People understand the world by constructing explanations for what they observe. It is thus important to identify the cognitive processes underlying these judgments. According to a recent proposal, everyday explanations are often constructed heuristically: Because people need to generate explanations on a moment-by-moment basis, they cannot perform an exhaustive search through the space of possible reasons, but may instead use the information that is most easily accessible in memory (Cimpian & Salomon 2014a, b). In the present research, we tested two key claims of this proposal that have so far not been investigated. First, we tested whether—as previously hypothesized—the information about an entity that is most accessible in memory tends to consist of inherent or intrinsic facts about that entity, rather than extrinsic (contextual, historical, etc.) facts about it (Studies 1 and 2). Second, we tested the implications of this difference in the memory accessibility of inherent versus extrinsic facts for the process of generating explanations: Does the fact that inherent facts are more accessible than relevant extrinsic facts give rise to an inherence bias in the content of the explanations generated (Studies 3 and 4)? The findings supported the proposal that everyday explanations are generated in part via a heuristic process that relies on easily accessible—and often inherent—information from memory.
Avoid common mistakes on your manuscript.
The ability to generate explanations allows people to abstract meaning from everyday experiences. From a young age, we make sense of the world by seeking to explain what we observe and what we hear from others (e.g., Carey, 1985; Gelman, 2003; Gopnik, 1998; Keil, 2006; Lombrozo, 2012; Wellman, 2011). What underlies the ability to explain? The literature on this topic has provided many insights into the workings of everyday explanations, including their typical structure (i.e., the “ingredients” that go into a satisfying explanation; e.g., Lombrozo, 2007; Lombrozo & Carey, 2006; Sloman, 2005) and their influence on learning (e.g., Chi, Bassok, Lewis, Reimann, & Glaser, 1989; Legare & Lombrozo, 2014; Lombrozo & Gwynne, 2014; Murphy & Allopenna, 1994).
A few predictions of this inherence heuristic proposal have already found support. For example, as with many heuristic processes, reliance on this explanatory shortcut is stronger in participants with lower cognitive resources or lower motivation for effortful thought (e.g., Hussak & Cimpian, 2015; Salomon & Cimpian, 2014; Sutherland & Cimpian, 2015). However, the research to date has provided no direct test of the key claims that (1) inherent information is highly accessible in memory and (2) due to this accessibility advantage, inherent information is overused when constructing explanations. The present studies tested these claims. Before detailing these studies, we briefly review the inherence heuristic proposal, which we then situate in the context of relevant prior research.
Heuristic reasoning is characterized by a reliance on information that comes to mind readily, even on occasions when this information is not appropriate—or at least not sufficient—for generating a correct judgment (e.g., Evans, 2006, 2008; Evans & Stanovich, 2013; Kahneman, 2011; Kahneman & Frederick, 2002; Shah & Oppenheimer, 2008; Stanovich & West, 2000). People tend to use such heuristic reasoning as a means of generating quick, approximate solutions to a wide range of complex problems that arise in daily life (for reviews, see Gilovich, Griffin, & Kahneman, 2002; Kahneman, 2011; Stanovich & West, 2000).
To be more precise, the distinguishing feature of heuristic reasoning is its reliance on effort-reducing steps such as substituting a simpler, easier-to-access attribute (e.g., present mood) for a more complex one (e.g., general life satisfaction) or considering only the subset of attributes that immediately come to mind to form a judgment that would otherwise require integrating more information (e.g., Kahneman & Frederick, 2002; Schwarz & Clore, 1983; Shah & Oppenheimer, 2008). Most of the “classic” heuristics described in the judgment and decision making literature, such as availability (A. Tversky & Kahneman, 1973, 1974), representativeness (Kahneman & Tversky, 1973; A. Tversky & Kahneman, 1974), and anchoring and adjustment (A. Tversky & Kahneman, 1974), are nowadays understood as particular instantiations of such effort-reducing processes (Kahneman, 2011). For example, when people erroneously judge the probability that Linda is a feminist bank teller to be higher than the probability she is a bank teller (the representativeness heuristic; Tversky & Kahneman, 1983), they are—without realizing it—substituting a simple judgment (Linda’s similarity to the stereotypical feminist) for a more complicated one concerning probabilities. This spontaneous tendency to fall back on simple, easy-to-access information is a basic feature of human reasoning.
