Further evidence for the memory state heuristic: Recognition latency predictions for binary inferences

Abstract

According to the recognition heuristic (RH), for decision domains where recognition is a valid predictor of a choice criterion, recognition alone is used to make inferences whenever one object is recognized and the other is not, irrespective of further knowledge. Erdfelder, Küpper-Tetzel, and Mattern (2011) questioned whether the recognition judgment itself affects decisions or rather the memory strength underlying it. Specifically, they proposed to extend the RH to the memory state heuristic (MSH), which assumes a third memory state of uncertainty in addition to recognition certainty and rejection certainty. While the MSH already gathered significant support, one of its basic and more counterintuitive predictions has not been tested so far: In guessing pairs (none of the objects recognized), the object more slowly judged as unrecognized should be preferred, since it is more likely to be in a higher memory state. In this paper, we test this prediction along with other recognition latency predictions of the MSH, thereby adding to the body of research supporting the MSH.

1 Introduction

The recognition heuristic (RH) is a fast and frugal decision strategy proposing that, for binary decisions, if one object is recognized and the other is not, one should infer that the recognized object scores higher on the criterion under consideration (Goldstein and Gigerenzer, 2002). This simple decision rule has gained a lot of attention, and a large body of research was dedicated to it (see Gigerenzer and Goldstein, 2011; Pachur et al., 2011, for reviews). However, one key concept of the RH has often been neglected: recognition. While literally at the core of the heuristic, only a modest amount of research has focused on understanding the role of recognition in use of the RH (e.g., Erdfelder, Küpper-Tetzel and Mattern, 2011; Pachur and Hertwig, 2006; Pleskac, 2007; Castela et al., 2014; Castela and Erdfelder, 2017). Notably, Erdfelder, Küpper-Tetzel and Mattern proposed a framework that extends the RH by accommodating the role of recognition memory, the memory state heuristic (MSH). Their framework was later also supported by Castela et al. (2014) and Castela and Erdfelder (2017) using formalizations of the MSH in the framework of multinomial processing tree models (Batchelder and Riefer, 1999; Erdfelder et al., 2009). In this paper, we primarily aim to test a crucial and counterintuitive prediction that has not been directly addressed before and conflicts with the popular notion that processing fluency – or cognitive fluency in general – boosts preference for a choice option (Schooler and Hertwig, 2005; Zajonc, 1968). In addition, we provide support for the MSH through conceptual replications of predictions previously tested by different researchers (Erdfelder, Küpper-Tetzel and Mattern, 2011; Hertwig et al., 2008; Hilbig, Erdfelder and Pohl, 2011; Schooler and Hertwig, 2005; Pohl et al., 2016). In this way, we aim at closing a gap in previous research on the MSH and provide converging evidence on the importance of memory strength (rather than recognition judgments) in recognition-based decision making.

This paper will be organized as follows: First, we will introduce the RH and discuss how recognition memory has so far been understood in its context. Second, we will introduce the MSH and describe the evidence relevant to it. Third, we will report our two new studies that complement the body of work on the MSH, each consisting of a re-analysis of previously published data and a new experiment.

1.1 The Recognition Heuristic

To better understand how recognition memory has been (or can be) integrated in the RH, it is first essential to describe more precisely how the heuristic has been proposed. To simplify that process, we will refer to the most prominent paradigm associated with the RH as an illustrative example. This is the city size paradigm, which involves a paired comparison task and a recognition task. In the paired comparison task, participants must infer which of two cities has a larger population. In the recognition task, participants are asked to indicate for each city involved whether they have heard of it (yes) before the study or not (no). With the data from this recognition task, all pairs of cities in the comparison task can be categorized into three types: knowledge cases (both objects are recognized), recognition cases (one object is recognized and the other is not), and guessing cases (none of the objects is recognized). The RH applies only to recognition cases, for the obvious reason that it cannot discriminate between objects in the other two types of pairs.

Importantly, Gigerenzer and Goldstein (2011) specified additional preconditions for use of the RH. First, there should be a strong correlation between recognition and the decision criterion. In our example, recognition should be strongly correlated to the size of a city (which, indeed, it is). Additionally, further cues should not be readily available. This means that, for example, when comparing the sizes of Berlin and Mannheim, the information that Berlin is the capital of Germany, or that it has an international airport, should not be presented to the participant simultaneously (whereas, of course, it could be retrieved from memory). Finally, they asserted that the RH applies only to natural recognition, that is, artificially inducing recognition in the laboratory (by, for example, presenting objects to the participants several times) should not necessarily lead to use of the RH.

1.2 Recognition memory in the context of the RH

In the previous section we outlined the basic concepts surrounding the recognition heuristic, but it is still unclear how the memory processes underlying recognition influence use of the heuristic. In its original definition, the RH was not related to recognition memory, but only to recognition judgments. Goldstein and Gigerenzer (2002) assumed that the RH works on the output of the recognition process, and that the process itself can be disregarded. In other words, they assumed the RH operates on yes or no recognition judgments, and whatever underlies that judgment can be ignored for the purpose of investigating the heuristic. This assumption implies that the frequency with which an object has been encountered does not affect use of the RH, and only the final all-or-none process of remembering any encounter or not will matter. It follows that the RH will treat objects with different memory strengths equally, as long as they are both recognized or unrecognized. Erdfelder, Küpper-Tetzel and Mattern (2011) challenged the notion that memory strength should not influence use of the RH. Specifically, they argued that “Showing that the RH is an ecologically rational and well-adapted choice strategy obviously requires a formal theoretical link between (1) the memory strengths of choice option names — a latent variable which is affected by environmental frequency and previous processing — and (2) binary recognition judgments for choice option names — an empirical variable which is assumed to affect decision behavior.”

