Appendix 1: Description of the weighted compositional regression method for projection of multi-category BMI prevalence, including estimation of prediction intervals, goodness of fit and sensitivity analyses Each step of the weighted compositional projection method will be described in this Appendix. Vector and matrix notation will be used when possible for conciseness, and familiarity with standard multivariate regression and GLS (generalized least squares) estimation theory is presumed. 1. Construct measured BMI prevalence compositions and covariance matrix time series from survey data Let πΜ ππ (π‘) = {ππππ (π‘), π = 0 … 3} be the set of 4 BMI prevalences specific age group ′π′, sex ′π′ and measured from survey timepoint ′π‘′; the index ′π′ corresponds to BMI category. πΜ ππ (π‘) is said to form a 4-dimensional ‘composition’ (Aitchison J 1986; Mills T 2009). Let πΜΏππ (π‘) be the 4x4 measured covariance matrix of πΜ ππ (π‘). The diagonal elements of πΜΏππ (π‘) represent the variances of the prevalences of each BMI category, while the off-diagonal elements represent the covariances between different BMI categories. In the current study, time series in πΜ ππ (π‘) and πΜΏππ (π‘) comprise 14 survey cycles corresponding to timepoints π‘ ∈ (1987, 1992.5, 1994.5, 1996.5, 1998, 1998.5, 2000.5, 2002, 2003, 2005, 2007, 2008, 2009, 2010). 2. Transformation of 4-dimensional prevalence compositions to 3-dimensional real space Statistical modelling or analyses of πΜ ππ (π‘) must account for its particular numerical properties that include bounded support (each prevalence component is bounded between 0 and 1), positivity (prevalence components cannot be negative), summation constraint (∑3π=0 ππππ (π‘) = 1), and correlation between the components (including but not limited to, the effect of the common denominator). As shown by Aitchison (1986) and Aitchison and Shen (1980) (Aitchison J & Shen SM 1980; Aitchison J 1986), if πΜ ππ (π‘) has a multivariate logistic-normal distribution, the additive log-ratio transformation, π¦Μ ππ (π‘) = πππ (πΜ ππ (π‘)), can be applied to map πΜ ππ (π‘) to a 3-dimensional real space such that π¦Μ ππ (π‘) is multivariate normal. The wide range of inference tools developed for multivariate normal data (including multivariate regression) can then be applied to π¦Μ ππ (π‘); results are then backtransformed to recover the original prevalence space. This compositional approach has distinct advantages over the use of simple linear regression; the latter method can result in biased projection estimates (Mills T 2009) as well as estimated values that violate the numerical properties described. The alr-transformation is shown below: Μ ππ (π) = {ππππ (π), π = π … π} = {πππ ( π π·πππ (π) ) , π = π … π} π − ∑ππ=π π·πππ (π) The prevalence component π0ππ (π‘) is termed the ‘fill-up’ value and was assigned to the underweight BMI category; however the results of the analysis are in fact invariant to the choice of category for π0ππ (π‘) (Aitchison J 1982). 3. Calculation of the transformed covariance matrices ΜΏππ (π‘) = πππ (π¦Μ ππ (π‘)) denote the 3x3 covariance matrix of the transformed BMI composition corresponding to Let Ω ΜΏππ (π‘) is estimated from πΜΏππ (π‘) using the delta method: age category ′π′, sex ′π′ and timepoint ′π‘′. Ω ππ¦Μ (π‘) ππ¦Μ (π‘) ΜΏππ (π‘) = ( ππ ) πΜΏππ (π‘) ( ππ ) Ω ππΜ ππ (π‘) ππΜ ππ (π‘) π 4. Projection analysis using weighted multivariate regression 1 Multivariate regression is used to fit and extrapolate trends in π¦Μ ππ (π‘): ΜΏ π¦ΜΏππ = πΜΏπ½ππ For the current study π¦ΜΏππ is the 14x3 matrix representing the time series in each component of π¦Μ ππ (π‘), πΜΏ is the 14x2 ΜΏ is the 2x3 matrix of regression coefficients. To covariate matrix consisting of intercept and time columns, and π½ππ account for the covariance matrix unique to each survey timepoint, the preceding regression equation is reduced to a univariate form by ‘stacking’ the matrix terms (Jobson JD 1991): ∗ Μ ∗ , π¦Μ ππ = πΜΏ ∗ π½ππ ∗ ΜΏ∗ Μ ∗ are the corresponding 42x1, 42x6 and 6x1 stacked matrices. where π¦Μ ππ , π and π½ππ The regression coefficients can then be estimated using the standard GLS estimator (Kutner M et al. 2005): ∗−1 ΜΏ ∗ ΜΏ ∗π ΜΏ ∗−1 ∗ ∗ Μ ∗ = (πΜΏ ∗π Ω ΜΏππ ΜΏππ π½ππ π )π Ωππ π¦Μ ππ , where Ω is the ‘stacked’ covariance matrix that includes the transformed covariances for all 14 timepoints. 