Bayesianism and Inference to the Best Explanation | Leah Henderson, British Journal of Philosophy of Science, 65 (2014), 687-715.

@article

IBE and Bayesianism are very influential theories of scientific inference. (The emphasis on "scientific" makes me wonder to what extent both are also theories of everyday abductive reasoning).

Henderson argues against Van Fraassen's position that IBE and Bayesianism are incompatible, because "any probabilistic version of IBE would violate Bayesian conditionalization" (687).

She however argues that the standard compatibilist account is also not satisfactory.

Henderon's position can be called emergent compatibilism, as opposed to the standard compatibilist position that aims to make Bayesianism compatible with IBE by putting (explanatory) constraints on priors and likelihoods, which Henderson calls constraint-based compatibilism.

She instead aims to show how Bayesianism already facilitates explanatory considerations strictly on its own terms, rather than by "auxiliary hypothesis" and "without explicit guidance by IBE" (687).

Main claim:

The aim is to show that IBE and Bayesianism may be compatible, not because they can be amalgamated, but rather because they capture substantially similar epistemic considerations. (687)

Henderson makes her case using a case study on the Copernican Revolution.

Introduction

Main constraint-based compatibilist position: "the Bayesian either does or should make use of explanatory considerations in assigning probabilities (priors and/or likelihoods) to the hypothesis in the hypothesis space" (688).

Cited here as examples of constraint-based position:

  • Okasha 2000
  • Lipton 2004
  • Weisberg 2009

Arguments in this paper:

  1. Neutralize Van Fraassen's argument (section 5)
  2. Case study: "constraint-based compatibilism lacks independent motivation" (688) (section 6)
  3. Argument for emergent compatibilism (section 7)

Terminology:

An "explanationist" refers "both to an agent who performs IBE and to someone who characterizes scientific inference in terms of IBE." Ibid. for a "Bayesian". (688)

Epistemic attitudes

Henderson makes an interesting and important distinction in what's the particular type of epistemic attitude that IBE and Bayesianism address when it comes to assessing hypotheses.

  1. "IBE tells us what hypothesis to select for belief (of perhaps acceptance)" (689)
  2. "whereas Bayesianism operates in terms of degrees of belief" (689)

The potential problem here: "If one regards IBE as wedded to belief rather than degrees of belief, then on some views of the relationship between beliefs and degrees of belief, IBE and Bayesianism may be incompatible from the outset" (689). Henderson does not fully work out this issue, but refers to some articles in footnote 1. The relevant distinction here is between degrees of belief, and full belief.

Following Van Fraassen, the compatibilist debate instead finds "that it is reasonable for an explanationist to have degrees of belief in different hypotheses" (689).

Related point: I've read several arguments (can't recall where exactly, but I think Glass) that point out that Bayesianist are in no way committed to actually accepting any inference (I think this is what is meant by "full" belief) based on their degrees of belief.

Inference to the Best Explanation

Main accounts of IBE:

  • Harman 1965
  • Thagard 1978, 1992
  • Niiniluoto 1999
  • Lipton 2004

Common denominator: "It is often suggested that a theory provides the best explanation if it possesses the optimal combination of explanatory virtues in relation to the phenomena, where explanatory virtues include considerations such as simplicity, unification, scope, and fruitfulness." (689-690)

Henderson argues that this is a too shallow understanding of IBE to fully understand how Bayesianism should render it.

Bayesianism (minimal, subjective, objective)

Minimalistic version includes:

  • Coherence postulate: "degrees of belief in the different hypotheses should be represented by a :clist 3 Bayesians can be placed between two extremes.

Subjective Bayesians:

  • The minimalistic version is the complete story
  • Consequently, any probability assignment over the hypothesis space "represents a reasonable state of degrees of beliefs" assuming the coherence and conditionalization requirement are met.

Authors:

  • Jeffreys

Objective Bayesians:

  • There is in principle "just one rationally acceptable probability function, given the available background belief" (690)
  • As a consequence, objective Bayesians usually impose some normative constraint on the used probabilities.

Authors:

  • Jaynes

In practice, many Bayesians are fine with putting some constraints on priors and likelihoods, without necessarily admitting there's only a single rationally acceptable probability assignment.