Following from this understanding of a necessary link between memory strengths and recognition judgments, Erdfelder, Küpper-Tetzel and Mattern (2011) proposed to integrate a model of recognition memory with the RH theory. To do so, they relied on one of the most well-supported models of recognition memory available — the two-high-threshold (2HT) model (Kellen, Klauer and Bröder, 2013; Snodgrass and Corwin, 1988). Importantly, besides being one of the most successful models of recognition memory, the 2HT model has the added advantage of being easily combinable with the RH (Erdfelder, Küpper-Tetzel and Mattern, 2011).

The 2HT model belongs to the class of multinomial processing tree models (Batchelder and Riefer, 1999; Erdfelder et al., 2009). Like other multinomial processing tree models, the 2HT model is based on the assumption that observed categorical responses emerge from a defined set of discrete states and that the probability of such states being entered depends on the probability of certain cognitive processes occurring or not. The basic premise of the 2HT model is that there are three possible memory states underlying recognition judgments — recognition certainty, uncertainty, and rejection certainty. The probability of those states being entered depends on the probability of two thresholds being exceeded (Figure 1). Specifically, for objects experienced before, if the memory strength exceeds the first threshold with probability r, the object will be in the recognition certainty state and a yes recognition judgment will be given. If, with complementary probability 1−r, the memory strength lies below this threshold, the object will be in the uncertainty state, and the recognition judgment will depend on a second process of guessing, resulting in a yes judgment with probability g and a no judgment with probability 1−g. For objects not experienced before, if the memory strength lies below the second threshold with probability d, the object will be in the rejection certainty state and a no recognition judgment will be given. With complementary probability 1−d, the memory strength lies above this second threshold and the object will be in the uncertainty state, just like unrecognized objects experienced before. Again, the recognition judgment will depend on guessing yes or no with probabilities g and 1−g, respectively.¹

Figure 1: Graphical representation of the two-high-threshold model. Parameter r denotes the probability of old objects exceeding the recognition threshold. Parameter d denotes the probability of new objects falling bellow the rejection threshold. Parameter g denotes the conditional probability of guessing yes in the uncertainty state.

To combine this model with the RH theory, Erdfelder, Küpper-Tetzel and Mattern (2011) suggested a new framework — the memory state heuristic (MSH). The MSH is a straightforward extension of the RH, which mainly replaces recognition judgments by memory strengths. That is, it assumes that memory strengths, and not recognition judgments per se, affect decision behavior. This simple extension enriches both the predictions that can be drawn and the explanatory scope of the heuristic. Whereas the RH has predictions for recognition pairs (i.e., recognition cases) only, the MSH has predictions for any pair that involves objects in different memory states. These predictions can be summarized by two simple premises: (1) if objects are in different memory states, there should be a preference for the object in a higher state; (2) the larger the discrepancy between the memory states of objects in a pair, the higher should be the probability of choosing the object in a higher state. By implication, the probability of choosing the object in the higher state should be larger for pairs of one object in the recognition certainty state and the other in the rejection certainty state than for pairs where one of the objects is in the uncertainty state. Based on these two principles, Erdfelder, Küpper-Tetzel and Mattern could both explain previous results that challenged the RH and also draw and test new predictions. To do so, they relied on the fact that multinomial processing tree models like the 2HT model can be interpreted as probabilistic serial processing models (Batchelder and Riefer, 1999; Heck and Erdfelder, 2016). By implication, the number of cognitive processing stages in a given branch of the model will influence its total processing time. Specifically, in the case of the 2HT model, whenever an object reaches the memory state of uncertainty and a second cognitive process — guessing — is required, the response time distribution should be stochastically larger than when an object reaches one of the two certainty memory states (Heck and Erdfelder, 2016). Following from this interpretation of the 2HT model, a clear prediction can be made: “The larger the recognition judgment latencies, the more likely it is that the judgment originates from guessing and the less likely it is that it originates from memory certainty” (Erdfelder, Küpper-Tetzel and Mattern, 2011, p. 13).

The MSH improves on the RH by offering a straightforward explanation to results that have challenged the latter. Specifically, the observation that factors beyond recognition (e.g., speed of recognition, availability of further knowledge) seem to affect the preference for the recognized objects has suggested that there is no non-compensatory use of recognition. However, Erdfelder, Küpper-Tetzel and Mattern (2011) argued that those findings can be easily accommodated by the MSH. For example, the observation that recognized objects that are recognized faster are preferred over those recognized more slowly (Hertwig et al., 2008; Marewski et al., 2010; Newell and Fernandez, 2006) has been interpreted as suggesting that retrieval fluency is also driving the inference. However, the MSH would predict this preference because objects recognized faster are more likely to be in the recognition certainty state than the ones recognized more slowly (Erdfelder, Küpper-Tetzel and Mattern, 2011). Another example discussed by Erdfelder, Küpper-Tetzel and Mattern (2011) concerns the finding that RH accordance rates (proportion of times the recognized object is chosen in recognition pairs) are larger when RH-consistent inferences are correct than when they are incorrect (Hilbig and Pohl, 2008). While this observation may suggest that further knowledge is involved in the inferential process, a different interpretation follows from the MSH: Recognition pairs originating from recognition and rejection certainty states should lead to correct inferences more often than the ones associated with at least one object in the uncertainty state, simply because recognition validity (i.e., the correlation between recognition and the criterion) is higher for these pairs.

The fact that the MSH offers an explanation for these controversial results suggests that it can be an important extension of the RH. Furthermore, the MSH has found a considerable amount of other support. First, Erdfelder, Küpper-Tetzel and Mattern (2011) tested seven predictions of the MSH, focused on RH accordance rates and decision latencies, both as a function of recognition and rejection latencies. The first three predictions, which state that RH accordance rates should increase with decreasing recognition and rejection latencies, and that their effect is additive, were supported in their study. Additionally, they tested whether the decision latency in recognition pairs increases with both the recognition latency of the recognized object and the rejection latency of the unrecognized object, and that their effect is additive. These further three predictions were also supported by their data. Finally, they found support for their seventh prediction, which stated that response bias manipulations (aimed at selectively affecting the guessing probability) in the recognition test should affect recognition judgments but not performance in the comparison task. Since the RH theory assumes that recognition judgments per se influence decisions, it would predict that a bias manipulation will also affect inferences. The MSH, in turn, predicts the observed result, since memory-states rather than recognition judgments should influence decisions, that is, since biasing the guessing probability does not alter the memory-states distribution, inferences should be left unaffected.

