Bayesianism and IBE: The Case of Individual vs. Group Selection | Leah Henderson, in "Best Explanations" ed. Kevin McCain and Ted Poston, 248-261

Given are "two paradigmatic ways of talking about scientific inference" (248).

  1. Bayesian: scientific inference means determining the most probable theory in light of the evidence
  2. IBE: many actual cases of "theory evaluation in the history of science" can be accurately described as cases of inference to the best explanation.

On the matter of whether these two are compatible, Henderson refers to her own work on this topic topic which poses an alternative to 'explanationist Bayesianism', see Henderson-Bayesianism_and_Inference_to_the_Best_Explanation.md .

Summary of Henderson's position called 'emergent compatibilism' (as opposed to 'constraint-based compatibilism'):

The idea here is that, when the structure of scientific theories is taken properly into account, a Bayesian model of the preference for better explanations can be provided, making only assumptions which are quite natural and independent of any explanatory concerns. The key is to recognize that scientific theory evaluation occurs at multiple levels of generality or abstraction. A general level or framework theory is often regarded as providing a better explanation to the extent that it explains the phenomena without too much reliance of fine-tuning of auxiliary hypotheses. (249)

This article applies Henderon's account to a case study in evolutionary biology. George C. Williams argues against a group selection hypotheses for explaining altruistic behavior, instead arguing an explanation on the individual level was more parsimonious (249).

Henderson's approach is pitted against an existing "Bayesian explanationist" account of Elliot Sober, which sees this as a reason "to assign a higher prior to the individual selection hypothesis" (249). (Typical example where explanatory considerations guide/constrain the choice of priors).

Henderson's counter:

  • The claim to parsimony is already taken care of naturally in the Bayesian likelihood! (See Henderson 2014, where I elaborated on this).
  • "There is no need then for explanatory considerations to govern the assignment of priors." (249)

Henderson makes an additional claim about Sober's "'local' view on the epistemic significance of parsimony" (249), but it is at this point not clear to me what this means.

IBE and Bayesianism

How are the explanatory and confirmatory aspects of scientific theories related?

  1. We think that comparison of scientific theories should be done based on how much a theory is confirmed in relation to some evidence, in comparison with competing theories (cf. 249)
    • This is where the Bayesian approach fits in
  2. Scientist often appeal to "explanatory advantages" (250) of a theory
    • E.g. Darwin: natural selection better explanation of 'large classes of facts' compared to special creationism.
    • Copernicus: his theory more harmonic than Ptolemaic theory
    • IBE

Motivation: Why is it desirable that they are comparable? They both seems to quite accurately describe and apply to historical cases of scientific reasoning, which suggests they are compatible.

Criticism on constraining priors

Ted Poston, for example, argues that if more explanatory hypotheses are not given higher priors, inductive scepticism follows (Poston 2014). Others have advocated that simpler hypotheses should be given higher priors (e.g. Jeffreys 1998 [1939]), though there has also been serious opposition to this view (e.g. Kelly 2007). The basic problem, critics urge, is that we have no reason to think the world is more likely to be simple than complex, so no grounds for making that assumption in the prior. (250)

The 'local' constraint variant of Williams

Another approach is to adopt a more local strategy. The idea is that perhaps even without a general connection between parsimony and priors, in certain specific cases, we can justify giving higher priors to more parsimonious hypotheses. (250)

Capture general and specific level with Hierarchical Bayes

My suggestion is that a large part of what makes for better explanation is a feature of the relationship between the specific and framework levels of theory evaluation, in particular how much tuning of specific parameters or auxiliaries is required in order for the higher level framework to produce a good empirical fit to the data (Henderson 2014).

Henderson's suggestion is that such a distinction can be accounted for using a hierarchical Bayesian model.

When we build a Bayesian model, it is helpful to capture this distinction between frameworks and specific (or fitted) theories. Rather than the standard approach of placing all candidate-specific theories in one big hypothesis space, we can use a hierarchical Bayesian model, in which the hypothesis space is structured into levels (Henderson et al. 2010). Frameworks can compete with one another at a high level of the model, whereas the fitted theories occupy a lower level. Bayesian updating then is carried out at each level of the model. (251)

The reasoning why the likelihood punishes theories making less assumptions is the same as in Henderson 2014, but there I actually did not notice that clearly that she used hierarchical Bayes.

See 251 for details.

"As long as the prior probability is spread roughly evenly over the hypotheses of the model, the average likelihood penalizes less simple, or more fine-tuned explanations." (251-252).

Bayesian model selection favors simpler hypotheses:

That is because in such cases we have to average over a relatively large number of ill-fitting specific hypotheses which have very low likelihoods. The principles at work here are those that underpin the Bayesian approach to ‘model selection’, which is known to favour simpler hypotheses (MacKay 2003). (252)

Cf. Bayesian Information Criterion (BIC)

Henderson's "stokpaardje":

There are some assumptions about how the prior probabilities are spread over the specific hypotheses in a framework, but these need not be motivated by anything to do with explanatoriness. Rather they appear to be quite ‘natural’ assumptions that one would make on independent grounds.

