Introduction

The capacity to learn from experience is critical for survival and therefore ubiquitous across most organisms. An important type of learning, called statistical learning, involves sensitivity to structured patterns in the environment1,2,3,4. Patterns come in many varieties, such as how certain objects typically appear (e.g., animals, trees, and human-made objects each have a particular form and shape) to the regularities found in temporal input streams such as music and language (e.g., hearing random notes on a piano would not be considered music, just as a random concatenation of speech syllables would sound meaningless).

Decades of research on statistical learning, and its related cousin, implicit learning5,6, has demonstrated that humans (and other organisms) are sensitive to spatial and temporal patterns across many perceptual, motor, and cognitive domains7,8. This type of learning appears to occur relatively automatically, without intention to learn, and results in behavioral, cognitive, and/or perceptual facilitation that is often (though not always) accompanied by a lack of conscious awareness of the learned patterns. Because of the apparent reach of statistical learning to explain functioning across a myriad of domains and situations, it is perhaps not surprising that this area of research has been exploding in recent years9,10.

However, recent critiques have highlighted challenges facing statistical learning research. Namely, the construct of statistical learning itself is underspecified in terms of its underlying neurocognitive mechanisms, how these mechanisms relate to various laboratory tasks commonly used to measure statistical learning, and what is the relationship between statistical learning and other forms of learning, memory, and cognition9,11,12,13,14. While some theoretical frameworks do exist, including the extraction-integration framework15,16,17,18, chunking approaches9,19,20, frameworks favoring the role of brain plasticity and modality constraints2,7,21, and other multi-component models22,23,24,25, there is no universally-agreed upon theory that can adequately address these issues.

We contend that these challenges are not unique to statistical learning research but in fact are difficulties confronting the psychological sciences more broadly. It has been argued that the field of psychology faces a “theory crisis”; that is, psychological theories are often poor at explaining the relevant phenomena of interest and no amount of improved statistical techniques or replication studies will improve the situation26,27,28. Whereas efforts have been made to encourage psychologists to create more formal or precise theories29, Eronen and Bringmann30 have recently argued that “the core of the problem is that developing good psychological theories is extremely difficult” (p. 780). Therefore, understanding the obstacles facing theory building in psychology is the first necessary step for making progress in addressing the theory crisis.

In this paper, we use Eronen and Bringmann’s30 and others’ insights about the theory crisis as a foundation for moving the field of statistical learning forward, one that has seen an explosion of growth in recent years9. Statistical learning, and even implicit learning, have historically been challenging to define, as different tasks may elicit different cognitive processes and engage distinct neuroanatomical systems (for a review, see ref. 14). Therefore, for the purposes of this paper, we define statistical learning prototypically, as encompassing situations that tend to have a specific set of characteristics, in particular, those that involve the learning of patterns in the environment, through multiple exposures, and under incidental conditions7. Note that this definition can be thought of as a higher-level description about what problem or goal the system has evolved to address—i.e., to learn environmental patterns over multiple exposures under incidental conditions—rather than a definition predicated on identifying the learning processes or neural mechanisms involved. This higher-level description is similar to Marr’s31 top level of analysis (what he calls the “computational” level), whereas descriptions of the cognitive and neural mechanisms fall under Marr’s 2nd (“algorithmic”) and 3rd (“implementational”) levels, respectively (see Conway et al.32, for elaboration of Marr’s framework in relation to statistical learning). The advantage of starting with such a higher-level definition is that it is based on the characteristics of the learning situation, rather than on the characteristics of the learner, which are “unseen” cognitive and neural processes that can be difficult to identify14. Thus, this definition provides (what we believe is) a relatively uncontroversial description of the types of situations that statistical learning typically encompasses and that clearly specifies to what extent any given task or situation is relevant to this area of research.

With this definition in place, the ultimate aim should be to derive a theoretical framework that provides meaningful and valid conceptualizations of the various phenomena that statistical learning touches upon. To work towards this goal, in the following sections we first describe three key issues as laid out by Eronen and Bringmann: lack of robust phenomena, issues with construct validity, and issues with establishing causality. Next, we discuss each issue in terms of how it relates to statistical learning by focusing on several research phenomena that have been examined in some detail. In the final section of this paper, we offer suggestions for ways to address the theory crisis in statistical learning research to move the field forward in fruitful directions.

