Perceived danger associated with a property modulates cross category generalization

Abstract

Identifying projectable predicates is a key issue in understanding inductive inference. It is proposed that looking into the evolutionary psychology literature for adaptive properties may be one useful approach. One hypothesis that emerges from this literature is that properties that signal danger or harm should be more salient than properties that do not. Two studies are carried out to test this hypothesis. In study 1 participants were presented with a scenario involving the discovery of novel animals, for which there was incomplete information. Three types of properties (a harmful property, a neutral property, a beneficial property) were associated with animals in one (base) category and participants were asked to indicate strength of generalization of the property to a target within the category, and to a target across to another category. In the second experiment, the procedure was repeated, but in addition, subjects were also explicitly asked to indicate whether the base and target belonged to the same or different categories. Study 1 showed that the harmful property was more projectable compared to the beneficial and neutral properties. Study 2 reconfirmed this and further showed that it also promoted excessive generalization across categories. The results suggest that examination of adaptations identified by evolutionary psychologists may be a useful source of insight in the study of inductive inference.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11571-022-09793-3.

Introduction

The central issue in inductive inference is the differential projection of predicates (Goodman 1955). For example, if we dig up a mammoth from the permafrost in Siberia, and discover sage, thyme, and fir cones in its stomach, we will happily draw some conclusions about the dietary habits of Siberian mammoths. However, the mammoth we found also has some broken bones. In fact, as far as we know, all mammoths discovered to date in Siberia, have had some broken bones. But we do not draw the conclusion that all mammoths had broken bones. Notice, the evidence for both conclusions is identical. But the first seems reasonable, the second does not. In this example, we are happy to accept dietary habits as a projectable property, but not the property of having broken bones. So, not all properties are created equal when it comes to inference. Why are some predicates more projectable than others? There is no satisfying answer to this question. However, psychological studies have been identifying factors such as similarity, causation, and centrality as relevant in determining projectability.

Consider the following examples (Heit and Rubinstein 1994):

• (A) Hawks have a liver with two chambers;

  \ Chickens have a liver with two chambers.

• (B) Hawks have a liver with two chambers;

  \ Tigers have a liver with two chambers.

A is considered a stronger inference than B, with the explanation that Hawks are more similar to chickens than they are to tigers, thus the property is more likely to generalize to chickens than tigers.

But now consider the following two examples (Heit and Rubinstein 1994):

• (C) Hawks prefer to feed at night;

  \ Tigers prefer to feed at night.

• (D) Hawks prefer to feed at night;

  \ Chickens prefer to feed at night.

In this example, C is considered a stronger inference than D. The explanation is that, inference strength is not simply a matter of similarity but also a function of the underlying causal stories that subjects believe or assume. In the former example, subjects are focusing on the biological properties of chickens, hawks and tigers and concluding that chickens and hawks are more closely related than hawks and tigers, in terms of anatomy. In the latter example, subjects are focusing on the fact that hawks and tigers are hunters and carnivores, while chickens are not.

A third proposal revolves around centrality of properties, defined as follows (Hadjichristidis et al. 2004, p. 48): “All else being equal, features that are central for a concept are more projectable to other concepts than features that are less central.” For example a hormone that controls many functions in a dolphin (i.e. is a central property of dolphins) will be more projectable to seals than a hormone that controls few functions (i.e. less central hormone). But the influence of centrality is a function of shared structural features of the two animals or categories. If there are a few shared features (for example in mapping from a dolphin to a frog or a banana) it is less likely that the central properties of the base category retain their central status in the target category.

In this study we propose and test another possible factor that may influence projectability of properties. Evolutionary psychologists have argued that reasoning is shaped by adaptive mechanisms that facilitated survival and reproduction in the Plesistocene environment (Cosmides 1989; Tooby and Cosmides 2003; Pinker 2005; Buss 2012). Identifying and avoiding harm is one of the most basic mechanisms of survival. If we see a tiger, we run. If we believe there is a tiger in a certain area, we will avoid that area, even if the probability of the tiger being present is very small. On such an account, we should be much more sensitive to properties that pose a risk or threat to us than properties that do not. This should impact the saliency and projection of these properties.

This has been demonstrated in the case of deduction with the Wason Card Selection Task. Cosmides (1989) convincingly showed that the content effect in the Wason card selection task, was not simply a function of familiar content, but specific to knowledge related to catching cheaters. She further argued cheater detection was an adaptive, built-in instinct necessary for survival in social groups, and not part of a general purpose reasoning system. Here we raise the issue whether this line of thought can help us identify more projectable predicates from less projectable predicates in inductive inferences.

