We used two sample
t tests to compare mean baseline values of continuous variables in people who developed diabetes and those who did not. Where appropriate, we log transformed variables and present geometric means and approximate standard deviations. We used the χ
2 test to compare categorical variables. We assessed the association of each genotype with risk of diabetes by logistic regression analysis and summarised the data by odds ratios and 95% confidence intervals. We used published regression coefficients to calculate the Cambridge type 2 diabetes risk score and Framingham offspring study type 2 diabetes risk score for each participant.7 (
link)
8 (
link) In addition, we calculated two genetic scores. In the first, we assigned each person a score based simply on the number of risk alleles carried. Thus for
CDKAL1,
CDC123/CAMK1D,
FTO,
HNF1A,
IGFBP2,
KCNJ11,
NOTCH2, TCF2,
TCF7L2,
TSPAN8/LRG5, and
VEGFA, we coded genotypes “0” for common allele homozygotes,11 (
link) “1” for heterozygotes, and “2” for rare allele homozygotes,22 (
link) and for
ADAMTS9,
BCL11A,
CALPN10,
CDKN2A/2B,
HHEX,
JAZF1,
PPARG,
SLC30A8, and
THADA, coding was “2” for common allele homozygotes and “0” for rare allele homozygotes,11 (
link) as the rare allele is reported to be protective (see web table B). In the second score, we calculated a genetic risk function by using weights derived from the risk coefficient for each gene based on odds ratios reported in previous meta-analyses (web table A).15 (
link)
16 (
link)
18 (
link)
32 (
link) Risk estimates for each allele were available for 18 genes, and we multiplied these coefficients by 0, 1, or 2 according to the number of risk alleles carried by each person. Where effect estimates were reported for carriage of either one or two copies of each risk allele as a single group (
CALPN10 and
HNF1A), we multiplied risk coefficients by a score of 0 or 1. We assumed genetic and clinical variables to be independent and added the weighted genetic score to each of the risk algorithms to provide a combined phenotypic and genetic score.
We assessed discrimination with the detection rate, which is equivalent to sensitivity and defined as the proportion of all cases detected for a pre-specified false positive rate, as well as the area under the receiver operating characteristics curve. We assessed the calibration of the Cambridge risk score and Framingham offspring risk score in the estimation of the absolute risk of type 2 diabetes by comparing the difference between observed and expected event rates in different categories of risk with the Hosmer-Lemeshow test, with Akaike’s information criterion and the likelihood ratio test as global measures of model fit.33 (
link) We used the net reclassification improvement measure to assess the extent to which adding the genetic variables reassigned people to risk categories that better reflected their final outcome.34 (
link)
Talmud P.J., Hingorani A.D., Cooper J.A., Marmot M.G., Brunner E.J., Kumari M., Kivimäki M, & Humphries S.E. (2010). Utility of genetic and non-genetic risk factors in prediction of type 2 diabetes: Whitehall II prospective cohort study. The BMJ, 340, b4838.