Additionally, Castela et al. (2014) also found support for the MSH. They tested the predictions of (a) the RH, (b) knowledge integration accounts, and (c) the MSH regarding the proportion of RH-use for cases where the recognized object is said to be only merely recognized versus recognized along with further knowledge. While the RH predicts there should be no difference, since recognition alone should drive inferences in recognition pairs, knowledge integration accounts predict that RH-use should be lower when there is knowledge, since when there is further information on which we could base the inferences then we should not rely only on recognition. Notably, the MSH predicts the opposite pattern, since recognized objects for which there is further knowledge are more likely to have originated from recognition certainty than objects that are merely recognized. Reliance on the memory state should therefore be higher for the former. Through a reanalysis of 16 published data sets, Castela and collaborators showed that RH use is in fact more frequent when there is knowledge about the recognized object, a result that is predicted by the MSH and at odds with the other two accounts.

Finally, Castela and Erdfelder (2017) comprehensively tested the MSH by developing a formal model that incorporates its predictions for all possible memory-state combinations. We showed that restricting this model to hold the core prediction of the MSH, namely, that MSH-use is higher when the distance between memory-states is highest, leads to no significant increase in model misfit, thereby suggesting that such a model is consistent with the data. This is, to our knowledge, the most thorough and elaborated test of the MSH so far. Given all these results, it appears that the MSH is a well-supported framework which should be seriously considered as an important extension of the RH.

It is at this point clear that there is considerable evidence supporting the MSH. However, when we want to advocate the MSH, we must ensure that the support for it does not depend on the decision domain employed, on testing a limited number of predictions, on using specific methods of evaluation, or on referring to a limited set of proxy measures for underlying memory states. Probably most importantly, especially the bold and surprising predictions of the MSH need to be tested exhaustively, as these predictions most likely allow us to discriminate between the MSH theory and other theories of inferential decision making. Therefore, the primary aim of the present work is to address a previously untested counterintuitive prediction of the MSH regarding choices between pairs of objects, both of which are unknown to the decision maker. Moreover, we also aim at conceptually replicating and extending results previously tested in different decision contexts or with different measures of MSH use. In this way, we hope to close existing gaps and provide converging evidence that solidifies the body of research on the MSH.

The focus of Erdfelder, Küpper-Tetzel and Mattern (2011) has been on testing predictions for recognition pairs, but as explained before, the MSH also makes predictions for guessing and knowledge pairs, as long as the objects under comparison are in different memory states. This will be the focus of our first study. As for recognition pairs, the predictions follow from the basic premise of the MSH: If objects are in different memory states, there should be a preference for the one in a higher state. Therefore, in this study we will test two predictions:

1. 1. In knowledge pairs there should be a preference for the object recognized faster (as this one is more likely in the memory certainty state)

2. 2. In guessing pairs, there should be a preference for the object recognized more slowly (since this one is more likely in the uncertainty state, which is the highest possible state for unrecognized objects).

However, as outlined above, the MSH also predicts that the preference for the object in a higher state should be stronger, the higher the discrepancy between the memory states. While in recognition pairs the maximal memory state distance can be observed (one object in recognition certainty and the other in rejection certainty), in both knowledge and guessing pairs this is assumed never to occur, since (to a close approximation) objects will either be in the same state or in adjacent states (recognition certainty and uncertainty or rejection certainty and uncertainty, respectively). For this reason, we expect weaker effects of recognition latency differences on choice probabilities than those found for recognition cases. Additionally, we will also test whether effects on choice probabilities are stronger when the differences in latencies are higher, therefore increasing the probability of the objects being in adjacent states versus in the same state.

Note that the MSH prediction regarding knowledge cases overlaps with what is called the fluency heuristic (Hertwig et al., 2008; Schooler and Hertwig, 2005). The fluency heuristic states that, in knowledge cases, the fastest retrieved option should be chosen. Its premise is that the fluency with which an object is retrieved from memory (indexed by the latency of the recognition judgment) can be used as a single cue and determine inferences. They measured the accordance rate of the fluency heuristic by computing, for each participant, how many times the object retrieved faster is chosen in knowledge pairs (pairs with differences in recognition latency smaller than 100 ms were excluded),² and found it to be reliably higher than the chance probability of .50. Furthermore, they observed that accordance rates increase with the difference in latencies between objects. While the fluency heuristic can accommodate these results, its empirical scope is limited: It applies only to knowledge pairs, and within those, to pairs where the fluency difference is larger than 100ms. The MSH, in contrast, predicts these and other results, including predictions for guessing and recognition cases. It is, therefore, a framework with a wider scope and more parsimonious than a combination of different heuristics for knowledge, recognition, and guessing cases (Erdfelder, Küpper-Tetzel and Mattern, 2011). Importantly, the MSH also predicts that the preference for the faster recognized object in knowledge cases should be considerably weaker than the preference for recognized objects in recognition cases, simply because the memory-state discrepancy for knowledge pairs can only be small (i.e., recognition certainty and uncertainty) or even nonexistent (i.e., when both objects are in the same state). The fluency heuristic, in contrast, is silent about the predicted effect size for knowledge cases as compared with recognition cases. Notably, this MSH prediction has already found some support in previous research (e.g., Hilbig, Erdfelder and Pohl, 2011; Marewski and Schooler, 2011; Pohl et al., 2016; Schwikert and Curran, 2014).

While the predictions for knowledge cases seem plausible and straightforward, the core prediction for guessing cases is more surprising and counterintuitive as it conforms to the expectation of a preference for less fluent objects. To the best of our knowledge, no framework other than the MSH makes or can accommodate such a prediction. Therefore, the most important focus of our first study lies in the novel and apparently counterintuitive prediction for guessing cases.

