Study population

From 2000 to 2008, ATP recruited 29 878 Albertans aged 35–69 years with no history of cancer, other than non-melanoma skin cancer, at the time of enrolment by telephone random digit dialling.17 After enrolment, participants were invited to complete two follow-up questionnaires: Survey 2004 and Survey 2008. Survey 2004 was the first follow-up survey completed in 2004 for participants that were enrolled in 2000–2003. The mean time from baseline to this follow-up survey was 2.5±0.3 years; median was 2.5 years. Survey 2008 was a second follow-up survey conducted in 2008; this survey was conducted for all participants enrolled in 2000–2007. The mean time from initial baseline enrolment for participants to this survey was 4.2±2.2 years; median was 3.8 years. The anthropometrics self-reported at the follow-up surveys were used to determine changes in anthropometrics from the baseline. Incidence of diabetes was determined from the Alberta Health (AH) administrative healthcare data from time of enrolment to 31 March 2015, with mean follow-up time 10.4±2.7 years; median was 10.2 years. Figure 1 shows inclusion and exclusion criteria for this longitudinal study. After exclusion, the study population was 19 655 (n=19 628 for analysis of weight changes, n=19 616 for analysis of BMI changes, n=19 594 for analysis of WC changes, n=19 513 for analysis of WHR changes) diabetes-free participants with anthropometrics measured at the baseline and at least one follow-up and with completed information on key covariates, such as ethnicity, household income and lifestyle factors (smoking, physical activity and dietary intake). All ATP participants provided written consents to participating in ATP and allowing healthcare data linkage and long-term follow-ups at the time of enrolment.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Flowchart for the inclusion and exclusion of ATP cohort (2000–2008). ATP, Alberta’s Tomorrow Project; BMI, body mass index.

Anthropometric measurement

Basic anthropometrics, including weight (kg), height (cm), WC (cm) and hip circumference (HC, cm) were self-reported by participants at baseline and follow-ups. Similarly to our previous study,15 weight and height were corrected using sex-specific correction factors to reduce self-reporting bias.18 BMI was calculated by dividing body weight in kilograms by the square of participant’s height in metres. WHR was calculated as the ratio between WC and HC.

Anthropometric changes (∆) were calculated for weight, BMI (overall obesity indicator), WC (central obesity indicator) and WHR (body shape indicator). Absolute changes in anthropometrics (eg, ∆WC) were calculated by subtracting the baseline measures from follow-up measures. For easy interpretation and comparison, absolute changes were standardised with 1 SD of absolute changes. Relative changes in anthropometrics (eg, ∆WC%) were quantified as the percentage of change from baseline. Study participants were further classified based on relative changes in anthropometrics (ie, ±5%, >5% gain and >5% loss). All these three measures of anthropometric changes (absolute, relative and by categories) were used in this study to evaluate the potential impact of measurement bias on the results.

Diabetes case definition

Clinical information was obtained from the linked deidentified AH administrative health records (1 October 2000 to 31 March 2015). Diabetes cases were identified using the Canadian National Diabetes Surveillance System (NDSS) algorithm based on International Classification of Diseases (ICD)-9 and ICD-10 codes as previously described.15 19 An additional algorithm was developed to exclude prevalent cases based on self-report data, that is, self-reporting diabetes at enrolment, plus any of the following conditions: (i) ever hospitalised for diabetes, (ii) ever having physician claims for diabetes or (iii) ever taking diabetes medication with Anatomical Therapeutic Chemical Classification (ATC) code for insulin (A10A) or other glucose-lowering drugs (A10B).15 The index date of diabetes was determined by the earliest date of health records that contribute to the case definition. Cases were identified as ‘incident’ only if the index date was >6 months after enrolment to ensure true incident cases were identified. Incidence rate (IR) of diabetes was defined as number of incident cases per 1000 person-years (PY).

Other covariates in analyses

Based on risk factors for diabetes identified in the literature,20 participants were further categorised by the following variables at the baseline: parental history of diabetes (yes/no), whether or not (yes/no) physically active (defined as >210 min of moderate-intensity to vigorous-intensity recreational physical activities per week in the past year);21 tertiles (low, medium, high) of Healthy Eating Index 2005 Canada (HEI-2005-Canada), reflecting the overall diet quality;22 Elixhauser comorbidity index (0, 1, >2) calculated for the overall disease burden of participants;23 24 location of residence (rural/urban areas);25 smoking status (never, occasional smoker, former daily smoker, current daily smoker); annual household income (<$20k, $20k–29k, $30k–49k, $50k–79k, $80k+) and educational attainment (high school or less, some postsecondary, postsecondary).26

Statistical analyses

Means and SD were calculated for continuous variables; proportions were calculated for categorical variables. The normality of a continuous variable was checked using Shapiro-Wilk test. Pearson’s correlation (r) was used to examine the correlation between different anthropometric changes. Student t-test (for mean of continuous variables) and χ2 test (for proportion of categorical variables) were used to examine the difference between men and women in the baseline characteristics and anthropometric changes. Statistical analyses were performed using STATA software (Stata 2007, Release 14).

The sex-specific association between anthropometric changes and incidence of diabetes was examined using multivariable Cox proportional hazard models, with anthropometric changes as time-varying exposure variable and incidence of diabetes as outcome variable.27 Survival time in Cox regression models was defined as the time to incidence of diabetes or the end of study (31 March 2015). Time-varying exposures were handled by splitting time-to-event data into multiple observations based on the dates of follow-up measurement of anthropometrics in 2004 and 2008.28 Other covariates were not considered as time-varying variables in the Cox regression models, as the covariates did not vary over time (eg, ethnicity) or information was not reassessed in follow-ups beyond the baseline measurements (eg, physical activity and dietary intake). The magnitude of association was estimated by hazard ratio (HR) of diabetes associated with per SD change, per 5% change or per categories of anthropometric changes (±5% as reference group).

