Spline regression in clinical research

The objective of this article is to illustrate the statistical technique of spline regression, an under-utilized tool in clinical research. Spline regression was used to assess the dose-response association between serum albumin and hospital mortality.

Data from a previous study of patients hospitalized throughout Florida, United States of America (U.S.A.), for invasive group A streptococcal disease were accessed. For the current analysis, serum albumin (SA) at the time of admission was the risk factor of interest. The outcome was unadjusted hospital mortality among 117 patients. First, a traditional, suboptimal approach was employed by breaking SA into three groups and calculating the crude hospital mortality rate in each SA category. The second approach involved the creation of a curve using a quadratic spline model.

The traditional approach yielded only three points of information: the hospital mortality rate for the three SA groups. Among patients whose SA upon admission was < 2.5 g/l00 mL, 2.5 to 3.4, and 3.5 or greater, the hospital mortality rate was 40.7%, 14.8%, and 8.3%, respectively. The spline model, however, resulted in a smooth curve which was more clinically plausible.

The goal of this paper is to expose clinicians to splines. Spline regression, unlike categorical analysis, does not impose the unrealistic assumption of a homogenous risk within categories. Another disadvantage of categorical analyses is that they allow large changes in risk between categories,