Building on this dual-process (heuristic vs. analytic) framework, the inherence heuristic account (Cimpian & Salomon, 2014a, b) makes a further claim about the constraints imposed by memory retrieval on the process of generating explanations. The claim is that the most accessible facts—those that are likely to be called up from memory most often and most quickly when constructing explanations—are inherent or intrinsic facts about the entities in the explanandum (e.g., men’s and women’s presumed cognitive abilities).
Simply put, inherent or intrinsic facts about a thing are facts that a reasoner represents as being “entirely about that thing,” whereas extrinsic facts involve other entities in some way (e.g., Lewis, 1983, p. 197; see also Barr & Caplan, 1987; Caplan & Barr, 1991; Weatherson & Marshall, 2014). For example, the claim that women generally lack genius-level intelligence posits an inherent fact about them. In contrast, the claim that they are subject to societal stereotypes posits an extrinsic fact about them—a fact that involves entities other than women themselves.
A simple rule of thumb might help readers distinguish between inherent and extrinsic information: Inherent facts are those that, if changed, would lead to a change in the entity itself (Weatherson & Marshall, 2014). Changing a person’s cognitive abilities would induce an actual change in the person him- or herself, but changing cultural stereotypes about the person’s group would not. Of course, reasoners need not know this rule of thumb, nor do they need to distinguish between inherent and extrinsic facts while generating explanations. The inherence heuristic claim is not that people purposely select for their explanations inherent facts about the entities in the explanandum. Rather, these facts are unwittingly used to construct heuristic explanations because, we propose, they are so readily retrieved.
Why might inherent features of the entities in the explanandum be particularly accessible in memory? To begin with, these entities are in the focus of attention when the search for an explanation is triggered, so they are likely to serve as the first, and perhaps only, retrieval cues (e.g., Legrenzi, Girotto, & Johnson-Laird, 1993; Spiller, 2011; Weber et al., 2007). This attentional spotlight on the entities in the explanandum (e.g., women) may prevent other relevant entities (e.g., the historical and societal context) from coming to mind and may thereby lower the probability of retrieving extrinsic facts that relate the explanandum to phenomena beyond it.
Extrinsic facts may also be slower to come to mind because they tend to be less prominent in our interactions with the relevant entities. For instance, the influence that stereotypes or other societal constraints exert on women’s behavior is seldom apparent to a naive observer, resulting in a weaker associative link between such extrinsic factors and the concept woman in semantic memory. Even when they are noticed, extrinsic forces may be more cumbersome to describe in language, whereas inherent ones can often be expressed in a few words (e.g., via generic statements about what the relevant entities are or have, as in “Women aren’t brilliant”; e.g., Cimpian, Brandone, & Gelman, 2010). Any such difference in how complex it is to talk about inherent and extrinsic facts—and thus in how often they are actually talked about—could contribute to an accessibility difference between them.
This accessibility bias favoring the retrieval of inherent information about the entities in the explanandum is in turn hypothesized to constrain the content of the explanations generated. The facts retrieved from memory serve as building blocks for an explanation. Thus, if inherent facts are overrepresented among these building blocks (relative to extrinsic facts that would otherwise be as relevant), it is likely that they will also be overused in the explanations ultimately generated. Footnote 3
To clarify, inherence is distally, rather than proximally, related to accessibility. Inherent facts are accessible because of the specific ways in which they are represented in semantic memory, such as their greater associative relatedness with a target concept, their lower relational complexity, and so on (see foregoing discussion). When these mediating variables are taken into account, inherence may not, in and of itself, explain any unique variance in accessibility. Note, however, that identifying the specific variables that mediate the relationship between inherence and accessibility is beyond the scope of the present work. Our aim here is more modest: to test whether facts that are easily retrieved also tend to be inherent rather than extrinsic, and whether this accessibility difference in turn biases the content of heuristic explanations.
