Open Daily: 10am - 10pm | Alley door pickup 10am - 7pm 3038 Hennepin Ave Minneapolis, MN
612-822-4611

  • Open Daily: 10am - 10pm
    Alley-side Pickup: 10am - 7pm

    3038 Hennepin Ave Minneapolis, MN
    612-822-4611

Open Daily: 10am - 10pm | Alley-side Pickup: 10am - 7pm
3038 Hennepin Ave Minneapolis, MN
612-822-4611
The Probable Error of a Mean

The Probable Error of a Mean

Student (William Sealy Gosset)

Paperback

Series: Frontier Model Bibliophile Science Source Editions

Probability & Statistics

Currently unavailable to order

ISBN13: 9798259504042
Publisher: Frontier Model Bibliophile
Pages: 62
Weight: 0.21
Height: 0.13 Width: 6.00 Depth: 9.00
Language: English
In 1908 a brewer at Guinness, forbidden to publish under his own name, signed himself Student and worked out how to draw honest conclusions from a few observations. The t-distribution he introduced underlies the significance tests and confidence intervals used across science today. This source edition resets the text from a locked primary witness and reproduces the original tables and diagrams as source crops. By the closing years of the nineteenth century, the mathematical treatment of measurement error was a mature subject. The Gaussian theory of errors, developed for astronomy and geodesy, told an observer how the average of many measurements clustered around a true value, and how to attach a probable error to that average. The whole edifice rested on the assumption that the number of observations, n, was large. With hundreds or thousands of readings, the standard deviation of the sample could be treated as if it were the standard deviation of the underlying population, and the distribution of the mean was, to an excellent approximation, normal. This was a reasonable assumption for an astronomer who could photograph a star on many nights. It was useless for a brewer. William Sealy Gosset, writing as Student, confronted exactly this gap. At the Guinness brewery in Dublin he was asked to quantify the variability of barley, malt, hops, and yeast, and to decide whether one batch, one field, or one process genuinely differed from another. His experiments were small by necessity: a trial of two barley varieties might rest on a handful of plots, and a brewing comparison on a few runs. When n is small, the sample standard deviation is itself a poor and unstable estimate of the population value, and the classical machinery breaks down. Treating a small-sample standard deviation as if it were known with certainty produces probable errors that are too optimistic; the experimenter is misled into confidence the data do not warrant. Gosset's contribution was to ask the right question and answer it honestly. Rather than assume the population standard deviation was known, he derived the distribution of the quantity now written as t: the deviation of the sample mean from the true mean, measured in units of the sample standard deviation rather than the unknown population standard deviation. The resulting distribution is not normal. It is symmetric and bell-shaped, but with heavier tails, and its precise shape depends on the sample size. A small sample yields a broad distribution with substantial probability far from the center; as the sample grows, the distribution narrows and approaches the normal curve. The parameter governing this behavior is tied directly to n, and in the modern formulation is expressed through the degrees of freedom, n - 1. The heavy tails are the mathematics being honest: with little data, large apparent deviations are more probable than the normal curve would suggest, and any test of significance must allow for that. To derive the distribution, Gosset relied on a combination of analysis and intuition, supplemented by a now-famous numerical experiment. He took Macdonell's measurements of the height and the left-middle-finger length of three thousand criminals, wrote each on a card, shuffled the cards, and dealt them into small samples to study how the mean and standard deviation of those samples actually behaved. This shuffling exercise was an early and concrete instance of sampling simulation, performed by hand with playing cards because no other method was available.

Also in

Probability & Statistics

View All