Ask any individual about his or her navigational abilities, and the answer might range dramatically from statements such as “I get lost all of the time” to “I have a great sense of direction.” More critically, people will describe using very different methods for navigating the same everyday situations. Some individuals aim to establish effective (sometimes idiosyncratic) routes that can be consistently traversed. Others tend to develop a more map-like understanding of the global structure of their environment that allows the use of different routes to reach the same destination, depending on the current goals and circumstances. Some may also report the use of beaconing strategies that take them from one salient landmark to another. Self-report measures have been useful in defining and identifying navigational styles (Lawton, 1994; Pazzaglia & De Beni, 2001; Schmitzer-Torbert, 2007), helping to classify the types of cues and strategies that participants prefer to use. Although these preferences for different sources of information and strategies likely play into people’s perceptions of their navigational abilities, they also represent different effective solutions to the critical problem of navigation. Understanding individual differences in these navigational styles, as well as the degree to which individuals can flexibly engage different styles and strategies, will offer substantial insights into how humans accomplish the difficult task of learning about environments and responding to navigational challenges.

Recent empirical investigations have begun connecting these explicitly articulated navigational styles to underlying learning mechanisms (Marchette, Bakker, & Shelton, 2011; see Shelton, Marchette, & Furman, 2013, for review). This learning-mechanisms approach was derived from the extensive literature on rodent learning and memory, and specifically the distinction between place and response learning (e.g., Packard & McGaugh, 1996; Restle, 1957; Tolman, 1948). According to this model, rats learn space in (at least) two different ways, which can be revealed using a dual-solution T-maze. In the basic preparation, rats are trained to locate a food reward in one arm of a T-maze. During encoding, the start location and reward location are stable with respect to the cues in the environment, such that a rat could learn the location of reward relative to those cues (place learning) and/or the specific path that will lead to reward (response learning). Place and response learning can be distinguished using critical probe trials in which the T-maze is rotated 180º with respect to the environmental cues. If the rat returns to the location of the reward relative to the environmental cues, the behavior reflects place learning. By contrast, if the rat makes the same set of turns that led to the goal location during learning, the behavior reflects response learning.

A wealth of evidence suggests that rats may have a bias toward place or response learning, but can manifest both behaviors under different conditions (Packard & McGaugh, 1996; Tolman, Ritchie, & Kalish, 1946). For example, Packard and McGaugh observed that rats had a tendency to demonstrate place learning early in learning, with a tendency to shift to demonstrating response learning after substantial repetition. In addition, they found that lesions of the hippocampus disrupt place but not response learning, whereas lesions of the striatum disrupt response but not place learning (Packard & McGaugh, 1996). Moreover, a rat that has been showing response-learned behavior will still manifest place-learned behavior when the striatum is temporarily inactivated with lidocaine, supporting the simultaneous acquisition of place and response. Mirroring these results, a number of human neuroimaging studies have reported engagement of the hippocampus and caudate nucleus, an analogue of the rodent dorsal striatum, across a variety of spatial memory and navigation tasks (Hartley, Maguire, Spiers, & Burgess, 2003; Iaria, Petrides, Dagher, Pike, & Bohbot, 2003; Iglói, Doeller, Berthoz, Rondi-Reig, & Burgess, 2010; Marchette et al., 2011).

The apparent parallel between rats and humans affords an opportunity to consider how these mechanisms might provide a rich context for understanding the roots of human navigational style. Using place and response learning as a framework, Marchette et al. (2011) developed a spatial memory paradigm that emulates the core features of the T-maze, referred to as the dual-solution paradigm (DSP). In this task, participants are taught a novel environment, and the locations of objects within it, by watching a video of a circuitous fixed path through the space. The to-be-learned environment has a stable spatial structure, thereby making it possible to learn the positions of objects with respect to the fixed path during learning and/or the environment’s spatial structure. To determine what has been learned, participants are placed at a point along the previously viewed path and instructed to find one of the objects in the environment. On critical trials, the start location is selected, such that one could proceed along the familiar path or take a novel shortcut that could be inferred from the information available along the fixed path during learning. In theory, an individual who has a bias toward representing the habit-like behaviors obtained through repetition (response-like learning) would be more likely to use the familiar path that had been repeated throughout the initial learning, whereas an individual with a bias toward representing global structure and organization (place-like learning) would be more likely to use an available shortcut. As such, the data are quantified into the relative use of familiar paths and shortcuts, and this solution index can be used to test whether the variability in solutions across individuals can be associated with differential engagement of underlying learning mechanisms.