In addition to these predictions for knowledge and guessing cases, we focus on a prediction of the MSH for recognition cases in a second study. Erdfelder, Küpper-Tetzel and Mattern (2011) already showed that larger recognition and rejection latencies are associated with weaker preferences for the recognized object in recognition cases. In our second study, we aim to conceptually replicate this result in a more refined way using a better measure of MSH-use. The proportions of choices of the recognized objects used by Erdfelder, Küpper-Tetzel and Mattern are biased measures of MSH-use because counting the number of times choices are in line with MSH use does not take into account what led to that choice. An option might have been chosen because it was in a higher memory state, or because other information, which points in the same direction, was used. For example, when comparing supposed population sizes of Berlin and Mannheim, a non-european person might chose Berlin because she recognizes it with certainty and does not recognize Mannheim, or because she knows Berlin is the capital of Germany, and therefore likely to be a large city.

For this reason, Hilbig, Erdfelder and Pohl (2010) developed a multinomial processing tree model which estimates RH-use in a more sophisticated way. The r-model (Figure 2) consists of three trees, which correspond to the three types of pairs. For knowledge and guessing pairs, the trees have only a single parameter that accounts for the accuracy for knowledge and guessing pairs, respectively. For recognition pairs, on the other hand, the model considers the possibility that a recognized option is chosen through use of further knowledge, and provides in this way an unbiased estimate of RH-use (which corresponds to parameter r in the model; see Hilbig, Erdfelder and Pohl (2010) for addditional details about the r-model).

By adopting this model to measure MSH use for recognition cases, we can assess in a more precise way how recognition and rejection latencies are associated with noncompensatory reliance on recognition. Additionally, we can test whether in the most extreme cases, when the recognition judgment latencies are very short (so that both objects are most likely in recognition and rejection certainty states), people always rely on memory-states only, or whether even then the probability of choosing the recognized object is significantly smaller than one, suggesting that other processes such as integration of further knowledge are involved in at least some of the cases where conditions for relying on memory strength are optimal.

Figure 2: Graphical representation of the r-model: Parameter r denotes the probability of applying the recognition heuristic as originally proposed, that is, by ignoring any knowledge beyond recognition. a = recognition validity (probability of the recognized object representing the correct choice in a recognition case); b = probability of valid knowledge; g = probability of a correct guess; rec. = recognized; unrec. = unrecognized.

In general, since Horn, Pachur and Mata (2015) observed correlations above .90 between r parameter estimates and RH accordance rates, we expect similar results for our second study as previously reported by Erdfelder, Küpper-Tetzel and Mattern. However, the additional possibility to assess the hypothesis r = 1 (i.e., perfect reliance on recognition) for objects in memory certainty states renders use of the r-model particularly attractive for our current study.

2 Study 1: MSH predictions for guessing and knowledge cases

We first tested whether choices for guessing and knowledge cases are in accordance with the MSH prediction that there is a preference for the object in a higher state. Specifically, as outlined above, we used recognition and rejection latencies as proxies for underlying memory states. Therefore, we predicted that in knowledge pairs there is a preference for the object with a shorter recognition latency (and therefore a higher probability of being in a recognition certainty state) while in guessing pairs there is a preference for the object with the longest rejection latency (and therefore a higher probability of being in the uncertainty state).

2.1 Reanalysis of published data

We reanalyzed the data of 14 published data sets from our lab (Table 1), in order to look for preliminary evidence for our hypotheses. As shown in Figure 3, we observed that for all 14 data sets the proportion of choosing the object recognized faster in knowledge cases was significantly larger than .5 (smallest t(21) = 2.78, all p < .01). Regarding guessing cases, in 12 of the 14 data sets the proportion of choosing the object recognized more slowly was significantly larger than .5 (smallest significant t(63) = 2.08, p = .02). For comparison purposes, the accordance rates for the RH (applying to recognition cases) are also included in Figure 3, showing that choice preferences for recognized objects in recognition cases are much stronger than choice preferences in the other two cases.

Table 1: Source and description of the 14 reanalyzed data sets.

Figure 3: Proportion of choices of the fastest or slowest recognized or unrecognized object for knowledge and guessing cases, respectively, and of the recognized object for recognition cases, for all 14 reanalyzed datasets. Error bars represent standard error of the mean.

Clearly, these results are in line with our expectations. However, the studies included in the reanalysis were not conducted with our hypotheses in mind. In order to collect further evidence, we designed a new experiment specifically tailored to our hypotheses. With this new experiment, we primarily aimed at optimizing the proportion of knowledge and guessing cases in order to achieve more powerful tests of the MSH predictions for these cases. Moreover, we were also interested in generalizing the results across different decision domains beyond city-size comparisons.

2.2 Experiment 1

2.2.1 Material and procedure

The paradigm we used resembles the city-size paradigm outlined in the section The Recognition Heuristic but involves different types of decisions. This paradigm includes two tasks: (1) a recognition test, where objects are presented and participants must judge whether they have seen them before or not; (2) a comparison task, where participants see pairs of the objects and must infer which scores higher on a given criterion. Since the objects are paired exhaustively, the relative proportion of knowledge, recognition, and guessing cases will depend on the proportion of objects recognized. Therefore, in order to optimize the proportion of knowledge and guessing cases, it is important to include in the experiment a condition for which the proportion of recognized objects across participants is larger than .50 (resulting in many knowledge cases) and a different condition in which the proportion of recognized objects is clearly less than .50 (resulting in many guessing cases). A third condition should involve a recognition rate of about .50, resulting in (almost) equal frequencies of knowledge and guessing cases. Moreover, since we also wanted to generalize our findings across different domains, we made use of different types of objects and inference criteria in the three conditions. Specifically, all participants were presented with objects from three domains: largest world cities (with over 3 million inhabitants; http://en.wikipedia.org/wiki/List\_of\_cities\_proper\_by\_population), most successful celebrities (100 most successful celebrities according to the Forbes list of 2015; http://www.forbes.com) and longest rivers in the world (over 1900 km long; https://en.wikipedia.org/wiki/List\_of\_rivers\_by\_length). According to pre-tests conducted in our lab, we know that for the domain of world cities normally 50% of the objects are recognized. We included this domain for purposes of generalization, and also because it is one of the most often used domain in the study of the RH and should serve as benchmark. For the domain of celebrities, normally 65% of the objects are recognized. Therefore, this domain is ideal to test the hypothesis regarding knowledge cases. Finally, the rivers domain is ideal for testing the hypothesis regarding guessing cases, since usually only 35% of the objects are recognized. The experiment included three blocks, each consisting of the recognition test and the comparison task for each domain.