Before describing the four studies we conducted to test these claims, we briefly review some of the prior evidence on the links between memory and explanation.
Memory processes have largely been overlooked in the recent literature on explanation. Footnote 4 The most direct insights into the role of memory processes have come not from the literature on explanation per se, but from the neighboring literature on hypothesis generation. In particular, Thomas, Dougherty, and their colleagues recently proposed a comprehensive model of hypothesis generation (termed HyGene) that covers the various stages of this process, including the retrieval of plausible hypotheses from long-term memory and the evaluation of these hypotheses in working memory (e.g., Thomas, Dougherty, & Buttaccio, 2014; Thomas, Dougherty, Sprenger, & Harbison, 2008; see also Mehlhorn, Taatgen, Lebiere, & Krems, 2011). This elegant framework is compatible in many ways with our proposal of an inherence heuristic in explanation. For instance, the evidence reviewed by Thomas, Dougherty, and colleagues (2014, 2008) suggests that people do not perform an exhaustive memory search, but rather retrieve only a small set of hypotheses. In addition, this subset of retrieved hypotheses are those that have been most frequently associated with the explanandum in past experience and are thus most easily activated from long-term memory (e.g., a sore right abdomen was often found to be caused by appendicitis). In other words, HyGene posits that people tend to retrieve those hypotheses that are most accessible.
Although HyGene and the inherence heuristic framework share these basic mechanistic assumptions (i.e., that reasoners rely on a small set of easily retrieved explanatory facts), they also differ in a number of crucial respects. Most importantly, unlike HyGene, the inherence heuristic model makes a claim about the content of the information that is most accessible in memory, proposing that this information often consists of inherent facts about the entities in the explanandum. Moreover, the inherence heuristic model does not assume that reasoners retrieve fully formed hypotheses or explanations from long-term memory. Although this is a reasonable simplification that works well when modeling diagnostic reasoning in certain contexts (e.g., medical diagnosis), many of the observations that people seek to explain in everyday life are ones for which they do not have ready-made explanations stored in memory (e.g., why more men than women have been recognized as geniuses). As a consequence, people are often forced to retrieve various facts that may be relevant to an explanation but may not initially constitute an explanation in and of themselves.
In summary, although the HyGene model and the inherence heuristic model share several foundational assumptions about the ways in which memory constrains explanation, they also diverge in a number of equally foundational respects.
The proposal that inherent information is highly accessible in memory, and therefore likely to dominate the content of spontaneous explanations, licensed two specific predictions:
To provide a comprehensive test of these predictions, the stimuli for the present studies were selected from a range of domains (social conventions [Studies 3a and 4], historical events [Study 3b], and scientific phenomena [Study 3c]) and were varied systematically along a number of dimensions (e.g., concreteness [Study 2]).
In Study 1, we investigated the prediction that inherent facts would be frequently called to mind when considering a given entity (Prediction 1). Preliminary evidence for the accessibility of inherent facts can be found in work on concepts and semantic memory—particularly in studies using feature-listing and -verification paradigms. Although none of these prior studies coded features for their inherence (as defined here), a close examination of their materials and results suggests that inherent features tend to be (1) listed often and quickly when people are prompted with a concept (e.g., Ashcraft, 1978; Hampton, 1979, 1981; McRae, Cree, Seidenberg, & McNorgan, 2005; Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1976; B. Tversky & Hemenway, 1984; Vinson & Vigliocco, 2008; Wu & Barsalou, 2009) and (2) verified quickly as being true of that concept (e.g., Collins & Quillian, 1969; Jorgensen & Kintsch, 1973). However, because these studies did not actually classify features on the basis of whether they were inherent, the support they provide for our prediction is necessarily tentative.
To provide a direct test of the accessibility of inherent information, we used an existing feature-generation dataset (McRae et al., 2005) and tested for a positive relationship between the frequency with which a feature occurred in this dataset and its likelihood of being inherent.
The participants in this and all subsequent studies were recruited online via Amazon’s Mechanical Turk platform. Study 1 involved two independent samples of participants: one consisting of 85 participants (41 male, 44 female), and another consisting of 91 participants (37 male, 54 female). The procedures were identical across these two samples; only the items differed (see below).