Challenges facing SL research: lack of robust phenomena

Eronen and Bringmann30 contended that there is a lack of robust phenomena that can be used to constrain and formulate psychological theories. To understand this problem fully, it is necessary to describe the difference between data, phenomena, and theory29,30 (see Fig. 1). Data are the raw measurements collected in psychology experiments (i.e., measures of behavior or neurophysiology). Phenomena are a description of the effects compiled across the observed data; an example in psychology might be the Stroop effect33,34 or implicit bias/stereotype effect35, which are descriptions of consistently occurring patterns within data. Finally, theories are created to explain the phenomena; and yet, phenomena also constrain theories30. It is this last point that Eronen and Bringmann30 claimed has received little attention in psychology. Without the existence of robust—i.e., consistent and stable—phenomena, then theories are on shaky ground.

Fig. 1: The theory of evolution as a model for good theory building.
figure 1

Eronen and Brigmann30 emphasize the distinction between data, phenomena, and theories. Given the lack of universally accepted theories within Psychology, we turn instead to the theory of evolution to provide concrete examples of the interplay between data, phenomena, and theory. Charles Darwin did not set out with the goal of developing the theory of evolution. Rather, he initially made a wide range of different, at the time disparate, observations, including, for example, that the large ground finch primarily ate seeds (which happened to have a large, powerful beak), while the green warbler finch (with a narrower, pointed beak) fed upon insects. These observations represent data (individual images in the figure). Based on many pieces of data, more general phenomena emerged (e.g., that animals are well adapted to their environments; that offspring inherit the traits of their parents; and that differential survival rates leads to changes in how a given phenotype is represented in a population). Finally, Darwin advanced the Theory of evolution by natural selection151, which explained and unified these phenomena. Critically, a good theory, like evolution, must be able to explain all the relevant phenomena, and remains open to falsification if counter examples are found. In this way, not only do theories explain phenomena, but phenomena tightly constrain the space of possible theories. Note, the theory of evolution was selected as an example as it is extremely well developed and supported by data. The development of any new theory, including a theory of statistical learning, is likely to be less clear and much noisier. For example, it is yet to be determined whether all of the phenomena outlined in this article (or other phenomena, not discussed here), are sufficiently robust to be included in an eventual theory. Moreover, it is possible that a theory may arise that explains some, but not all of these phenomena. These difficulties represent some of the key challenges of theory building, but we hope that the proposals made in this article will help advance such theories in the future. The images of animals within the figure were created using BioRender.com. Wilson, B. (2025) https://BioRender.com/5iaoy0f.

Full size image

Is there a lack of robust phenomena in statistical learning research? What are the relevant phenomena in the first place, and what is the metric for evaluating whether they are robust? Robustness of a phenomenon is at least partly related to the notion of replicability; if a particular pattern of data is not observed consistently across studies, then it is not a robust finding. However, as pointed out by Eronen and Bringmann, it is just as important that for a phenomenon to be robust, it should be “verifiable and detectable in several independent ways and not dependent on a specific theoretical framework or observation method”30,36,37 (p. 780).

Based on this conceptualization of robustness, we suggest that the overall phenomenon of statistical learning itself, that is, the general finding that people are sensitive to patterns in the environment under incidental conditions, is robust, demonstrated across many studies in a multitude of situations using various laboratory tasks and stimuli38. While this general phenomenon appears robust, it is also important to delineate a number of additional sub-phenomena that have been examined in the statistical learning literature, including but not limited to: modality effects39, effects of input complexity40, age effects41, species differences42, role of attention in learning43, competition between executive functions and statistical learning44, the contribution of statistical learning to language acquisition13, and atypical learning in developmental disorders45 (see also7, for a review of some of these phenomena). In subsequent sections we provide an evaluation of the robustness of several of these key phenomena.

The point is that a good theoretical framework of statistical learning should not only encompass the general phenomenon of statistical learning itself, but also these additional sub-phenomena (e.g., the presence of modality effects)—but only those that are considered robust. If a phenomenon is not a robust one, then it may be counter-productive to devise theoretical frameworks to explain it. Attempting to devise theories to explain phenomena that are not robust could lead to the unfortunate situation where “correct” theories are rejected because they fail to account for these (unrobust) phenomena. We return to this point at the end of the paper when discussing recommendations. It is therefore imperative to identify which phenomena appear to be robust and which ones need additional empirical support or testing. We discuss the robustness of several such phenomena, in Section “Evaluating Statistical Learning Phenomena”, below.

Challenges facing SL research: issues with construct validity

The second challenge facing the psychological sciences, highlighted by Eronen and Bringmann30 concerns issues with construct validity. A simple definition of construct validity is that a measurement measures what it is intended to measure46. Whereas reliability can be easily quantified and tested, construct validity is more nebulous. As such, psychological scientists typically report little evidence for validity, instead focusing on more concrete psychometric properties such as reliability47,48. For instance, statistical learning researchers have made strides to improve the reliability of the measurements used49,50,51,52. This is an important and necessary step. However, while issues of validity have been mentioned in the statistical learning literature, including the use of vague or circular definitions for statistical learning and/or the use of different terms and a multitude of ways to measure it9,11,14,53, no formal attempts at establishing construct validity have been made.