5. Estimation of projected prevalence The projection estimate for future time π‘′ ∈ (2011, 2012, … , 2030) is determined by extrapolation of the regression model, for example using the stacked notation: ∗ Μ ∗ , π¦ππ (π‘ ′ ) = π¦Μ ππ = π₯ΜΏ ∗ π½ππ where π₯ΜΏ ∗ is a 3x6 matrix that specifies the intercept and particular future time π‘ ′ for each component of π¦ππ (π‘ ′ ). Backtransformation πΜ ππ (π‘ ′ ) = πππ −1 (π¦Μ ππ (π‘ ′ )) is used to recover the projected BMI prevalences. The back transformation equations are shown below: ππππ (π‘ ′ ) = π0ππ (π‘ ′ ) = ππ₯π (π¦πππ (π‘ ′ )) 1 − ∑3π=1 ππ₯π (π¦πππ (π‘ ′ )) , for π = 1 … 3 1 1 − ∑3π=1 ππ₯π (π¦πππ (π‘ ′ )) References Aitchison J 1982. The Statistical Analysis of Compositional Data. J. R. Statist. Soc. B 44: 139-177. Aitchison J 1986. The Statistical Analysis of Compositional Data. Chapman and Hall, London. Aitchison J and Shen SM 1980. Logistic-normal distributions: Some properties and uses. Biometrika 67: 261. Jobson JD 1991. Applied Multivariate Data Analysis: Regression and Experimental Design, Volume 1. Springer-Verlag. Kutner M, Nachtsheim C, Neter J, and Li W 2005. Applied Linear Statistical Models. Fifth ed. McGraw-Hill Irwin, New York. Mills T 2009. Forecasting obesity trends in England. J. R. Statist. Soc. A 172, Part 1: 107-117. 2 Appendix 2: Prediction Intervals, Goodness of Fit and Sensitivity Analyses 1. Estimation of Prediction Intervals The prediction variance of a projected BMI prevalence at time π‘ ′ is the sum of variance in estimated mean response and future measurement error (Kutner M et al. 2005): ππππ(ππππ (π‘ ′ )) = ππππ(ππππ (π‘ ′ )) + ππππ(ππππ (π‘ ′ )), where the prevalence specific to BMI category ′π′, age group ′π′, sex ′π′ and measured from survey timepoint ′π‘′ is considered. The covariance matrix of the estimated mean value of π¦Μ ππ at future time π‘ ′ (represented by π₯ΜΏ ∗ ) is calculated using Μ ∗ ) = ΜΏππ (π‘ ′ ) = π₯ΜΏ ∗ πΜΏπ½∗ π₯ΜΏ ∗π , where πΜΏπ½∗ = πππ(π½ππ standard GLS (generalized least squares) theory: πππ (π¦Μ ππ (π‘ ′ )) = Ω −1 ∗−1 ΜΏ ∗ ΜΏππ π ) (πΜΏ ∗π Ω is the covariance matrix of the estimated regression coefficients. The covariance matrix of the corresponding backtransformed prevalence composition at time π‘ ′ , is then estimated using the delta method: ′ Μ ′ Μ π ππππ (π‘ ) ΜΏππ (π‘ ′ ) (ππππ (π‘ ′ )) . The variance in estimated mean response, ππππ(ππππ (π‘ ′ )), πππ (πΜ ππ (π‘ ′ )) = πΜΏππ (π‘ ′ ) = (ππ¦Μ (π‘ ′ )) Ω ππ¦Μ (π‘ ) ππ ππ is obtained from the diagonal elements of πππ (πΜ ππ (π‘ ′ )). ππππ(ππππ (π‘ ′ )) is estimated by the variance structure of the 2010 CCHS, under the approximation that the survey design and sampling size of a future survey at time π‘ ′ will be similar: ππππ(ππππ (π‘ ′ )) ≈ πππ (ππππ (t =2010)). The prediction intervals for the projected prevalence of BMI category ′π ′ , age category ′π′, sex ′π′ at future time ′π‘ ′ ′ are then estimated using a Normal approximation, by: ππππ (π‘ ′ ) ± 1.96 × √ππππ(ππππ (π‘ ′ )) + ππππ(ππππ (π‘ ′ )). Prediction intervals represent estimates of the statistical uncertainty caused by variability in the measured data and the model estimation process. Prediction intervals in projection analyses are always underestimates of the actual error since they do not account for the structural error component, and thus they should not be interpreted as reliable bounds on estimated future values (Hakulinen et al. 1986; Moller et al. 2005). Prediction intervals are useful however for indicating the statistical stability of the projection model as well as providing a lower bound on the actual error. Prediction intervals for age aggregated BMI prevalence projections are shown in Figure A2 below. 3 Figure A2 Projections (2011 to 2030) of age-aggregated prevalence by BMI category, for men and women. The linear scenario is indicated by the black line (β), the deceleration scenario is indicated by the gray line (β), and the historical BMI time series data are indicated by the open circles (β). The dotted black (…) and gray (…) lines indicate prediction intervals for the linear and deceleration scenarios. The blue (β) and red (β) lines indicate the sensitivity analysis results for the linear and deceleration scenarios. 