Not Null Enough 1 Not Null Enough: Causal Null Hypotheses in Community Ecology and Comparative Psychology William Bausman and Marta Halina Philosophy of Biology at Madison 2014

1. Introduction We have three aims in this paper. First, we characterize a common research strategy that we see operating in community ecology and comparative psychology. This research strategy involves the application of what we call a "causal null hypothesis" and is used as if it were an extension of statistical hypothesis testing. Second, we challenge this research strategy by arguing that it is relevantly different from statistical hypothesis testing. Whereas the null hypotheses used in statistical hypothesis testing are hypotheses of no effect, the causal null hypotheses employed in our case studies are both positive causal hypotheses. Third, we reevaluate the use of causal null hypotheses in light of our challenge. Recognizing that a causal null is not a statistical null hypothesis entails taking away the epistemic privileges that it previously enjoyed. We argue that the causal nulls in our case studies should be treated on a par with their competitors. In the end, we hope that our analysis will stimulate critical discussions on the role that neutral models play in other cases in biology. We begin in section 2 by characterizing the causal null hypothesis strategy and showing how it is employed in community ecology and comparative psychology. In section 3, we critique this strategy. We argue that the causal null hypothesis cannot be justified as an extension of statistical null hypothesis testing because the causal nulls that it posits are relevantly different from the null hypotheses used in statistical hypothesis testing. In the section 4, we respond to an objection to our argument and consider one important consequence of our analysis—namely, that the causal nulls in our case studies should not be privileged over alternative hypotheses. It is important to emphasize that we are interested only in presenting and critiquing the causal null hypothesis strategy, not the specific hypotheses that this strategy supports. There may be a better argument or alternative line of evidence that supports the hypotheses featured in our case studies. We do not want to deny that this is the case. Instead, our focus is on critiquing the

Not Null Enough 2 causal null hypothesis strategy as it is employed in these areas of research.

2. The Causal Null Hypothesis Strategy In this section, we characterize the causal null hypothesis strategy and show how it is employed in two areas of contemporary scientific research: community ecology and comparative psychology. In outline form, the strategy consists of the following three steps: first, constructing a causal hypothesis that is taken to be simpler than some alternative hypothesis; second, maintaining that one must reject the simpler hypothesis before accepting the alternative hypothesis; third, holding that, if one fails to reject the simpler hypothesis, then it should be accepted as the best explanation for the phenomenon under investigation. In what follows, we present these three steps in greater detail and show how they apply in our two cases studies.1 Before introducing the causal null strategy, it is important to note the broader context in which it is employed—namely, an explanatory one. In both of our two case studies, researchers employ this strategy in an attempt to determine the best explanation for a phenomenon of interest. Moreover, the explanations presented are broadly causal. That is, explaining the phenomenon of interest is taken to involve identifying the causes or mechanism that gives rise to that phenomenon. The causal null hypothesis strategy is employed in order to determine which of two proposed causes or mechanisms is in fact responsible for the phenomenon. The causal null hypothesis strategy begins with an already existing causal hypothesis. This hypothesis is taken by the research community to provide an adequate explanation of the phenomenon of interest. The reasons for accepting this hypothesis are various. For example, the hypothesis might account for the existing empirical data and researchers might have independent theoretical reasons for finding the hypothesis plausible. Or the hypothesis might have led to a series of empirical predictions, which were tested and confirmed in the laboratory. Whatever the case might be, researchers take this hypothesis to explain the phenomenon of interest. We will call this hypothesis the “causal alternative hypothesis” or “causal-alternative.” The causal null hypothesis strategy next involves constructing a second hypothesis 1 While we will not discuss these papers here, we would be remiss to not acknowledge some of the many papers that have been written on this topic (e.g., Roughgarden 1983, Quinn & Dunham 1983, Beatty