A further challenge to assessing the validity of measures in statistical learning research is that different types of methods have been used to assess statistical learning (see refs. 11,53), including the following tasks: artificial grammar learning (AGL)5, serial reaction time (SRT)54, Hebb repetition55, triplet segmentation or embedded pattern tasks4, cross-situational learning56, and contextual cueing57. Not only this, but even the same paradigm (e.g., AGL) can vary in terms of the type of input used or how learning is measured. The challenge with using such an array of task methodologies is that, while all of them on the face of it appear to measure statistical learning, each one may emphasize a different type or aspect of such learning, and potentially be supported by different underlying neurocognitive mechanisms.

As pointed out by Eronen and Bringmann30, weak construct validity impacts robustness of phenomena. Because phenomena depend upon inferences made over data, if those data are not well-understood or well-validated, then the phenomena themselves are unlikely to be robust.

Challenges facing SL research: issues with establishing causality

The final challenge raised by Eronen and Bringmann30 is that determining causality among psychological variables is extremely challenging. As they note, good theories should “capture causal relationships between psychological variables” (p.783). However, doing so is extremely difficult58. While it is well-understood how to establish causality by manipulating external, non-psychological variables using sound experimental design (e.g.,59 psychological theories need to understand the causal relationships between psychological variables, which are much more difficult to directly observe and manipulate. Eronen and Bringmann30 refer to this as the “fat-handed” problem: to manipulate a psychological variable requires manipulating an external variable, but the manipulated external variable might be affecting other psychological processes/variables instead of or in addition to the specific psychological variable of interest.

To take an example from statistical learning research, one ongoing question is to what extent statistical learning depends upon attention (e.g.,7,60). One common way to manipulate participants’ attention is to use task instructions to direct selective attention to some stimuli but not others and to test if statistical learning still occurs for the unattended stimuli (e.g.,43). While such an approach is commendable and logical, the question is whether other psychological variables apart from attention may also be affected. For instance, it is possible that variables such as motivation, confidence, and/or effort may also be affected by the task instructions, and if so, it is difficult to know to what extent attention itself (or lack thereof) causally affected learning. Note that the problem of causality is exacerbated by the fact that different statistical learning tasks may depend on different underlying cognitive or neural processes. In sum, because psychological variables can only be measured indirectly, it is hard to verify which variables have specifically been changed and manipulated in any given paradigm.

Evaluating statistical learning phenomena

As we have seen, the challenges facing statistical learning research make it difficult to formulate accurate and reliable theories. To further illustrate how these issues pertain to statistical learning research, we focus on four phenomena that have been examined to varying degrees in the research literature: modality effects; age effects; species differences; and atypical statistical learning in developmental disorders, such as dyslexia. For each we consider (1) the robustness of the phenomena, (2) potential issues with construct validity, and (3) challenges with establishing causality. With this approach, we are advocating a “bottom-up” stance for theory-building. That is, before creating theories to explain particular phenomena, it is necessary to have a solid understanding of the robustness of the phenomena themselves, which includes assessing their validity and potential causal relationships between psychological variables.

Modality effects

Statistical learning has been observed across a range of scenarios, from auditory61 and visual scene analysis10 to language processing62 and motor learning63,64. However, comparing across modalities has indicated that the conditions that optimize statistical learning vary depending on both modality (e.g., visual vs. auditory) and domain (e.g., linguistic vs. nonlinguistic)39,65,66,67,68,69, for a summary, also see ref. 70. These modality-specific representations may exist alongside modality-independent representations and domain-general mechanisms2,71, although this remains unclear72.

Robustness

A key theory2 of statistical learning posits that statistical learning exists across modalities but has modality-specific constraints. This is supported by consistent evidence for differences between modalities; for example, performance in the auditory modality is typically superior to the visual modality for the serial presentation of materials39,67, especially at faster rates65,66. On the other hand, the auditory advantage seems to disappear when the visual stimuli are simultaneously presented65,67. There are also asymmetries across modalities for the same domain, for example, visual tasks show a non-linguistic advantage while auditory tasks show a linguistic advantage73. Furthermore, modalities appear to operate with relative independence, showing limited transfer74 (though also see ref. 75 for a possible exception and Milne, Wilson & Christansen70 for further discussion). Additionally, there is a lack of correlation across modalities52,76,77 as well as different developmental trajectories for statistical learning in different modalities78. The presence of modality differences across a broad range of tasks, stimuli and research groups speaks to the robustness of this phenomenon. However, the same theory2 suggests the presence of a unitary general statistical learning mechanism, such as the hippocampus, yet the existence and nature of such a mechanism is still unclear (see Bogaerts et al.72 for a discussion on whether a “good statistical learner” exists).