Suppose we have learned about two new animals. One is benign and the other is dangerous. On the evolutionary account, in the case of the dangerous animal, we should avoid not just this particular animal, but all related animals as a safety precaution. That is, we will more readily generalize the dangerous property to a broader category of animals than the benign property. Not only does perceived danger mark a projectable property, it results in over projection. Such over generalizations are often encapsulated in Proverbs: “Once bitten by a snake, one shies at a coiled rope for the next ten years” (in Chinese, “”1).

To test this hypothesis we carried out two studies using a category induction task based on taxonomic relations. In the first study we presented participants with a scenario involving the discovery of novel animals, for which we have incomplete information. The animals belonged to two main subcategories, with some unaccounted variation within categories. Three types of properties (a harmful property, a neutral property, a beneficial property) were associated with animals in one (base) category and participants were asked to indicate strength of generalization of the property within the base category and to another (target) category. In the second experiment, this procedure was repeated, but in addition, subjects were also explicitly asked to indicate whether the base and target animals belonged to the same or different categories. We hypothesized that a property that signaled danger or harm would be more salient than a beneficial or neutral property and result in over-generalization across categories.

Study 1

Method

Participants

One hundred and eight paid healthy undergraduate students (freshmen and sophomores), recruited from Southwest Minzu University, participated in this experiment. Equal numbers were randomly assigned to three property groups, with 36 subjects (19 female; 19.78 ± 0.96 years) to the neutral property group, 36 subjects (17 female; 19.90 ± 1.28 years) to the beneficial property group, and 36 subjects (20 female; 20.03 ± 1.45 years) to the harmful property group. All subjects were right-handed and had normal or corrected-to-normal vision. None of the subjects reported any history of neurological or psychiatric diseases. Written informed consent was obtained from each participant.

Materials and procedure

The task involved immersion in the following scenario about the discovery of some new animals on a South Pacific island and subsequent inferences:

A biologist visited an island in the South Pacific Ocean, where he discovered some novel biological species. He classified them according to the following four characteristics: (i) antenna (straight or curved), (ii) body shape (fat or thin), (iii) body markings (large or small spots, or stripes) and (iv) tail (forked or not), and provided the diagrams in Fig. 1a–c.

Fig. 1.

a Mossiva-Gubee. Notice the variability among the members, even though they are not further differentiated by subcategories. b Lamoodu-Poogusa. Notice the variability among the members, even though they are not further differentiated by subcategories. c Pigi-Poogusa. d The implicit concept hierarchy based on the given experimental scenario

He also classified animals into two major categories Gubee and Poogusa. Gubee further contains the subcategory mossiva-Gubee which includes the following varied members identified in Fig. 1a. Poogusa is further subdivided into two categories including lamoodu-Poogusa identified in Fig. 1b and pigi-Poogusa identified in Fig. 1c. The records of the biologist are incomplete. We know some of the properties of mossiva-Gubee (Fig. 1a). Beyond this we will have to speculate by drawing inferences.

Thus, a concept hierarchy as shown in Fig. 1d was implicitly offered to subjects (but not given explicitly). Given certain properties of mossiva-Gubee in Fig. 1a, participants were required to make inferences about the presence of these properties, both, in the other animals within the category of mossiva-Gubee, and to the animals in the Poogusa categories in Fig. 1b and c. This results in a 3 × 2 mixed factorial design as illustrated in Table 1. The first factor (between subjects) manipulates the nature of the property. This factor has the following three levels: a harmful property (HA in abbreviation; e.g., is very poisonous and aggressive), a neutral property (NE in abbreviation; e.g., produces the hormone Tg3-HWb), and a beneficial property (BE in abbreviation; e.g., produces substance that can treat cobra venom and save lives). The second factor (within subjects) is categorical distance, either within the mossiva-Gubee category or across to the Poogusa category. Each condition results in 3 trials by varying 1, 2, or 3 characteristics as shown in the supplementary material.

Table 1.

Illustration of the factorial design in Study 1 as well as example tasks

The task for subjects is to make a judgment of the strength of projectability of each type of property. That is, given that a base category has the property X, subjects are instructed to evaluate the likelihood that the target category also has the property X on a 21 point Likert scale. The task was administered with paper and pencil.

Results

Figure 2 displays the inductive strength as a function of property type (i.e., BE, NE and HA) and categorical distance (within category versus across category) between the base category and the target category.

Fig. 2.