In the initial study using the DSP, Marchette et al. (2011) observed a wide range of solution indices, ranging from using only the familiar path to using only shortcuts, with most individuals using a mixture of the two. Functional magnetic resonance imaging (fMRI) revealed that the relative activation of the hippocampus and caudate nucleus during initial learning was correlated with the solution index observed during subsequent navigation such that greater hippocampal activation was associated with more shortcut use, whereas greater striatal activation was associated with more familiar path use. Given that the presentation during initial encoding was constant across individuals, these differences in relative activation of the hippocampal and striatal systemsFootnote 1 are taken as evidence for different approaches to learning the same material. Beyond showing that navigational style can be experimentally quantified in a laboratory setting, these findings have laid the groundwork for thinking about place and response learning as a mechanistic framework for understanding why individuals differ in the ways that they navigate.

The ability to predict one’s solution index during subsequent navigation from the brain activation during encoding suggests that biases during learning will affect what information might be available. However, from the rodent literature, it is also clear that the engagement of these specific systems during the manifestation of behavior in later navigation is also informative (e.g., Packard & McGaugh, 1996). In the DSP, many participants use familiar paths and shortcuts on different trials, suggesting a dynamic relationship between what has been learned and what is happening during actual navigation. As such, it is important to consider what is happening during subsequent navigation in the brain regions known to predict performance. At the broadest level, confirming that navigation-related activity in these networks is also related to solution use will strengthen the relationship between putative learning mechanisms and navigational style. At a more fine-grained level, collecting brain activity during subsequent navigation will offer an opportunity to understand how these mechanisms relate to the selection of a path to the target (i.e., shortcut or familiar) on any given trial.

In the present study, we examine brain activation during both initial encoding and subsequent navigation using the DSP. The overarching prediction from the previous study and the rodent literature is that the engagement of the relevant networks during subsequent navigation should also be related to the types of solutions used (e.g., Packard & McGaugh, 1996). For example, in rodents, post-training injections of glutamate into the hippocampus or striatum can bias rats to exhibit place or response learning, respectively (Packard, 1999), implying that solutions reflect not only what was learned but also how the systems are engaged during subsequent navigation. Overall, this predicts that a given individual’s balance of activation in the putative place and response regions might be preserved during subsequent navigation to support their observed solutions. That is, if one has greater activation in the hippocampus than the caudate during encoding, he/she might also show this bias during subsequent navigation in accordance with use of shortcuts. The first goal of the present study is to evaluate whether the balance of activation is indeed stable or whether relative activation in these systems might vary as a function of the specific navigational tasks one encounters. Given that our participants knew only that they would need to navigate, and not the conditions of that navigation, it was possible that a different pattern of individual differences might emerge in response to subsequent navigation. To foreshadow the results, we observed a preservation of the relative activation across individuals, allowing us to ask a deeper set of questions about different models for the relationship between learning mechanisms and navigational styles.

Does a preserved relationship between activation in the hippocampus and caudate from encoding to subsequent navigation reflect the relative frequency with which an individual is engaging each mechanism to support selection of a specific solution, or does it reflect a stable balance of power between learning mechanisms that shapes how all spatial information is encoded and reasoned about? In the former scenario, the putative place and response networks would be modulating trial-by-trial as a function of the particular solution being used. This would naturally result in a net balance of activation that is consistent with the solutions selected; people who use more shortcuts would have more trials on which hippocampal activation was greater, whereas people who use more familiar paths would have more trials on which caudate activation was greater. This model would suggest that one’s preferential bias stems from the frequency with which the two systems tend to be engaged during navigation. Alternatively, the relative activation of the putative place and response systems may reflect a participant’s stable bias to engage these two different learning mechanisms during any navigational activity. During encoding, this bias might be acting to advantage the representation of certain types of information that will subsequently make certain solutions more or less readily available. At retrieval, the same bias might be present, reflecting the relative availability of the learned information. According to this account, the dynamic shifting among these mechanisms is not what gives rise to a mixture of solution use during navigation, but rather the relative availability of the different information coming from the two systems that allows flexible selection. That is, if someone has a strong bias to engage the putative response-learning system, solutions that depend on the information from the putative place system are simply less likely to be available.