Case study: William's against group selection

Problem statement: how can we explain altruistic behavior that favors the fitness of other individuals and decreases one's own, from a evolutionary account on the individual level. (253)

One potential explanation: there is also selection between groups.

Williams shows how altruistic behavior can also be explained on the individual level (cf. 253) and also holds that we should prefer the more parsimonious theory, i.e. only consider group level selection when the individual level fails.

Henderson quotes Sober:

Williams suggests that the hypothesis of group selection, if true, would explain the observations, and that the same is true for the hypothesis of individual selection that he invents. This means ... that the two hypotheses have identical likelihoods. If so, the hypotheses will differ in overall plausibility only if they have different priors. (Sober 1990, p. 144)

Henderson does not comment on this, but does the fact that both hypotheses explain the observations imply they explain it to the exact same degree (i.e. that they have identical likelihoods).

Alternative with hierarchical Bayes

When we ask whether the individual selection theory or the group selection theory provides the better explanation of biological phenomena, we are asking which of the two general frameworks IS or GS provides the better explanation. What Williams takes to be less explanatory about GS is that it needs to be combined with various very specific auxiliary hypotheses in order to account for the phenomena. If it is true, as Williams alleges, that the group selection framework needs to be highly fine-tuned, this will be reflected in a lower Bayesian likelihood for that theory, since the theory will be automatically penalized for fine-tuning by the averaging process. The Bayesian then favours the more explanatory theory without assigning any higher prior probability to the individual selection theory. Indeed it is quite possible for the priors for the competing frameworks to be equal: p(IS) = p(GS). (255)

This penalty on the likelihood depends on the distinction between the general theory/framework and specific fitted theories (where parameters are filled in).

The regular Bayesian approach instead assumes a fully fitted theory, i.e. it's always a specific theory with all parameters specified.

In this case: "Sober’s model works with fitted hypotheses and assigns a higher prior to the individual selection hypotheses." (256)

Note that you can always compute the posterior for a specific fitted hypothesis by marginalizing out the general frameworks (256).

Main advantage: when constraining the prior on specific fitted theories, how does one successfully 'justify why the prior is higher for more explanatory specific hypotheses' (256)?

For Henderson's account, this follows more naturally, from two components:

  1. How supported is the framework given the data?
  2. How well is a specific hypothesis supported by the evidence within a specific framework?

Less arbitrary / opaque than constraining priors:

The first component, as we saw above, naturally reflects the explanationist concern with fine-tuning since it involves an averaged likelihood. By contrast, in Bayesian reconstructions which work with fitted models, explanatory considerations can be grafted onto the Bayesian machine by having them constrain the priors, but it remains rather opaque why they, rather than some other considerations, should be what determines the priors. (256)

What assumptions on priors do we still make?

You can still make assumptions on the priors, for example apply the Principle of Indifference. (So emergent compatibilism is consistent with objective Bayesianism!)

Hendersons stresses multiple times that for equal priors, hierarchical Bayes can perfectly fine account for an explanatory difference between theories due to the amount of fine-tuning required for fitting a general theory.

But the main point is: compatibilism of IBE and Bayesianism does no longer depend on the (possibly arbitrary) choice of priors!

The important point is that, although the results do depend on assignments of priors to the specific hypotheses, there are generally reasons for these assignments, which can be motivated independently of the desire to ensure compatibility with IBE. (256)

How general or local should we apply IBE?

Aside: more motivation for compatibilism:

Part of the purpose of giving a Bayesian model of IBE is to elucidate the epistemic significance of explanatory considerations by connecting them to an established account of evidential support. (256-257)

Explanations view is relatively local:

The explanationist Bayesian model is compatible with, and indeed encourages the view that the epistemic significance of explanatory considerations such as parsimony is highly case-specific. (257)

See p. 257: Sober in fact argues it's a mistake to pose global characteristics for constraining priors, because reasons for doing so differ wildly from case to case. It may also be more appropriate to let parsimony be reflected in the likelihood than the prior, depending on the specific context.

Henderson, op het scherpst van de snede:

However, Williams’ argument only appears to be a very case-specific use of parsimony if we fail to identify and model the key structural feature that it has in common with other cases of IBE. (257)

Of course, precisely the idea that more fine-tuning indicates a worse fit of the general theory.

Henderson gives various examples on p. 257-258.

She also mentions that scientists themselves often see it as an explanatory strength if observations are explained from basic principles:

The concern to avoid fine-tuning also explains why scientists often claim an explanatory advantage for their theory based on its ability to explain from its basic principles, rather than requiring a number of arrangements of auxiliary hypotheses. (258)

E.g. Lavoisier (quote on p. 258)

The main argument is that IBE is not entirely local, even though there are also local assumptions involved; there is a common characteristic in IBE that can be accounted for by emergent Bayesianism:

it is not that there are no ‘subject-matter-specific’ assumptions or judgements at work. In particular, the identification of the framework and its parametrization are important background assumptions in any given case. However, there is a structural commonality to the explanatory considerations involved, in terms of the relationship between different levels of theory evaluation, which makes possible a more unified connection to Bayesianism than Sober has suggested. (259)