Monthly Archives: May 2020

Birthday of Allan Birnbaum: Foundations of Probability and Statistics (27 May 1923 – 1 July 1976)

27 May 1923-1 July 1976

27 May 1923-1 July 1976

Today is Allan Birnbaum’s birthday. In honor of his birthday, I’m posting the articles in the Synthese volume that was dedicated to his memory in 1977. The editors describe it as their way of  “paying homage to Professor Birnbaum’s penetrating and stimulating work on the foundations of statistics”. I had posted the volume before, but there are several articles that are very worth rereading. I paste a few snippets from the articles by Giere and Birnbaum. If you’re interested in statistical foundations, and are unfamiliar with Birnbaum, here’s a chance to catch up. (Even if you are, you may be unaware of some of these key papers.)


Synthese Volume 36, No. 1 Sept 1977: Foundations of Probability and Statistics, Part I

Editorial Introduction:

This special issue of Synthese on the foundations of probability and statistics is dedicated to the memory of Professor Allan Birnbaum. Professor Birnbaum’s essay ‘The Neyman-Pearson Theory as Decision Theory; and as Inference Theory; with a Criticism of the Lindley-Savage Argument for Bayesian Theory’ was received by the editors of Synthese in October, 1975, and a decision was made to publish a special symposium consisting of this paper together with several invited comments and related papers. The sad news about Professor Birnbaum’s death reached us in the summer of 1976, but the editorial project could nevertheless be completed according to the original plan. By publishing this special issue we wish to pay homage to Professor Birnbaum’s penetrating and stimulating work on the foundations of statistics. We are grateful to Professor Ronald Giere who wrote an introductory essay on Professor Birnbaum’s concept of statistical evidence and who compiled a list of Professor Birnbaum’s publications.


Table of Contents

SUFFICIENCY, CONDITIONALLY AND LIKELIHOOD In December of 1961 Birnbaum presented the paper ‘On the Foundations, of Statistical Inference’ (Birnbaum [19]) at a special discussion meeting of the American Statistical Association. Among the discussants was L. J. Savage who pronounced it “a landmark in statistics”. Explicitly denying any “intent to speak with exaggeration or rhetorically”, Savage described the occasion as “momentous in the history of statistics”. “It would be hard”, he said, “to point to even a handful of comparable events” (Birnbaum [19], pp. 307-8). The reasons for Savage’s enthusiasm are obvious. Birnbaum claimed to have shown that two principles widely held by non-Bayesian statisticians (sufficiency and conditionality) jointly imply an important consequence of Bayesian statistics (likelihood).”[1]
INTRODUCTION AND SUMMARY ….Two contrasting interpretations of the decision concept are formulated: behavioral, applicable to ‘decisions’ in a concrete literal sense as in acceptance sampling; and evidential, applicable to ‘decisions’ such as ‘reject H in a research context, where the pattern and strength of statistical evidence concerning statistical hypotheses is of central interest. Typical standard practice is characterized as based on the confidence concept of statistical evidence, which is defined in terms of evidential interpretations of the ‘decisions’ of decision theory. These concepts are illustrated by simple formal examples with interpretations in genetic research, and are traced in the writings of Neyman, Pearson, and other writers. The Lindley-Savage argument for Bayesian theory is shown to have no direct cogency as a criticism of typical standard practice, since it is based on a behavioral, not an evidential, interpretation of decisions.

[1]By “likelihood” here, Giere means the (strong) Likelihood Principle (SLP). Dotted through the first 3 years of this blog are a number of (formal and informal) posts on his SLP result, and my argument as to why it is unsound. I wrote a paper on this that appeared in Statistical Science 2014. You can find it along with a number of comments and my rejoinder in this post: Statistical Science: The Likelihood Principle Issue is Out.The consequences of having found his proof unsound gives a new lease on life to statistical foundations, or so I argue in my rejoinder.