Not Null Enough 3 against the background of this already existing causal alternative hypothesis. For reasons that will become clear, we call this second hypothesis the “causal null hypothesis” or “causal-null.” Like the causal-alternative, the causal-null is advanced as an account aimed at explaining the phenomenon of interest. However, the causal-null differs from the causal-alternative in one important respect. Namely, it is taken by its proponents to be simpler than the causal-alternative. Proponents of the causal-null hold that it is simpler than the causal-alternative because it excludes one or more causal factors that the causal-alternative includes. The causal-null is taken to be null with respect to these causal factors. Let us now illustrate the above with our two case studies. Our first case study comes from community ecology. A long-standing problem within community ecology is to understand the patterns of abundance and diversity of communities, and the mechanisms underlying these patterns. The particular pattern of interest in this case is the relative species abundance distribution or the number and population size of the species found within a single trophic level. Explaining relative species abundance distribution requires knowing the causes or mechanisms that produce this distribution. The traditional and still dominant explanation of this phenomenon is that it results from the competition between species for resources and the trade-offs between how different species utilize resources. The competitive exclusion principle is an extreme example of this approach. This principle holds roughly that, if species coexist, then there must be differences between how the species utilize resources. Mechanisms that depend upon species differences, as interspecific competition does, are called selection mechanisms by analogy to natural selection. Hypotheses following this approach function as causal alternative hypotheses and we will refer to them generally as "the selection hypothesis". Stephen Hubbell's neutral theory of ecology has challenged the selection hypothesis as an explanation of observed relative species abundance distributions (Hubbell 2001). The neutral theory of ecology holds that communities are structured entirely by three mechanisms: ecological drift, random immigration, and random speciation. It does this by assuming neutrality, that all individuals regardless of species have identical chances of giving birth and dying, of immigrating, and of speciating. This makes the neutral theory null with respect to species differences. Hubbell is explicit that this is the only sense of neutrality invoked, that it is not neutral is the sense of 'nothing going on'. The neutral theory of ecology predicts that a

Not Null Enough 4 community's observed relative species abundance distributions will be a particular statistical distribution, the zero-sum multinomial distribution, whose particular shape is determined by parameters interpreted as community size, metacommunity size, immigration rate, and speciation rate. We refer to this causal null hypothesis as "the neutrality hypothesis". The neutrality hypothesis is explicitly constructed to exclude the influence of species differences. Hubbell writes that the value of constructing the neutral theory of ecology is that, "we obtain a quantitative null hypothesis against which to test when, to what extent, and for which species demographic differences among species are necessary to explain observed community patterns." (Hubbell 2006, 1387) Because the neutral theory excludes species differences, the neutral theory plays the role of the "null" hypothesis and can be used to test the causal responsibility of species differences in producing observed relative species abundance distributions. Our second case study comes from comparative psychology. In the last 15 years, there has been a large amount of research devoted to determining whether chimpanzees are capable of mindreading. On the basis of the results of this research, most comparative psychologists have concluded that they have this cognitive ability (Tomasello & Call 2006, Call & Tomasello 2008). Broadly, mindreading is the ability to reason about the mental states of other agents. Researchers currently maintain that chimpanzees are capable of some, but not all, forms of mindreading. For example, they appear to reason about the goals and perceptions of other agents, but not their beliefs. The phenomenon to be explained in this case is chimpanzee behavior. Namely, researchers aim to determine the cognitive mechanisms that produces this behavior.2 The “mindreading hypothesis” holds that the best explanation for why chimpanzees behave as they do in mindreading experiments is that they are capable of reasoning about the mental states of other agents. Povinelli and colleagues have advanced a causal-null hypothesis that they take to explain chimpanzee and other nonhuman animal mindreading behavior (Povinelli & Vonk 2003, Povinelli & Vonk 2006, Penn & Povinelli 2007, Penn et al. 2008, Penn & Povinelli 2009). This

2 It is worth noting that this explanatory hypothesis is not as concrete as positing a neural or biological mechanism. However, this hypothesis can be considered causal in the sense that researchers take it to correspond to some neural mechanism that will eventually be discovered. For the purposes of this paper, we adopt Piccinini and Craver’s (2011) view of psychological explanations as mechanism sketches.