Van Fraassen: IBE contradicts Bayesianism

Henderson's summary of Van Fraassen's claims (691):

  • "IBE and Bayesianism have incompatible updating rules"
  • This is because IBE's explanatory considerations add something on top of the probability conditionalization of Bayesianism
  • As a consequence, the explanatory success (a term from Van Fraassen) of a hypothesis boosts the degree of belief in its correctness, thereby violating the requirements to be interpreted as a probability distribution.

His main argument here, that everyone cites, is that IBE falls victim to Dutch book arguments. For Van Fraassen this is an argument against IBE as a rule of inference (i.e. it is irrational, whereas Bayesian conditionalization is perfectly rational here).

However, is it really true that explanatory considerations only play the role of boosting, in terms of probabilities? Isn't this a mischaracterization of IBE in terms of probabilities?

He thus presents a kind of dilemma: either explanatory considerations play this boosting role and produce an updating rule that conflicts with conditionalization, or they play no role at all in a Bayesian context. (692)

However, this dilemma may be a false one. Can we find an interpretation of IBE in a probabilistic context (so other than the "boosting" interpretation) such that IBE and Bayesianism are not incompatible?

Constraint-Based Compatibilism

One alternative where explanatory considerations do not boost probabilities but nevertheless shape then, is one where they "constrain how Bayesian probabilities are assigned." (693)

Proponents view IBE and Bayesianism as mutually beneficial.

N.B. the only feedback on my TFCA_project_proposal.md was that the motivation for making IBE and Bayesianism compatible could be more clear. Refer to this.

Benefit to IBE: Bayesianism provides a formal and widely useful framework for studying inference (cf. Weisberg 2009)

Benefit to Bayesianism: IBE may fill in some gaps that the formal theory does not strictly cover. At this point Henderson does not elaborate, but later on (696) she for example mentions that Bayesianism is silent on what hypotheses are selected for the hypothesis space in the first place.

Is IBE descriptive or normative?

The issue is "whether the role of explanatory considerations is seen as purely descriptive or also normative" (Henderson 2014, 693).

  1. Descriptive -> IBE heuristic for normative Bayesianism; compatible with subjective Bayesianism.

Lipton argues that "people actually do find candidate hypotheses and assign priors and likelihoods to them based on explanatory considerations" (693; cf. Lipton 2004)

Precisely because we are so bad at doing probabilistic calculations, explanatory considerations may be considered a heuristic to approximate outcomes of probabilistic reasoning. Cf. Lipton 2004, p. 120.

IBE is thus a description of how people actually reason, that ideally approximates the normative outcomes as dictated by Bayesian conditionalization.

  1. Normative -> Objective Bayesianism

If on the other hand you argue that explanatory considerations are normative, then you almost certainly have to commit to objective Bayesianism because then IBE will have to impose normative constraints on priors and likelihoods (meaning: for a correct application of Bayesianism these are the right priors and likelihoods).

Example: Weisberg, 2009.

Subjective or Objective Bayesianism?

Within the constraint-based camp, we can thus ask for each proponent "whether they are trying to marry IBE with a subjective or with a more objective form of Bayesianism." (693)

It is good to note that in the case of an objective Bayesian, explanatory considerations are constraints in addition to the restrictions that were already in place on the choice of priors and likelihoods.

This also means that in the case of an objective Bayesian, there are already guidelines for choosing a prior. According to Henderson, Lipton "tends to argue that the Bayesian would be very much at a loss regarding the assignment of probabilities, without explanationist help" (693). This argument thus mostly holds for subjective Bayesianism.

Weisberg is "open on the issue" according to Henderson, which I do not fully understand because Weisberg seems clearly committed to seeing IBE as normative and to objective Bayesianism. Also, the quote Henderson cites for support, assumes objective Bayesianism as a default in my opinion. Cf.: "He says that explanationist thinking should be used to ‘fix the prior probabilities either in conjunction with, or in place of, existing objectivist principles. Ideally, explanationist considerations would complement existing objectivist principles like the Principle of Indifference’ (Weisberg [2009], p. 137)." (693-694).

Constraining probabilities. What probabilities?