In both of these accounts of preserved overall bias, the balance of activation in the putative place and response systems will influence solution selection, but the accounts are critically distinct with respect to how the selection is related to the bias. In the trial-modulation account, whichever system is more active on a given trial will drive the behavior, much like a race model in which the systems themselves are competing. The stable-bias account suggests a less directly competitive interaction in which both systems are offering information, but to varying degrees, and the availability of information is what allows a solution to emerge. By examining how activation in the putative place and response systems changes as a function of the solution that is selected, we can evaluate which of these accounts is supported by the patterns of activation.

Method

Participants

A group of 28 participants, 18–32 years of age, participated in return for financial compensation. Four participants were excluded because they did not complete a sufficient number of trials to allow for an accurate calculation of their navigational style (two females, two males), resulting in 24 participants included in all analyses (12 female, 12 male; mean age = 23.32 ± 2.87 years). All participants were right-handed, had normal or corrected-to-normal hearing and vision, and self-reported no history of either neurological/psychiatric illness or any contraindications to the magnetic resonance environment. All consenting and study procedures were conducted in accordance with the Johns Hopkins Homewood and the Johns Hopkins Medicine Institutional Review Boards.

Materials

All environments and videos were created using the free Source SDK Hammer Editor platform (Valve Software, Bellevue, WA, www.valvesoftware.com/) and run using the commercially available Portal game (Valve Software, Bellevue, WA, www.valvesoftware.com/).

Virtual environments and initial encoding

Two desktop virtual environments (VEs) were constructed using 11 × 11 units (approximately 22 × 22 m) in virtual space. Twelve different objects were placed at different locations throughout the space, and a 62-s tour was constructed to allow a complete path through the environment. The path took the participant past all 12 objects and all choice point hallways, to allow full appreciation of the structure of the environment (see the example in Fig. 1).

Fig. 1
figure 1

Top panels: Schematics of one of the environments. The left side shows path during initial encoding, and the right shows possible classifiable paths during one subsequent navigation trial. Bottom panels: Screen shots of the initial encoding (left) and subsequent navigation (right) as they appeared to the participants. Participants never viewed the schematics. The top panels are from “Cognitive Mappers to Creatures of Habit: Differential Engagement of Place and Response Learning Mechanisms Predicts Human Navigational Behavior,” by S. A. Marchette, A. Bakker, and A. L. Shelton, 2011, Journal of Neuroscience, 31, p. 15265. Copyright 2011 by the Society for Neuroscience. Adapted with permission

Full size image

Encoding control

Additional environments were created for control trials in order to mimic the optic flow and tracking of objects in the VEs. The assignment of textures and the structure of the environment changed from trial to trial to limit what participants could learn from these environments. Participants were given a 1-s cue indicating that it was a control trial, followed by a 30-s video tour through these random hallways. Throughout the environment, blue and red spheres appeared at randomly spaced locations. At the end of each control video, participants were asked to indicate whether the color of the first sphere that appeared matched the color of the last sphere that had appeared.

Navigation trials

For subsequent navigation, each trial was designed to have a start location along the familiar path and a goal location that could be reached via either the familiar path or a novel shortcut. All shortcuts used hallways that were visible to the participant during the initial encoding. These were the same navigation trials that had been used by Marchette et al. (2011) in their critical shortcut-available condition.

Navigation control trials

To control for the motoric commands and optic flow of the navigation trials, empty grid environments were created, and participants were given an opportunity to practice navigating. The infinite grid structure of the environment prevented participants from being able to garner any structure or learn any spatial layout. (At most, they could discern that they were in a very, very large repeating grid.)