Construct validity

Modality effects may be greatly influenced by construct validity. For example, what is the best way to measure statistical learning in each modality to provide a “fair” comparison between them? That is, how can we make methods comparable across modalities, while recognizing that there are some fundamental differences across sensory systems? At a minimum, discriminability of the individual items should be comparable, as Conway & Christiansen39 ensured in a pre-training phase before exposure to the statistical input stream. Even when such methodological controls are used to equate stimuli across modalities, it is difficult to isolate statistical learning itself from other cognitive processes. For example, if modality differences are observed, is it due to statistical learning differences, or the role of working memory79,80,81 or attention mechanisms across modalities (e.g., object-based attention82)? Alternatively, many participants impose a linguistic code on visual and auditory stimuli during statistical learning tasks68, which makes it difficult to isolate pure auditory and visual learning. Relatedly, if participants are using explicit strategies during training and testing, these could both mask or exacerbate differences in modalities; this may particularly be the case for commonly used two-alternative forced-choice tasks6,72. Finally, many publications comparing across modalities study individual differences; therefore ensuring the retest-reliability of the methods, and that they capture sufficient variance, is especially critical51.

Causality

If modality effects exist, we also need to establish what is driving them to inform any theoretical framework. Establishing causality relates closely to the issue of construct validity. While construct validity asks us to ensure we are measuring what we think we are measuring, delineating the related perceptual factors, cognitive functions and potential strategies will also inform us on how the differences emerge and their relationship with statistical learning. In this process, we should also consider the fundamental reason why statistical learning exists. In most cases, this is to influence the way we interact with the world, be that in terms of the immediate allocation of cognitive resources (e.g.,83) or the long-term acquisition of speech regularities15. Better appreciating why the brain may have adapted to use statistics in a given modality could help to understand the differences or similarities in the causal relationships between statistical learning and other psychological variables.

Age effects

There are three ideas that have been advanced regarding how statistical learning abilities may interact with age. First, it has been suggested that statistical learning is age-invariant, with some studies showing that learning was unaffected by age18,41,84,85. Second, like many other cognitive abilities, statistical learning might improve with age86,87,88,89,90,91 (however, such an effect may be due to changes in other cognitive abilities such as working memory, which are known to improve with age92 : see “Causality” below). Third, given the importance of statistical learning in language acquisition, it has been argued that infants and children show better statistical learning compared to adults. Infancy and early childhood are where the majority of language learning typically takes place, with some suggestion that language learning abilities decline with age93,94,95.

Robustness

Longitudinal studies are especially important for understanding potential age effects in statistical learning; indeed, a recent longitudinal study provides support for better statistical learning in younger compared to older children between the ages of 7 and 1496. However, there is at least some evidence supporting all three possibilities for how statistical learning and age interact (age-invariance, childhood superiority, and adult superiority), effectively weakening the robustness of all of them. Complicating matters further, there is also evidence that any age-related changes in statistical learning may be influenced by modality effects: Raviv and Arnon41 found improvements in visual statistical learning between 5- and 12-year-olds, but no changes in statistical learning in the auditory domain. However, in a similar study comparing visual statistical learning with auditory statistical learning of non-linguistic stimuli, improvements in both visual and auditory statistical learning were found between 5- and 12-year-olds97. This suggests that any modality-specific differences may be further complicated by differences between stimuli, both of which may affect the conclusions that can be drawn from the literature regarding the developmental trajectory of statistical learning. In sum, the nature of developmental changes across age is not a robust phenomenon, further complicated by the apparent influence of modality- and stimulus-specific effects on the development of statistical learning. On the other hand, there is good evidence that humans of varying ages (infants through adults) are all capable of statistical learning, so the presence of statistical learning across ages appears to be a robust finding.

One critical methodological consideration that specifically impacts our ability to determine changes in statistical learning across development relates to the “more room to improve” effect98. This refers to the difficulties that occur when trying to compare different age groups who have different baseline reaction times. While high variability between age groups is often controlled for by using standardized scores, in learning experiments specifically, this variability is actually a critical aspect of the learning process. Specifically, individuals who show greater learning effects typically show greater variability relative to poorer learners. Therefore, using standardized scores may obscure key learning differences, rendering comparisons between age groups uninformative. More recently, efforts have been made to overcome such challenges, for example, by controlling for average speed when using reaction time-based measures of learning98.

