Anders Hald A History of Parametric Statistical Inference from Bernoulli to Fisher, 1713 to 1935

211 Pages·2005·1.65 MB·English

Anders Hald
A History of Parametric
Statistical Inference from
Bernoulli to Fisher,
1713 to 1935
department of applied mathematics and statistics
university of copenhagen
department of applied mathematics and statistics
university of copenhagen
universitetsparken 5
dk-2100 copenhagen ø
c Anders Hald
°
2004
ISBN 87-7834-628-2
Contents
Preface v
Chapter 1. The three revolutions in parametric statistical inference 1
1.1. Introduction 1
1.2. Laplace on direct probability, 1776-1799 1
1.3. The first revolution: Laplace 1774-1786 2
1.4. The second revolution: Gauss and Laplace 1809-1828 3
1.5. The third revolution: R. A. Fisher 1912-1956 5
Part 1. BINOMIAL STATISTICAL INFERENCE
The three pioneers: Bernoulli (1713), de Moivre (1733) and
Bayes (1764) 9
Chapter 2. James Bernoulli’s law of large numbers for the binomial, 1713, and
its generalization 11
2.1. Bernoulli’s law of large numbers for the binomial, 1713 11
2.2. Remarks on further developments 13
Chapter 3. De Moivre’s normal approximation to the binomial, 1733, and its
generalization 15
3.1. De Moivre’s normal approximation to the binomial, 1733 15
3.2. Lagrange’s multivariate normal approximation to the multinomial and
his confidence interval for the binomial parameter, 1776 19
3.3. De Morgan’s continuity correction, 1838 21
Chapter 4. Bayes’s posterior distribution of the binomial parameter and his
rule for inductive inference, 1764 23
4.1. The posterior distribution of the binomial parameter, 1764 23
4.2. Bayes’s rule for inductive inference, 1764 25
Part2. STATISTICALINFERENCEBYINVERSEPROBABILITY.
Inverse probability from Laplace (1774), and Gauss (1809) to
Edgeworth (1909) 27
Chapter 5. Laplace’s theory of inverse probability, 1774-1786 29
5.1. Biography of Laplace 29
5.2. The principle of inverse probability and the symmetry of direct and
inverse probability, 1774 30
5.3. Posterior consistency and asymptotic normality in the binomial case,
1774 33
i
ii
5.4. The predictive distribution, 1774-1786 35
5.5. Astatistical model andamethodofestimation. Thedoubleexponential
distribution, 1774 36
5.6. The asymptotic normality of posterior distributions, 1785 38
Chapter 6. A nonprobabilistic interlude: The fitting of equations to data,
1750-1805 43
6.1. The measurement error model 43
6.2. The method of averages by Mayer, 1750, and Laplace, 1788 44
6.3. The method of least absolute deviations by Boscovich, 1757, and
Laplace, 1799 45
6.4. The method of least squares by Legendre, 1805 47
Chapter 7. Gauss’s derivation of the normal distribution and the method of
least squares, 1809 49
7.1. Biography of Gauss 49
7.2. Gauss’s derivation of the normal distribution, 1809 50
7.3. Gauss’s first proof of the method of least squares, 1809 52
7.4. Laplace’s large-sample justification of the method of least squares, 1810 53
Chapter 8. Credibility and confidence intervals by Laplace and Gauss 55
8.1. Large-sample credibility and confidence intervals for the binomial
parameter by Laplace, 1785 and 1812 55
8.2. Laplace’s general method for constructing large-sample credibility and
confidence intervals, 1785 and 1812 55
8.3. Credibility intervals for the parameters of the linear normal model by
Gauss, 1809 and 1816 56
8.4. Gauss’s rule for transformation of estimates and its implication for the
principle of inverse probability, 1816 57
8.5. Gauss’s shortest confidence interval for the standard deviation of the
normal distribution, 1816 57
Chapter 9. The multivariate posterior distribution 59
9.1. Bienaymé’s distribution of a linear combination of the variables, 1838 59
9.2. PearsonandFilon’sderivationofthemultivariateposteriordistribution,
1898 59
Chapter 10. Edgeworth’s genuine inverse method and the equivalence of
inverse and direct probability in large samples, 1908 and 1909 61
10.1. Biography of Edgeworth 61
10.2. The derivation of the t distribution by Lüroth, 1876, and Edgeworth,
1883 61
10.3. Edgeworth’s genuine inverse method, 1908 and 1909 63
Chapter 11. Criticisms of inverse probability 65
11.1. Laplace 65
11.2. Poisson 67
11.3. Cournot 68
iii
11.4. Ellis, Boole and Venn 69
11.5. Bing and von Kries 70
11.6. Edgeworth and Fisher 71
Part 3. THE CENTRAL LIMIT THEOREM AND LINEAR
MINIMUM VARIANCE ESTIMATION BY LAPLACE AND
GAUSS 73
Chapter 12. Laplace’s central limit theorem and linear minimum variance
estimation 75
12.1. The central limit theorem, 1810 and 1812 75
12.2. Linear minimum variance estimation, 1811 and 1812 77
12.3. Asymptotic relative efficiency of estimates, 1818 79
12.4. Generalizations of the central limit theorem 81
Chapter 13. Gauss’s theory of linear minimum variance estimation 85
13.1. The general theory, 1823 85
13.2. Estimation under linear constraints, 1828 87
13.3. A review of justifications for the method of least squares 88
13.4. The state of estimation theory about 1830 90
Part 4. ERROR THEORY. SKEW DISTRIBUTIONS.