Categories: Birnbaum, Likelihood Principle, Statistics, strong likelihood principle | Tags: | 3 Comments

Graduate Research Seminar: Current Controversies in Phil Stat: LSE PH 500: 21 May – 18 June 2020


Ship StatInfasST will embark on a new journey from 21 May – 18 June, a graduate research seminar for the Philosophy, Logic & Scientific Method Department at the LSE, but given the pandemic has shut down cruise ships, it will remain at dock in the U.S. and use zoom. If you care to follow any of the 5 sessions, nearly all of the materials will be linked here collected from excerpts already on this blog. If you are interested in observing on zoom beginning 28 May, please follow the directions here

For the updated schedule, see the seminar web page.

Topic: Current Controversies in Phil Stat
(LSE, Remote 10am-12 EST, 15:00 – 17:00 London time; Thursdays 21 May-18 June)

Main Text SIST: Statistical Inference as Severe Testing: How to Get Beyond the Statistics Wars CUP, 2018):

I. (May 21)  Introduction: Controversies in Phil Stat:  

SIST: Preface, Excursion 1
Excursion 1 Tour I
Excursion 1 Tour II

Notes/Outline of Excursion 1
Postcard: Souvenir A

II. (May 28) N-P and Fisherian Tests, Severe Testing:

SIST: Excursion 3 Tour I (focus on pages up to p. 152)

Recommended: Excursion 2 Tour II pp. 92-100

Optional: I will (try to) answer questions on demarcation of science, induction, falsification, Popper from Excursion 2 Tour II

Handout: Areas Under the Standard Normal Curve

III. (June 4) Deeper Concepts: Confidence Intervals and Tests: Higgs’ Discovery:

SIST: Excursion 3 Tour III

Optional: I will answer questions on Excursion 3 Tour II: Howlers and Chestnuts of Tests 

IV. (June 11) Rejection Fallacies: Do P-values exaggerate evidence?
      Jeffreys-Lindley paradox or Bayes/Fisher disagreement:

SIST: Excursion 4 Tour II

           SIST: Excursion 4 Tour II

          Recommended (if time)Excursion 4 Tour I: The Myth of “The Myth of Objectivity” 

V. (June 18) The Statistics Wars and Their Casualties:

SIST: Excursion 4 Tour III: pp. 267-286; Farewell Keepsakepp. 436-444
-Amrhein, V., Greenland, S., & McShane, B., (2019). Comment: Retire Statistical Significance, Nature, 567: 305-308.
-Ioannidis J. (2019). “The Importance of Predefined Rules and Prespecified Statistical Analyses: Do Not Abandon Significance.” JAMA. 321(21): 2067–2068. doi:10.1001/jama.2019.4582
-Ioannidis, J. (2019). Correspondence: Retiring statistical significance would give bias a free pass. Nature, 567, 461.
-Mayo, DG. (2019), P‐value thresholds: Forfeit at your peril. Eur J Clin Invest, 49: e13170. doi: 10.1111/eci.13170


Information Items for SIST

-References: Captain’s Bibliography
-Summaries of 16 Tours (abstracts & keywords)
Excerpts & Mementos on Error Statistics Philosophy Blog (I will link to items from excerpted proofs for interested blog followers as we proceed)
Schaum’s Appendix 2Areas Under the Standard Normal Curve from 0-Z

DELAYED: JUNE 19-20 Workshop: The Statistics Wars and Their Casualties

Categories: Announcement, SIST | Leave a comment

Final part of B. Haig’s ‘What can psych stat reformers learn from the error-stat perspective?’ (Bayesian stats)


Here’s the final part of Brian Haig’s recent paper ‘What can psychology’s statistics reformers learn from the error-statistical perspective?’ in Methods in Psychology 2 (Nov. 2020). The full article, which is open access, is here. I will make some remarks in the comments.

5. The error-statistical perspective and the nature of science


As noted at the outset, the error-statistical perspective has made significant contributions to our philosophical understanding of the nature of science. These are achieved, in good part, by employing insights about the nature and place of statistical inference in experimental science. The achievements include deliberations on important philosophical topics, such as the demarcation of science from non-science, the underdetermination of theories by evidence, the nature of scientific progress, and the perplexities of inductive inference. In this article, I restrict my attention to two such topics: The process of falsification and the structure of modeling.