Construct validity

There are issues relating to the construct validity of statistical learning that are particularly important to consider when measuring people of different ages. While statistical learning in infancy was originally measured using preferential-looking paradigms – which offer an implicit method of measuring learning when the participants are not able to provide explicit judgements—when measuring learning in adults and children, learning is typically evidenced through above chance performance in a grammaticality judgement or familiarity task. In these tasks, participants are asked to provide explicit judgements on whether a sequence fits or breaks a pattern to which they have previously been exposed. Because these tasks measure learning directly by asking participants to classify sequences (often based on ‘gut instinct’), performance may be affected by additional cognitive abilities, such as understanding the task instructions and decision-making skills, which are known to improve across development99. As an example, “yes” biases—in this case a preference for categorizing sequences as grammatical—are commonly reported in AGL tasks with children100,101, which may suggest that children have a lower threshold for classifying sequences as grammatical, that they pick up and use irrelevant features of the sequences to judge their grammaticality, or that they simply do not understand how to complete the task. In any case, it is likely that such explicit measures of statistical learning may underestimate children’s abilities, which makes comparing statistical learning across development using these tasks difficult.

More recently, alternative tasks have been developed that assess statistical learning indirectly, by measuring other variables that are facilitated by statistical learning. For example, SRT tasks102,103, serial recall tasks104,105,106, and tasks measuring neural entrainment107 can assess statistical learning without requiring explicit reflection on what has been learned. Indirect measures of learning also typically provide additional benefits over direct measures in that they can measure learning over the course of the task, while learning is taking place. Therefore, indirect measures of learning may provide a useful method of comparing statistical learning abilities across age, by avoiding the requirement for more explicit processes which may not accurately represent children’s performance. Additionally, these tasks provide information about the trajectory of learning across the task, which could reveal less obvious differences in learning patterns between children and adults.

Causality

As we have discussed, the developmental trajectory of statistical learning, from infancy to adulthood, is already somewhat unclear, both due to a lack of robust age effects and the issues surrounding the validity of the experimental measures used. These issues are compounded by (or possibly caused by, at least in part) challenges with understanding causal relationships between different psychological variables. Aside from statistical learning, myriad cognitive processes change and develop as we age, including attention, memory, executive functions, and language. It is very likely that these abilities contribute to performance on statistical learning tasks, and each of these, and their interactions, could impact the results of developmental studies. While establishing the causal relationships between different psychological processes is a challenge for all statistical learning research, this issue is compounded in developmental research, due to the co-development of so many other aspects of cognition. However, examining changes in other statistical learning phenomena across development using longitudinal studies may actually aid attempts to establish causality. By measuring these other related cognitive functions alongside statistical learning at various timepoints in development, we may be better placed to draw conclusions regarding which aspects of cognition – or other environmental factors such as education – influence the development of statistical learning.

Species differences

Over the last two decades, much comparative research has investigated the extent to which statistical learning abilities might be shared by species other than humans. This research has used a wide range of methods (see Construct Validity, below), but here we will primarily focus on those that assess learning under incidental conditions, particularly in nonhuman primates, as opposed to explicit, operant training (which is common in birds and rodents). These approaches have been used to demonstrate statistical learning abilities in a wide range of primates (tamarins108,109,110, marmosets21,111, squirrel monkeys112, rhesus macaques68,111,113,114, baboons115,116,117, and chimpanzees118,119). Beyond the general existence of statistical learning in nonhuman animals, considerable amounts of comparative research have focused on the specific types of statistical regularities that animals are able to learn, with some studies suggesting there may be differences across species.

Robustness

The general phenomenon that statistical learning is conserved in other primates (and likely more broadly in the animal kingdom), appears to be highly robust42,120. However, while all animals tested learn relatively simple ‘adjacent dependencies’ (where stimuli co-occur immediately in space or time), ‘nonadjacent dependencies’ (in which intervening elements separate these stimuli), may prove more difficult to learn (for a review, see121). Moreover, more complex classes of stimuli (e.g., governed by supra-regular grammars, e.g.,108) have typically proved too complex for nonhuman primates to learn, at least under incidental conditions108,122 (although for a successful demonstration using operant approaches, see123). Therefore, the central phenomenon, that core statistical learning abilities exist across the primate taxa (and likely more broadly in the animal kingdom), appears to be highly robust. However, as the complexity of these statistical relationships increases, species differences become pronounced. Here, it is important to note that these results may not represent differences in statistical learning per se, but may instead relate to other species differences (e.g., perception, attention, motivation, etc.) as will be discussed in ‘Causality’, below.