The data totaled 24,880 ratings, which exceeds the typical power recommendations for multilevel models (e.g., Snijders, 2005; West, Ryu, Kwok, & Cham, 2011). Power simulations for all models were conducted using the SIMR package in R (Green & MacLeod, 2016), and each model was determined to have at least 80% power to detect the predicted effect.
To examine whether a fact’s accessibility is related to its inherence, we presented our participants with facts from an existing dataset (McRae et al., 2005) and asked them to code these facts as either inherent or noninherent. Importantly, McRae et al. also recorded the frequency of each fact in their dataset (e.g., how often “has keys” was generated for the item accordion), enabling us to examine quantitatively the relationship between accessibility (operationalized as frequency of mentions) and inherence.
This dataset lists the facts that were generated about 514 items, along with the frequency of each fact. Each item was presented to exactly 30 participants, so the maximum frequency was 30. A fact had to have been produced by at least five participants in order to be included in the dataset, so the minimum frequency was 5. The elicitation instructions were worded as follows: “On the following pages, there are words that each denote a concept, with each being followed by 14 blank lines. Please fill in as many of these lines as you can with properties of the concept to which the word refers” (p. 556). The participants were then given a few examples of properties they might generate, which included both inherent (e.g., “physical properties”) and extrinsic (e.g., “things that the concept is related to,” “where it comes from”) examples. McRae et al.’s instructions were thus neutral with respect to the inherent-versus-extrinsic dimension.
We randomly sampled two nonoverlapping sets of 100 items (and the facts that accompanied them) from the 514 items in McRae et al.’s dataset and administered them to our two samples of participants. The participants in Samples 1 and 2 rated 1,414 and 1,404 facts, respectively (see the Appendix Table 13 for a list of the items).
The study was administered using the Qualtrics survey platform (Qualtrics, Provo, UT). The procedure consisted of two phases: a training phase and a coding phase.
Participants first read a definition of inherence (“an inherent feature of an object is one that refers to the object itself, without making reference to external objects or forces, or historical events”) and saw several examples of inherent and noninherent facts. Participants were then given four practice trials that required them to decide whether or not a fact was inherent. A “not sure” option was also available, both here and in the coding phase (see below). Two of the practice facts were inherent, and two were noninherent. After each practice trial, participants saw the correct answer as well as the rationale for it. Participants who did not code the last two practice facts correctly (“used in Nike campaigns” [noninherent] and “is round” [inherent] about a soccer ball) were omitted from further analyses. Fourteen participants (beyond those reported above) were excluded on the basis of this criterion in Sample 1, and ten in Sample 2. Footnote 5
After training, each participant was presented with ten items randomly selected from the set of 100 we had pulled from McRae et al.’s (2005) dataset. Participants coded all facts that accompanied those ten items as inherent or noninherent, without receiving any feedback. The order of the features was randomized. To ease the memory load, we reiterated on each trial what is meant by an inherent feature. Each item’s facts were coded by 6–11 participants (medians = 8 in Sample 1 and 9 in Sample 2).
In all studies, the data were analyzed with mixed-effects models using the lme4 package in R (Bates, 2007). Unless otherwise noted, these mixed-effects models included random intercepts and slopes for both items and subjects. Also, because no analytical (i.e., precise) methods exist for calculating degrees of freedom—and thus p values—for mixed-effects models (Bates, Mächler, Bolker, & Walker, 2014), we bootstrapped 95% confidence intervals (CIs) for all our coefficients and used whether these CIs crossed 0 to determine significance.
The raw data and analytic syntax for this and all subsequent studies are available on the Open Science Framework (https://osf.io/pju7r/?view_only=ad8f1a0e51524ffaa5cf0553093c977f).
Participants judged 59.5% of the facts to be inherent and 36.1% to be noninherent. The predominance of inherent facts is broadly consistent with our claim that such facts are easily accessible. Also of interest, participants were seldom unsure of how to code a fact (4.4% of cases). Agreement among participants was high: On average, 80.5% of participants agreed about whether or not a fact was inherent, despite the minimal instructions.