Priors:

  • Lipton: "explanatory considerations of unification, simplicity, and their ilk" (Lipton 2004, p. 115)
  • Lipton: priors depend on current empirical evidence in the field, and because explanatory considerations had a role to play here, they also have an indirect influence of the priors.
  • Lipton 2004, 116-148-151: when it comes to this indirect role, IBE may fill the gap in Bayesianism of what hypotheses provide potential explanations in the first place, i.e. the hypothesis-generation aspect of abductive reasoning.
    • Background beliefs based on previous empirical work is important for determining relevant hypotheses; IBE in turn had a role to play in previous empirical findings, similar to the argument in the previous point.
  • Weisberg:
    1. Higher probability to "more systematizing theories" (balance between simplicity and strength). See Weisberg 2009, p. 139. Cf. Henderson 2014, 694.
    2. Constraints on conditional probabilities indirectly impose constraints on priors. If a hypothesis provides a better explanations, we expect it's posterior to be higher. This could be accomodated by giving it a higher prior, in combination with whatever constraints on the likelihood that were already in place. REVIEW, Henderson is a bit vague here. See Weisberg 2009, p 138-139.
    3. Assistive to the principle of indifference.
      • "There are traditional objections to the principle of indifference that stem from inconsistency on apparently equally valid choices of hypothesis space, such as those related by reparametrization. Weisberg’s thought is that these objections might be overcome, to some extent, if explanatory considerations help to determine the appropriate hypothesis space to which we should apply the principle. Specifically, we should choose the hypothesis space where the hypotheses count as ‘potential explanations’ of the relevant data." (Henderson 2014, 694)
  • TODO Huemer 2009a makes a similar point as Weisberg according to Henderson, but based on on explanatory priority: the explanans must be in some sense prior to the explanandum (and I guess that justifies constraints on priors when we are trying to explain some evidence).
    • "Explanatory priority, according to Huemer, may arise due to causal priority, temporal priority, a part–whole relation, an in-virtue-of relation, or supervenience (Huemer [2009a], pp. 352–3)."

Likelihoods:

  • Lipton: 'one way we judge how likely D is on H is by considering how well H would explain D' Lipton 2004, p. 114.
    • This sentence is a bit baffling to me: "In particular, the idea is that considering the mechanism linking cause and effect in a potential causal explanation may help us to form a judgement of how likely the cause would make the effect (Lipton [2004], p. 114)." (Henderson 2014, 694)

Hypothesis space: what hypotheses provide potential explanations?

  • Lipton:

Criticism of constraint-based compatibilism

For subjective Bayesians, the constraints imposed by explanatory considerations do not conflict with pre-existing constraints, like the ones objective Bayesians have.

It is not obvious how explanatory considerations can constrain the prior

Henderson's reasoning is very elegant.

  1. Priors "are supposed to represent the epistemic situation before taking the new data into account" (696)
  2. If explanatory considerations are to guide the priors, they must be determinable a priori, or be "based on background beliefs from previous investigations" (696). So far so good.
    • This implies that it's all about how hypotheses can explain past data.
  3. "However, explanatory considerations like simplicity and unification also depend on the new data." (696)
    • E.g. what a good hypothesis unifies is both past and present data.

This argument is for example geared towards Lipton.

Henderson's criticism here is that this move lacks "independent motivation" (696). She asks whether IBE is "merely accommodated in, rather than explicated by, the Bayesian framework." (696). Subjective Bayesianism is permissible, but we should still ask why exactly the Bayesian should care about IBE.

Weisberg instead argues "that we should constrain priors in such a way as to agree with IBE" (696), which amount to a form of objective Bayesianism.

In the case of objective Bayesians, IBE has to compete with other (independently motivated) constraints. So here the question is: why precisely a form of objective Bayesianism that uses IBE, over the possible alternatives?

Does IBE really fill a gap in the Bayesian framework?

The motivation could be that IBE fills "gaps" in the Bayesian framework (cf. 697). Above we already mentioned discovery.

Interesting point: the arguments used in favor of IBE, may also apply to the Bayesian framework. Henderson thus argues that some arguments about a "gap" in the Bayesian framework are not fair.