4 2. Goodness of Fit Visual Assessment Both linear and log models appeared to fit the age-aggregated (Figure 1 of the main text) and age-category specific (Appendix 4) measured prevalence trends well. The fitted models followed trends that were clearly apparent in the measured data, and residuals did not indicate the presence of any systematic bias. Quantitative Assessment 2 Buse 1973 (Buse A 1973) derived the π πΊπΏπ statistic to assess goodness of fit for a GLS model that accounts for the weighting matrix in computing the residual sums of squares and the mean response: π (π − πΜ) π −1 (π − πΜ) =1− , (π − πΜ π)π π −1 (π − πΜ π) where π −1 is the weighting matrix, π π = (1, … ,1), and πΜ is the weighted mean of the response variable: π = 2 π πΊπΏπ π π π −1 π . π π π −1 π 2 This measure is analogous to the OLS (ordinary least squares) π ππΏπ , and measures the proportion of the 2 2 generalized sums of squares attributable to the regression model. π πΊπΏπ ranges from 0 to 1 and reduces to the π ππΏπ 2 when the weighting matrix is identity. π πΊπΏπ was computed for each of the eight age by sex regression analyses, and for both linear and deceleration scenarios, as shown in Table A2.1 below. All values were found to be > 0.96 indicating goodness of fit. Table A2.1 However as each regression analysis comprises fitting to all 3 transformed BMI categories simultaneously, this can 2 lead to large and uninformative values of π πΊπΏπ as the between BMI category variation can mask within BMI category 2 variation. Thus π πΊπΏπ was further applied to assess goodness of fit within the obese BMI category only, using subsets of the measured and fitted backtransformed data and covariance matrices. Results are shown in Table A2.2 below. 2 It can be seen that π πΊπΏπ > 0.6 for fitted obesity trends in all age and sex categories, except for 18-24 year old women. 2 As can be seen in Appendix 3, Figure A3.1 however, the low values of π πΊπΏπ for this stratum are due to the ‘flatness’ of the obesity trend and consequent low percentage of variability explained by the fitted trend, rather than poor goodness of fit. Table A2.2 5 3. Sensitivity Analyses Sensitivity analyses were performed to assess the robustness and validity of the projection results. These analyses comprised estimations of projected age-aggregated prevalence trends that used only the seven CCHS general health surveys from 2000-2001 to 2010. The purpose of the analyses was to assess primarily if projected trends might change if only the more recent historical data were used, and thus if the older historical data might be exerting undue influence on the projections. The analysis was further used to examine if additional stabilisation or ‘levellingoff’ in recent obesity prevalence trends might be present, as well as if the analysis of a more homogeneous time series might yield different results. Results of the sensitivity analyses are shown superimposed on projection results, in Figure A2. It can be concluded that the projection results are robust to the use of more recent historical data. Projected trends using the data subset follow the full dataset projections. In particular, the projected obesity prevalence spanned by the linear and deceleration scenarios using the data subset lie within the span projected by the full dataset, for both men and women. References Buse A 1973. Goodness of Fit in Generalized Least Squares Estimation. The American Statistician 27: 106-108. Hakulinen,T., Teppo,L., and Saxen,E. 1986. Do the predictions for cancer incidence come true? Experience from Finland. Cancer. 57: 2454-2458. Kutner M, Nachtsheim C, Neter J, and Li W 2005. Applied Linear Statistical Models. Fifth ed. McGraw-Hill Irwin, New York. Moller,B., Weedon-Fekjaer,H., and Haldorsen,T. 2005. Empirical evaluation of prediction intervals for cancer incidence. BMC. Med Res Methodol. 5: 21. 