5.1. Falsificationism

The best known account of scientific method is the so-called hypothetico-deductive method. According to its most popular description, the scientist takes an existing hypothesis or theory and tests indirectly by deriving one or more observational predictions that are subjected to direct empirical test. Successful predictions are taken to provide inductive confirmation of the theory; failed predictions are said to provide disconfirming evidence for the theory. In psychology, NHST is often embedded within such a hypothetico-deductive structure and contributes to weak tests of theories.

Also well known is Karl Popper’s falsificationist construal of the hypothetico-deductive method, which is understood as a general strategy of conjecture and refutation. Although it has been roundly criticised by philosophers of science, it is frequently cited with approval by scientists, including psychologists, even though they do not, indeed could not, employ it in testing their theories. The major reason for this is that Popper does not provide them with sufficient methodological resources to do so.

One of the most important features of the error-statistical philosophy is its presentation of a falsificationist view of scientific inquiry, with error statistics serving an indispensable role in testing. From a sympathetic, but critical, reading of Popper, Mayo endorses his strategy of developing scientific knowledge by identifying and correcting errors through strong tests of scientific claims. Making good on Popper’s lack of knowledge of statistics, Mayo shows how one can properly employ a range of, often familiar, error-statistical methods to implement her all-important severity requirement. Stated minimally, and informally, this requirement says, “A claim is severely tested to the extent that it has been subjected to and passes a test that probably would have found flaws, were they present.” (Mayo, 2018, p. xii) Further, in marked contrast with Popper, who deemed deductive inference to be the only legitimate form of inference, Mayo’s conception of falsification stresses the importance of inductive, or content-increasing, inference in science. We have here, then, a viable account of falsification, which goes well beyond Popper’s account with its lack of operational detail about how to construct strong tests. It is worth noting that the error-statistical stance offers a constructive interpretation of Fisher’s oft-cited remark that the null hypothesis is never proved, only possibly disproved.

5.2. A hierarchy of models

In the past, philosophers of science tended to characterize scientific inquiry by focusing on the general relationship between evidence and theory. Similarly, scientists, even today, commonly speak in general terms of the relationship between data and theory. However, due in good part to the labors of experimentally-oriented philosophers of science, we now know that this coarse-grained depiction is a poor portrayal of science. The error-statistical perspective is one such philosophy that offers a more fine-grained parsing of the scientific process.

Building on Patrick Suppes’ (1962) important insight that science employs a hierarchy of models that ranges from experimental experience to theory, Mayo’s (1996) error-statistical philosophy initially adopted a framework in which three different types of models are interconnected and serve to structure error-statistical inquiry: Primary models, experimental models, and data models. Primary models, which are at the top of the hierarchy, break down a research problem, or question, into a set of local hypotheses that can be investigated using reliable methods. Experimental models take the mid-positon on the hierarchy and structure the particular models at hand. They serve to link primary models to data models. And, data models, which are at the bottom of the hierarchy, generate and model raw data, put them in canonical form, and check whether the data satisfy the assumptions of the experimental models. It should be mentioned that the error-statistical approach has been extended to primary models and theories of a more global nature (Mayo and Spanos, 2010) and, now, also includes a consideration of experimental design and the analysis and generation of data (Mayo, 2018).

This hierarchy of models facilitates the achievement of a number of goals that are important to the error-statistician. These include piecemeal strong testing of local hypotheses rather than broad theories, and employing the model hierarchy as a structuring device to knowingly move back and forth between statistical and scientific hypotheses. The error-statistical perspective insists on maintaining a clear distinction between statistical and scientific hypotheses, pointing out that psychologists often mistakenly take tests of significance to have direct implications for substantive hypotheses and theories.