The idea (cf. Lipton) was that IBE plays a role in hypothesis generation and prior determination based on background beliefs and prior empirical investigation. However, any Bayesian would likewise generate hypotheses and priors based on the posterior probabilities of previous empirical work.

Thus, one might say that Bayesian conditionalization plays a role in the generation of hypotheses for consideration, just as much as IBE does. In fact, in any two-stage process comprising generation of alternatives and selection among them, where the generation process depends on some products of previous selection, one might say that the selection criteria are to some extent involved in the generation process. (697)

Apparently, Huemer and Weisberg point out some missing aspects in other forms of objective Bayesianism (check Weisberg later).

"They suggest that explanatory considerations are to complement objective Bayesianism by helping to determine the hypothesis space to which the principle of indifference should be applied." (697)

Fair enough. But here Henderson points out that they still do not fully meet Van Fraassen's original point for non-compatibilism, because for that we would have to show how these constraints in objective Bayesianism "amount to a probabilistic version of IBE" (698)

Important reasoning step, that justifies Henderson's own approach to compatibilism:

Van Fraassen’s challenge concerns the relationship between IBE and Bayesian conditionalization in the context of justification, taking a certain hypothesis space as a starting point. Even if explanatory considerations do help to determine that hypothesis space, we still need an argument to establish that the assessment of the explanationist on that hypothesis space will be fully taken into account by a reasonable Bayesian who conditionalized on the evidence. (698)

Emergent Compatibilism

Description of emergent compatibilism:

In my view, a better approach is to ask whether a Bayesian who adopts constraints on her probabilities, which are reasonable on her own terms, would end up favouring more explanatory theories. If so, this would give us an alternative conception of the relationship between IBE and Bayesianism, on which IBE would ‘emerge’ without the Bayesian doing anything differently from what she would anyway. I call this kind of position ‘emergent compatibilism’. (698-699)

So this addresses Henderson's own criticism of constraint-based formalism.

Addresses...

1) ... that using IBE i.c.w. subjective Bayesianism lacks independent motivation * In this case, no such motivation is needed because the Bayesian does inference on her own terms 2) ... that IBE i.c.w. objective Bayesianism does not successfully show how a probabilistic version of IBE looks like * Because Henderson claims the Bayesian would not do anything else than normally.

IBE

Lipton: inference to the most lovely explanation (leading to the most understanding)

Loveliness includes explanatory virtues, like simplicity or unification.

Of course, being a lovely explanation is not sufficient; likeliness always matters as well. A lovely explanation that is highly implausible is worse than one that is lovely and plausible.

So lovely explanations always have these two components of loveliness and being likely enough.

On what type of hypotheses/theories does IBE operate?

  1. "the hypothesis space consists of specific hypotheses that do not leave any details or parameters unspecified", i.e. most discussions focus on easy examples like these.
  2. general theories

IBE on specific hypotheses

Claim: "In judging which hypothesis renders the data most understandable, we consider nothing more than which hypothesis renders it most expected." (700)

Thus, IBE would involve two considerations in cases like this: a hypothesis provides a better explanation than another if i) it would make the explanandum more expected if true, and/or ii) it is initially more plausible. (700)

Henderson tries to give an account of probabilistic IBE. However, I wonder to what extent Henderson's proposal is susceptible to Lipton's arguments that we should not make IBE and Bayesianism compatible by stripping out IBE. So I should ask myself, in what way does Hendersons' proposal not reduce loveliness to likelihood?

In this section Henderson gives a straightforward account of IBE (doctor, symptoms, infer cause etc.) where she shows how information about probability influences how good a given explanation is (relative to another scenario where the information is different). This is the straightforward Bayesian story, as far as I can see.

Henderson does not require a better explanation to have a high probability in absolute terms. This is clear, because she clearly phrases the matter of explanation in a contrastive sense, i.e. for a given hypothesis space, we determine which one is better.

Although an explanation need not make the explanandum expected in absolute terms, nonetheless, an explanation would be better if it did make the explanandum more expected. If the patient were given two treatments, the better explanation of her recovery would still be the treatment that made her recovery more expected, everything else being equal. (701)

IBE on general theories

Ah great! Henderson herself addresses that typical explanatory virtues do not seem to play a role in the previous example.

