Short answer
Based on the experimental restrictions provided in the question and comments, my best bet would be to include 6 catch trials and to randomly present them from trial 6 and onward. Trials 1 & 2 are reserved for training purposes, trials 3 - 5 I would reserve for above-threshold stimuli.
Background
First a problem breakdown as I understand it from the question and the comments:
- A Yes/No task is performed using method-of-constant stimuli;
- Dynamic stimulus range is known and expected outcomes are available;
- The detection rate is important, whereas a 50% threshold or psychometric curve are not of interest;
- Zero-stimulus ('catch') trials are a necessary component to validate the method in the normal-group;
- No pilot experiment can be performed to validate the methods in a small group of controls;
- Two practice trials are needed (stimulus-tone alone; noise-alone);
- Critical bottle neck: 20 trials per run
So - here is my suggestion, based on my experiences.
First off, in a yes/no task there's anything possible from correct rates from 0% to 100%. To have a reasonably estimate of the correct rate at a certain constant stimulus, multiple trials are necessary. With one trial you only get 0% or 100% correct - such a binary approach may make sense, but it will add substantial jitter. With four trials, you basically end up with correct-rate increments of 25%. Not too bad, but not pretty either. One subject error and the outcome will be substantially affected. With 8 trials, you are left with steps of 12.5%. From my point of view, 8 trials is the least you should have to have a reasonable estimate with some precision and to allow the subject to make an error without too much of an affect on the outcome.
From a more practical standpoint, you are left with 18 trials per run per subject (?), namely 20 trials minus two practice runs. At the minimum, from my stand point, you should use 3 stimulus levels: zero-intensity (catch trials) and two above detection-threshold stimuli to spread your chances of finding a statistical difference. I would recommend one far-above detection threshold ('High' condition), where also your impaired group will reach correct rates >50% and one low level ('Low) where your impaired group will score <50% to have a condition that will make a positive outcome (difference control-impaired) guaranteed.
This means 18 trials for three conditions: 'Catch'; 'Low'; 'High'. That means 6 trials per condition, leaving you with correct-rate increments of 16.7%. Not too bad.
Now, where to put those catch trials into the picture?
I would personally enforce the paradigm to start with the two practice trials (first the tone, then noise) and then have at least the first two or three trials above threshold; best thing to do is have the 'High' condition first for two times, then the 'Low' and start randomizing from there. Again, this is all personal opinion, based on my experience that dis-ambiguating the test protocol is paramount. There's nothing worse than a confused subject. By having 2 well-above detection threshold stimuli (the same as training trial #1 I guess) the task is clear cut. If you start off with a catch trial after training they will likely be confused ('I didn't hear a tone!?' ; 'Should I press the button?').