To test our first prediction—that frequently generated facts are likely to be inherent—we modeled whether a fact was judged as inherent on the basis of the frequency with which it occurred in McRae et al.’s (2005) dataset (range = 5 to 30; median = 9). As predicted, facts that occurred more often in McRae et al.’s dataset were also significantly more likely to be judged as inherent by the participants in our study, b = 0.03, SE = 0.01, bootstrapped 95% CI: [0.01, 0.05] (see Table 1 for the full model). Concretely, this result indicated that, with each additional mention by McRae et al.’s participants, a fact’s odds of being judged as inherent (vs. noninherent) increased by 3.0%. Alternatively, a 1 SD increase in frequency was accompanied by a 16.7% increase in the odds of being judged as inherent.
Table 11 Mixed-effects logistic regression predicting the likelihood of generating an explanation based on condition, explanation content, and the interaction of these two variables (Study 4)
To clarify, the fact that participants in the nonspeeded condition relied on inherent and extrinsic information equally often doesn’t necessarily mean that they overcame the inherence bias in explanation. The term “bias” signifies a departure from the normatively correct response. While we cannot be certain of the correct answers to the ten “why?” questions in this study, chances are that virtually all of them involve an extrinsic element; after all, these questions concern social conventions, such as that engagement rings have diamonds or that women wear long hair, that have likely been shaped by countless historical and cultural forces. Thus, we suspect that participants’ explanations in the nonspeeded condition (even though balanced between inherent and extrinsic information) were still systematically biased relative to the correct explanations, in that they probably underused extrinsic information and overused inherent information.
Our hypothesis concerning the differential accessibility of inherent and extrinsic information during the process of explaining makes an additional pair of predictions: First, the longer that participants take to think about their response, the more likely they might be to retrieve relevant but less-accessible extrinsic information, and thus the more likely they might be to generate an explanation containing such information. Second, deliberation time should not be related to the likelihood of relying on inherent information; this information comes to mind quickly (see Studies 1 and 2) and should thus be available to the explainer throughout the deliberation period.
We explored these predictions with the data from the nonspeeded condition only, largely because the variability in response times in the speeded condition (in which everyone was cut off after 15 s) was minimal. Specifically, we conducted two mixed-effects logistic regressions that modeled the probability that an explanation would contain inherent and, separately, extrinsic information as a function of the number of seconds participants took to generate that explanation. Footnote 10 Both the intercept and the slope of this relationship were allowed to vary randomly across subjects and across items. As predicted, the longer that participants spent thinking about their response on a certain trial, the more likely they were to generate an explanation that invoked extrinsic information, b = 0.018, SE = 0.004, bootstrapped 95% CI: [0.010, 0.026] (see Table 12 for the full results). More concretely, each additional 10 s spent thinking about an explanation was associated with a 17.8% rise in the odds that the explanation would contain extrinsic information. In contrast, latencies were not related to the likelihood of generating an explanation containing inherent information, b = −0.002, SE = 0.005, bootstrapped 95% CI: [−0.011, 0.008]. Because inherent facts are easily accessible, they are equally likely to be included in explanations with short and long latencies.
Table 12 Mixed-effects logistic regression predicting the likelihood of generating an explanation containing extrinsic information as a function of explanation latency in the nonspeeded condition of Study 4
The findings from Study 4 provide evidence for an inherence heuristic in explanation. Explanations containing inherent information about the entities in the explanandum were more prevalent when participants generated answers quickly, which suggests that inherent information is easily retrieved and used to construct off-the-cuff explanatory intuitions. In contrast, when participants were allowed (and took) ample time to come up with an explanation, extrinsic information was retrieved as well and was incorporated into their explanatory judgments. Thus, when reasoners are motivated to arrive at an accurate explanation—and also have the cognitive resources to dedicate to this task—they can partly overcome the strong inherence bias that characterizes quicker explanations.
