These averages strike me as almost entirely useless! If only half of the people taking the survey are lesswrong participants then the extra noise will overwhelm any signal when the probabilities returned by the actual members are near to either extreme. Using averaging of probabilities (as opposed to, say, log-odds) is dubious enough even when not throwing in a whole bunch of randoms!

(So thankyou for providing the data!)