean and SD were calculated by treatment arm on the Y4 change in the BMD of lumbar spine and total hip and the Y4 change in continuous measures of interest: weight, fitness, physical activity, and A1C. These means were tested by treatment arm for differences using t tests. To examine the relationship between Y4 and Y1 changes in lumbar spine and total hip BMD, the Y4 changes were examined as a function of the Y1 changes. The within-arm mean baseline-to-Y4 change was tested for equality to zero using t tests for the outcome measures. An ANOVA was used to test for differences in change between the ILI and DSE arms. The categorical measures (smoking status, alcohol use in the past year, menopausal status, proteinuria) and arm of the study completed the list of independent variables of interest.

Generalized linear modeling (GLM) procedures, adjusting for site, were used to examine the relationship between Y4 BMD change (response variables) and predictor variables identified as correlated with the outcome measures. As noted, fewer paired Y1 scans were available than for the Y4 cohort, and to take this into account, an available data method as opposed to a complete case method was used. No imputed data (e.g., last observation carried forward) were used.

We examined correlations to identify potential confounding or colinear variables. The variables change in weight and change in fitness were identified as predictor variables based on a visual examination of the bivariable scatterplots, biological plausibility, and significant change over the 4-year interval. Change in A1C was added to the list because of its clinical significance as a marker of diabetes disease control. All models were adjusted for age, diagnosed diabetes duration, baseline BMI, site, race, baseline insulin, thiazolidinedione (TZD) use, menopausal status (for women), proteinuria, smoking status, and baseline level of the response and predictor measures. Repeated-measures (mixed) modeling was used to assess the relationship between time and change in BMD and was more fully adjusted for Y1 changes in weight and BMD. For the GLM and the repeated-measures modeling, the by-arm interaction was evaluated for all significant main effects among the independent measures. Rate of change (β) estimates are reported for continuous measures, and adjusted means are reported for categorical measures. Significance was determined using a fixed α= 0.05. All analyses were performed using SAS 9.2 (SAS Institute, Cary, NC) software. The significance of the total hip BMD results was also assessed using a calculation of Cohen's d for estimating the effect size (15) for the ILI arm compared with the DSE arm.

Sensitivity analyses were performed to examine the impact of age and glucocorticoid and TZD use because they have been shown to be deleterious to bone (16–21). Additional analyses examined the impact of age, bariatric surgery, and insulin use versus nonuse on the current results.

The baseline characteristics of subjects in the two study groups are summarized in Table 1, as are the variables describing the DXA subsample and the full Look AHEAD cohort (Supplementary Data). Randomization was generally successful in distributing subjects equally for demographic variables in the overall trial; however, weight differed significantly across treatment groups (P = 0.049). There was a very low prevalence of osteoporosis (defined as a T-score <−2.5) at any site of measurement at baseline and no difference in prevalence between groups.

The demographic variables of the subsample of the Look AHEAD cohort participating in the bone substudy differed from the total Look AHEAD cohort (Supplementary Data), reflecting the Latino population at these participating centers. The DXA cohort was shorter, weighed less, had a lower BMI, was more fit, was more likely never to have smoked, was treated less frequently with TZDs and steroids, and comprised fewer blacks and non-Latino whites.

Men randomized to ILI had a twofold greater rate of bone loss at the hip (−0.78% per year) than those randomized to DSE (−0.36% per year, P < 0.01) at Y4, with an effect size of 0.42. However, no intervention effect on lumbar spine BMD (0.34% vs. 0.33% per year, P = 0.99) was found. The changes in male total hip BMD were also reflected in similar changes at the femoral neck. Women in both treatment arms had decreases in all measures of BMD (0.26–1.31% per year), and in contrast to men, there were no significant differences (P ≥ 0.24) in mean BMD loss between treatment groups by Y4 at any site.

