Objectives: The risk for disease or a bad prognosis can be calculated by means of prediction or classification models that take into account multiple variables. Different methods exist to construct such models. Some of those methods, including the likelihood ratio (LR) product method neglect dependency between variables. We aimed to evaluate the effect of neglecting dependency between variables in prediction or classification models.
Patients and methods: Population I consisted of 1003 consecutive patients with a new diagnostic problem for which RA was included in the differential diagnosis and final diagnoses (RA or non-RA) were established after 1 year. The baseline variables included in the model are rheumatoid factor, anti-citrullinated protein/peptide antibodies and the HLA-shared epitope. Population II consisted of 847 patients with definite ankylosing spondylitis (AS). Six variables (psoriasis, inflammatory bowel disease, uveitis, HLA-B27 status and latest available CRP) were evaluated. Here, specificities of the features were derived from literature and different scenarios of association between variables in controls and diseased are estimated.
Results: When two features are similarly associated in cases and controls, risks for disease will be overestimated by neglecting dependency between variables. In the presented datasets, this resulted in a up to 12% overestimation of the risk.
Conclusions: We showed how the height of over- or underestimation of risks can be evaluated when dependencies between two variables are neglected. This is important to evaluate the predictive value of combinations of features in cases where no data are available on associations in controls.