Procedures

Prior to the 1-h neuroimaging session, all participants completed a 30-min behavioral testing session in which a number of individual difference measures were collected. The behavioral testing preceded the neuroimaging session by varying lags from 1 h up to eight days.

Behavioral testing session

During the behavioral testing session, participants completed a battery of individual difference measures that focused on spatial abilities and navigational preferences. Of critical interest here were the Spatial Perspective Taking test (SPT; Kozhevnikov & Hegarty, 2001) and the Questionnaire on Spatial Representation (QSR; Pazzaglia & De Beni, 2001).

SPT

The SPT measures the ability to flexibly reason about different perspectives in a two-dimensional layout by indexing angular error in pointing judgments. In this task, participants view a two-dimensional image of a layout of common objects and are presented with a series of problems. For each trial, the participant must imagine standing at one object and facing another to then point to a third object (e.g., “Imagine you are at the flower, facing the tree. Point to the cat.”) On each page, a response circle represents the imagined heading as facing forward (or up on the paper) along a line that runs from the center to the top of the circle (e.g., flower at the center of the circle and tree at the top of the circle). Participants respond by drawing another line on a circle to indicate the direction of the target object relative to the specified heading. The angle of pointing is measured and compared to the veridical direction. Performance is measured as the 180º absolute average angular error; higher scores on this measure indicate more accurate pointing judgments. Previous work from our lab has shown the measure to be associated with behavior in the DSP (Marchette et al., 2011).

QSR

The QSR is a self-report measure of the types of information that an individual prefers to use when orienting and navigating in an environment. Participants are asked to rate items from 1 to 5 on the basis of whether the item applies to their navigational preferences. The score can be broken down into preference for route-based information (e.g., “Are you a person who orients him/herself by remembering routes connecting one place to another?”), preference for survey-based information (e.g., “Are you a person who tries to create a mental map of the environment?”), and preference for landmark information (e.g., “Are you a person who orients him/herself by looking for well-known landmarks?”). Self-ratings on this measure have been associated with how well an individual can complete different navigation tasks that require different sources of environmental information (Pazzaglia & De Beni, 2001). In the present experiment, this measure was used as a self-report of navigational style using a relative preference score, taken as the difference between the survey and route scores on the assessment. This QSRS-R score provides an index ranging from –8 (route preference) to +8 (survey preference). All data were also verified against the individual subscales (QSRSurvey and QSRRoute), where appropriate.

Additional measures

In addition to the SPT and QSR, participants completed the Mental Rotation Test (MRT; Vandenberg & Kuse, 1978), the Santa Barbara Sense of Direction scale (SBSOD; Hegarty, Richardson, Montello, Lovelace, & Subbiah, 2002), and the Spatial Anxiety Scale (SAS; Lawton, 1994). These measures were included to allow an exploratory analysis of possible factors affecting different aspects of performance; however, we did not observe any relationships with the critical measures of navigational success or style (behavioral or brain).

Neuroimaging session

Participants completed the environmental learning task used previously (Marchette et al. 2011; see Fig. 1). Prior to entering the scanner, participants were seated at a laptop computer and shown images of each object that would appear in the environment. The participant had to correctly name the object (e.g., table, box, telephone booth, barrel, chair, bush, wheelbarrow, traffic barrier, fridge, cart, satellite dish, and trash bin). The experiment then proceeded in stages.

Initial encoding

At the start of the initial encoding phase, participants were told that they were to learn an environment so that they would later be able to navigate to the different objects in the environment. Each experimental run included three of the encoding tours and four of the control videos, in which participants tracked the colored spheres. These were presented in random order within each run, with the restriction that neither the tour nor the control could appear more than twice in a row. A total of three initial encoding runs were presented, for a total of nine exposures.

Practice

Following initial encoding, participants were given the opportunity to practice navigating in a novel virtual environment with no objects and no complex structure to acclimate them to navigation with the MRI-compatible button boxes.

Subsequent navigation

In the subsequent navigation phase, participants completed both navigation and navigation-control trials. For navigation trials, the participant was first given the name of an object to locate and then transported to a start location along the familiar path in the environment facing a wall. To enable participants to view the environment before initiating navigation, the participant was rotated 360º. Following reorientation, the participant was given control of navigation and had 39 s to reach the goal before the trial would time out. In control trials, the participant was transported to the sparse control environment and given 18 s to practice navigating. Four scan runs were presented, with four navigation trials and four navigation control trials per run in random order.