CORRELATION. SAMPLING DISTRIBUTIONS 93
Chapter 14. The development of a frequentist error theory 95
14.1. The transition from inverse to frequentist error theory 95
14.2. Hagen’s hypothesis of elementary errors and his maximum likelihood
argument, 1837 96
14.3. Frequentist error theory by Chauvenet, 1863, and Merriman, 1884 97
Chapter 15. Skew distributions and the method of moments 101
15.1. The need for skew distributions 101
15.2. Series expansions of frequency functions. The A and B series 102
15.3. Biography of Karl Pearson 107
15.4. Pearson’s four-parameter system of continuous distributions, 1895 109
15.5. Pearson’s χ2 test for goodness of fit, 1900 111
15.6. The asymptotic distribution of the moments by Sheppard, 1899 113
15.7. Kapteyn’s derivation of skew distributions, 1903 114
Chapter 16. Normal correlation and regression 117
16.1. Some early cases of normal correlation and regression 117
16.2. Galton’s empirical investigations of regression and correlation,
1869-1890 120
16.3. The mathematization of Galton’s ideas by Edgeworth, Pearson and
Yule 125
16.4. Orthogonal regression. The orthogonalization of the linear model 130
Chapter 17. Sampling distributions under normality, 1876-1908 133
iv
17.1. The distribution of the arithmetic mean 133
17.2. The distribution of the variance and the mean deviation by Helmert,
1876 133
17.3. Pizzetti’s orthonormal decomposition of the sum of squared errors in
the linear-normal model, 1892 136
17.4. Student’s t distribution by Gosset, 1908 137
Part 5. THE FISHERIAN REVOLUTION, 1912-1935 141
Chapter 18. Fisher’s early papers, 1912-1921 143
18.1. Biography of Fisher 143
18.2. Fisher’s “absolute criterion”, 1912 147
18.3. The distribution of the correlation coefficient, 1915, its transform,
1921, with remarks on later results on partial and multiple correlation148
18.4. The sufficiency of the sample variance, 1920 155
Chapter 19. The revolutionary paper, 1922 157
19.1. The parametric model and criteria of estimation, 1922 157
19.2. Properties of the maximum likelihood estimate 159
19.3. The two-stage maximum likelihood method and unbiasedness 163
Chapter 20. Studentization, the F distribution and the analysis of variance,
1922-1925 165
20.1. Studentization and applications of the t distribution 165
20.2. The F distribution 167
20.3. The analysis of variance 168
Chapter 21. The likelihood function, ancillarity and conditional inference 173
21.1. The amount of information, 1925 173
21.2. Ancillarity and conditional inference 173
21.3. The exponential family of distributions, 1934 174
21.4. The likelihood function 174
Epilogue 175
Terminology and notation 176
Books on the history of statistics 177
Books on the history of statistical ideas 177
References 178
Subject index 195
Author index 199
Preface
This is an attempt to write a history of parametric statistical inference. It may
be used as the basis for a course in this important topic. It should be easy to read
for anybody having taken an elementary course in probability and statistics.
The reader wanting more details, more proofs, more references and more in-
formation on related topics may find so in my previous two books: A History of
Probability and Statistics and Their Applications before 1750, Wiley, 1990, and A
History of Mathematical Statistics from 1750 to 1930, Wiley, 1998.
The text contains a republication of pages 488-489, 494-496, 612-618, 620-626,
633-636, 652-655, 670-685, 713-720, and 734-738 from A. Hald: A History of Math-
ematical Statistics from 1750 to 1930, Copyright c 1998 by John Wiley & Sons,
°
Inc. This material is used by permission of John Wiley & Sons, Inc.
I thank my granddaughter Nina Hald for typing the first version of the manu-
script.
September 2003 Anders Hald
I thank Professor Søren Johansen, University of Copenhagen, for a thorough
discussion of the manuscript with me.
I thank Professor Michael Sørensen, Department of Applied Mathematics and
Statistics, University of Copenhagen for including my book in the Department’s
series of publications.
December 2004 Anders Hald
v
vi
James Bernoulli (1654-1705) Abraham de Moivre (1667-1754)
Pierre Simon Laplace (1749-1827)
vii
Carl Frederich Gauss (1777-1855)
Ronald Aylmer Fisher (1890-1962)