Not Null Enough 5 causal-null posits that nonhuman animals employ a set of abstract behavioral rules that allows them to reason in sophisticated ways about observable behaviors. It is this set of rules that produce the behavior in which chimpanzees engage. We will refer to this causal-null as the “behavior-reading hypothesis.” Proponents of the behavior-reading hypothesis hold that it is simpler than the mindreading hypothesis because it explains chimpanzee behavior without invoking mindreading. The mindreading hypothesis explains the behavior of chimpanzees in particular experimental tasks by holding that they 1) observe a situation involving an agent, 2) attribute an appropriate mental state to that agent, given the situation, and 3) predict how that agent will behave, given his or her mental state. For example, a chimpanzee might observe an experimenter directing her eyes at a tool as it is hidden in a container and, on this basis, attribute to the experimenter the mental state of “knows where the tool is hidden.” From there, the chimpanzee can predict that the experimenter will search in the correct location for that tool if she needs it to obtain a reward. In contrast, the behavior-reading hypothesis holds that a chimpanzee need not attribute a mental state to an agent in order to predict how that agent will behave, but rather, can make such predictions on the basis of the observable situation alone. For example, the hypothesis holds that a chimpanzee will predict that an experimenter will search in the correct location for a tool on the basis of the fact that the experimenter had her eyes directed at that tool during the hiding process. Thus, proponents of the behavior-reading hypothesis take it to be simpler from the mindreading hypothesis because it does not posit the additional cause of reasoning about mental states (see Povinelli & Vonk 2003, Penn et al. 2008, Penn & Povinelli forthcoming). The next significant move in the causal null hypothesis strategy is to hold that researchers must reject the causal-null before they are justified in accepting the causal-alternative. This move establishes an epistemic asymmetry between the two hypotheses that is not based on a differential ability to account for the empirical data. The causal-null strategy holds that researchers must reject or falsify the causal-null before they are justified in accepting the causalalternative as the best explanation for the phenomenon of interest. If the causal-null does not account for the empirical data, then the causal-null is rejected and the causal-alternative is sustained. If the causal-null does account for the empirical data, however, then the causalalternative is shown to be unnecessary and so judged unwarranted, and the causal-null is accepted. Thus, in order to determine which of the two hypotheses best explains the phenomenon

Not Null Enough 6 under investigation, research must determine whether and how they can reject the causal-null. The above inference is justified by those who use it on the grounds that the causal-null is simpler than the causal-alternative. Given this, although both hypotheses might be individually sufficient for explaining the phenomenon of interest, the causal-alternative is not necessary for explaining this phenomenon. It is not necessary because it posits everything that the causal-null posits and more. Because the causal-null can account for the empirical data, these additional causes are deemed unnecessary for explaining the phenomenon. Thus, proponents of the causalalternative need to show that the “and more” included in their hypothesis is necessary by demonstrating that the causal-null is insufficient for explaining the phenomenon on its own. The neutral theory aims to provide a general framework for understanding the mechanisms and causal factors underlying ecology diversity, but the neutrality hypothesis is evaluated for each community independently. Thus, the neutral theory might fit one tropical forest's data very well and another's less well. The conclusions for these two communities will then be different. The informative question for illustrating the method is, “what is inferred when the neutral theory fits the data to a high enough standard?” Ecologists have pointed out that diversity data can not in general function as a crucial experiment for the selection and neutrality hypotheses and that both hypotheses are in general capable of accounting for the observed data. To this Hubbell has responded, "I agree, but obtaining acceptable fits from neutral models shifts the burden of proof to those who would assert that more complex theory is required to explain nature and with what level of detail and generality" (Hubbell 2006, 1387). Because the neutral theory is simpler than any theory that assumes species differences, the burden of proof is on proponents of the more complex theory to reject the neutral theory. If they are unable to do so, they ought to conclude that neutral processes are dominant in producing observed diversity patterns for that community. Hubbell asks, "why should the UNT [the neutral hypothesis] ever get an acceptable answer—even approximately—for the wrong reason?" He answers, "I believe that the theory simply could not do so unless it accurately captures some fundamental statistical-mechanical characteristic of the behavior of biodiversity in aggregate and at large spatial and temporal scales" (Hubbell 2006, 1388). On the basis of the superior simplicity and empirical adequacy of the neutrality hypothesis, Hubbell argues that we should conclude that neutral processes relying upon species similarities are the dominant causal factors underlying diversity.

Not Null Enough 7 Proponents of the behavior-reading hypothesis in comparative psychology also hold that this hypothesis must be falsified before accepting the causal-alternative that chimpanzees mindread. As Penn and Povinelli (2007) write: in order to produce experimental evidence for an fToM [theory of mind function], one must first falsify the null hypothesis that the agents in question are simply using their normal, first-person cognitive state variables… One must, in other words, create experimental protocols that provide compelling evidence for the cognitive (i.e. causal) necessity of an fToM in addition to and distinct from the cognitive work that could have been performed without such a function. (734, emphasis original) They go on to write: From a scientific stance... we are only warranted in attributing an ms variable [the representation of another agent’s mental state] to the subject if we can specify why an fToM of some kind is computationally necessary in order to perform the given behavior (734). Thus, proponents of the behavior-reading hypothesis hold that researchers must reject this causal-null hypothesis of behavior-reading before accepting the mindreading hypothesis. Until they do so, that have failed to show that the mindreading hypothesis is necessary to explain the behavior of interest and thus have “no evidence for theory of mind in animals” (732). Moreover, until researchers reject the behavior-reading hypothesis, they should conclude with Penn and Povinelli (2007) that, “the available evidence suggests that chimpanzees, corvids and all other non-human animals only form representations and reason about observable features, relations and states of affairs from their own cognitive perspective” (737). We have presented the conclusions of the causal null strategy as all or nothing, either accepting or rejecting the causal-null or causal-alternative. However there is also a quantitative extension of the strategy that allows apportioning relative responsibility of the causal-null and causal-alternatives. This is what Hubbell means when he says that the neutrality hypothesis provides a 'test when, and to what extent' species differences are necessary to explain observed patterns. In other words, if the causal null hypothesis can fit the data to degree X, then the conclusion should be that responsibility for the pattern is X:1-X (causal-null:causal-alternative).