To make sense of everyday life, people must retrieve relevant information from memory to construct explanations. Although the role of memory in the process of generating explanations is widely acknowledged, little research has actually investigated how the dynamics of memory retrieval shape the explanations that people come up with (e.g., Thomas et al., 2008). The inherence heuristic account (Cimpian & Salomon, 2014a, b) elucidates one potential constraint that memory processes impose on explanation. Specifically, this account proposes that everyday explanations tend to be constructed on the basis of information that is most easily retrieved from memory, which often concerns the inherent, constitutive features of the entities involved. Although previous research on the inherence heuristic has already provided support for several predictions of this account (e.g., Cimpian & Steinberg, 2014; Hussak & Cimpian, 2015; Salomon & Cimpian, 2014), two key claims have (until now) remained untested: namely, that (1) inherent facts are generally more accessible than noninherent ones, and (2) this accessibility differential gives rise to an inherence bias in the content of spontaneous explanations. The present research has provided evidence for these two claims across four studies that employed diverse methodologies: Not only did inherent facts come to mind often (Study 1) and quickly (Study 2) for a wide range of entities, but these facts were also readily retrieved in the service of explanation across a wide range of explananda (Studies 3 and 4). Together, these results support the proposal that the explanations generated during ordinary cognitive activity overuse inherent facts about the entities in the explanandum because these facts are more easily accessible in memory than equally relevant extrinsic facts.
Given that explanations form the basis for how people understand much of the world, a bias in the process by which these judgments are generated is likely to have far-reaching consequences. To illustrate, this inherence bias in explanation has been shown to shape sociopolitical reasoning—people’s understanding of, and attitudes toward, the workings of their society (Hussak & Cimpian, 2015). Consider, for instance, a pervasive feature of human societies: status hierarchies between groups (e.g., that Group A is richer than Group B). When reasoning about such hierarchies, relying on accessible inherent information about the groups involved often results in their disparities appearing natural and legitimate (e.g., Group A is richer because they are inherently smarter, so they deserve their higher status). As a result, people tend to accept the status quo more readily than they might otherwise. Because of the immediate accessibility of inherent information, people likewise overestimate the stability and immutability of current societal structures (Cimpian & Steinberg, 2014). If Group A is understood to have high status because of their smarts, it may also appear to a reasoner that this group has always been, and will always be, at the top of the socioeconomic ladder—and perhaps even that they ought to be there (Tworek & Cimpian, 2016). These considerations illustrate the profound impact that the inherence bias in explanation is likely to have on how individuals understand and interact with the world.
The present findings raise several interesting questions and potential directions for future work. In particular, it would be worthwhile to investigate what moderates the inherence bias in explanation. We consider two possible moderators here. First, a potential moderator is the scope of the explanandum (narrow vs. broad). Information about circumstances, external constraints, and so forth, may be more easily called to mind when explaining a particular event (e.g., why Sally got a bad grade on her math test) than when explaining a broad pattern (e.g., why women tend to underperform in math; Cimpian & Markman, 2009, 2011). If extrinsic information is indeed more accessible when people are explaining specific events, the inherence bias may be somewhat attenuated for these explananda (Cimpian & Salomon, 2014b). However, as has been suggested by the literature on the correspondence bias (the tendency to overattribute others’ behaviors to dispositional factors; e.g., Gilbert & Malone, 1995) and by the results of Studies 3b and 3c (in which many of the stimuli concerned specific events), inherent information often remains highly accessible even when explaining individual cases/events, and as a result, this information is still overrepresented in the explanations generated. Consistent with this possibility, ongoing work in our lab suggests that the correspondence bias relies on the same cognitive mechanisms as the inherence heuristic investigated here, and thus may be a particular instantiation of this phenomenon (Storage & Cimpian, 2015).
Second, cultural factors might also moderate the memory accessibility of inherent facts, and thus their likelihood of being used in explanations. For instance, research on the correspondence bias has found that in East Asian cultures situational attributions (i.e., extrinsic explanations) are more common than dispositional attributions (i.e., inherent explanations) when interpreting behavior (e.g., Choi, Nisbett, & Norenzayan, 1999; Morris & Peng, 1994). This phenomenon suggests that culture-specific factors might influence how likely one is to retrieve and rely on inherent information when constructing an explanation. To speculate, it seems unlikely to us that such cultural influences would completely undo the accessibility advantage for inherent facts, which is more or less a function of how our cognitive systems are set up; rather, culture may simply attenuate this basic advantage for inherent information. At this point, however, this is an open empirical question.