6. The philosophy of statistics

A heartening attitude that comes through in the error-statistical corpus is the firm belief that the philosophy of statistics is an important part of statistical thinking. This emphasis on the conceptual foundations of the subject contrasts markedly with much of statistical theory, and most of statistical practice. It is encouraging, therefore, that Mayo’s philosophical work has influenced a number of prominent statisticians, who have contributed to the foundations of their discipline. Gelman’s error-statistical philosophy canvassed earlier is a prominent case in point. Through both precept and practice, Mayo’s work makes clear that philosophy can have a direct impact on statistical practice. Given that statisticians operate with an implicit philosophy, whether they know it or not, it is better that they avail themselves of an explicitly thought-out philosophy that serves their thinking and practice in useful ways. More particularly, statistical reformers recommend methods and strategies that have underlying philosophical commitments. It is important that they are identified, described, and evaluated.

The tools used by the philosopher of statistics in order to improve our understanding and use of statistical methods are considerable (Mayo, 2011). They include clarifying disputed concepts, evaluating arguments employed in statistical debates, including the core commitments of rival schools of thought, and probing the deep structure of statistical methods themselves. In doing this work, the philosopher of statistics, as philosopher, ascends to a meta-level to get purchase on their objects of study. This second-order inquiry is a proper part of scientific methodology.

It is important to appreciate that the error-statistical outlook is a scientific methodology in the proper sense of the term. Briefly stated, methodology is the interdisciplinary field that draws from disciplines that include statistics, philosophy of science, history of science, as well as indigenous contributions from the various substantive disciplines. As such, it is the key to a proper understanding of statistical and scientific methods. Mayo’s focus on the role of error statistics in science is deeply informed about the philosophy, history, and theory of statistics, as well as statistical practice. It is for this reason that the error-statistical perspective is strategically positioned to help the reader to go beyond the statistics wars.

7. Conclusion

The error-statistical outlook provides researchers, methodologists, and statisticians with a distinctive and illuminating perspective on statistical inference. Its Popper-inspired emphasis on strong tests is a welcome antidote to the widespread practice of weak statistical hypothesis testing that still pervades psychological research. More generally, the error-statistical standpoint affords psychologists an informative perspective on the nature of good statistical practice in science that will help them understand and transcend the statistics wars into which they have been drawn. Importantly, psychologists should know about the error-statistical perspective as a genuine alternative to the new statistics and Bayesian statistics. The new statisticians, Bayesians statisticians, and those with other preferences should address the challenges to their outlooks on statistics that the error-statistical viewpoint provides. Taking these challenges seriously would enrich psychology’s methodological landscape.

*This article is based on an invited commentary on Deborah Mayo’s book, Statistical inference as severe testing: How to get beyond the statistics wars (Cambridge University Press, 2018), which appeared at It is adapted with permission. I thank Mayo for helpful feedback on an earlier draft.

Refer to the paper for the references. I invite your comments and questions.


Categories: Brian Haig, SIST | 3 Comments

Part 2 of B. Haig’s ‘What can psych stat reformers learn from the error-stat perspective?’ (Bayesian stats)


Here’s a picture of ripping open the first box of (rush) copies of Statistical Inference as Severe Testing: How to Get Beyond the Statistics Wars*, and here’s a continuation of Brian Haig’s recent paper ‘What can psychology’s statistics reformers learn from the error-statistical perspective?’ in Methods in Psychology 2 (Nov. 2020). Haig contrasts error statistics, the “new statistics”, and Bayesian statistics from the perspective of the statistics wars in psychology. The full article, which is open access, is here. I will make several points in the comments.


4. Bayesian statistics

Despite its early presence, and prominence, in the history of statistics, the Bayesian outlook has taken an age to assert itself in psychology. However, a cadre of methodologists has recently advocated the use of Bayesian statistical methods as a superior alternative to the messy frequentist practice that dominates psychology’s research landscape (e.g., Dienes, 2011; Kruschke and Liddell, 2018; Wagenmakers, 2007). These Bayesians criticize NHST, often advocate the use of Bayes factors for hypothesis testing, and rehearse a number of other well-known Bayesian objections to frequentist statistical practice.