However, she argues that this is because very simple example with clear unstructured hypotheses obscure the role of these virtues in real scientific practice. Instead, we should consider things we know from philosophy of science, such that some theory has some core claims with various variations and auxiliaries ("free parameters") and so on (e.g. Lakatos, Kuhn etc.)

Henderson's claim:

I will argue that a central factor in the explanatory considerations comprising ‘loveliness’ is the relative responsibility of the core and the auxiliaries in providing an explanation of the phenomena, and that this provides the basis for virtues like simplicity and unification.

Henderson assumes that the hypothesis space has already been determined, including core aspects of theories. The main relevant insight here is that you need auxiliary hypotheses to fit some general theory to a specific domain (e.g. Henderson mentions Newton's second law applied to planets).

Henderson respecs the concept of lovely etc. in terms of these notions of "core" and "auxiliary".

How good an explanation a theory provides, if true, is not just determined by the special case of the theory that makes the phenomena most expected. In many cases, several general theories may be capable of making the phenomena highly expected, when combined with certain auxiliary hypotheses. My proposal is that in such situations, one theory provides a lovelier explanation than another the more the core claims of the theory by themselves allow the phenomenon to be explained and the less the explanation is reliant on auxiliary hypotheses. (702)

Case study: Copernicus vs. Ptolemy

However, the two cases differ in the extent to which the additional hypotheses are essential to the explanation. In the Copernican explanation, the observation of retrograde motion and the coincidence between retrograde motion and opposition for superior planets appear as a direct consequence of placing the sun at the centre. The observed Equation 3 also follows directly from this core claim together with the auxiliary hypothesis that outer planets move more slowly. In the Ptolemaic case, on the other hand, the explanation of the occurrence of retrograde motion necessitates extra hypotheses of planetary epicycles. Without the epicycles, retrograde motion would be unexplained by the Ptolemaic theory. (707)

This is fairly straightforward, and indeed fits what people usually understand to be a lovely explanation; one that relies lies on auxiliary assumptions.

Thus, the Copernican theory provides the lovelier explanation because accepting its core claim that the sun is at the centre of the planetary orbits allows the phenomenon of retrograde motion to be explained with less dependence on auxiliary hypotheses. (707)

Explanatory virtues

So how does that fit with IBE?

Simplicity -> fewer auxiliary hypotheses

Thagard claims that this is the meaning of simplicity anyways. Henderson cites many authors that would support the idea that adding parameters to a theory makes its explanations less simple. In other words, filling in adjustable parameters to explain some fact can be seen as a way of adding auxiliary hypotheses. Indeed, this seems very convincing and straightforward to me.

Unification -> Because the Copernican theory correlates various phenomena by explaining them by core assumptions, rather than accommodating them with auxiliary hypotheses, it can be said to be a more unifying theory.

So simplicity and unification can both be framed in terms of "core" and "auxiliary". But note that unification concerns bringing together phenomena from different domains.

Scope/consilience: trade-off against simplicity and unification; i.e. if you try to make a theory of everything you'll end up making more and more auxiliary assumptions.

Fruitfulness: provide more simple explanations for increasingly many phenomena.

Summary of IBE

I suggest that there is a certain conceptual ‘skeleton’ of IBE that governs how it assigns degrees of belief. There are two considerations involved in IBE: first, the ‘loveliness’ of the hypothesis with respect to the phenomena to be explained— that is, how well the hypothesis would explain the phenomena, if it were true—and, second, the initial plausibility of the hypothesis. In cases where the hypotheses in question are simple and unstructured, how lovely the hypothesis is simply amounts to how expected it would make the evidence if it were true, whereas in cases where the hypotheses are general theories comprising core claims and auxiliaries, how lovely a hypothesis is depends also on the extent to which it provides an explanation in terms of its core claims without relying on auxiliary hypotheses.

I'm not sure yet what Henderson's account adds to Bayesianism in the former case. That's hopefully clear in the next section.