Not Null Enough 8

3. The Causal Null Hypothesis Strategy is Not Justified In section 2, we introduced the causal null hypothesis strategy and showed how it operates in ecology and comparative psychology. In this section, we determine whether the causal null hypothesis strategy justified. One plausible source of justification on which those engaged in the causal null hypothesis strategy might be drawing is the claim that the causal null hypothesis strategy is an extension of statistical null hypothesis testing. We show that this claim does not withstand criticism, however. Given this, and the fact that there is no other apparent source of justification for the causal null hypothesis strategy, we conclude that it is unjustified and thus should not be used as a strategy for deciding between causal hypotheses. Instead, causal-null and causal-alternative hypotheses should be treated as on a par. 3.1. Justification by Extension from Statistical Null Hypothesis Testing One plausible justification for the causal null hypothesis strategy is that it is an extension of statistical null hypothesis testing. The general rhetoric used by those engaged in the causal null hypothesis strategy suggests that they view it as such an extension. As we saw above, proponents of the neutral hypothesis in ecology and the behavior-reading hypothesis in comparative psychology refer to their hypotheses as “null hypotheses” and contrast it with an alternative hypothesis. In addition, they hold that one must reject this null hypothesis before accepting the alternative hypothesis. That is, they take the evaluation of the two hypotheses to be asymmetrical. Statistical hypothesis testing was developed in the 1920-1930s most importantly by Fisher, and Neyman and Pearson, under the frequentist interpretation of probability. While their theories are distinct and arguably mutually inconsistent, modern statistics has merged them and others into a toolbox of hybrid methods. The most relevant features of each theory are the number of hypotheses being tested and the conclusions that can be drawn from that test. Fisher's version of hypothesis testing tests one hypothesis, the null hypothesis, and the possible outcomes of the test are the rejection of the null hypothesis and failure to reject the null hypothesis. The null hypothesis can never be accepted.

Not Null Enough 9 The Neyman-Pearson version of hypothesis testing tests two hypotheses, the null hypothesis and the alternative hypothesis, and the possible outcomes of the test are the acceptance of one hypothesis and the rejection of the other. Here then the null hypothesis can be accepted. The causal null hypothesis strategy aligns more closely with Neyman-Pearson statistical null hypothesis in these two respects. Another way in which the causal null hypothesis strategy is like Neyman-Pearson testing is in its asymmetrical treatment of the null and alternative hypotheses. Proponents of the neutral and behavior-reading hypotheses argue that one must reject a causal-null before accepting the causal-alternative, but they do not hold that one must reject the causal-alternative before accepting the causal-null. Indeed, they hold that failure to reject the causal-null is sufficient reason to accept it. This asymmetrical criterion for evaluating hypotheses is reflected in NeymanPearson testing. As Godfrey-Smith (1994) points out, the null hypothesis in Neyman-Pearson testing “typically gets the benefit of the doubt” (280). In the context of Neyman-Pearson testing, a type I error occurs when one rejects a true null hypothesis and accepts a false alternative hypothesis, while a type II error occurs when one accepts a false null hypothesis and rejects a true alternative hypothesis. In other words, NeymanPearson testing favors accepting a false null hypothesis over a false alternative and rejecting a false alternative hypothesis over a false null because it allows for more type II errors than type I errors. As Godfrey-Smith observes, the Type I error rate, α, is usually set at some standard value such as 0.05, 0.01 or 0.001. There are no widely recognized standard values for β [the Type II error rate]. Instead is usually allowed to vary, within broad limits, as a consequence of the choice of α. Acceptable values of β can easily be around 0.3 or 0.4. It is rarely asked that β be as low as α” (281). In this way, the null hypothesis in Neyman-Pearson testing is privileged over the alternative hypothesis: ceteris paribus, it presumes that it is better to accept the null over the alternative and better to reject the alternative over the null. This asymmetry between null and alternative hypotheses is a practical feature of Neyman-Pearson hypothesis testing. That is, it is part of the practice of Neyman-Pearson testing that type I errors are more stringently policed than type II errors. The question of how this differential treatment of error is justified is a different question. One way to pose this

