Mathematical contributions to the theory of evolution. IV. On the probable errors of frequency constants and on the influence of random selection on variation and correlation

Karl Pearson, Louis Napoleon George FilonProceedings of the Royal Society of London 1898 · DOI

TL;DR

Pearson and Filon establish the foundational mathematics for determining the "probable errors" (modern-day standard errors) of statistical parameters like correlation and skewness. Their breakthrough reveals that these errors are themselves normally distributed and—most importantly—correlated, meaning that selecting for one trait in a population inevitably reshapes the entire statistical architecture of related traits.

What Problem Does This Paper Address?

In the late 19th century, the field of "biometry" was emerging, attempting to turn the study of evolution into a rigorous mathematical science. While researchers could calculate means, standard deviations, and correlations for biological traits, they lacked a formal way to measure the reliability of these numbers. If you measure the correlation between the wing length and leg length of 100 birds, how much of that result is due to true biology versus random sampling noise?

Furthermore, Pearson and Filon wanted to understand how "selection"—the engine of evolution—affects a population. If nature selects for larger individuals, does that change only the average size, or does it also fundamentally alter the population's variability and the way different organs correlate with one another?

Methods

The authors employ a mix of geometric proofs and the binomial theorem to derive a general theorem for the distribution of errors. Key methodological highlights include:

  • Error Distribution Modeling: They prove that the errors in calculated constants (like the mean or correlation coefficient) follow a normal distribution, regardless of whether the original data distribution was normal or "skew" (asymmetric).
  • Covariance of Errors: They introduce formulas for the "mean product of deviations," effectively calculating the covariance between errors of different parameters.
  • Mechanical Computation: Interestingly, the authors note that while the arithmetic for these formulas is "lengthy," it was made practical through the use of the "Brunsviga" calculator, an early mechanical pinwheel calculator. This highlights the early intersection of statistical theory and hardware acceleration.

Datasets

As a theoretical memoir, this paper does not introduce a new primary dataset. However, it references:

  • Barometric Data: The formulas were tested in a concurrent study by Pearson and Alice Lee on the distribution of barometric heights at various stations.
  • Zoological Frequency Distributions: The authors discuss the "second, or tangent curve" type of distributions common in zoology to justify the need for formulas that handle skewness.

Key Insights and Ideas

  • The Interconnectedness of Statistics: The "aha moment" here is that statistical errors are not independent. For example, the error in calculating a standard deviation is mathematically linked to the error in calculating the correlation coefficient ($r_{12}/\sqrt{2}$).
  • Selection is Holistic: The paper argues that selection is never isolated. Selecting for a specific size (mean) will, in all probability, alter the variability (standard deviation) of the population. Specifically, they note that "increased size means decreased variability."
  • Stability of Skewness: One surprising finding is that a random selection that shifts the mean value does not necessarily alter the skewness of the distribution (the correlation of errors between mean and skewness is zero).
  • Universality of the Normal Law: They demonstrate that the "Normal Law of Frequency" is a powerful meta-tool; it describes the behavior of the errors of constants even when the constants themselves describe non-normal data.

Key Results

The paper provides a "cheat sheet" of formulas that became standard in 20th-century statistics. If $\lambda = 0.67449$ and $n$ is the sample size:

  • Probable error of a correlation coefficient ($r_{12}$): $\lambda \frac{1-r_{12}^2}{\sqrt{n}}$.
  • Correlation of errors in two standard deviations: Exactly equal to the correlation between the traits themselves ($r_{12}^2$).
  • Regression Coefficients: Formulas are provided for both two-organ and three-organ systems, allowing researchers to predict how selection on one trait propagates through a system of three correlated organs.
  • Modal Frequency: The probable error of the modal frequency ($y_0$) is shown to be consistently less than the probable error of the mean, suggesting the mode is a more "stable" measure in certain skew distributions.

Limitations and Future Work

  • Computational Complexity: The authors admit the results for complex skew frequency curves are "too long to be cited" in the abstract and involve grueling arithmetic.
  • Random vs. Directed Selection: This paper focuses primarily on the effects of random selection. The authors explicitly state that the more complex case of "directed selection" (where specific traits are targeted by environmental pressures) is reserved for a separate, forthcoming memoir.
  • Approximation: The authors note that the system follows the normal law approximately, leaving room for further refinement in small-sample contexts (a problem later tackled by "Student's" t-distribution).

Summary updated

Abstract

Abstract This memoir starts with a general theorem, by which the probable errors made in calculating the constants of any frequency distribution may be determined. It is shown that these probable errors form a correlated system approximately following the normal law of frequency, whatever be the nature of the original frequency distribution, i. e.,whether it be skew or normal. The importance of this result for the theory of evolution is then drawn attention to.
BibTeX
@inproceedings{pearson1898mathematical,
  author = {Karl Pearson and Louis Napoleon George Filon},
  title = {{Mathematical contributions to the theory of evolution. IV. On the probable errors of frequency constants and on the influence of random selection on variation and correlation}},
  booktitle = {Proceedings of the Royal Society of London},
  year = {1898},
  doi = {10.1098/rspl.1897.0091},
}
Figure