On the statistical significance of electrophysiological steady-state responses

Abstract

Steady-state stimulation is a useful paradigm in many physiologic and clinical situations, for ERG, Pattern-ERG and VEP. One of the advantages is the easy evaluation of the response via Fourier analysis. However, the question whether a given response is statistically significant or not has received little attention so far, although it is especially relevant in high noise, low amplitude recordings, as often occur in pathologic conditions. A given response is statistically significant if it is unlikely that its value is due to noise fluctuations. Thus appropriate estimates of noise and response are required. We have analytically derived formulas for the statistical significance of a given signal-to-noise-ratio s, based on two different estimates of noise: (1) Noise estimate by a `no stimulus' recording, or by a `±average'. The former needs an additional recording, the latter can simultaneously be calculated as the standard average. (2) Noise is estimated as the average of the two neighboring spectral lines (one below, and one above the response frequency). Analytical solutions were obtained for both noise estimates that can easily be evaluated in all appropriate recordings. Noise estimate (1) performs much poorer than noise estimate (2), as can be seen from the following landmark values: Typical significance levels of 5%, 1%, and 0.1% require s values of 4.36, 9.95, and 31.6 (1), and 2.82, 4.55, and 8.40 (2). The noise estimate based on the neighboring frequencies can be easily applied after recording, provided that the noise spectrum is reasonably smooth around the response and frequency-overspill was avoided. It allows a quantitative assessment of low responses in physiological threshold analyses and pathological conditions, e.g., `submicrovolt flicker-ERG'.