Abstract
OBJECTIVES: Audiologists regularly use serial monitoring to evaluate changes in a patient's auditory function over time. Observed changes are compared with reference standards to determine whether further clinical action is necessary. Reference standards are established in a control sample of otherwise healthy subjects to identify the range of auditory shifts that one might reasonably expect to occur in the absence of any pathological insult. Statistical approaches to this seemingly mundane problem typically invoke 1 of 3 approaches: percentiles of the cumulative distribution, the variance of observed shifts, and the "standard error of measurement." In this article, the authors describe the statistical foundation for these approaches, along with a mixed model-based alternative, and identify several necessary, although typically unacknowledged assumptions. Regression to the mean, the phenomenon of an unusual measurement typically followed by a more common one, can seriously bias observed changes in auditory function and clinical expectations. An approach that adjusts for this important effect is also described. DESIGN: Distortion product otoacoustic emissions (DPOAEs) elicited at a single primary frequency, f2 of 3175 Hz, were collected from 32 healthy subjects at baseline and 19 to 29 days later. Ninety percent test-retest reference limits were computed from these data using each statistical approach. DPOAE shifts were also collected from a sample of 18 cisplatin patients tested after 120 to 200 mg of cisplatin. Reference limits established according to each of the statistical approaches in the healthy sample were used to identify clinically alarming DPOAE shifts in the cisplatin patient sample. RESULTS: Reference limits established with any of the parametric methods were similar. The percentile-based approach gave the widest and least precisely estimated intervals. The highest sensitivity for detecting clinically alarming DPOAE shifts was based on a mixed model approach that adjusts for regression to the mean. CONCLUSIONS: Parametric methods give similar serial monitoring criteria as long as certain critical assumptions are met by the data. The most flexible method for estimating test-retest limits is based on the linear mixed model. Clinical sensitivity may be further enhanced by adjusting for regression to the mean.