Not Null Enough 10 justificatory question is, “what justifies treating one hypothesis as the privileged null and the other as the alternative?” For Neyman, the answer to this question was determined on pragmatic grounds (Godfrey-Smith 1994). According to Neyman, we should treat as the null hypothesis the one that we believe will lead to fewer bad consequences, if accepted. Today, this justification is controversial. Moreover, there is another justification that is more widely accepted in the sciences. This alternative justification holds that the null hypothesis should be one that posits “no effect” or “nothing going on” (Godfrey-Smith, 1994, 282). Godfrey-Smith refers to this justification as a “semantic” justification because it concerns the content of the hypothesis, rather than merely the pragmatic implications of accepting it. As Godfrey-Smith explains, If hypotheses of "no effect" are nulls, then the asymmetry between α and β operates as the wielder of Occam's razor. The more serious error is multiplying effects beyond necessity, rather than not recognising enough effects. The asymmetry establishes a bias in favor of the simpler hypothesis (282). Thus, the semantic approach justifies the differential treatment of error found in NeymanPearson testing by appealing to simplicity. The simplicity invoked is of a particular kind, however. A hypothesis is allowed to occupy the privileged place of null if it is a hypothesis of no effect. This appeal to simplicity is motivated by the idea that it is worse to hold that something is going on in the world (when in fact it is not) than it is to hold that nothing is going on (when in fact something is). Ceteris paribus, we should prefer to accept the hypothesis holding that there is nothing going on over the hypothesis that posits a cause, process, or effect operating in the world. In summary, the causal null hypothesis strategy is analogous to Neyman-Pearson in several respects. First, it relies on two hypotheses, a null and an alternative. Second, it couples the rejection of one hypothesis with the acceptance of the other. Rejecting the alternative entails accepting the null and vice versa. Third, it privileges the null hypothesis over the alternative. Falsely accepting the null is viewed as less problematic, everything else equal, than falsely accepting the alternative. Given these similarities, it may seem that the causal null strategy is justified as an extension of statistical null hypothesis testing. We do not think that this is the case, however. In the next section, we argue that the causal null hypothesis strategy differs from Neyman-Pearson testing in several crucial respects.

Not Null Enough 11

3.2. The Causal Null Hypothesis Strategy and Neyman-Pearson Testing: Relevant Disanalogies The first respect in which the causal null hypothesis strategy and Neyman-Pearson testing differ is in the content of their null hypotheses. Neyman-Pearson testing relies on null hypotheses of no effect; however, neither of the causal-null hypotheses in our two cases studies takes such a form. A hypothesis of no effect is one in which there is no cause, process, or effect operating in the world. What this means exactly will depend on one’s alternative hypothesis. For example, in psychology, it is common to test whether subjects are sensitive to some feature of the world. In order to do this, researchers design two conditions (a test and a control) that vary only with respect to the feature under question. In this case, the alternative hypothesis is that subjects will differentially respond across these two conditions, while the null hypothesis is that there will be no detectable difference across conditions. For example, if both conditions require that subjects make a choice between two options, then, according to the null hypothesis, they will choose randomly between those two options. Neither of the causal null hypotheses in our case studies are hypotheses of no effect. Hubbell (2001) precisely defines what he means by "neutral", implicitly contrasting it with "null". He writes: Before proceeding, I need to be more precise about the meaning of neutrality as used in this book. Despite its moniker, the concept of neutrality actually has many meanings in the literature. To most people, the word neutral congers up the qualitative notion of "nothing going on." But exactly what people mean by this phrase often turns out to differ from one person to the next. I use neutral to describe the assumption of per capita ecological equivalence of all individuals in a trophically defined community. This is a very unrestrictive and permissive definition of neutrality because it does not preclude interesting biology from happening or complex ecological interactions from taking place among individuals. (6-7, original emphasis) The behavior-reading hypothesis is also more than a claim of no effect. It holds that chimpanzees possess a set of behavioral rules or heuristics that allow them to make behavioral predictions on the basis of observable situations. According to proponents of the behaviorreading hypothesis, the mindreading hypothesis makes a very different causal claim. For example, Penn and Povinelli (2013) characterize the difference between nonhuman behavior-