Construct validity

A central challenge of studying animal cognition is the requirement that tasks are both feasible in nonhuman animals and also accurately measure the constructs of interest, in this case, statistical learning. Early research in nonhuman primates made use of paradigms originally developed to test preverbal infants, such as preferential looking or habituation/dishabituation paradigms measuring head turn responses21,108,109,110,112, or equivalent responses using eye-tracking68,111,114. While these approaches allow reasonable comparisons of monkey and infant data, most adult human statistical learning research uses alternative methods (e.g., grammaticality judgement tasks), which are not feasible in nonhuman animals. A second popular method for testing nonhuman primate statistical learning is the use of SRT tasks, which have been combined with a range of different artificial grammars (e.g.,113,115,116), in which animals make a series of rapid responses with a joystick or on a touchscreen. These tasks can allow direct comparisons with humans (as in ref. 116). However, SRT tasks may rely on somewhat different, more procedural learning processes compared to preferential looking paradigms, raising further challenges in assessing construct validity and particularly to drawing comparisons across studies and experimental designs. Additionally, adaptations to these designs allow the investigation of how primates use statistical regularities to encode sequences of stimuli into increasingly large ‘chunks’, over the course of training124,125. Given this range of approaches, many cross-species comparisons (either within a single comparative study (e.g.,111,114), or those aggregating across multiple studies) are based on disparate methods, which may, at least to some extent, measure different constructs. In cases where similar patterns of learning are observed, despite such methodological differences, it may be reasonable to conclude that not only do those similarities in statistical learning exist, but that they are also robust even to these methodological differences. However, when differences are observed, this leads to difficulties in concluding whether it is animals’ statistical learning abilities that differ, or whether different methods and approaches used across species may actually measure somewhat different psychological constructs.

Causality

The clearest overarching conclusion to be taken from the comparative literature on statistical learning is that many species are capable of learning at least relatively simple statistical relationships. However, when statistical regularities become more complex, species differences appear to emerge. This raises the challenging question of determining whether these differences result from more limited statistical learning abilities in nonhuman animals (as has often been concluded) or whether other differences in perception, motivation, attention, or other psychological variables, may contribute to these effects. One interesting demonstration of this may be the recent move to testing nonhuman primates using more operant training approaches. Using traditional, incidental learning approaches monkeys have not been shown to learn complex, supra-regular grammars. However, using an explicit, operant training approach in which rhesus macaques were rewarded for correct responses over many thousands of trials, monkeys were capable of learning these types of grammars123,126. It is highly likely that these tasks recruit different learning systems, but these task differences also introduce many other differences, including perceptual differences as well as different levels of motivation and attention, due to the promise of reward for correctly completed trials. This highlights the interplay between different psychological variables and the challenges of isolating statistical learning from other factors. While this specific example applies to learning in nonhuman primates, the same issues of perception, motivation, attention, etc., are likely relevant to human studies of statistical learning, and represent one of the major challenges to theory building in statistical learning.

Atypical statistical learning in developmental dyslexia

Finally, there has been much interest in examining whether statistical learning is atypical in different developmental disorders related to language and communication. Because statistical learning appears to be crucial for learning spoken and written language in typical development11, it is possible that certain language and communication disorders might arise from atypical statistical learning127. Developmental dyslexia is one such disorder that has been examined through the lens of statistical learning (e.g.,128,129,130,131). Note that while we focus here on developmental dyslexia, some of the same questions about robustness, validity, and causality are likely relevant for many other developmental disorders.

Robustness

Based on several meta-analytic reviews, which can be useful for systematically assessing robustness across studies (e.g.,38), the extent of a statistical learning impairment in dyslexia remains unclear. On the one hand, Lum et al.’s132 systematic review showed evidence of a statistical learning impairment in those with dyslexia, based on studies using SRT tasks in adults and children. Likewise, van Witteloostuijn et al.’s133 meta-analysis also found evidence of impaired learning in dyslexia, based on studies using AGL paradigms. However, their analysis suggested the presence of publication bias, which makes it harder to gauge the robustness of the relationship between statistical learning and dyslexia. Similarly, Schmalz et al.’s134 systematic review of statistical learning in dyslexia using both SRT and AGL tasks also noted the presence of publication bias and concluded that there is “insufficient high-quality data to draw conclusions.” Finally, Singh and Conway135 reviewed recent studies that were not included in the three prior meta-analytic and systematic reviews and concluded that there was some evidence for a statistical learning impairment in adults with dyslexia (with 11 of 19 studies showing impairment), but that the evidence for an impairment in children with dyslexia was weaker (with only 6 of 14 studies showing impairment). Singh and Conway135 proposed that some of the inconsistent findings could be due to heterogeneity in the tasks and methods used to assess statistical learning, the heterogeneity of dyslexia itself, and the role of publication bias. Thus, the finding of impaired statistical learning in developmental dyslexia does not appear to be a robust phenomenon.