Bayesian account

Claims:

  1. prior takes into account prior plausibility in Henderson's account of IBE (but granted, her account was basically Bayesian; ok, that's a good thing I guess, since her account was still fair)
  2. "the likelihood takes into account the same considerations as the 'loveliness' component of IBE"

In simple cases: "We saw that in such cases loveliness involves little more than expectedness of the evidence given the hypothesis, which is exactly what the Bayesian likelihood also represents." (709)

In the general case: "loveliness does not simply depend on a theory having the best-fitting special case." (709)

There is literature claiming that Bayesian likelihood can reflect explanatory virtues.

  • Rosenkrant 1977
  • Jefferys and Berger 1992
  • MacKay 2003
  • Huemer 2009b
  • Henderson et al. 2010

The likelihood of a general theory is not just that of a special and specific case, "but is an average of the likelihoods across all the specific models that they theory allows" (709)

As such, for free parameters theta (709):

$$Pr(D|T) = \int{}^{}Pr(D|\theta,T)Pr(\theta|T)\delta\theta $$

Where T is a general theory and theta free parameters that can be used to fit the general theory to a particular domain.

Again, note that the likelihood of the theory is expressed in terms of how expected it makes the data.

The likelihood of the theory given the data, Pr(T|D), is usually defined in terms of the likelihood of Pr(D|T), i.e. if this theory were true, how well would it explain the data? In this sentence, "explain" is interpreted in terms of likelihood, where the best core theory would basically be argmax Pr(D|T) or argmax Pr(D|T,theta) for more specific theories.

  • The Bayesian should choose the prior as usual ("prior assignments that may be thought natural" (710)), independent of explanatory considerations

Depending on how the prior over parameters Pr(theta|T) is chosen, the Bayesian likelihood for the Copernican theory may be higher or lower than the likelihood for the Ptolemaic theory. However, there is a strong intuition that, in the absence of any information that parameter values fall in a particular region of the parameter space, it will be most natural to spread the prior probability density out over the whole parameter space. The thought is, in part, that it would be unreasonable to concentrate the prior in any particular subregion of the parameter space without good reason to do so, as in so doing, one might miss out on assigning probability to the correct value. (709)

What is less clear to me is this sentence: "Depending on how the prior ... is chosen, the Bayesian likelihood for the Copernican theory may be higher or lower than the likelihood for the Ptolemaic theory." (my emphasis)

I sometimes have the feeling she talks about the posterior probability and not what is commonly know as the likelihood. Because indeed, to get the posterior you multiply likelihood with prior pointwise (+ normalization, of course).

Okay, I think this is what is going on.

  1. Henderson expresses the likelihood of the observed data with respect to the core theory T. Because we are interested in expressing the value of e.g. the Copernican in general i.e. without all auxiliary hypotheses. So we need to marginalize over these auxiliary hypotheses expressed with theta.
    • In other words, we assume we have the likelihood Pr(D|theta,T), but we want the likelihood of the core theory.
  2. Then the question is, how do we obtain the likelihood? I think Henderson assumes here we get it with a posterior predictive distribution, i.e. if we would predict data given our core theory and a particular auxiliary hypothesis how likely would the data then be?
    • This is the crucial step because here we utilize the priors of the theory combined with the specific auxiliary hypotheses!
  3. If we have this likelihood expresses w.r.t. core theory T, we can compute a posterior probability of core theory T given the data. But in this procedures, the priors over auxiliary hypotheses did influence the used likelihood.

So, in summary:

In the average, the likelihoods of specific models Pr(D|theta, T) are weighted by the prior probability density conditional on the theory, Pr(theta|T).

So we are dealing with a prior over the parameters theta, not over the overall theory.

  • Henderson shows in the Copernicus vs Ptolemy case study that nevertheless Bayesians favor Copernican theory for the same reason as explanationists
  • Having more free parameters is penalized compared to explaining data with core theory T.

Because the data do in fact conform to the relationship expressed in Equation 3, most of the additional possibilities allowed by the Ptolemaic theory have poor fit and serve only to bring down the average likelihood of the Ptolemaic theory compared to the Copernican theory.

Note that complexity penalization is an inherent characteristic of Bayesian model comparison (Henderson does not mention that explicitly here). Review why.