Of course, there are challenges for Bayesians from the error-statistical perspective, just as there are for the new statisticians. For example, the frequently made claim that p values exaggerate the evidence against the null hypothesis, but Bayes factors do not, is shown by Mayo not to be the case. She also makes the important point that Bayes factors, as they are currently used, do not have the ability to probe errors and, thus, violate the requirement for severe tests. Bayesians, therefore need to rethink whether Bayes factors can be deployed in some way to provide strong tests of hypotheses through error control. As with the new statisticians, Bayesians also need to reckon with the coherent hybrid NHST afforded by the error-statistical perspective, and argue against it, rather than the common inchoate hybrids, if they want to justify abandoning NHST. Finally, I note in passing that Bayesians should consider, among other challenges, Mayo’s critique of the controversial Likelihood Principle, a principle which ignores the post-data consideration of sampling plans.

4.1. Contrasts between the Bayesian and error-statistical perspectives

One of the major achievements of the philosophy of error-statistics is that it provides a comprehensive critical evaluation of the major variants of Bayesian statistical thinking, including the classical subjectivist, “default”, pragmatist, and eclectic options within the Bayesian corpus. Whether the adoption of Bayesian methods in psychology will overcome the disorders of current frequentist practice remains to be seen. What is clear from reading the error-statistical literature, however, is that the foundational options for Bayesians are numerous, convoluted, and potentially bewildering. It would be a worthwhile exercise to chart how these foundational options are distributed across the prominent Bayesian statisticians in psychology. For example, the increasing use of Bayes factors for hypothesis testing purposes is accompanied by disorderliness at the foundational level, just as it is in the Bayesian literature more generally. Alongside the fact that some Bayesians are sceptical of the worth of Bayes factors, we find disagreement about the comparative merits of the subjectivist and default Bayesianism outlooks on Bayes factors in psychology (Wagenmakers et al., 2018).

The philosophy of error-statistics contains many challenges for Bayesians to consider. Here, I want to draw attention to three basic features of Bayesian thinking, which are rejected by the error-statistical approach. First, the error-statistical approach rejects the Bayesian insistence on characterizing the evidential relation between hypothesis and evidence in a universal and logical manner in terms of Bayes’ theorem. Instead, it formulates the relation in terms of the substantive and specific nature of the hypothesis and the evidence with regards to their origin, modeling, and analysis. This is a consequence of a strong commitment to a piecemeal, contextual approach to testing, using the most appropriate frequentist methods available for the task at hand. This contextual attitude to testing is taken up in Section 5.2, where one finds a discussion of the role different models play in structuring and decomposing inquiry.

Second, the error-statistical philosophy also rejects the classical Bayesian commitment to the subjective nature of prior probabilities, which the agent is free to choose, in favour of the more objective process of establishing error probabilities understood in frequentist terms. It also finds unsatisfactory the turn to the more popular objective, or “default”, Bayesian option, in which the agent’s appropriate degrees of belief are constrained by relevant empirical evidence. The error-statistician rejects this default option because it fails in its attempts to unify Bayesian and frequentist ways of determining probabilities.

And, third, the error-statistical outlook employs probabilities to measure how effectively methods facilitate the detection of error, and how those methods enable us to choose between alternative hypotheses. By contrast, orthodox Bayesians use probabilities to measure belief in hypotheses or degrees of confirmation. As noted earlier, most Bayesians are not concerned with error probabilities at all. It is for this reason that error-statisticians will say about Bayesian methods that, without supplementation with error probabilities, they are not capable of providing stringent tests of hypotheses.

4.2. The Bayesian remove from scientific practice

Two additional features of the Bayesian focus on beliefs, which have been noted by philosophers of science and statistics, draw attention to their outlook on science. First, Kevin Kelly and Clark Glymour worry that “Bayesian methods assign numbers to answers instead of producing answers outright.” (2004, p. 112) Their concern is that the focus on the scientist’s beliefs “screens off” the scientist’s direct engagement with the empirical and theoretical activities that are involved in the phenomenology of science. Mayo agrees that we should focus on the scientific phenomena of interest, not the associated epiphenomena of degrees of belief. This preference stems directly from the error-statistician’s conviction that probabilities properly quantify the performance of methods, not the scientist’s degrees of belief.