The order of blocks was randomized for all participants. In each block, the recognition test always preceded the comparison task. In the recognition test participants saw all 20 objects (randomly selected from each domain, but the same for all participants) and had to decide whether they have heard of them before or not. Objects were presented one at a time, in random order, and a 500 ms interstimulus fixation-cross followed each response. Response times were recorded along with the recognition judgments. After each recognition test, a comparison task followed. In the comparison task, participants saw 190 pairs, consisting of the exhaustive pairing of the 20 objects, and had to infer which one scored higher on the criterion. Each pair was presented at a time, in random order, and a 500 ms interstimulus fixation-cross followed each response. Response times were recorded along with the responses. For the world cities, the criterion was city-size; for celebrities, the criterion was how successful they were³; and for the rivers, the criterion was their length.

2.2.2 Participants

To provide an appropriate balance between type-1 and type-2 error rates in χ ² model tests (Erdfelder, 1984; Moshagen and Erdfelder, 2016), we recruited 75 students (50 women) from the University of Mannheim aged between 19 and 46 years (M = 22.00, SD = 5.04). Participation was compensated monetarily as a function of performance in the comparison task. Every participant received at least two euros, and they could earn up to 7.70. They gained one cent for each correct answer, and lost one cent for each wrong one.

2.2.3 Results

One participant had to be removed from the analysis for all domains, because he indicated that he did not recognize any object in any domain. Furthermore, one participant was removed from the guessing analysis of the cities domain because he recognized 19 out of the 20 cities, therefore having no guessing pairs. Finally, two additional participants were removed from the knowledge analysis of the rivers domain because they only recognized one river and therefore had no knowledge pairs. For the remaining participants, the proportion of recognized items was on average .68 for celebrities, .58 for the world cities, and .36 for rivers. This was in line with the pre-tests, although a bit higher than what we expected for the world cities domain.

Since our hypotheses refer to the preference for the object recognized faster in knowledge pairs, and the one judged unrecognized more slowly in guessing pairs, we first calculated per participant the proportion of times their choices were in line with those hypotheses (accordance rate). We then performed one-sample t-tests to assess whether the mean accordance rates were larger than .50. As can be seen in Table 2, we found support for both hypotheses in all three domains assessed. For comparison purposes, the accordance rates for the RH are also included in Table 2, replicating the previous result that choice preferences for recognized objects in recognition cases are much stronger than the predicted choice preferences for knowledge and guessing cases.

Table 2: Experiment 1. Results of one-sample t-tests testing if the mean of the individual proportion of choices in accordance with our hypotheses is higher than .50. For knowledge cases, accordance means choosing the fastest recognized object, for guessing cases accordance means choosing the slowest unrecognized object, and for recognition cases accordance means choosing the recognized object. ^* significant at the .05 α level.

In addition to testing for an above chance preference for the items more likely to be in a higher state, we also wanted to assess whether this preference would increase with an increasing difference in recognition latencies (i.e., latencies of yes judgments) or rejection latencies (i.e., latencies of no judgments) between objects in a pair (and therefore an increasingly higher probability of being in adjacent states). To do so, we ran a multilevel logistic regression⁴ (level 1: choices per participant; level 2: participants) with Accordance as a dependent variable. Accordance is essentially a binary variable which takes the value one if choices are in line with our hypotheses, and zero when they are not. Specifically, for knowledge pairs, Accordance will be one whenever the fastest recognized object is chosen, and zero otherwise. Conversely, for guessing pairs, Accordance will be one whenever the slowest unrecognized object is chosen, and zero otherwise. As predictors, we included both the main effects and the interactions of the RT difference (difference in recognition or rejection latencies between the objects in a pair) with Case (knowledge or guessing) and with Domain (celebrities, cities or rivers). Additionally, the model includes a random intercept for each participant and a random slope for the effect of RT difference within each participant. Our hypothesis would be that RT difference has a positive effect on Accordance for both cases and in all domains. We find support for our hypothesis.

As can be seen in Table 3, RT difference has a significant positive effect on Accordance. Additionally, there are no differences in Accordance between the domains.⁵ Moreover, while the effect is present for both knowledge and guessing cases (Figure 4), we find that it is significantly stronger for knowledge cases. While this was not directly predicted, it does not compromise our findings. This will be addressed in more detail in the Discussion section.

Figure 4: Experiment 1. Marginal effect of RT difference on accordance for guessing and knowledge cases. Error bars represent 95% confidence intervals.

Table 3: Experiment 1. Summary of fixed effects results in multivel logistic regression showing how the difference in latencies between two objects in a pair (RT difference) predicts the accordance. Accordance is defined as choosing the fastest recognized object in knowledge cases, and the slowest recognized object in guessing cases.

For discrete predictors, information in parentheses clarifies the levels of the predictor which are being compared. The RT difference is scaled in seconds. ^* significant at the .05 α level.

