Philosopher Elliott Sober is well known for his trenchant critique of design arguments, that is, arguments that aim to infer the existence of God or, more modestly, an intelligent designer from features of the universe allegedly indicative of design. Sober has focused specifically on organismic versions of the design argument, which postulate an intelligent designer to explain physical features of organisms such as complex adaptation. Sober’s main objection to such arguments, whether in their classical or contemporary forms, is that the design hypothesis is untestable. Understood more precisely, his argument is that the testability of a hypothesis depends on the testability of the auxiliary statements on which the predictive consequences of the hypothesis depend, the auxiliary hypotheses required for the design hypothesis to have predictive consequences are not testable, and therefore, the design hypothesis is not a testable hypothesis.
I don’t want to discuss whether Sober is correct in his criticism of design arguments. What interests me is the similarity between Sober’s criticism of design arguments and one of my criticisms of empirical arguments for life after death from the data of psychical research e.g., arguments for survival from alleged communications with the dead (associated with mediumship and apparitions of the dead), cases of the reincarnation type, and near-death experiences. In my more recent publications, I’ve drawn attention to what I call the problem of auxiliary hypotheses (hereafter, PAH).
PAH may be concisely stated as follows.
(1) The evidential force of the relevant data (drawn from psychical research) depends on the survival hypothesis having predictive power.
(2) The survival hypothesis has no predictive power unless it’s supplemented with auxiliary hypotheses.
(1) and (2) jointly entail what I call an auxiliary hypothesis requirement: empirical arguments for survival are necessarily dependent on various auxiliary statements or assumptions. This in turn results in a problem for empirical survival arguments because on my view:
(3) The auxiliary hypotheses required for the survival hypothesis to have predictive power are not independently testable.
It may not be immediately obvious why (3) is a problem, but I think it’s a serious problem for empirical arguments for survival. In fact, I consider it the nub of the problem for empirical survival arguments. I think it’s far more serious than the traditional objections from the alleged low antecedent or prior probability of survival and the appeal to various alternative non-survival explanations of the data.
I’ve decided to devote three blogs to this topic, which will also allow me to sketch arguments that will be more thoroughly developed in my book in progress.
- The first installment (below) looks at Elliott Sober’s critique of design arguments and provides some preliminary observations on its implications for the empirical survival debate. Sober’s critique of design arguments supplies us with both a general conceptual background and a relevant analogue to the survival hypothesis and empirical arguments for survival.
- In the second installment, to be published next month, I’ll look more carefully at empirical arguments for survival in the light of Sober’s critique. I’ll provide an account of the kinds of auxiliary hypotheses that must be, and typically are at least implicitly, relied upon in classical empirical arguments for survival. I’ll also explain why most, if not all, of them are not independently testable, as well as why this is a problem for empirical arguments for survival.
- In the final installment I’ll consider some possible survivalist responses to PAH, but I’ll argue that these maneuvers are ultimately unsuccessful. Possible means of circumventing PAH saddle empirical arguments for survival with epistemically toxic residue in the form of strengthened antecedent probability and alternative explanation objections that are immune to traditional survivalist criticisms of these objections. Hence PAH places empirical survivalists on the horns of a significant dilemma.
1. Sober’s Critique: What’s Not Wrong with Design Arguments
In Evidence and Evolution (Cambridge, 2008) Sober considers the force of the hypothesis of intelligent design as an ostensible explanation of the complex adaptive features of organisms. Since the latter part of the nineteenth century, most biologists have explained such features by appealing to Darwinian evolution, but the intelligent design hypothesis postulates the agency of an intelligent being as the explanation of complex physical adaptations. Sober argues that classical and contemporary versions of the design argument fail, but not for the reasons normally encountered among critics of such arguments.
First, Sober does not argue that the design argument is defeated on the grounds that the intelligent design hypothesis has a low prior probability (E&E, p. 121). (“Prior probability” here refers to a hypothesis’s probability independent of, or prior to, considerations drawn from present data adduced in support of the hypothesis.) Sober is actually explicit that the evaluation of the intelligent design hypothesis, as well as the rival Darwinian evolution hypothesis, must do without considerations of prior probability. Why? Simply this: there’s no objective way to assign prior probabilities to these kinds of hypotheses. If we assign prior probabilities, it would amount to little more than an expression of our personal or subjective belief predilections. Science has more rigorous aims. Since prior probabilities play a role in judgments about the net plausibility or overall probability of a hypothesis, Sober does not propose to render a verdict on this with respect to either intelligent design or Darwinian evolution. His aim is modest. Assess whether the relevant evidence favors the design hypothesis over its main rival, Darwinian evolution.