Construct validity

As mentioned earlier, statistical learning research uses a variety of task methodologies, and it is currently unclear to what extent they measure the same construct. This in turn can lead to discrepancies across studies intending to measure statistical learning in dyslexia. For instance, if two studies examine statistical learning in dyslexia using different tasks, and the findings are inconsistent with each other, it is unclear whether such differences are due to differences in the use of the tasks or some other factor11,12,135. The solution, of course, is not to focus on only one statistical learning task, but instead to carry out systematic investigations of dyslexia using a variety of tasks, chosen based on theoretical reasons53, and used across multiple studies. And as is true for the other phenomena reviewed above, construct validity requires increasing our understanding of how different tasks engage different brain systems and therefore different underlying cognitive processes (e.g., the SRT task is heavily dependent on the basal ganglia136,137, while the triplet segmentation task instead appears to rely on a combination of neocortical sensory networks and the hippocampus2,7).

Causality

One issue for research investigating statistical learning in dyslexia is that the findings are essentially correlational in nature: observing a statistical learning deficit in individuals with dyslexia compared to a control group is consistent with the idea that the learning deficit caused reading problems, but it is not the only possible conclusion. Some other cognitive function (i.e., a “third variable” such as attention or phonological processing) may impact both statistical learning performance and reading ability. Moreover, statistical learning impairments may not cause developmental dyslexia but rather result from it. To establish the causal nature of a statistical learning-dyslexia link would require experimentally influencing statistical learning ability—through the use of neurostimulation (e.g., TMS, tDCS138,139 or possibly dual-task interference tasks140,141)—and then observing how such manipulations impact reading ability (i.e., inducing a “virtual lesion”142). However, such attempts run into Eronen and Bringmann’s30 aforementioned “fat-handed” problem: where direct attempts to experimentally manipulate statistical learning ability may instead affect related cognitive processes such as attention or working memory. Further complicating attempts at causality is the heterogeneity within dyslexia in terms of variability in reading143 and spelling ability profiles30,144. There are also often co-occurring conditions with developmental dyslexia (e.g., attention, executive function, and motor skills143,145,146,147). The heterogeneity and possible comorbidity of developmental dyslexia presents a further challenge to establishing causality, because it is possible that even if statistical learning has a causal influence on reading ability and disability, it may only be the case for certain individuals.

Discussion and recommendations

In sum, the statistical learning phenomena reviewed here vary in their robustness and have similar though also unique issues related to construct validity and causality. Some issues are generalizable to much of statistical learning research (e.g., separating statistical learning abilities from other psychological processes such as memory, attention and motivation), whereas others are more specific to each area (e.g., unavoidable methodological differences when testing different ages or species). As discussed above, the general phenomenon of statistical learning—incidentally acquiring information from environmental stimuli—is highly robust. Based on the present evidence, modality effects, specifically the auditory-serial and visual-spatial advantages for learning, also appears to be a relatively robust finding. Likewise, the presence of statistical learning across age and across nonhuman species also appear to be robust, though it is likely that not all species learn all types of patterns equally well. The extent to which statistical learning changes with age, as well as to what extent statistical learning is impaired in developmental dyslexia, do not seem to be robust phenomena. For all four areas of research, crucial questions about construct validity and causality must be addressed to ensure an accurate understanding of the phenomena in question. While we focused on these four phenomena as ones that have received substantial coverage in the research literature, we acknowledge that many other relevant statistical learning phenomena also exist, and may face similar challenges related to robustness, validity, and causality.

While each statistical learning research phenomenon varies in the extent and nature of these challenges, and thus may require unique solutions, we believe the following recommendations are general enough to broadly deal with many of the most important and common issues facing statistical learning research. We also note that some of these recommendations may be useful for other areas of psychological research that face similar challenges.

There may be value in taking a “bottom-up” approach to theory building

Eronen and Bringmann30 (see also29,148,149,150) suggested that psychology needs more “phenomenon-driven” research, with the focus on identifying what are the robust phenomena, so that we know what findings theories must encapsulate. Discovering new (and robust) phenomena will help us constrain the possible theory space so that we avoid attempting to develop theoretical frameworks to explain findings that turn out to not be robust. Furthermore, as our understanding of robust phenomena increases, this allows one to see abstract patterns across them that would not be otherwise possible without such knowledge. Note, a bottom-up approach to theory-building does not preclude a top-down approach to experimental design or hypothesis-driven research; in fact, these are complementary, not antagonistic, approaches.

