Background Many factors influence survival from colorectal cancer, including stage of diagnosis, treatment centre, and associated risk factors such as age, sex and socioeconomic background (SEB). Previous studies have typically considered stage of disease as a potential confounder. Stage however may lie on the causal path between SEB and survival. Statistical adjustment for stage as a confounder can therefore introduce interpretational difficulties due to the reversal paradox. Classification of stage may also be imprecise and incomplete, leading to bias. Latent Class Analysis (LCA) may minimise these problems by including stage as a ‘class predictor’, accommodating uncertainty associated with stage explicitly via the latent class part of the model.

Methods We used a dataset of patients in a large UK regional population diagnosed with colorectal cancer between 1998 and 2004. The outcome was death within three years. Diagnostic centre was the NHS Trust where the latest staging took place. Following exclusions, 24,640 records were available for analysis. We constructed multilevel latent class models to allow for the hierarchical data structure: patients nested within NHS Trusts. The optimum number of latent classes at the patient and Trust level was determined with reference to likelihood-based model-fit criteria and classification error.

Results The three-patient five-Trust class multilevel LCA model was chosen, and provided a better fit to the data than standard multilevel models. Patients were apportioned into either a good, reasonable or poor prognosis class. The patient classes had a graduated survival status analogous to that observed for different stages of disease. More deprived SEB and older age were associated with increased odds of death in all classes (SEB OR 1.33 [95% CI 1.26, 1.41]; age OR 1.46 [1.33, 1.60] per 5-year increase in the good prognosis class, for example). Females had significantly decreased odds of death compared with males in the good prognosis class only (0.59 [0.40, 0.87]).

Discussion Multilevel LCA improves upon the standard multilevel approach by producing models that are better fitting to the data and provide an enhanced interpretation of the data, while minimising interpretational difficulties due to the reversal paradox and avoiding bias due to incomplete data. Trust classes identified outlying Trusts, indicating that the standard multilevel model would not have been sufficient to model these data. The Trust classes differ only due to patient survival status and the relationship between survival and the covariates, potentially enabling us to highlight differences in patient care that might explain the differences and so be worthy of further investigation.