Not Null Enough 12 reading and human mindreading as follows: we believe that human and nonhuman animals possess a variety of mechanisms for recognizing those relations that are causally relevant to predicting the goal-directed behavior of other intentional agents. These heuristics enable both human and nonhuman animals to pick out the causally relevant relations between ‘what’ an agent is ‘looking’ at and how that agent is likely to behave in the near future… However, only humans cognize the higher-order analogical similarities between perceptually disparate behaviors and thus only humans possess the ability to reinterpret other agent’s goal-directed relations in terms of abstract mental state relations disembodied from any particular task context. (20) Although Penn and Povinelli do not state the type of heuristics that are employed in behaviorreading, they are clear that they take the mechanisms involved in behavior-reading to be causally distinct from those employed in mindreading. The hypothesis that nonhuman animals behaviorreading is not the mere claim that they lack “the ability to reinterpret other agent’s goal-directed relations in terms of abstract mental state relations” but also the claim that they possess a set of heuristics, which allows them to make the correct behavioral predictions even though they lack the reinterpretational ability that human mindreaders possess.3 In addition to not being a null of no effect, the causal null hypothesis differs from a Neyman-Pearson null hypothesis with respect to whether its content overlaps with that of the alternative hypothesis. In Neyman-Pearson hypothesis testing, the null and alternative hypotheses must be exclusive and exhaustive. This difference was pointed out in an earlier episode in ecology by Peter Sloep (1986). The hypotheses being exclusive and exhaustive is what allows the acceptance of one to entail the rejection of the other. However, the causal null hypotheses in our case studies are not mutually exclusive. Neutral hypotheses and selection hypotheses can be acting concurrently in a system to differing degrees. Some ecologists have

3 At times, such as in the quote above, Povinelli and colleagues suggest that both human and nonhuman animals behavior-read and that human mindreading is merely a reinterpretation of what is at bottom causally achieved through behavior reading. At other times, however, they explicitly deny that mindreading is causally superfluous in this way. For example, Penn and Povinelli deny that “an fToM has no functional, adaptive value or, worse, may by [sic] a figment of our folk psychological imagination,” arguing instead that, “it is clear to us that the ability to cognize the world from the cognitive perspective of another agent would provide an animal with enormous advantages over and above the ability to reason in terms of observable first-person relations alone” (741). Insofar as mindreading is not causally superfluous in humans, the cognitive mechanisms underlying human and nonhuman behavior (in those mindreading tasks in which the behavior of human and nonhuman subjects is the same) must be different.

Not Null Enough 13 called for the synthesis of neutral and non-neutral, selective mechanisms (for example Hubbell 2001 and Vellend 2010) and mixed models incorporating neutral and non-neutral mechanisms have been and continue to be developed (for example Tilman 2004). Similarly, the behaviorreading and mindreading hypotheses are not viewed as mutually exclusive. Indeed, proponents of the behavior-reading hypothesis hold that there is significant overlap between the two. For example, in the case of reasoning about perceptual states, both mindreaders and behavior-readers process information about the orientation of another agent’s gaze and make behavioral predictions on the basis of this information. The main difference between a mindreader and a behavior-reader is that the inferential process employed by the former involves reasoning about unobservable mental states, while the inferential process employed by the latter involves reasoning about observable features of the world alone. The causal null hypothesis strategy shares several features with Neyman-Pearson hypothesis testing. However, the differences between these two methods mean that the former is not a straightforward extension of the latter. Most importantly, however, the feature of NeymanPearson hypothesis testing that is invoked in order to justify the asymmetrical treatment of the null and alternative hypotheses is precisely the feature that the causal null hypothesis strategy lacks—namely, a null hypothesis of no effect. Given this, we suggest that this strategy be eschewed until its proponents develop a justification for its use.

