Editorial: Decisions! Decisions!

complexities, is still poorly understood in many circles. We tell students that there is no free lunch, that if you want to use a screen to identify everyone, or almost everyone, who might have a given abnormality, then you will have to accept that you will falsely identify many individuals who later turn out to be normal. In the jargon of decision analysis theory, high hit rate means high false-alarm rate, and low false-alarm rate means low hit rate. And both numbers are critically dependent on the pass/fail criterion. Nothing illustrates these concepts better than the data provided in the paper “Transient Evoked Otoacoustic Emissions in Adults: A Comparison between Two Test Protocols” in this issue of JAAA. Authors Joseph Kei, Ravi Sockalingam, Clive Holloway, Alan Agyik, Craig Brinin, and Doreen Baine of the University of Queensland, Brisbane, Australia, set out to compare two TEOAE screening protocols in 115 adult ears. The two procedures, offered on a commercially available device, differed only in the time window over which a response is measured and the rate at which clicks are presented. The gold standard was the puretone audiometric threshold at each test frequency. The failure criterion was a threshold in excess of 15 dBHL. The authors asked how successfully the TEOAE signal-to-noise ratio (SNR) predicted the presence of such a loss, as a function of the SNR pass/fail criterion. Basic findings were straightforward. The protocol employing a shorter time window and a higher click rate (Quickscreen) yielded slightly better results than the protocol employing a longer time window and slower click rate (Default). Of particular interest to audiologists engaged in any type of screening, however, are the data that the authors have carefully catalogued in the appendices. Here are found the raw data underlying their clinical decision analysis: hit rate, false-alarm rate, efficiency, predictive value of a positive result, and predictive value of a negative result. They are presented systematically for virtually every possible pass/fail criterion. As an example, consider the data in Appendix 3. Here are listed all of the results for predicting a loss of more than 15 dB at 4000 Hz. In the upper section (Quickscreen) the first column lists possible TEOAE pass/fail criteria for SNRs ranging from -3 dB to 25 dB. The remaining columns list the various indices of screening outcomes based on each SNR criterion. If we enter the row for SNR = 3 dB (a commonly used pass/fail screening criterion), we see that the hit rate (the percentage of ears correctly identified as having a loss greater than 15 dB) is 55.6 percent. The false-alarm rate (the percentage of ears incorrectly identified as having a loss greater than 15 dB) is 11.3 percent. Not very impressive performance since, at this SNR criterion, almost half of the people we are seeking to identify are missed. But the false-alarm rate is not too bad. Only slightly more than 1 in 10 ears is incorrectly identified as having a loss. We can improve the hit rate by simply changing the TEOAE SNR criterion. Let us try an SNR of 19 dB. Now the hit rate is much better, 88.9 percent. We have only missed about 1 in 10 persons with an actual loss. But at what price? The false-alarm rate is now 85.8 percent. Here we have incorrectly identified almost as high a percentage of ears as we have correctly identified. As you study the HR and FA columns, you will see vividly how hit rate and false-alarm rate move up and down together like reluctant mates. The principle is dramatically illustrated, not only in theory but, here, in actual real-world data. But there are more columns to explore. Consider efficiency. This refers to the overall accuracy