  • So even if you assign Copernican theory and Ptolemaic theory the same "natural" prior (i.e. following more of less the objective style of maximizing indifference), "Ptolemaic theory would still be penalized in the likelihood." (710)

Why is that exactly?

This is the crux:

The relation is expected on the basis of the core claims of the theory, and not on the basis of any particular parameter settings. In the Ptolemaic theory, on the other hand, the relation in Equation 3 only holds for some parameter settings but not for others, as it is not determined by the core of the theory. In the Copernican theory, the prior probability for specific models is only distributed over cases in which Equation 3 holds, whereas the Ptolemaic prior is spread out over a larger parameter space that includes many possibilities that do not fit the relation. This is because the Ptolemaic theory has extra free parameters corresponding to the periods of the epicycles that can be tuned to account for observations of any synodic periods. Thus, it can account for any relationship between the synodic and sidereal periods. Because the data do in fact conform to the relationship expressed in Equation 3, most of the additional possibilities allowed by the Ptolemaic theory have poor fit and serve only to bring down the average likelihood of the Ptolemaic theory compared to the Copernican theory. Thus, with prior assignments that may be thought natural, the Bayesian likelihood favours the Copernican theory because it does not rely on auxiliaries in the form of adjustable parameters to as great a degree as the Ptolemaic theory. (710)

So even if you assign the same prior to both core theories (T), the average likelihood of the theory is affected by the amount of free parameters. When core theory T is better, it explains the data relatively independent of auxiliary parameters and requires less auxiliary parameters. If instead you rely on many parameters with a bad fit ( lower Pr(D|theta,T)), these extra parameters are not very helpful, but even more, they do damage because the few thetas for which Pr(D|theta,T) is decent, are now scaled down by the prior over the specific assumption, Pr(theta|T), because the probability mass is distributed over more parameters.

The issue whether IBE is normative

Not entirely clear yet:

I will not attempt to discuss this further here, but what I have said so far gives an indication of how emergent compatibilism may provide the basis for some specific ways of tackling the problem of IBE’s normative standing. (711)

Constraint-based accounts assume that IBE is normative; otherwise it wouldn't be able to impose normative constraints on priors etc.

Henderson's compatibilism boils down to saying that IBE and Bayesian are two ways of talking about the same, but on a different descriptive level

Emergent compatibilism would not make IBE redundant. IBE and Bayesianism might be regarded as different levels of description of the inference process, where IBE describes the gross phenomena of inferential thinking, whereas the quantitative detail that gives rise to these explanatory features is captured by the Bayesian model. (711)

Hmmm, one may argue that if they really are different levels of description, the higher level of description (IBE) does also not necessarily emerge from the lower one.

So I'm not sure how apt the term "emergent compatibilism" is.

Summary

The first task was to determine what these considerations are. I suggested the following conceptual skeleton of IBE: a theory provides a better explanation of certain evidence than another theory to the extent that a) the theory itself has more initial plausibility; and b) it is ‘lovelier’, that is, compared with the other theory, it would explain the evidence better if it were true. I then showed how in certain realistic examples, the Bayesian, making assignments of probabilities on her own terms, takes these considerations into account. If the theory is more plausible, it gets a higher prior; if it is lovelier, it gets a higher likelihood. Cases of the latter connection fell into two categories. For cases of specific hypotheses, there appeared to be little more to the loveliness of a hypothesis than how expected it makes the phenomena. Thus, the Bayesian would in such cases assign the higher likelihood to the lovelier hypothesis. For cases of general theories, such as those of Copernicus and Ptolemy, I proposed that the lovelier theory is that whose explanation relies more on the core of the theory, as opposed to the auxiliaries. With certain independently motivated choices of priors, the Bayesian likelihood favours the lovelier theory because of the way it averages over specific cases that the theory allows. It was not necessary to assume, as the constraint-based compatibilist does, that explanatory considerations help to determine priors or likelihoods. (712)

What is there to mutually gain for IBE and Bayesianism?

Emergent compatibilism would allow the two methods to be mutually illuminating. IBE provides a way to describe in qualitative terms what is going on in Bayesian inference, with the Bayesian account providing a systematic way to explain how explanatory virtues are preferred in IBE. (712)