(FORTUNE Magazine) – Dear Mr. Statistics: As a chap who took the liberal arts course and still keeps forgetting what a standard deviation is, I was recently unnerved to read in the Journal of Educational Psychology (Vol. 74, No. 3) that Japanese school children study statistical sampling in the fifth grade. Not having a Nipponese 10-year-old standing in the closet, I am writing for your help with my latest problem: explaining those mysterious formulations in the political news about margins of sampling error and whether it is or isn't a big deal that Ross Perot leads George Bush among registered voters by 34% to 31%, as stated in a recent Washington Post-ABC poll (whose interpreters in the Post disparaged this gap as ''statistically insignificant'' because the margin of sampling error was 4%). -- Numbed by Numbers

Dear Numnum: To understand political survey data, you need to know several things the pollpersons never tell you. Well, almost never. The first is that they are aspiring to a ''confidence level'' of 95%. This means that if the sample of people polled is representative of voters generally, and the candidates are separated by more than the margin of error, then it is 19 to 1 (i.e., 95 to 5) that voters will in fact go for the candidate rated highest. The 95% confidence level is entirely arbitrary and somewhat unsatisfactory, since it means that the pollpersons can do everything right and still lose to a longshot, just like all the favorite players who glumly watched Lil E. Tee win the Kentucky Derby at 17 to 1 not long ago. The pollsters could in principle offer 99% confidence levels (i.e., odds of 99 to 1). But this would typically require them either to (a) increase the sampling levels from the usual 1,300 or so to around 2,250, which is expensive, or else (b) increase their margins of error from the usual 3% to around 4%, which would result in more matchups in which the differences between candidates were within the margin of error. Now about the Post-ABC poll you have cited: It is not at all an insignificant matter that Perot led Bush by three percentage points. In this case, the sample size was smaller than usual (784 registered voters instead of the usual 1,300 or so) and the margin of error was therefore larger than usual (4% instead of 3%). But it is fallacious to say that leads within the margin of error are meaningless. The correct statement would have been that Perot's advantage could be postulated only at a confidence level lower than 95% -- in this case, it would have been about 91%. Even if the poll had shown Perot ahead by only two percentage points, he still would have been better off than Bush: The confidence level would then have been around 74%, meaning he was still a 3-to-1 favorite. Now for the main thing the pollsters never tell you: It is very, very hard to produce a sample that talks the way voters vote. The pros are pretty good at producing samples that are kosher with respect to age, race, sex, income, and registration status; what they don't know how to control for is ''response meaningfulness.'' Many people have no more ideas than a Valley Girl about the pending election, but are ashamed to admit this to pollsters. Far more serious, many folks limply broach whatever views they think the pollster will approve of. Since pollsters look to many citizens like members of the liberal media elite, there is a special tendency for conservative views to be understated. A week before Reagan's landslide election in 1980, the AP reported that ''national polls say the race between Jimmy Carter and Ronald Reagan is too close to call.'' In the wake of the Tory upset in the April elections, Britain's Market Research Society has demonstrated that support for Labour has been overstated in polls in eight of the past ten national elections, a finding we rate significant at the 99% confidence level.