Measure of inductive strength (on a 21 point Likert scale) as a function of task type (i.e., BE, NE and HA) and categorical distance between the premise category and the conclusion category (i.e., within and across categories) for Study 1. HA = Harmful; NE = Neutral; BE = Beneficial

The main effect of property type (F(2,105) = 56.372, p = 0.000, ${\eta }_p^2=0.518$) and categorical distance (F(1,105) = 40.285, p = 0.000, ${\eta }_p^2=0.277$) are both significant. More importantly, there is a type by category distance interaction (F(2,105) = 3.894, p = 0.023, ${\eta }_p^2=0.069$) driven by the fact that the difference in within category and across category generalizations are significantly reduced for the harmful property compared to the other property types. The inductive strength within category inference was significantly higher than that of cross category inference for BE (t(35) = 4.198, p = 0.000) and NE (t(35) = 4.595, p = 0.000), but not for HA (t(35) = 1.780, p = 0.084). In the within category condition, there were significant effects of BE < NE (t(35) = − 4.197, p = 0.000), BE < HA (t(35) = − 7.195, p = 0.000), and NE < HA (t(35) = − 3.593, p = 0.001). In the cross category condition, there were similar patterns, including BE < NE (t(35) = − 3.211, p = 0.003), BE < HA (t(35) = − 9.470, p = 0.000), and NE < HA (t(35) = − 7.509, p = 0.000).

Discussion experiment 1

The robust main effect of categorical distance was consistent with previous reports (Osherson et al., 1990; Heit et al. 2000, 2008) that the inductive strength would diminish with increasing categorical distance between the base and target categories. As expected, the main effect of property type, as well as the interaction effect between task type and categorical distance, were significant. Participants judged the inductive strength of property HA as stronger than the inductive strength of property NE, but weaker than the inductive strength of property BE, irrespective of categorical distance between the premise and conclusion. The interaction effect indicated that only the inductive strength of properties BE and NE were significantly reduced when the premise-conclusion categorical distance changed from within to across categories.

These results demonstrated that the harmful property, as compared to the beneficial and neutral properties, was considered more projectable from the base to the target. The beneficial property was treated as the least projectable property. In contrast to the neutral property, the beneficial and harmful properties differentially modulate saliency of properties in inductive inference. The harmful property condition accentuated while the beneficial property condition attenuated property saliency leading to narrower or broader generalization of categories, respectively. However, no explicit categorization ratings were collected in this experiment, which could directly test this possible explanation. This will be considered in the following study.

Study 2

Method

Participants

One hundred and twenty-six paid healthy undergraduate students (freshmen and sophomores) recruited from Southwest Minzu University participated in this experiment. These subjects were allocated into three groups as follows: 43 subjects (24 female; 19.98 ± 1.04 years) to the neutral property group, 41 subjects (20 female; 20.34 ± 1.39 years) to the beneficial property group, and 42 subjects (23 female; 19.98 ± 0.95 years) to the harmful property group. All subjects were right-handed and had normal or corrected-to-normal vision. None of the subjects reported any history of neurological or psychiatric diseases. Written informed consent was obtained from each participant.

Materials and procedure

To further test whether the type of property actually modulates the scope of target categories, the inductive strength judgment task was repeated, followed by an explicit categorization task. Figure 3 illustrates the experimental task in study 2. After administration of the task from study 1, a follow up categorization task was administered. Subjects were instructed to evaluate the likelihood of the base category and the target category belonging to the same category on a 5 point Likert scale. The task was administered on pencil and paper.

Fig. 3.

Illustration of experimental tasks in study 2

The cognitive capacity of participants was evaluated by the number series task (Harrison et al. 2013; Thurstone 1938). Subjects were shown a series of number sequences following an arithmetic rule, and they were instructed to choose the next number in the series from five possible answer options. The test required the completion of 10 trials within 5 min. The dependent measure was the total number of correctly completed problems. There were no differences in accuracy across groups (F(2,125) = 0.048, p = 0.953) (7.451 ± 2.507 for the neutral property group, 7.384 ± 2.368 for the beneficial property group, 7.286 ± 2.484 for the harmful property group).

