Aims/IntroductionLimited data are available regarding the performance of non‐high‐density lipoprotein cholesterol (non‐HDL) in predicting incident diabetes. We aimed to analyze the association between non‐HDL and development of diabetes, and to estimate the cut‐off point of non‐HDL for discriminating incident diabetes in people with normal glucose tolerance.Materials and MethodsOf 3,653 middle‐aged and elderly Chinese with normal glucose tolerance at enrollment, 1,025 men and 1,805 women returned to the 3‐year follow up and were involved in the final analysis. Logistic regression analysis was used to test the association between cholesterol indices and incident diabetes, and receiver operating characteristic analyses were used to identify the optimal cut‐off of each cholesterol variable for incident diabetes.ResultsNon‐HDL was an independent risk factor for diabetes for women, but not for men. In women, a 1‐standard deviation increment in non‐HDL was associated with a 1.43‐fold higher risk of diabetes (95% confidence interval 1.14–1.79; P = 0.002), whereas odds ratios for total cholesterol and low‐density lipoprotein cholesterol were 1.33 (95% confidence interval 1.06–1.67; P = 0.015) and 1.30 (95% confidence interval 1.04–1.64; P = 0.024), respectively. The discriminatory power and the optimal cut‐off value of non‐HDL for incident diabetes increased across body mass index categories. For women with obesity, the threshold of non‐HDL for screening of diabetes was estimated as 3.51 mmol/L.ConclusionsNon‐HDL had better performance than traditional cholesterol indices in predicting diabetes in women, but not in men. A body mass index‐specific threshold value for a non‐HDL‐controlling target is required in the prevention of type 2 diabetes.