3 Study 2: The influence of removing items with longer recognition/rejection judgment latencies on reliance on recognition

As mentioned above, in our second study we wanted to test the MSH predictions regarding recognition judgment latencies for recognition cases. Similar predictions were previously tested by Erdfelder, Küpper-Tetzel and Mattern (2011), but by relying on accordance rates only. The core question of our second study is whether we can replicate their results using the r parameter of the r-model as a more refined proxy for MSH use (Hilbig, Erdfelder and Pohl, 2010). Specifically, we aimed to test whether there is an increase in r when we sequentially remove items with longer recognition and rejection latencies and fit the r-model to those subsets of data.⁶ The rationale behind this is that by removing those “slow” items we reduce the subset mostly to objects in recognition certainty and rejection certainty states. By doing so successively, we artificially create the perfect preconditions for relying uniquely on recognition, which should lead to increasingly higher r estimates. If this prediction holds, an interesting further question to pursue is to what degree r estimates approximate 1 (i.e. perfect reliance on recognition in paired comparison judgments) if the subset is reduced to objects with the fastest yes or no recognition judgments only.

3.1 Reanalysis of published data

To address these questions, we first reanalyzed the data for the 14 published data sets that we used in our previous reanalysis (Table 1). For each data set, we first identified for each participant which items where in the first, second, third or fourth quartile of their individual recognition and rejection latency distributions. In a second step, we created (at the aggregate level)⁷ four subsets of pairs that consisted only of objects with latencies in each of the quartiles of the latency distributions.⁸ Next, we fitted the r-model simultaneously to these four disjoint subsets of data by replicating the r-model trees four times, that is, for each subset of pairs. By implication, we ended up with four r estimates. At the level of parameters, our hypothesis can be described as an order restriction such that the r parameters decrease from r ₁ to r ₄, with the index 1 corresponding to the first quartile of the distribution (only the fastest recognized and unrecognized objects are included) and 4 the last quartile of the distribution (only the slowest recognized and unrecognized objects are included).

All model-based analyses were performed with MPTinR (Singmann and Kellen, 2013) in R (Team, 2015). We first fitted the model without any restrictions; this baseline model fits the data well for 9 of the 14 data sets (Table 4). To test our hypothesis, we excluded the 5 data sets that were associated with misfit.⁹ To evaluate our order restriction we need two tests. First, we test the order restriction, r ₁ ≥ r ₂ ≥ r ₃ ≥ r ₄, against the baseline model (with no restriction on the four r parameters). Second, we test the model with order restrictions, r ₁ ≥ r ₂ ≥ r ₃ ≥ r ₄, against a model imposing equality restrictions, r ₁ = r ₂ = r ₃ = r ₄. If the order restriction corresponds to the most suitable version of the model, the first test should fail to reach statistical significance, while the second test should lead to statistically significant results.

Table 4: Goodness-of-fit statistics, corresponding degrees of freedom, and p-values for all reanalyzed data sets and Experiment 2.

* indicates that the baseline model does not fit the data well, leading to statiscally significant misfit.

Since our hypothesis involves an order restriction between four parameters, the sampling distribution of the likelihood-ratio test statistic Δ G ² does not follow a default χ ² distribution with the appropriate degrees of freedom. Given the challenge involved in determining the appropriate distribution, we opted for using a double bootstrap method (Van De Schoot, Hoijtink and Deković, 2010) to compute p-values. For example, when we want to test the order restrictions, r ₁ ≥ r ₂ ≥ r ₃ ≥ r ₄, against the baseline model, the double bootstrap consists of the following steps: (1) a non-parametric bootstrap sample is obtained from a given data set; (2) the model imposing the null hypothesis, r ₁ ≥ r ₂ ≥ r ₃ ≥ r ₄, is fitted to that data set; (3) those parameter estimates are used to obtain a parametric bootstrap sample; (4) both models under scrutiny (i.e., the model imposing the order restriction r ₁ ≥ r ₂ ≥ r ₃ ≥ r ₄ and the baseline model) are fitted to that sample and the difference in fit is calculated; (5) steps 1 to 4 are repeated many times (we repeated it 1000 times). We then compute the p-value by assessing how many times the difference in fit obtained with the bootstrapped samples is equal or more extreme than the difference in fit obtained with the original data set, and reject the null hypothesis if this proportion is smaller than .05. Additionally, we also compare the models through the model selection measure FIA (Fisher Information Approximation), which takes complexity into account.¹⁰

The results are shown in Table 5 and Figure 5. We find a clear support for the order-restricted model both with the goodness-of-fit test and the FIA comparison.¹¹ In all except one data set (Data Set 14) the order restriction did not lead to significant misfit, while the equality restriction did. In line with these results, FIA was smaller for the order restricted model than for the baseline or the equality restricted model. Only for Data Set 14, in line with the results from the goodness-of-fit test, the difference in FIA between the baseline and the order restricted model is not sufficient to support the former.

Table 5: Maximum likelihood parameter estimates of all r parameters and p-values and differences in FIA for comparisons between the baseline model and the order-restriced model (BO) and between the order-restricted and the equality-restricted model (OE) for all reanalyzed data sets and Experiment 2.

Additionally, to test whether r approaches one, we looked at the 95% confidence interval of the r ₁ probability estimates. For all 9 data sets this confidence interval does not include 1, suggesting that even under ideal conditions for use of memory state information alone people still sometimes rely on other strategies, like use of further knowledge.

While these results lend support to our hypothesis, the re-analyses are not ideal because, when creating the subsets of pairs, we necessarily limit the data points available for analysis (Table 5). Therefore, we designed Experiment 2 with the goal of testing our hypothesis with greater power.

Figure 5: r probability estimates in all four quartiles of recognition and rejection latency distributions for all reanalyzed datasets and for Experiment 2. Error bars represent standard errors.

3.2 Experiment 2

3.2.1 Participants

To provide an appropriate balance between type-1 and type-2 error rates in χ ² model tests (Erdfelder, 1984; Moshagen and Erdfelder, 2016), we recruited 52 students (35 women) from the University of Mannheim aged between 18 and 45 (M = 22.38, SD = 5.49). Participation was rewarded either with a monetary compensation (2 euros) or with study participation credits. Additionally, for each correct response in the comparison task, participants gained 2.5 cents, and for each incorrect response they lost 2.5 cents.