Relevant concepts need to be clearly and transparently defined

As mentioned earlier, statistical learning is sometimes defined differently across studies and assessed in different ways11,53, potentially causing serious issues whereby a single term might be used to refer to several quite different cognitive or neural processes14. Until we have a clear and accurate understanding of relevant constructs and how best to measure them, it will be impossible to know what the robust phenomena are and how constructs causally relate to one another. Thus, while much recent focus has been on the reliability of statistical learning measures50,51, and rightly so, the issue of construct validity is also critical to consider. Relatedly, it is important to point out that using different measures for statistical learning is not inherently counterproductive (see next recommendation); however, we should be aware that using different methodological approaches for studying statistical learning may mean we are not measuring the same processes.

The field will benefit from a diversity of statistical learning measures

While it is important to be mindful of methodological differences across statistical learning tasks, a wide range of methods are needed to fully understand the relevant phenomena. As Eronen and Bringmann30 pointed out, robustness is improved if a given phenomenon is observed across multiple methods and measures. Thus, we suggest that future research investigating statistical learning phenomena—even within individual studies—should incorporate multiple measures of statistical learning, to determine the robustness and generalizability of any such findings, rather than focusing on a single task9. A more complete understanding of statistical learning requires a deep exploration of how differences in the type of task, nature of the input, individual differences, and other factors affect the learning process and the resulting knowledge that is acquired. Even so, a diversity of measures is only valuable if we know what they are measuring (i.e., again, good construct validity is necessary).

There is value in methods that can be commonly used across different modalities, ages, species, and populations

While we recommend incorporating a diversity of (valid) statistical learning measures, some areas of research will benefit from the use of common methods. For instance, using equivalent methods and measures may be important when attempting to examine potential commonalities and/or differences across modalities, age, species, or populations. As pointed out above, this may be especially true when interpreting observed learning differences (e.g., across age, or species, etc.). If a difference across age or species is observed when using different methods or tasks, then it is not clear if the differences are due to differences in methods or actual differences due to age or species. For instance, some methods such as the SRT paradigm have been used successfully across species and across much of the human lifespan94,116; similarly, neural measures can also be adapted across different ages, species, in addition to acquiring behavioral data.

Statistical learning should not be studied in a vacuum

To better understand statistical learning phenomena, it is also important to measure as many other relevant processes and variables as possible to see what other factors may be associated with statistical learning9. For instance, several lines of research suggest that there may be a competitive relationship between executive functions and statistical learning71,141. Consequently, if a difference in statistical learning is uncovered between species, or between ages, it is important to determine if such differences are due to statistical learning per se or to differences in executive function, or some other psychological variable that may also relate to statistical learning, such as motivation or attention. To solve this issue requires both incorporating well-validated measures not just of statistical learning but of other relevant psychological constructs and considering the interactions between them.

Interactions among phenomena should be explored

For convenience, the four sample phenomena evaluated above were presented largely independent of one another; however, in reality, the various phenomena interact in potentially complex ways. For instance, as noted above, there is some suggestion that the developmental trajectory of statistical learning is affected by modality, with improvements in visual but not auditory statistical learning being found across age. Likewise, modality effects observed in humans may or may not generalize to nonhumans68. There is also evidence that statistical learning impairments in developmental dyslexia may depend at least somewhat on age and possibly sensory modality as well (see135). Thus, to fully evaluate and understand each phenomenon may require understanding their interactions with one another.

Conclusion

Despite the surge of interest in statistical learning, there are a number of critical challenges facing this area of research9,11,12,13. We have argued that these challenges are similar to those facing the psychological sciences more generally. Relying on recent conceptual advances in contending with the theory crisis30, we have evaluated a sample of statistical learning research areas in terms of the robustness of each phenomenon, issues with construct validity, and difficulties with establishing causality. We have also offered recommendations for how to deal with these issues in a way that we believe will lead to promising areas for theory-building, specifically by focusing on ways to better establish robustness, validity, and causality in statistical learning research.

We hope this analysis and the recommendations contained within will improve future attempts at constructing theories to explain various statistical learning phenomena. Such theories would need to be pitched at both a higher level—i.e., what neural mechanisms and cognitive processes underlie statistical learning itself32 and how they relate to other psychological and cognitive constructs—but also within the embedded sub-phenomena reviewed here—i.e., theories to explain how and why differences exist (or not) across modality, age, species, and developmental disorders. The development of such theories will be crucially dependent upon the continued identification and exploration of the robust phenomena, the refinement of a diverse set of valid measures, and an understanding of the causal relationships relevant to statistical learning.