4. A Call for Parity In the previous section, we argued that the causal null hypothesis might be justified as an extension of Neyman-Pearson testing. The main motivation for that proposal was the idea that if the causal null hypothesis strategy were relevantly similar to Neyman-Pearson testing, then it might be justified on the same grounds that Neyman-Pearson testing is justified. We also noted that a common justification of the asymmetrical treatment of hypotheses in Neyman-Pearson testing involved an appeal to simplicity. That is, in the context of Neyman-Pearson testing, a hypothesis is often justified as the null if it is simple in the sense of being a hypothesis of no effect. The question of whether this appeal to simplicity in Neyman-Pearson is justified is not

Not Null Enough 14 something that we addressed.4 Instead, we granted it for the sake of argument, allowing that, if the causal null hypothesis strategy was relevantly similar to Neyman-Pearson testing, then perhaps it could draw on the same justificatory arguments employed in statistical hypothesis testing. However, we showed that the causal null hypothesis strategy was not relevantly similar to Neyman-Pearson testing and thus could not avail itself to these justificatory grounds. As Sober (1994) writes, When a scientist uses the idea [of parsimony], it has meaning only because it is embedded in a very specific context of inquiry. Only because of a set of background assumptions does parsimony connect with plausibility in a particular research problem. What makes parsimony reasonable in one context therefore may have nothing in common with why it matters in another. (Sober, 1994, 77) Phrased in terms of our case, there are certain commitments that make privileging the simpler hypothesis in Neyman-Pearson reasonable to the community that employs it. However, these commitments are not shared with those who employ the causal null hypothesis strategy. Given this, and the fact that we see no other grounds for justifying the asymmetrical, privileged treatment that the causal nulls in our case studies receive, we hold that they should be treated on a par with those alternative hypotheses with which they are competing. In other words, we should reject the use of the causal null hypothesis strategy and view the neutral and behaviorreading hypotheses as on a par with the selection and mindreading hypotheses. At this point, proponents of the neutral and behavior-reading hypotheses might argue that there are general reasons for preferring the simpler hypotheses to complex ones. That is, they might hold that there is no need to draw on Neyman-Pearson testing or any other specific method of hypothesis evaluation in order to justify the claim that a simpler hypothesis should be preferred to a complex one. Instead, the causal null hypothesis strategy is justified on the general grounds that simplicity is an epistemic virtue. There are several ways to respond to this objection. The first is to challenge the idea that simplicity is a general epistemic virtue. For example, one could follow Sober (1994) and argue that simplicity is not a virtue that is a priori and subject-invariant, but rather one that must be evaluated on a case-by-case basis. Determining whether simplicity favors the neutral and 4 Indeed, the present authors disagree about whether an appeal to simplicity is justified in the context of Neyman-Pearson testing.

Not Null Enough 15 behavior-reading hypothesis, under this view, would require examining context-specific issues such as the likelihood of the data, given the hypotheses. Another response, which we pursue here, is to point out that the causal null hypothesis strategy claims something more than the fact that simplicity is an epistemic virtue. Instead, it claims that simplicity should trump all other epistemic virtues. Recall that the strategy holds that the null hypothesis is the best explanation for a phenomenon insofar as it cannot be rejected. This claim holds regardless of the epistemic virtues achieved by the alternative hypothesis. The alternative hypothesis might be predictively successful, fruitful, general, and more, but these virtues will not count in its favor unless researchers succeed in rejecting the causal null hypothesis. Richard Levins made the tradeoffs between different scientific virtues among different models and modeling practices well known to scientists and philosophers of science (Levins 1966). We enlist the idea of tradeoffs between different virtues to make our point. One could choose simplicity as the criterion for adjudicating between causal hypotheses, but one could also choose another virtue such as generality for this task. Levins discussed tradeoffs between generality, realism, and simplicity. But he was not being exhaustive. There are many more virtues than these and they all exist in complex tradeoff relations. The important issue is whether a given strategy, such as focusing on simplicity, is best suited to the end at hand. In our cases of the causal null hypothesis strategy, and we expect others, there is clearly a tradeoff between simplicity and generality. By generality we mean the ability of one hypothesis to account for a wide variety of phenomena. Take the case of ecology. There are many empirical patterns that can be explained by appealing to species differences. For example, correlations between species traits and abiotic conditions, such as the fact that certain types of trees tend to be found in the same temperature belts. These explanations are not challenged by the empirical success and simplicity of the neutral theory. The neutral theory challenges whether this type of explanation, which is successful in many areas of ecology, is necessary and correct for particular patterns of biodiversity and biogeography. It threatens to fracture the generality of the selection hypothesis by valuing simplicity. This fracturing means that particular explanations will be simpler, but that researchers will have to commit to more types of mechanisms operating in a given ecosystem. One of the usual goods of simplicity is that one does not have to pay the cost for many different mechanisms. Since selection mechanisms and species differences are already invoked as causal factors for other patterns, they are in some sense free for the taking and do not

Not Null Enough 16 reduce the overall cost. It is then an open possibility to reject the strategy of dividing up empirical patterns and independently developing the simplest models for each. It is also an open possibility to decide to group phenomena together and require common explanations. We do not think that these tradeoffs can be avoided and so do not mean to challenge the choice of simplicity by arguing that simplicity is always unwarranted. Rather, given these tradeoffs between virtues, the choice of one over another requires justification. It is the assumption of simplicity as a virtue that trumps all others—masked in the language of null hypothesis testing—that we challenge.