3.2.2 Material and procedure

The experiment consisted of the city-size paradigm, involving two tasks. First, participants had a recognition task, where they saw 60 city names and had to indicate whether they recognize them or not. Naturally, response times were recorded along with the recognition judgments. The 60 cities were a random selection from the largest world cities (with over 3 million inhabitants; http://en.wikipedia.org/wiki/List\_of\_cities\_proper\_by\_population). After the recognition task, cities were paired according to their recognition and rejection latencies, with the cities having similar recognition or rejection latencies assigned to the same set. More precisely, there were four subsamples of pairs, created according to the corresponding four bins of individual recognition and rejection latencies. Whenever the number of recognized objects (or the corresponding number of objects judged unrecognized) was not divisible by four, it was randomly decided which bin(s) would have one object more than the other(s). After the pairs were created (the number of pairs across the four sets varied between participants, being either 420, 421 or 422), participants saw them and had to decide for each pair which city was more populous. As in the analysis of published data sets, a pair of cities was considered for subsequent analysis only when (a) one city was recognized and the other was not and (b) the corresponding individual recognition and rejection latencies fell in the same quartile of the response time distribution.

3.2.3 Results

Before fitting the model, we removed one participant because he recognized only one of the 60 cities, while the remaining participants recognized on average 57% of the objects. With the data from the remaining 51 participants, we determined the frequencies for each category of the model, separately for the four bins of data. Then, we fitted the r-model to the four bins of data. The model performed well in describing the data (G ²(4) = 7.44, p = .11, FIA = 65.49). We repeated the same analysis that we performed with the published data sets, with the goal of testing our order hypothesis on the parameters r ₁ to r ₄. As can be seen in Table 5 and Figure 5, we again found support for our hypothesis.¹² Additionally, the 95% confidence interval of the probability estimates of r ₁ did not include 1, which again shows that even under ideal conditions for reliance on memory states alone, other strategies than mere reliance on memory strength take place.

4 Discussion

When they introduced the MSH, Erdfelder, Küpper-Tetzel and Mattern (2011) contributed to the RH literature by providing an extension of the heuristic that parsimoniously links it with the recognition memory literature. The MSH not only explains a lot of previously problematic results but also provides a set of new predictions. While Erdfelder, Küpper-Tetzel and Mattern (2011), Castela et al. (2014), and Castela and Erdfelder (2017) tested several of these predictions and already gathered some support for the MSH, a crucial additional empirical prediction regarding decisions between pairs of unrecognized objects (“guessing cases”) has not been addressed so far. Our primary aim was to close this gap and, in addition, to provide further evidence on MSH predictions regarding pairs of recognized objects (“knowledge cases”) and mixed pairs of one recognized and one unrecognized object (“recognition cases”), conceptually replicating and extending previously published results on recognition judgment latency effects in binary decisions. We addressed both of these issues in two studies by reanalyzing previously published data sets and conducting two new experiments. In this way, we found strong converging evidence in line with the MSH.

In our first study, by relying on recognition and rejection latencies as a proxy for memory states — under the assumption that longer latencies are associated with the uncertainty memory state while shorter latencies are associated with certainty states — we found evidence for the MSH prediction that for knowledge and guessing cases people also have a preference for objects that are likely to be in a higher memory state. While for knowledge cases the MSH prediction overlaps with predictions of the fluency heuristic (Hertwig et al., 2008), the prediction regarding guessing cases cannot be accounted by any other framework we are aware of. Furthermore, that latter prediction is quite counterintuitive, since it maintains that objects recognized more slowly, and thus judged less fluently, should be preferred in guessing cases. Obviously, this prediction conflicts with the popular notion that cognitive fluency boosts choice preferences (e.g. Schooler and Hertwig, 2005; Zajonc, 1968). Nevertheless, we found unequivocal evidence for our prediction.¹³

It is also important to note that the MSH not only captures the preference effects for knowledge and guessing cases correctly, but also predicts they should be much smaller than the corresponding effects in recognition cases. This is due to the fact that, in knowledge and guessing pairs, the objects can only be either in the same memory state or in adjacent memory states. Therefore, the preference for the object in a higher state should be less marked than in cases where the distance between states is maximal (pairs of one object in recognition certainty and one object in rejection certainty), a combination that can only occur for recognition pairs. Note that this prediction cannot be derived from the fluency heuristic theory (Schooler and Hertwig, 2005), simply because the latter considers knowledge cases in isolation. Hence, the MSH theory not only makes more predictions than the fluency heuristic theory, it also makes more precise (or “specific”) predictions. In other words, the MSH theory has larger empirical content (in the Popperian sense) compared to the latter theory (Glöckner and Betsch, 2011, see, e.g.,). We thus believe the MSH presents itself as the most parsimonious framework for understanding how recognition is used in binary inferences, clearly outperforming other heuristic-based approaches, like the RH and the fluency heuristic, in its explanatory power, empirical content, and empirical scope.

One result worth noting is that the effect of latencies was stronger for knowledge cases than for guessing cases. While we had not predicted this explicitly, it fits nicely with previous results. Specifically, Castela and Erdfelder (2017) observed that MSH-use is higher for recognition pairs if one object is in recognition certainty and one object in the uncertainty state than for recognition pairs with one object in uncertainty and one object in the rejection certainty state. Since these are the memory state combinations that can underlie adjacent state cases within knowledge and guessing pairs, respectively, our current results seem to be exactly in line with what was found by Castela and Erdfelder — a stronger tendency to use the MSH in the former cases. Given the converging evidence concerning this effect, future studies should focus on testing possible explanations for it. One such explanation, already suggested by Castela and Erdfelder, is that the effective distance in memory strength between the recognition certainty and uncertainty memory states might be larger than the corresponding difference between the uncertainty and rejection certainty memory states. This would suggest that a simple ordinal description of the states might be insufficient.