Results

Figure 4 displays the result of inductive strength as a function of property type (i.e., BE, NE and HA) and categorical distance (i.e., within and across categories) for Study 2. The results are similar to study 1. The main effect of task type (F(2,249) = 27.01, p = 0.000, ${\eta }_p^2=0.181$), and categorical distance (F(1,249) = 19.54, p = 0.000, ${\eta }_p^2=0.074$), as well as the interaction effect between task type and categorical distance (F(2,249) = 3.431, p = 0.034, ${\eta }_p^2=0.027$), were all significant. The inductive strength of inferences were significantly higher within categories than across categories for BE (t(40) = 5.929, p = 0.000) and NE (t(42) = 4.656, p = 0.000), but not for HA (t(41) = 0.851, p = 0.399). In within category conditions, there were significant effects of BE < HA (t(40) = − 3.354, p = 0.002) and NE < HA (t(41) = − 4.371, p = 0.000), but there was no significant difference between BE and NE (t(40) = 0.042, p = 0.967). In across category conditions, there were significant effects of BE < HA (t(40) = − 9.617, p = 0.000), NE < HA (t(41) = − 7.005, p = 0.000), and trend level effect of BE < NE (t(40) = − 1.937, p = 0.060).

Fig. 4.

Measure of inductive strength (on a 5 point Likert scale) as a function of task type (i.e., BE, NE and HA) and categorical distance between the premise category and the conclusion category (i.e., within and across categories) for study 2. HA = Harmful; NE = Neutral; BE = Beneficial

Figure 5 shows the result of the categorization as a function of property type (i.e., BE, NE and HA) and categorical distance between the base category and the target category. The main effect of task type (F(2,249) = 1.110, p = 0.331, ${\eta }_p^2=0.009$) did not reach the level of statistical significance. While the main effect of categorical distance (F(1,249) = 31.602, p = 0.000, ${\eta }_p^2=0.115$) and the interaction effect between task type and categorical distance (F(2,249) = 4.335, p = 0.014, ${\eta }_p^2=0.034$) were significant. Except for HA (t(41) = 1.387, p = 0.173), the categorization likelihood within and between categories was significant for BE (t(40) = 5.241, p = 0.000) and NE (t(42) = 4.634, p = 0.000). In within category conditions, the categorization likelihood of BE versus NE (t(39) = 0.647, p = 0.521), BE versus HA (t(39) = 1.297, p = 0.202), and NE versus HA (t(41) = 0.813, p = 0.421), were not significant. In across category conditions, the categorization likelihood of BE versus HA (t(39) = − 3.848, p = 0.000) and NE versus HA (t(41) = − 2.311, p = 0.026) were significant, except for BE versus NE (t(39) = − 0.815, p = 0.420).

Fig. 5.

Measure of categorization, following inductive reasoning, (on a 5 point Likert scale) as a function of task type (i.e., BE, NE and HA) and categorical distance between the premise category and the conclusion category (i.e., within and across categories) for study 2. HA = Harmful; NE = Neutral; BE = Beneficial

We also undertook a signal detection theory (SDT) analysis of the categorization data (Fig. 6). Receiver operating characteristic (ROC) curves were generated to plot hits against false alarms. The ROC procedure provides another method to assess the classification accuracy (independently from response bias) across different conditions and is useful for data interpretation. The rating score of signal (i.e., the base category and the target category is of higher possibility to belong to the same category) was recorded as the hit rate, and the rating score of noise (i.e., the base category and the target category is of lower possibility to belong to the same category) was recorded as the false alarm rate. The subjects' ratings on signal and noise were then respectively converted into the response probabilities on four kinds of signal-to-noise ratios, so as to obtain the combination of four groups of hit rate and false alarm rate (Banks 1970). According to these four groups of hit rate and false alarm rate, a ROC curve can then be drawn.

Fig. 6.

ROC curve for the categorization after inductive reasoning. HA = Harmful; NE = Neutral; BE = Beneficial

ROCs were constructed separately for the trials with different property conditions, i.e., HA, NE and BE, as shown in Fig. 6. Visual inspection of the ROCs suggested three crucial things can be derived. First, the distance of ROCs under HA conditions is far from the upper left corner than ROCs under BE and NE conditions, suggesting lower sensitivity (independent from bias) under HA conditions. Second, the relative shift of ROCs under HA conditions to the right, in contrasted to ROCs under the other two conditions, indicates a more liberal response bias (for identical levels of sensitivity) under HA conditions. Finally, ROCs under BE and NE conditions are smoothly curved and appear consistent with the assumptions of the SDT model, implying that analysis of variance (ANOVA) are inappropriate analyses. While ROCs under HA conditions is closer to be linear rather than curved, which is compatible with ANOVA analyses.