Primary explanatory variables believed to affect change in BMD from baseline are summarized in Table 2, stratified by sex and intervention arm. There was a statistically significant effect of ILI to reduce body weight and BMI and increase fitness in both groups, but physical activity was only increased in men. No statistically significant effect of treatment arm on alcohol intake in the past year, smoking status, or dietary calcium intake in either men or women was found.

Among men, the magnitude of weight loss was significantly correlated with bone loss at the hip (β = 0.03, P < 0.001) but with some gain in spine BMD (β = −0.02, P = 0.005) (Table 3). In women, the magnitude of weight loss was also significantly correlated with bone loss at the hip (β = 0.05, P < 0.001) but not at the spine (β = −0.008, P = 0.251). Hip bone loss did not correlate with change in fitness in either sex. Improved glycemic control, as reflected by lower A1C levels, correlated with loss of bone in the lumbar spine in men and hip bone loss in women. A statistically significant −0.4% (4 mmol/mol, P < 0.01) decrease in HbA_1c in the ILI arm compared with the DSE arm was observed in men only in this DXA cohort, which can be estimated to account for 0.04% per year spine BMD loss.

For change in total hip BMD in men, there was no longer a significant difference by Y4 between the ILI and DSE arms (rate of BMD loss −0.53%, P = 0.93) after adjustment for postrandomization changes in weight. In women, after adjusting for these changes, ILI was associated with less (−1.01% per year) Y4 hip bone loss compared with DSE (−1.27% per year, P = 0.005).

In a sensitivity analysis that excluded subjects using any steroid, TZD, or bisphosphonate at baseline, the results were similar. Only 12.8% of the sample was using inhaled or topical steroids not generally believed to be detrimental to bone; 5.4% was using oral steroids known to cause bone loss; and 1.5% was using bisphosphonates, a medication known to increase BMD measures. The use of steroids was similar in the ILI and DSE groups. There also was no significant difference (P = 0.355) in TZD use across arms of the DXA substudy or by sex. Adjusting for age did not significantly change the results. For example, the rates of bone loss at the hip at Y4 in women for the ILI and DSE groups were −1.38 and −1.31% per year with adjustment for age vs. −1.33 and −1.27% per year after removing the age adjustment. Of the 1,309 participants included in these analyses (i.e., those who had baseline and at least one follow-up DXA scan), 19 reported having bariatric surgery by the time of the Y4 visit (6 in ILI, 13 in DSE). Only 2 of these 19 participants (1 in each arm) had bariatric surgery by the time of their Y1 visit. There are no significant differences at Y1 in weight loss and bone loss of those with and without bariatric surgery by Y4. By Y4, those in both the ILI and DSE arms with the bariatric surgery had significantly more weight loss and bone loss than those without the surgery. The GLM models representing the data in Fig. 1 were rerun with the 19 bariatric surgery participants removed, showing only slight changes in the estimate values and no change in significance. There was no significant difference in hip and spine BMD for insulin users versus nonusers stratified by randomization arm and visit.

The results of a fully adjusted repeated-measures GLM are summarized in Fig. 1. The purpose of this model was to determine whether the effect of the intervention on the annual rate of bone loss varied over time. A visual inspection of Fig. 1 shows marked differences in the pattern of bone loss and weight with time in men and women. There is a striking parallel between the changes in weight and hip BMD in men and a significant (P < 0.001) correlation between percent change in lean mass and percent change in hip BMD at both Y1 (r = 0.294) and Y4 (r = 0.339). There was also a significant (P < 0.001) correlation between percent change in total body fat and percent change in hip BMD at both Y1 (r = 0.341) and Y4 (r = 0.290). A significant (P ≤ 0.0001) time × arm interaction for hip BMD was found in men as well as in women. As a percentage of the initial BMD, these estimated hip losses for men and women, respectively, were −1.38% and −0.88% per year for ILI over the 1- to 4-year follow-up intervals. This model showed a greater rate of bone loss at Y1 compared with Y1–Y4 at the hip for both men and women. There was no significant effect on the lumbar spine (P value for time × arm interaction = 0.591 and 0.935 in men and women, respectively).