Poor Methods Used to Develop Diabetes Risk Prediction Models
Studies often use poor development methods and reporting for type 2 diabetes risk prediction models
MONDAY, Oct. 10 (HealthDay News) -- Methodological deficiencies and poor reporting of data are seen in studies which attempt to develop risk prediction models for type 2 diabetes, according to a review published online Sept. 8 in BMC Medicine.
Gary S. Collins, from the University of Oxford in the United Kingdom, and colleagues reviewed available literature published before May 2011 to evaluate the methodology and reporting quality of studies used to develop risk prediction models for prevalent and incident type 2 diabetes in adults. A total of 39 studies describing development of 43 risk prediction models were analyzed for study design, sample size and number of events, risk predictor selection and coding, missing data, definition of outcome, model-building strategies, and aspects of performance.
The investigators found that 17 studies described model development for predicting incident type 2 diabetes, and 15 reported how models were derived to predict prevalent type 2 diabetes. The number of events per variable was fewer than 10 in nine studies, and could not be calculated in 14 studies due to insufficient reported data. Between four and 64 candidate risk predictors were used but this number was not clearly reported in seven studies. Eight studies reported using statistical significance from univariate screening to select risk predictors for multivariable models, whereas the selection procedure was not clear in 10 studies. A method of categorizing all continuous risk predictors was used to develop 21 of the risk prediction models. Sixteen studies did not mention the method of treating and handling of missing data.
"We found widespread use of poor methods that could jeopardize model development," the authors write.