Explanation is a vital cognitive tool by which we abstract meaning from our experiences. The present research contributes to theories of explanation by investigating the role of memory retrieval in the process of constructing explanations. Consistent with key (but previously untested) assumptions of the inherence heuristic account (Cimpian & Salomon, 2014a, b), the four studies reported here suggest that inherent facts are more easily accessible in memory than noninherent facts, and that this accessibility difference in turn biases the content of everyday explanations toward inherence. This bias has deep, pervasive consequences for how people understand the world.
To be clear, we do not claim that all explanations involve effort-saving shortcuts. The answers to some “why?” questions will undoubtedly be generated via other processes. For example, there are entire classes of explananda (e.g., the interactions of physical objects) about which we reason via skeletal causal-explanatory frameworks that are conserved across ontogeny or even phylogeny (Spelke & Kinzler, 2007). Heuristic explanations are more likely in circumstances in which previous knowledge doesn’t supply preformulated answers, and thus reasoners would have to engage in effortful processing to construct an answer.
If extrinsic facts are more relationally complex, that might also be why they are more difficult to put into words.
The question of how a reasoner determines which of the retrieved facts constitute plausible explanations is outside the scope of the inherence heuristic account, which focuses specifically on the memory and attentional constraints on the process of generating explanations. However, other work provides some insight into this question (e.g., research on explanatory preferences; Lombrozo, 2007) and can be integrated with the inherence heuristic proposal for a more complete account of explanation.
Some evidence exists on the reverse causal link: namely, the influence of explanations on subsequent memory. In brief, explanations boost memory for the specific aspects of the explananda that are featured in, or otherwise relevant to, the explanations themselves (e.g., Walker, Lombrozo, Legare, & Gopnik, 2014)—consistent with the depth-of-processing effects widely documented in the literature (e.g., Craik & Lockhart, 1972). However, explanations also seem to decrease subsequent memory for aspects of the explananda that do not end up being relevant to the formulation of these explanations (e.g., Legare & Lombrozo, 2014; Walker et al., 2014; Williams & Lombrozo, 2010).
No significant results became nonsignificant when we included these participants in our analyses. In fact, the results with and without these participants were nearly identical.
Inherence was also positively related to accessibility in a model that did not adjust for plausibility, b = 0.26, SE = 0.05, bootstrapped 95% CI: [0.15, 0.35].
Inherence was also positively related to accessibility in a model that did not adjust for plausibility, b = 0.11, SE = 0.04, bootstrapped 95% CI: [0.03, 0.20].
The magnitude of the relationship between inherence and accessibility was attenuated in this study relative to Study 3a. One possible reason for the weaker relationship is our use of historical events as explananda. However, it is important to also keep in mind that participants were given twice as long as in Study 3a to think about each question (20 vs. 10 s). The more time people spend retrieving information from memory, the more likely they may be to retrieve less-accessible extrinsic information (see Study 4 for empirical evidence on this point), which might also explain why inherent information was not as dominant in participants’ retrieval in this study.
Inherence was also positively related to accessibility in a model that did not adjust for plausibility, b = 0.20, SE = 0.03, bootstrapped 95% CI: [0.14, 0.26].
This latency variable was positively skewed. Applying a natural logarithm transformation reduced the skew substantially, but the results of the mixed-effects logistic regressions with the transformed and untransformed latency variables showed the same pattern of significant relationships. For simplicity, we report the results with the untransformed variable.
We are grateful to our participants, to the Cognitive Development Lab at the University of Illinois for research assistance and helpful discussion, and to Zach Horne, John Hummel, Sunny Khemlani, and Brian Ross for their insightful comments on a previous draft of the manuscript. This research was supported by a Graduate Research Fellowship from the National Science Foundation (L.J.H.) and by research funds from the University of Illinois and New York University (A.C.).