Sober carries out this assessment on the basis of a Likelihoodist approach to confirmation theory (E&E, pp. 121-122). According to a widely discussed formulation of the Law of Likelihood (LL): evidence e favors some hypothesis h1 over hypothesis h2 if and only if the probability of e given h1 is greater than the probability of e given h2. Likelihoodism, then, ignores the prior probability of hypotheses, as well as the correlated attempt to justify claims about the probability or net plausibility of a hypothesis. It’s only interested in determining whether evidence favors, supports, or confirms a hypothesis, and this by virtue of assessing whether a hypothesis better leads us to expect the relevant data/evidence/observation than does some specific rival hypothesis. More technically stated, Likelihoodism focuses on the likelihood of a hypothesis, which is a technical way of referring to the probability of the evidence given the hypothesis. This probability, Pr(e/h), should be distinguished from the probability of the hypothesis given the evidence, Pr(h/e). The first might be high but the latter low. The hypothesis that there are gremlins bowling in my attic has a high likelihood because it renders the sounds I hear in my attic very probable, easily more probable than many different competing hypotheses, but the gremlin hypothesis has a low probability because it has a low prior probability.
Furthermore, notice that Likelihoodism is a contrastive approach to evidential support. It explicates the favoring or supports relation by comparing the likelihoods of hypotheses with each other. While Bayesian approaches to confirmation theory involve contrasting a hypothesis h1 with its negation ~h1, Likelihoodism contrasts a particular hypothesis h1 with some other hypothesis h2. For any two hypotheses, h1 and h2, and observational evidence e, it aims to assess whether the probability of e is greater given h1 than it is given h2, that is, formally whether Pr (e/h1) > Pr(e/h2). Note that if Pr (e/h1) > Pr(e/h2) this does not require that Pr(e/h2) be low or that Pr(e/h1) be high, only that Pr(e/h1) is greater than Pr(e/h2), though of course it might be “much greater.”
Sober takes the view that design arguments are best formulated as Likelihood arguments that compare the likelihood of the design hypothesis with that of a rival hypothesis. In this way, their aim is modest, the thorny problem of assessing prior probabilities is avoided, and the arguments circumvent some of the traditional skeptical criticisms, for instance some of David Hume’s criticisms in the eighteenth century that assume the argument is an argument from analogy that depends on a high degree of overall resemblance between organisms and human artifacts like watches.
So, for example, William Paley’s famous organismic design argument, which focuses on complex adaptive features of organisms (O), should be formulated as:
(1) Observation O favors the intelligent design hypothesis over the chance hypothesis if and only if Pr(Observations / Intelligent design) > Pr(Observations / Chance)
(2) Pr(Observations / Intelligent design) > Pr(Observations / Chance)
(3) Observation O favors the intelligent design hypothesis over Chance
However, an apparent weakness of Paley’s argument, and by implication all organismic design arguments, is that, even if Paley was correct that such evidence favors intelligent design over purely random natural processes, it may nonetheless still be the case that:
Pr(Observations / Darwinian evolution) > Pr(Observations / Intelligent design),
or even that
Pr(Observations / Darwinian evolution) >> Pr(Observations / Intelligent design).
Hence, if we contrast intelligent design with a hypothesis other than chance, which in the case of Darwinian evolution Paley himself could not have anticipated, we get a different result. And in fact, one of the responses to Paley-style design arguments is that Darwinian evolution has the upper hand since the relevant data are more probable, perhaps much more probable, given Darwinian evolution than intelligent design. Sober notes, for example, that Stephen J. Gould takes this approach (E&E, pp. 127-128). Gould has argued that imperfect adaptations in nature are very surprising if organisms have been designed by an intelligent being, but wholly expected if Darwinian evolution tells the correct story. For example, Gould argues that the panda’s “thumb” (that is, the spur bone extending from the panda’s wrist), which together with the panda’s paw is used to strip bamboo stalks for eating, is highly inefficient. While such inefficiencies are to be expected on the hypothesis of Darwinian evolution, they are not to be expected given the hypothesis of intelligent design. Therefore, the observation like the panda’s “thumb” (and many others could be provided) favors Darwinian evolution over the intelligent design hypothesis.
Sober, though, has a very different kind of criticism, and here’s where Sober’s approach gets interesting.
2. Sober’s Critique: What’s Wrong with Design Arguments
Just as Sober doesn’t think that the Achilles Heel of the design argument rests in the low prior probability of the intelligent design hypothesis, he also doesn’t think that the nub of the problem within a Likelihood framework is that Darwinian evolution has a higher likelihood than intelligent design. The problem is that we simply are not in position to as much as assess whether
Pr(Observations / Intelligent design) > Pr(Observations / Darwinian evolution),
much less whether
Pr(Observations / Intelligent design) >> Pr(Observations / Darwinian evolution),
These (weaker and stronger) intelligent design hypotheses have inscrutable likelihoods because we can’t really say what the empirical world should look like if the design hypothesis is true. While we can make claims about the likelihoods found on the right hand of the equation above, we cannot do so for the likelihood on the left side (E&E, pp. 141-147, 189).