For each trial, the participant’s position along the x- and y-axes of the VE was collected every fifth of a second. Using custom MATLAB scripts, the participant’s path on each navigation trial was reconstructed and compared against both the familiar path (shown in the tour video) and one or more preidentified shortcut path(s). Trials were first categorized as complete (i.e., the goal object was found) or incomplete. Completed trials were then classified as “shortcut,” “familiar path,” or “wandering.” For a trial to be classified as a shortcut, the path either contained fewer steps than the familiar path or was more compatible with a preidentified shortcut path(s) than with the familiar path. For a trial to be classified as the familiar path, the path had to be more consistent with the familiar path than with any of the available shortcut paths. Completed trials on which the participant used a longer path than the familiar path were identified as wandering trials. For incomplete trials, if 90 % of path steps used by the participant belonged to an identified shortcut or the familiar path, they would be classified accordingly. Incomplete trials that did not meet this criterion were classified as wandering trials.

fMRI acquisition

Data were acquired using a 3.0-T Philips scanner equipped with an eight-channel head coil. Functional images were collected using a T2*-weighted echo-planar single-shot pulse sequence with the following parameters: Acquisition matrix = 72 × 72, TR = 2 s, Echo time = 35 ms, Flip angle = 70º, Sense factor = 2, and In-plane resolution = 3 × 3 mm. Each volume consisted of 34, 3-mm-thick axial slices with no gap, aligned parallel to the line connecting the anterior and posterior commissures. The structural scan was acquired using an MP-RAGE T1-weighted sequence with 231 oblique slices with a field of view of 240 mm, yielding an isotropic resolution of 0.65 mm.

fMRI preprocessing and analysis

Data were preprocessed using the Statistical Parametric Mapping (SPM) 8 toolbox, run in the MATLAB environment (R2012b, The MathWorks, Natick, MA). Before the functional images were preprocessed, slice artifacts due to scanner noise were detected and repaired using the Artifact Repair Toolbox (Mazaika, Hoeft, Glover, & Reiss, 2009). Using the default parameters in SPM, functional images from both the initial encoding and navigation phases were realigned, and each participant’s anatomical scan was coregistered to their mean echo-planar imaging. Cross-participant alignment used an anatomical template generated by Marchette et al. (2011) based on methods described in detail by Stark and colleagues (Kirwan & Stark, 2007; Lacy, Yassa, Stark, Muftuler, & Stark, 2010; Yassa & Stark, 2009). Following affine registration of anatomical and functional images to the Talairach coordinate system (Talairach & Tournoux, 1988), the caudate, the hippocampus, and the perirhinal, entorhinal, and parahippocampal cortices were manually segmented bilaterally using methods described by Insausti et al. (1998). Advanced Normalization Tools (ANTS; Avants, Epstein, Grossman, & Gee, 2008) was then used to create a modal template. For the present study, ANTS was used to determine the 3-D vector field transformation needed to normalize each participant’s anatomical scan to this previously generated anatomical template. During normalization, the resulting transformation parameters were applied to all of the functional images, and the normalized images were then smoothed with a 4-mm isotropic Gaussian kernel.

At the individual level, functional images were high-pass filtered (128 s), and the initial encoding phase and subsequent navigation phase images were entered into separate general linear models. For the initial encoding model, the tour and control videos were included as regressors of interest. For the subsequent navigation model, the four types of navigation trials (shortcut, familiar path, ambiguous, and wandering) and the navigation-control trials were modeled as regressors of interest. Both models included six regressors of no interest to account for head motion. For both the initial encoding and subsequent navigation models, regressors of interest were convolved with the canonical hemodynamic response function. Using these models, single-subject contrasts maps were generated and taken to a second, group level to assess the relevant questions of interest. Unless stated otherwise, all contrasts were investigated with one-sample t tests corrected with a combined uncorrected height threshold of p < .001 and an extent threshold of 50 contiguous voxels, to yield an approximate corrected α = .05, on the basis of Monte Carlo simulations run in AlphaSim (http://afni.nimh.nih.gov/afni/doc/manual/AlphaSim; Ward, 2000).

