Pediatric Obesity · 2 authors, 3 centres
This summary was generated by AI from a single paper. It has not been reviewed by a clinician and is not clinical advice. Verify against the source before acting on it.
This study presents an improved algorithm that harmonizes child overweight and obesity prevalence rates across different BMI references (WHO, IOTF, CDC). By using paired prevalence data for overweight and obesity, the algorithm reduces residual variance from 11.8% to 3.3%, explaining 96.7% of baseline variance. This tool enables direct comparison of prevalence data from major global studies, addressing a key barrier in childhood obesity research.
**Background:** Child obesity prevalence rates vary depending on which BMI reference and cut-off are used (WHO, IOTF, CDC), making it impossible to directly compare results from major global studies like NCD-RisC and GBD. An earlier algorithm attempted to convert prevalence rates between references but left 12% of variance unexplained.
**Methods:** The improved algorithm uses paired prevalence rates of overweight (including obesity) and obesity from a target group based on one reference (e.g., WHO) to estimate the slope (b) of the QQ plot (z-score prevalence vs. z-score cut-off). This slope is then used to adjust z-scores for the other reference's cut-offs and back-transform to predicted prevalence. The algorithm was tested on 228 groups of children aged 6–17 years from 20 countries, using data from COSI (13 European countries, n=225,190), Deren et al. (Ukraine, n=18,144), and individual-level data from CHNS (China), ENSANUT/ENNViH (Mexico), NHANES (US), HSE (UK), RLMS (Russia), and Algerian data (total n=481,709). Prevalence rates were calculated for WHO, IOTF, and CDC references. The analysis involved weighted linear regression of Z on z across four cut-offs (two per reference) to estimate b and predict the unknown prevalence rates.
**Key Results:** The median correlation between z-score prevalence and mean z-score cut-off across the four points was 0.9994 (p=0.0006), with only 3 of 580 comparisons below 0.98. The regression coefficient b ranged from 0.4 to 1.5 (mean 0.81, SD 0.17). The overall SD of residuals (observed minus predicted prevalence) was 0.8% (n=2320 comparisons), compared to a baseline SD of 4.3% for the difference between original prevalence rates—meaning the algorithm explained 96.7% of baseline variance (vs. 88.2% for the original algorithm). Residual SD was 1.0% for overweight and 0.6% for obesity. Correlations between observed and predicted prevalence were 0.995 for obesity and 0.996 for overweight (both p<0.0001). The algorithm was unbiased (mean residual near zero). Four-fifths of obesity residuals were <0.6% in absolute value (max 2.5%), and four-fifths of overweight residuals were <1.0% (max 8.2%, in small groups with ≤224 children).
**Clinical Implications:** This algorithm enables direct comparison and combination of child obesity prevalence data from studies using different BMI references, eliminating the 'silo effect' that has hampered global obesity surveillance. It is reversible and generalizable across populations, requiring only that both overweight and obesity prevalence rates are available and non-zero. The algorithm is implemented in the CRAN sitar package (function ob_convertr2). Limitations include lack of validation for children under age 5 and the need for sufficiently large groups to ensure non-zero obesity prevalence.