It’s important to more clearly state and explore the nature of the problem here.
First, the problem is not that the hypothesis of intelligent design by itself has no predictive consequences. Sober emphasizes the Duhem-Quine thesis that predictive consequences emerge only when we consider sets of statements, a hypothesis + auxiliary statements (E&E, pp.144-145). So it’s not a problem that the supposition of an intelligent designer by itself has no predictive consequences. The same would be true for rival hypotheses. After all, natural selection only makes predictions if it’s supplemented with its own set of auxiliary hypotheses, e.g., about the “targets” of selection and “constraints” on selection processes.
Second the problem isn’t that we can’t find any statement that, once conjoined to intelligent design, has predictive consequences. As Sober further notes (E&E, p. 129-131), it’s monumentally easy to find auxiliary statements that will assist the intelligent design hypothesis in generating testable predictions. For example, postulate an intelligent designer, but further postulate that the designer would have wanted everything in the world to be purple. This generates, by deductive entailment, the prediction that every object in the world should be purple. Clearly this prediction is false. Therefore, the design hypothesis is falsified. We could also suppose that the intelligent designer gave vertebrates their eyes, which of course entails that vertebrates have eyes. This also results in a specific kind of prediction, so intelligent design turns out to be a falsifiable (though not falsified) hypothesis.
Notice that Gould does something similar to show that Darwinian evolution has a higher likelihood than intelligent design. He supposes, not that the designer would have wanted the world to consist of only purple objects, but rather that he would have wanted the panda’s “thumb” to be more efficiently constructed. Therefore, the panda’s “thumb” is very surprising given the design hypothesis. However, we could just as easily have picked an auxiliary hypothesis that would be favorable to intelligent design, like an intelligent designer who would have wanted humans to have eyes with the features our eyes actually have and pandas to have a spur bone extending from their wrists. So we can easily pick auxiliaries that result in the observational evidence having a probability of unity, zero, or anywhere in between, given the hypothesis of an intelligent designer and the chosen auxiliaries (E&E, pp. 142-144).
The problem should now be apparent. The predictive consequences in each of the above instances are derived from the hypothesis of intelligent design supplemented with an auxiliary assumption that attributes to the designer, if such a being should exist, abilities and desires/goals of a particular sort. But Gould is no more entitled to make an assumption unfavorable to the design hypothesis here than Paley and company are entitled to make assumptions favorable to the design hypothesis. Neither adopts an assumption that can be independently tested. Neither is justified in believing what the abilities and goals of an intelligent designer would be. (And Sober thinks that same conclusion follows if the intelligent designer is more robustly described as “God,” that is, an all-knowing, all-powerful, and all-good being). What is relevant for the testability of a hypothesis is that we derive predictions with assistance from independently testable auxiliary hypotheses. And this requires that our justification for believing auxiliary statements does not depend on our believing that either H1 or H2 (the hypotheses whose likelihoods are under consideration) is true, or even that the observational datum is true. (E&E, p. 152).
So, on Sober’s view, the problem with the design hypothesis is that it cannot be tested because we don’t know or have justified beliefs about what auxiliary hypotheses are true. We neither know nor justifiably believe (independent of the hypothesis of intelligent design) what the goals and abilities of the designer would be should such a being exist. As Sober says, “The problem with the hypothesis of intelligent design is not that it makes inaccurate predictions but that it doesn’t predict much of anything at all” (E&E, p. 154). Hence, we’re not in a position to justifiably claim that
Pr(Observations / Intelligent design) > Pr(Observations / Darwinian evolution),
much less that
Pr(Observations / Intelligent design) >> Pr(Observations / Darwinian evolution).
Therefore, we’re not in a position to say that the relevant evidence favors intelligent design over Darwinian evolution.
3. Sober’s Critique and Empirical Survival Arguments: Preliminary Considerations
Although I’m a philosopher of religion with a long-standing interest in arguments for the existence of God, at present Sober’s central criticism of intelligent design arguments interest me because of its implications for another species of empirical argument at the center of my current work, arguments for postmortem survival or life after death from the data of psychical research. The data here would be data collected from paranormal phenomena such as of out-of-body and near-death experiences, mediumistic communications, cases of the reincarnation type, and apparitions of the dead. I’ve argued in a few places that the survival hypothesis leads us to expect such data only if we adopt a significant number of auxiliary hypotheses whose epistemic credentials are at best questionable. Otherwise stated, the predictive power of the survival hypothesis depends on auxiliary hypotheses that lack the appropriate epistemic credentials. This I maintain constitutes a defeater for empirical arguments for survival.