A purposeful selection method was used to build statistical models, that is, the known risk factors for diabetes, such as age, social determinants (annual household income), parental history of diabetes and lifestyle behaviours (physical activity, HEI-2005-Canada categories), were forced into the regression model and other covariates, including ethnicity, the corresponding baseline anthropometrics, living in rural/urban areas, educational attainment, smoking status, comorbidity index and interaction terms, that were non-significant at p≥0.05, were removed from the final model. In this study, a priori interaction with sex was examined and the corresponding subgroup analyses were conducted. Wald’s χ² test was used to examine the significance of interaction terms (sex differences) of the association from the Cox regression models. Inverse logarithm (antilog) was applied to the beta-coefficient of interaction term to calculate sex-specific HRs and the differences.

We also undertook several sensitivity analyses: (i) excluding participants who had extreme change in anthropometrics (ie, 2% cut-off in tails of the normally distributed anthropometric changes) to estimate impact of these extreme outliers; (ii) excluding participants who had a high degree (>3%) of discrepancy in self-reported heights between baseline and follow-ups to examine the potential impact of self-report errors; (iii) including participants who had missing values in covariates using a ‘missing indicator’ category to ensure our primary analysis was not affected by these exclusions; (iv) subgroup analysis of sex differences in the association by age groups.

Patient and public involvement

Patients and/or public were not involved in development of research question, study design, conducting study and result dissemination.

Characteristics of study participants

The baseline characteristics of study participants are shown in table 1. Over 60% of participants were women. A greater proportion of men (83.0%) than women (64.6%) were considered being ‘overweight or obese’. On average, women had larger HC, but smaller WC and WHR than men (table 1). Compared with participants included in the analyses (n=19 655), participants who were excluded (n=8636) had similar age (49.4 vs 50.4 years), similar sex distribution (women: 60.5% vs 62.8%), similar levels of educational attainment (postsecondary or higher education: 71.2% vs 72.2%), slightly higher percentage of smokers (57.9% vs 54.0%) and slightly higher percentage of having multiple (≥2) comorbidities (16.6% vs 13.9%).

Anthropometric changes during follow-ups

On average, all four anthropometrics (weight, BMI, WC and WHR) increased from the baseline across follow-ups. Change in weight/BMI was moderately correlated (r=0.61) with changes in WC and weakly correlated (r=0.18) with changes in WHR. The majority (60%–70%) of participants had minimal to negligible (ie, ±5%) changes in anthropometric from their baseline measures, and a slightly greater proportion of participants had increased anthropometrics (∆>+5%) than those with a reduction (∆<–5%) (table 2). Women had a greater mean absolute gain in all four anthropometrics than men since the baseline, especially for WC and WHR (p<0.001). This sex-specific difference was more prominent in changes in WC and WHR than weight and BMI (table 2).

Sex difference in anthropometric changes and incidence of diabetes

In 203 683 PY (average follow-up time: 10.4±2.7 years), 1226 incidences of diabetes were identified, with an incident rate (IR) of 6.0 cases per 1000 PY. There was a significant higher (p<0.01) IR of diabetes in men (591 cases in 75 741.0 PY, ie, 7.8 cases per 1000 PY) than in women (635 cases in 127 941.6 PY, ie, 5.0 cases per 1000 PY). When comparing the crude IR of diabetes across different categories of anthropometric changes (±5%, >5% gain and >5% loss), participants with >5% gain in anthropometrics, especially in weight, BMI and WC, had a higher IR of diabetes compared with participants with >5% loss. This result was consistently observed in men and women (table 3).

Sex-stratified multivariable Cox regression analyses showed that on average, the positive association between anthropometric changes and risk of diabetes was greater in men than in women (table 4). Specifically, the increase in risk of diabetes associated with a per SD increase in anthropometrics was 20% (95% CI 14% to 27%) in men vs 13% (95% CI 8% to 19%) in women, respectively, for weight; 23% (95% CI 15% to 31%) vs 14% (95% CI 8% to 19%), respectively, for BMI; 25% (95% CI 18% to 34%) vs 18% (95% CI 12% to 24%), respectively, for WC and 17% (95% CI 11% to 22%) vs 8% (95% CI 5% to 11%) for WHR. Moreover, in both men and women, per SD (or per 5%) increase in WC was associated with the largest increase in risk of diabetes, followed by ∆BMI (or ∆weight) and ∆WHR (table 4).

In the regression models of the total population with an interaction term for sex, statistical tests showed that there were no significant (alpha=0.01) interactions with sex for the association between anthropometric changes and diabetes, except for changes in WHR (p<0.01). The risk difference between men and women for diabetes associated with 1 SD increase in anthropometrics was 0.07 (95% CI −0.02 to 0.14) for weight changes, 0.08 (95% CI −0.03 to 0.17) for BMI changes, 0.07 (95% CI −0.02 to 0.15) for WC changes and 0.09 (95% CI 0.03 to 0.13) for WHR changes. Similar results were observed for the sex difference in risk of diabetes associated with per 5% increase in anthropometrics and with different categories of anthropometric changes (table 4).

In sensitivity analyses of (i) excluding participants who had extreme anthropometric changes (2% cut-off in the normal distribution), (ii) excluding participants who had high degree (>3%) discrepancy in self-reported height between baseline and follow-ups, (iii) including participants who had missing values in categorical covariates and (iv) subgroup analysis of the sex difference by age groups, no significant changes were observed in beta-coefficient estimates of HRs for diabetes (online supplementary tables S1–S4).