NavList:
A Community Devoted to the Preservation and Practice of Celestial Navigation and Other Methods of Traditional Wayfinding
Re: Real accuracy of the method of lunar distances
From: Fred Hebard
Date: 2004 Jan 1, 15:23 -0500
From: Fred Hebard
Date: 2004 Jan 1, 15:23 -0500
Trevor, You are correct, Jan and I are discussing whether the normal distribution or the t distribution apply. The normal distribution is used when you're working with a single observation, the t when you're working with a mean of more than one. We're working here with a mean of six observations, and it makes sense that we can estimate this more precisely with six observations than one, and that the precision of the estimate depends on the number of observations. Jan was using the test statistic for the normal distribution, usually denoted z, with y equal to the value of a _single_ observation, u the actual or parametric mean and s the standard deviation is: z = (y - u) / s, whereas I was using the test statistic for the t distribution, denoted t, with Y equal to the observed mean of a set of observations, u their parametric mean, s the standard deviation and n the number of observations is: t = (Y - u) / (s / the square root of n). For the first statistic the error term is the standard deviation, for the second the standard error of the mean. I might note that implicit in the z statistic is that n=1. This is not a "paired t test." It is not even a t test for separating two means. To me, the only point in question was whether the standard deviation could be estimated from the mean of the observations, which it appears Jan and I agree can be done, and, equivalently, that the variation between a set of observations was the same size as the variation within. You can find a discussion of normal and t distributions in elementary statistics texts. As I indicated previously, the statistics are very robust to departures from assumptions and, in my opinion, it's likely that the data on which Jan is reporting conform fairly closely to the expected distribution. But I also agree with your points about outliers. Finally, perhaps I have managed to refute Jan's bleak picure, as he wished, and shown that, in the hands of a competent observer, 95% of the time a lunar will be accurate to within 25" of arc with 6 replicate observations, and 99.9% of the time accurate to within 44" of arc. Fred On Jan 1, 2004, at 7:26 AM, Trevor J. Kenchington wrote: > Fred & Jan, > > With apologies for butting into your exchange but you two seem to be > misunderstanding one another: > > Fred: What you are suggesting is, I think, a paired-comparison t-test > of > the null hypothesis that the mean of all possible positions estimated > by > lunars is not different from the mean of all possible positions > estimated by chronometer, using a sample of paired position estimates, > one each by lunar and chronometer (where, of course, each "estimate" is > actually based on multiple sextant observations on the same day). That > would be one way to see whether there was a bias in the lunar estimates > of position but it would not address Jan's interest in the average > absolute magnitude of the differences between the lunar and chronometer > estimates. > > For that, Jan already has someone else's estimate of the "probable > error", meaning the absolute difference that will be exceeded by 50% of > paired position estimates (on average). To get from that value to the > standard deviation of the errors in the lunars, he has assumed that the > chronometer positions are exact (which is likely close enough to the > truth) and has consulted the percentage points of the _Normal_ > distribution, _not_ the t-distribution which you turned to. Armed with > the standard deviation, he has again turned to the Normal distribution > to read off the error that should occur in 3 estimates in every > thousand > (i.e. 1-0.997 = 0.003 of lunar estimates). > > > The problem with that, as I have noted in earlier messages, is that the > errors in the lunars are unlikely to be exactly Normally distributed. > Otherwise, I think that Jan has followed the right path in trying to > estimate the magnitude of the errors that a skilled navigator might get > with lunars. He doesn't need hypothesis testing and hence he does not > need the t or F distributions. > > > Trevor Kenchington > > > -- > Trevor J. Kenchington PhD Gadus@iStar.ca > Gadus Associates, Office(902) 889-9250 > R.R.#1, Musquodoboit Harbour, Fax (902) 889-9251 > Nova Scotia B0J 2L0, CANADA Home (902) 889-3555 > > Science Serving the Fisheries > http://home.istar.ca/~gadus >