In the next two blogs I’ll develop this argument. Here I’ll offer some preliminary remarks.
As I see it, far too many empirical survivalists are either unconscious of the extent to which their arguments depend on auxiliary assumptions, or they are unconscious of the implications this has for the assessment of the evidential force of the relevant data. One contributing factor here is the refusal of empirical survivalists to rigorously develop the empirical argument for survival, for instance, by addressing some very basic issues in confirmation theory, or otherwise putting their principles of inductive inference on the table and clearly applying them to the survival hypothesis. The tendency is to pile up data, in much the same way that many eighteenth and nineteenth century theologians thought they could prove the existence of God by simply piling up alleged examples of design in the world. But a mass of data does not an argument make.
However, another reason for this degree of unconsciousness about the relevance of auxiliary assumptions is rooted in a particular strategy of argument adopted by a large number of empirical survivalists. It’s what Sober calls “lazy testing.”
“The lazy way to test a hypothesis H is to focus on one of its possible competitors H0, claim that the data refute H0, and the declare that H is the only hypothesis left standing. This is an attractive strategy if you are fond of the hypothesis H but are unable to say what testable predictions H makes.” (E&E, p. 353)
This sums up one of the central strategies of argument found in the bulk of survival literature. Empirical survivalists routinely think the survival hypothesis has acquired some sort of positive epistemic credential because they identify some particular datum that is allegedly improbable given an alternative non-survival hypothesis. For example, empirical survivalists think they’ve refuted appeals to psychic functioning among living persons by pointing to behavioral patterns or skills exhibited by trance mediums or young children, where the behavior or skills are characteristic of some deceased person. This sort of phenomenon is allegedly improbable or not to be expected if we adopt a living-agent psi hypothesis. Well, in the light of Sober’s critique of design arguments, it’s clear that such a tactic only facilitates distraction from the central issues, namely the extent to which the survival hypothesis renders the data probable, and what must be assumed about survival to determine this.
Notice also that when critics of survival arguments argue that more robust versions of the living-agent psi hypothesis (e.g., Stephen Braude’s motivated living-agent psi hypothesis) challenge the survival hypothesis, survivalists shift to a different debunking strategy. They try to rack up considerations that lower the prior probability of the counter-explanation. For example, living-agent psi explanations of the data are often said to be overly complex, or they fail to fit with our alleged background knowledge since they postulate psi of a potency, magnitude, or level of refinement for which there is no independent evidence, or they depend on psychodynamic hypotheses that stand in need of independent support. Again, the focus is on how competitors fail, not on how the survival hypothesis succeeds.
As I’ve argued, counter-explanations may indeed have a very low prior probability, but if the empirical argument for survival is construed as a likelihood argument, then it’s irrelevant that the prior probability of motivated living-agent psi, dandy psi, superman-psi, God-potent psi, or whatever, is low. As Sober emphasizes, Likelihood arguments don’t bring prior probabilities to bear on evidence assessment. Moreover, as far as prior probability assessments go, the relevant comparison must be between robust versions of all the explanatory candidates, including the survival hypothesis. So if we are die-hard Bayesians, and we wish to legitimately introduce considerations of prior probability, we can’t sensibly compare a simple survival hypothesis with a robust counter-explanation. We must compare the prior probability of robust versions of the competitors with robust versions of the survival hypothesis, because it’s only robust versions of the explanatory candidates that have any predictive consequences.
Sober’s observation, derived from Richard Royall, is instructive at this juncture. There are two kinds of questions that need to be distinguished. We can pose the question, “What does the present evidence say?” We can also pose the question “What should you believe?” The Likelihood approach addresses the first; the Bayesian approach the second. It’s best that survivalists more clearly distinguish these questions in relation to their assessments of the alleged evidence for survival. Accordingly, they need more clearly to distinguish between whether they want to defend modest likelihood claims or stronger claims about the net plausibility of the survival hypothesis, based on the joint consideration of likelihoods and priors.
That being said Sober’s critique illuminates what I consider the nub of the problem facing empirical arguments for survival, whether they are formulated along Likelihood or Bayesian lines. Empirical survivalists need to state the kinds of auxiliary assumptions that are required for the survival hypothesis to establish a genuine connection with the empirical world, specifically the range of data adduced in support of the survival hypothesis. And they need to show that the survival hypothesis does a better job vis-à-vis its predictive consequences than do the competitors. In my next blog, I’ll sketch some of the auxiliary assumptions needed for classical empirical arguments for survival, and I’ll also begin exploring why this is a problem. By the third installment, I hope it’s clear why I think PAH—the problem of auxiliary assumptions—poses the most fundamental kind of challenge to empirical arguments for survival.