Statistical Confidence

Experiment-wide false positive and false negative rates were determined from calibration curves, as described in Section 2.3.5.

Following the suggestion of a colleague (B. Tyler, personal communication), comparison-wide confidence levels were also established, by evaluating the confidence level for rejecting the null hypothesis that a sequence is from taxon $A$. This measure of confidence varies continuously with $t$, and is computed as $1-[A(t)-B(t)]/[A(t)+B(t)]$, where $A(t)$ and $B(t)$ are the calibration curves, or cumulative distribution functions of $t$, obtained from the resampling procedure described above. The confidence function was calculated from normal approximations to calibration curves. Parameters for the normal approximation were inferred from the calibration curves. This calculation made it possible both to determine the error rate for any comparison, and to establish a fixed confidence level and interpret as statistically significant only those test results that yield a value greater than the confidence function evaluated at that confidence level.