6 Appendix 3: Projection of type 2 diabetes: (1) separation of epidemiologic and demographic components, and (2) survey estimates of BMI, age and sex-specific prevalence of type 2 diabetes 1. Separation of epidemiologic and demographic components As described in the Methods (Projected impact of BMI on chronic disease prevalence) section of the manuscript, the projected BMI, age and sex specific numbers of cases of type 2 diabetes (ππππ (π‘)) is equal to the product of the type 2 diabetes prevalence in 2009-2010 (ππππ ) and projected numbers of individuals (ππππ (π‘)) corresponding to the same category. ππππ (π‘) itself is the product of the projected BMI prevalence (ππππ (π‘)) and population (πππ (π‘)). Thus, ππππ (π‘) = ππππ × ππππ (π‘) = ππππ × ππππ (π‘) × πππ (π‘) Total projected numbers of cases (ππ (π‘)) and prevalence (ππ (π‘)) by sex is obtained by aggregation over BMI and age categories as was described in the main manuscript (Methods, Projected impact of BMI on chronic disease prevalence). The equations for ππ (π‘) and ππ (π‘) are shown below, showing the respective dependence of these quantities on BMI prevalence (termed the epidemiologic component) and population projections (the demographic component): ππ (π‘) = ∑ ππππ × ππππ (π‘) × πππ (π‘) π,π ππ (π‘) = 1 ∑ ππππ × ππππ (π‘) × πππ (π‘) ππ (π‘) π,π Estimation of the demographic contribution to type 2 diabetes numbers (ππ (π‘)) and prevalence (ππ (π‘)) is done by holding the age and sex specific BMI prevalences at a fixed reference level ππππ (π‘πππ ) and allowing only the population to evolve. Thus: ππππ (π‘) = ∑ ππππ × ππππ (π‘πππ ) × πππ (π‘) ππππππ (π‘) = ∑ ππππ π,π π,π 1 1 ππππ (π‘) = ππππππ (π‘) = ∑ ππππ ∑ ππππ × ππππ (π‘πππ ) × πππ (π‘) ππ (π‘) ππ (π‘) π,π π,π The epidemiologic contribution (due to BMI related change) is estimated by subtracting the demographic effect from the overall trends for both numbers and prevalence. Thus πππ ππ (π‘) = ππ (π‘) − ππππππ (π‘) πππ ππ (π‘) = ππ (π‘) − ππππππ (π‘) πππ ππ (π‘) can be interpreted as the component of change in prevalence amenable to intervention on population BMI, πππ and ππ (π‘) as the estimated number of avoidable cases of type 2 diabetes. 7 2. 2009-2010 cross-sectional survey estimates of BMI, age and sex-specific prevalence of type 2 diabetes In the current projections, the BMI, age and sex specific prevalence of type 2 diabetes (ππππ ) is estimated from the 2009-2010 Canadian Community Health Survey (CCHS), and is assumed to remain constant over the projected time horizon. Figures A3a and A3b show plots of the values of ππππ for men and women respectively. BMI categories are indicated on the x-axis and the age categories are represented by separate curves. The marked increase in type 2 diabetes prevalence with BMI can be seen by the upward trends in the curves for both men and women. There is also a substantial increase in prevalence with age, as shown by the separation of the different curves in both plots. Thus it is expected that increasing BMI and population aging will likely combine to drive projected future increases in type 2 diabetes prevalence. Figure A3, Type 2 diabetes prevalence vs. BMI plotted by age, for (a) Men and (b) Women. 8 Appendix 4: Age-specific BMI prevalence projections Figure A4.1, Projections of obesity prevalence by age category and sex for men and women. The linear scenario is indicated by the black line (β), the deceleration scenario is indicated by the gray line (β), and the historical BMI time series data are indicated by the open circles (β). The dotted black (…) and gray (…) lines indicate prediction intervals for the linear and deceleration scenarios. 9 Figure A4.2, Projections of overweight prevalence by age category and sex for men and women. The linear scenario is indicated by the black line (β), the deceleration scenario is indicated by the gray line (β), and the historical BMI time series data are indicated by the open circles (β). The dotted black (…) and gray (…) lines indicate prediction intervals for the linear and deceleration scenarios. 10 Figure A4.3, Projections of normal weight prevalence by age category and sex for men and women. The linear scenario is indicated by the black line (β), the deceleration scenario is indicated by the gray line (β), and the historical BMI time series data are indicated by the open circles (β). The dotted black (…) and gray (…) lines indicate prediction intervals for the linear and deceleration scenarios. 11 Figure A4.4, Projections of underweight prevalence by age category and sex for men and women. The linear scenario is indicated by the black line (β), the deceleration scenario is indicated by the gray line (β), and the historical BMI time series data are indicated by the open circles (β). The dotted black (…) and gray (…) lines indicate prediction intervals for the linear and deceleration scenarios. 12