Discussion of experiment 2

A comparison of Figs. 2 and 4 reveals that the patterns of inference in study 2 replicated those in study 1, except for the fact that there were no significant differences across the beneficial and neutral property conditions, and within the harmful property condition, there were no differences in within category and across category inferences. This may be explained by the usage of the different Likert scales in the two studies. The 21 point scale in study 1 may have greater discrimination and precision than the 5 point scale in study 2.

Both types of analyses—ANOVA model which assumes linear ROCs, and the SDT model which assumes curved ROCs—lead to the same conclusion of significant difference of classification performance between the HA condition and the other two conditions (i.e., BE and NE). The observed agreement in these measures does not mean the confirmation of each other and is far from assured, as the underlying ROCs cannot be both curved and linear simultaneously (Rotello et al. 2018). Confounded by response bias, wrong conclusions are quite likely when inappropriate outcome measures are applied (Rotello et al. 2015). Despite all this, the ROCs shape of HA conditions has demonstrated a noticeable difference from BE and NE conditions, as the shape of an observed ROC curve provides information about the cognitive processes that support performance on a particular task.

The results of the explicit categorization task following the property inference task (Fig. 5) and ROC curve (Fig. 6) are consist with our expectations. In contrast with the neutral and beneficial property conditions, the category boundaries of the target category were enlarged in the harmful property condition. In other words, subjects were more likely to collapse the distinction between the base and the target in the case of the harmful property condition. The tendency to do this was much greater in the cross category condition, where there was no difference in generalization within and across categories (for harmful properties). These results suggest that enhanced projectability of adaptive properties, such as those that would cause us harm, results in overgeneralization of categories, to error on the side of caution.

General conclusion

The results of the two studies are consistent with our hypothesis that evolutionarily adaptive properties, like harm, are considered more salient and projectable than neutral or beneficial properties and result in overgeneralization, in accordance with a “better safe than sorry” strategy. Within the inductive reasoning literature they are consistent with a prior study of risk context effects (Sakamoto and Nakagawa 2007), in which participants tend to rate inductive reasoning arguments according to the direction of risk aversion. Within the deductive reasoning literature, these results are consistent with findings in the Wason card selection task (Cosmides 1989) where the adaptive property of cheater detection is thought to driv the reasoning effect.

What type of model can account for these results? There are two basic types of models in the literature. First, one can view these evolutionarily determined, adaptive properties as modulating the salience of properties within a general inductive reasoning system. Second, one can take a massive modularity approach and view each adaptive property as being associated with a self-contained “reasoning” module. We have clearly adopted the first approach in motivating and carrying out the study.

However, there is not much in the way of data to help us decide between the two models. For example, within the neuroimaging literature on inductive reasoning involving meaningful content, while there is some overlap in the recruitment of frontal, temporal, and parietal systems, the overall patterns are heterogeneous (e.g., Goel et al. 1997; Goel and Dolan 2000,2004; Liang et al. 2014, 2016; Jia et al. 2011, 2015). However, these differences are most parsimoniously accounted for in terms of differing modalities, contents, and tasks rather than distinct systems of reasoning.

While the domain specific reasoning module approach is possible, and does solve the major puzzle of induction, that is, selecting the relevant or salient properties from a large, even unbounded set (Cosmides and Tooby 1994), it is not unproblematic. In particular, it replaces the notion of the general-purpose reasoning system with many localized systems that that are causally triggered by specific features of the environment. The whole point of induction is that it allows us to take into consideration any relevant world knowledge in drawing inferences. Our results leave open the possibility of either model, but they do provide us with additional insight for gauging saliency or projectability of properties for the purpose of inference.

A comprehensive discussion of the neural mechanisms underlying the overgeneralization effect for the harmful property is beyond the scope of this study. However, it is worth highlighting the possibility of amygdala involvement. Amygdala, a phylogenetically older brain structure, is central to threat processing across species. One possibility, therefore is that the prioritization of harfmful/threatening stimuli compete with other characteristics of visual stimuli for perceptual processing resources, and affect the allocation of top-down attentional resources. This is consistent with attentional control theory (Eysenck et al. 2007), as well as the finding that threat affects the balance between stimulus-driven and goal-directed behaviors (Shackman et al. 2011), and supported by the fact that amygdala functionally links with perceputal, attentional and salience network (Sylvester et al. 2020; Kerestes et al. 2017). Future neuroimaging studies are required to test this potential explanation.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

KJ and PPL designed the study, KJ ran the experiment and did the data analysis, PPL, KJ and VG explained the results, PPL, KJ and VG wrote the manuscript.

Declarations

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.