Second, Henry Kyburg is puzzled by the Bayesian’s desire to “replace the fabric of science… with a vastly more complicated representation in which each statement of science is accompanied by its probability, for each of us.” (1992, p.149) Kyburg’s puzzlement prompts the question, ‘Why should we be interested in each other’s probabilities?’ This is a question raised by David Cox about prior probabilities, and noted by Mayo (2018).

This Bayesian remove from science contrasts with the willingness of the error-statistical perspective to engage more directly with science. Mayo is a philosopher of science as well as statistics, and has a keen eye for scientific practice. Given that contemporary philosophers of science tend to take scientific practice seriously, it comes as no surprise that she brings it to the fore when dealing with statistical concepts and issues. Indeed, her error-statistical philosophy should be seen as a significant contribution to the so-called new experimentalism, with its strong focus, not just on experimental practice in science, but also on the role of statistics in such practice. Her discussion of the place of frequentist statistics in the discovery of the Higgs boson in particle physics is an instructive case in point.

Taken together, these just-mentioned points of difference between the Bayesian and error-statistical philosophies constitute a major challenge to Bayesian thinking that methodologists, statisticians, and researchers in psychology need to confront.

4.3. Bayesian statistics with error-statistical foundations

One important modern variant of Bayesian thinking, which now receives attention within the error-statistical framework, is the falsificationist Bayesianism of Andrew Gelman, which received its major formulation in Gelman and Shalizi (2013). Interestingly, Gelman regards his Bayesian philosophy as essentially error-statistical in nature – an intriguing claim, given the anti-Bayesian preferences of both Mayo and Gelman’s co-author, Cosma Shalizi. Gelman’s philosophy of Bayesian statistics is also significantly influenced by Popper’s view that scientific propositions are to be submitted to repeated criticism in the form of strong empirical tests. For Gelman, best Bayesian statistical practice involves formulating models using Bayesian statistical methods, and then checking them through hypothetico-deductive attempts to falsify and modify those models.

Both the error-statistical and neo-Popperian Bayesian philosophies of statistics extend and modify Popper’s conception of the hypotheticodeductive method, while at the same time offering alternatives to received views of statistical inference. The error-statistical philosophy injects into the hypothetico-deductive method an account of statistical induction that employs a panoply of frequentist statistical methods to detect and control for errors. For its part, Gelman’s Bayesian alternative involves formulating models using Bayesian statistical methods, and then checking them through attempts to falsify and modify those models. This clearly differs from the received philosophy of Bayesian statistical modeling, which is regarded as a formal inductive process.

From the wide-ranging error-statistical evaluation of the major varieties of Bayesian statistical thought on offer, Mayo concludes that Bayesian statistics needs new foundations: In short, those provided by her error-statistical perspective. Gelman acknowledges that his falsificationist Bayesian philosophy is underdeveloped, so it will be interesting to learn how its further development relates to Mayo’s error-statistical perspective. It will also be interesting to see if Bayesian thinkers in psychology engage with Gelman’s brand of Bayesian thinking. Despite the appearance of his work in a prominent psychology journal, they have yet to do so. However, Borsboom and Haig (2013) and Haig (2018) provide sympathetic critical evaluations of Gelman’s philosophy of statistics.

It is notable that in her treatment of Gelman’s philosophy, Mayo emphasizes that she is willing to allow a decoupling of statistical outlooks and their traditional philosophical foundations in favour of different foundations, which are judged more appropriate. It is an important achievement of Mayo’s work that she has been able to consider the current statistics wars without taking a particular side in the debates. She achieves this by examining methods, both Bayesian and frequentist, in terms of whether they violate her minimal severity requirement of “bad evidence, no test”.

I invite your comments and questions.

*This picture was taken by Diana Gillooly, Senior Editor for Mathematical Sciences, Cambridge University Press, at the book display for the Sept. 2018 meeting of the Royal Statistical Society in Cardiff. She also had the honor of doing the ripping. A blogpost on the session I was in is here.

Categories: Brian Haig, SIST | 6 Comments

Blog at