References Beatty, John. 1987. "Natural Selection and the Null Hypothesis." in The Latest on the Best: Essays on Evolution and Optimality. ed. John Dupre Call, J., and Tomasello, M. (2008). Does the chimpanzee have a theory of mind? 30 years later. TRENDS in Cognitive Sciences, 12(5), 187–192. Cooper, Gregory. 1993. "The Competition Controversy in Community Ecology." Biology and Philosophy Godfrey-Smith, Peter. 1994. "Of Nulls and Norms", In PSA Proceedings, Vol. 1994, Vol. One: pp. 280-290. Chicago: The University of Chicago Press Gotelli, Nicholas J. and Gary R. Graves. 1996. Null Models in Ecology. Washington: Smithsonian Institution Press Gotelli, Nicholas J. and Brian McGill. 2006. "Null vs Neutral Models: What's the Difference?" Ecography. 29(5): 793-800. Hubbell, Stephen. 2001. The Unified Neutral Theory of Biodiversity and Biogeography. Princeton University Press. Hubbell, Stephen. 2006. "Neutral theory and the evolution of ecological equivalence." Ecology 87(6): 1387-1398. Nitecki, Matthew H. and Antoni Hoffman 1987. Neutral Models in Biology. Oxford: Oxford University Press. Penn and Povinelli (2007). On the lack of evidence that non-human animals possess anything remotely resembling a ‘theory of mind.’ Phil. Trans. R. Soc B, 362, 731-744.

Not Null Enough 17

Penn, D. C., & Povinelli, D. J. (2009). On becoming approximately rational: The relational reinterpretation hypothesis. In S. Watanabe, A. P. Blaisdell, L. Huber, & A. Young (Eds.), Rational Animals, Irrational Humans. Tokyo, Japan: Keio University Press. Penn, D. C., & Povinelli, D. J. (In press). The comparative delusion: Beyond behavioristic and mentalistic explanations for nonhuman social cognition. In H. S. Terrace & J. Metcalfe (Eds.), Agency and joint attention. New York, NY: Oxford University Press. Penn, D. C., Holyoak, K. J., & Povinelli, DJ. (2008). Darwin’s mistake: Explaining the discontinuity between human and nonhuman minds. Behavioral and Brain Sciences, 31(2), 109– 178. Piccinini and Craver (2011). Integrating psychology and neuroscience: Functional analysis as mechanism sketches. Synthese, 183, 283-311. Povinelli, D. J., & Vonk, J. (2003). Chimpanzee minds: Suspiciously human? TRENDS in Cognitive Sciences, 7(4), 157–160. Povinelli, D. J., & Vonk, J. (2006). We don’t need a microscope to explore the chimpanzee’s mind. In S. Hurley, & M. Nudds (Eds.), Rational Animals? (pp. 385–412). New York, NY: Oxford University Press. Roughgarden, Jonathan. 1983. "Competition and Theory in Community Ecology" The American Naturalist. 122(5) Sloep, P.B. (1984) Null Hypotheses in Ecology: Towards the Dissolution of a Controversy. In: PSA 1986, Volume 1, A. Fine and P. Machamer (eds.), Philosophy of Science Association, East Lansing, Michigan, pp.307-314. Tilman,David. 2004. "Niche tradeoffs, neutrality, and community structure: A stochastic theory of resource competition, invasion, and community assembly". PNAS. 101:30 Tomasello, M., & Call, J. (2006). Do chimpanzees know what others see—or only what they are looking at? In S. Hurley, & M. Nudds (Eds.), Rational Animals? (pp. 371–384). New York, NY: Oxford University Press. Vellend, Mark. 2010. "Conceptual Synthesis in Community Ecology". The Quarterly Review of Biology. 85:2 Quinn, James F. and Arthur E. Dunham. 1983. "On Hypothesis Testing in Evolution and Ecology." The American Naturalist. 122(5)