INCREASED EFFICIENCY OF ANALYSES - CUMULATIVE LOGISTIC-REGRESSION VS ORDINARY LOGISTIC-REGRESSION

The common practice of collapsing inherently continuous or ordinal var iables into two categories causes information loss that may potentiall y weaken power to detect effects of explanatory variables and result i n Type II errors in statistical inference. The purpose of this investi gation was to illustrate,using a substantive example, the potential in crease in power gained from an ordinal instead of a dichotomous specif ication for an inherently continuous response. Ordinary (OLR) and cumu lative logistic regression (CLR) modeling were used to test the hypoth esis that the risk of alveolar bone loss over 2 years is greater for s ubjects with poorer control of non-insulin-dependent diabetes mellitus (NIDDM) than for those who do not have diabetes or have better contro lled NIDDM. There were 359 subjects; 21 of whom had NIDDM. Analysis of main effects using OLR for the dichotomous outcome (no change in radi ographic bone loss vs any change) produced parameter estimates for bet ter control and poorer control that were not statistically significant . CLR analysis of main effects using a 4-category ordinal specificatio n for radiographic bone loss also produced a parameter estimate for be tter control that was not statistically significant, but which estimat ed poorer control to have a significant effect. The fit of this CLR mo del was significantly better at P<0.05 than that for the OLR. While an OLR model testing the interaction between age and control status did not converge after 100 iterations, the CLR interaction model converged without difficulty and estimated a significant effect for interaction between age and poorer control. Results from the CLR analysis, in con trast to the OLR model, would lead one to conclude that the risk for m ore severe bone loss progression after 2 years is greater in subjects with poorer controlled NIDDM and that subjects with better controlled NIDDM may not have greater risk of bone loss progression than those wi thout diabetes. The use of an ordinal instead of a dichotomous specifi cation for an inherently continuous response provided increased power, more precise parameter estimates, and a significant ly better fitting model. In estimating parameter estimates for odds ratios or risks, it is important to consider using ordinal logistic regression where the response is inherently continuous or ordinal.