With our second study, we aimed at further testing the effect of recognition and rejection latencies on choices for recognition pairs. While this is largely a conceptual replication of the test carried out by Erdfelder, Küpper-Tetzel and Mattern (2011), we relied on a different measure of RH-use, which we believe is more adequate. Erdfelder, Küpper-Tetzel and Mattern relied on accordance rates which, as explained above, are a confounded measure, since people might choose the recognized option for reasons other than reliance on recognition, namely because they rely on further knowledge. For this reason, Hilbig, Erdfelder and Pohl (2010) proposed the r-model, and specifically the r parameter of the model, as a better measure. The main advantage is that the r-model disentangles choices of the recognized option that originate from reliance on recognition from the ones stemming from use of further knowledge. Extending the scope of tests previously carried out by Erdfelder, Küpper-Tetzel and Mattern (2011), we additionally investigated the prediction that MSH-use as indexed by the r parameter should increase the shorter the recognition and rejection latencies of objects in a pair. We found support for this hypothesis by reanalyzing 9 data sets and, in addition, with a new experiment tailored exactly to this test. Furthermore, we assessed whether in the most extreme cases, that is, when the recognition and rejection latencies were shortest and therefore the probabilities that both objects are in recognition and rejection certainty states were highest, MSH-use would be the only strategy used. The 95% confidence intervals for the corresponding r ₁ parameter estimates did not include 1 in any of our data sets, suggesting that this is not the case. Hence, even under perfect conditions for relying on memory strength, people will sometimes resort to other inference strategies and integrate further knowledge. Overall, our results are in stark conflict with the recognition heuristic (RH) theory as originally proposed by Goldstein and Gigerenzer (2002) and in line with the MSH theory.

Recently, Heck and Erdfelder (2017) also criticized the RH framework, but from a different perspective than the MSH framework does. It thus seems worthwhile to consider their work in a bit more detail and compare it with our current work. Using an extension of Hilbig et al.’s r-model to response times as an innovative measurement model, Heck and Erdfelder (2017) showed that the decision latency predictions of the RH are in conflict with virtually all available data on RH use in natural decision domains. Only a small proportion of individual data sets could be adequately described by a serial RH theory according to which recognition vs. non-recognition is considered as the first cue in binary decisions with probability r, possibly followed by consideration of further knowledge about the recognized object with probability 1−r. The vast majority of individual decisions could be described better by an information integration account as formalized in the parallel constrained satisfaction (PCS) model advocated by Glöckner and Betsch ( 2008; see also Glöckner and Bröder, 2011, 2014, and Glöckner, Hilbig and Jekel, 2014). According to the PCS account, the recognition cue and further knowledge cues are always considered simultaneously, resulting in fastest decisions when all cues are congruent, that is, when both recognition and further knowledge suggest the choice of the same object.

What are the implications of Heck and Erdfelder (2017) work for MSH research as addressed in the current paper? An immediate implication is that the r parameter of the r-model should not be interpreted as the probability of applying a serial heuristic as presented by Goldstein and Gigerenzer (2002) or as part of the take-the-best heuristic that considers recognition always as the first cue in binary decisions (Gigerenzer and Goldstein, 1996). This is unproblematic for our current work since we consider r as a proxy for use of the MSH in which the order of cue processing is left unspecified. In our application, r just represents the probability of noncompensatory reliance on recognition in the sense that the influence of recognition dominates the joint influence of all further knowledge cues. Note that this is not in conflict with a parallel information integration account, as the weight of recognition in a PCS model can be so high that the influence of recognition cannot be overruled by any combination of other decision cues with (much) smaller weights. Thus, the r parameter of Hilbig, Erdfelder and Pohl’s (2010) r-model (and also the corresponding parameter of the r-s-model, cf. Hilbig et al., 2011) can still be interpreted as a measure of noncompensatory reliance on recognition if noncompensatory reliance on recognition is not confused with reliance on recognition alone (i.e., as a serial cognitive strategy predicted by the RH theory).

Now let us consider the reverse question: What are the implications of our current MSH research for the PCS model of recognition-based decisions advanced by Heck and Erdfelder (2017)? In fact, the latter model shares one weakness with the RH theory, namely, that recognition is considered as a binary cue only. Although this simple parallel model suffices to explain a number of results that the RH cannot explain (as shown by Heck and Erdfelder), it has difficulties in explaining some of the results that the MSH can account for. For example, we cannot see how to explain the choice preference for the object judged unrecognized more slowly in guessing pairs using a PCS model with only a single dichotomous recognition cue as assumed in Heck and Erdfelder (2017, p. 446, Fig. 2). If anything, then such a PCS model would need to be extended to include several nodes representing differences in recognition information. For the time being, however, the MSH appears to be the only model that captures the preference for the option judged unrecognized more slowly. Recall also that Castela et al. (2014) found a preference for choices of recognized objects for which participants reported having further knowledge as compared to objects judged merely recognized. This result is predicted by the MSH (assuming that virtually all objects with further knowledge are in the recognition certainty state), whereas it poses difficulties for information integration accounts like PCS. For fixed weights of the cues this model would predict more reliance on recognition when further knowledge other than recognition is not readily available.¹⁴

In sum, with our work we tried to answer questions left open by Castela et al. (2014), Castela and Erdfelder (2017), and Erdfelder, Küpper-Tetzel and Mattern (2011), thereby accumulating further support for the MSH. We believe we achieved this goal in two different ways: First and primarily, by finding support for its bold predictions for guessing cases and in this way showing how it can parsimoniously explain a much larger chunk of data than the RH or the fluency heuristic can; second, by conceptually replicating and finding converging support for its predictions regarding knowledge and recognition cases that have larger empirical content than those derived from the RH or the fluency heuristic. Finally, our results also show that while the MSH appears to be a more useful framework than the RH, it should not be understood in a deterministic way, since even when the objects are (likely to be) in the two extreme memory states — recognition certainty and rejection certainty — people sometimes resort to strategies other than choosing the option in a higher memory state.