The need for better analysis of observational studies in orthopedics

Abstract

Background and purpose The conventional statistical methods employed in observational studies in orthopedics require the fundamental assumption that the outcomes are independent. However, fractures treated by the same surgeon cannot be regarded as being independent of each other and should be nested in the statistical analysis. If the effect on outcome of early rather than delayed surgery depends on the severity of the fracture, we have a case of interaction. This is rarely considered in orthopedic research, but could affect the conclusions drawn. The aim of this paper is to describe the concepts of multilevel modeling and interaction in orthopedics. Patients and methods In a cohort of 112 patients with single supracondylar humerus fractures, 78 patients were examined clinically on average 4 years after surgery. The range of motion was measured and the global satisfaction was assessed. The results were used to compare traditional least-squares regression analysis with a 2-level model with interactions. Results We found that 25% of the variance in outcome could be attributed to between-surgeon variance. We identified an interaction between the surgeons' experience and the severity of the fractures that influenced the conclusions. The variable “number of pins” was not significant in the 2-level model (p = 0.07), while the ordinary least-squares analysis gave a result that was statistically significant (p = 0.01). Interpretation Researchers should consider the need for a 2-level model and the presence of interactions. Standard statistical methods might lead to wrong conclusions.