To quantify learning-relevant activity in the hippocampus and caudate, we used a priori functionally defined regions of interest (ROIs) within the hippocampus and caudate, taken from Marchette et al. (2011) The ROIs were derived in a combined functional–anatomical method using the two-tailed t test contrasting encoding and control conditions, with a liberal threshold of p < .05 and 50 contiguous voxels. The resulting activations were then combined with anatomical segmentation to isolate active voxels in the hippocampus and caudate. For each individual, we extracted the percentages of signal change using Marsbar (Brett, Anton, Valabregue, & Poline, 2002). In addition to these individual signal change values (∆), we calculated normed ratios of hippocampal to caudate activation. First, we reset the smallest percent signal change value across all ROIs to be 0 and rescaled the data to ensure that we were achieving the best estimate of relative activation (adj∆); we then calculated the normed ratio:

$$ normed\ ratio=\frac{\mathrm{adj}{\varDelta}_{\mathrm{hippocampus}}\hbox{--} \mathrm{adj}{\varDelta}_{\mathrm{caudate}}}{\mathrm{adj}{\varDelta}_{\mathrm{hippocampus}}+\mathrm{adj}{\varDelta}_{\mathrm{caudate}}.} $$

These normed ratios provide a metric for how much more or less active the hippocampus was relative to the caudate, while accounting for potential differences in the size of the ROIs and the vasculature contributing to signal magnitude.

In addition to these independently derived ROIs, we repeated these procedures to obtain ROIs defined by the activation during the initial encoding and subsequent retrieval in the present data. We use these ROIs to verify that the results of the primary analysis could not be strictly attributed to differences in the size and location of activation across different samples. For most comparisons, we have placed these results in the supplemental materials, since they are simply replicating the basic findings; however, for cases in which the replication is clearly addressing possible bias, we include the results below.

Results

Table 1 shows a summary of the means and standard deviations for all behavioral and brain measures. All of these measures were checked for sex differences, and none of the differences approached significance, even with uncorrected thresholds, all ps > .18. Throughout the results, we have used correlations individually and in various combinations. The critical information is in comparison to previous values or ranges. As such, we focus on the effect size (r value) rather than the inferential value of any given test. As such, probabilities are given in uncorrected form, unless otherwise specified. In areas in which multiple correlations were conducted in relevant subsets, we also provide an indicator as to whether the correlation meets a corrected criterion; however, we stress that the primary question was the replication of the coefficient magnitude.

Table 1 Means (and standard deviations) for all behavioral and brain measures
Full size table

Behavioral performance

Success rates and navigational style

Participants successfully completed an average of 67 % ± 21 %. On average, 90 % of the successful trials and 21 % of the unsuccessful trials were classifiable on the basis of whether the familiar path or a novel shortcut was used (the remaining trials were ambiguous/wandering), resulting in an average of 66 % of all active navigation trials being classified and used to calculate the subsequent solution index.Footnote 2

To quantify navigational style, we computed a solution index (SI) for each participant (Marchette et al. 2011) by determining the proportion of classified trials (shortcut or familiar path) that had been solved using an identifiable shortcut:

$$ SI=\frac{\# shortcuts}{\# shortcuts+\# familiar\_ paths} $$

A score of 1 on this index represents that an individual used a shortcut on every classifiable trial, and a score of 0 indicates that he/she used the familiar path on every classifiable trial.

Replicating prior findings (Marchette et al. 2011), solution indices ranged from 0 (all familiar paths) to 1 (all shortcuts), with many individuals using a mix of both solutions (Fig. 2). Similarly, we observed only a weak relationship between the SI and success rate, r = +.25, p = .24. Although the result is not significant, it is worth considering whether this correlation represents a tendency for more successful individuals to use more shortcuts. Notably, among individuals with perfect success rates, the SI values still covered the full range from 0 (all familiar paths) to 1 (all shortcuts), consistent with the broader claim that using either solution could support successful navigation.

Fig. 2
figure 2

Distribution of solution indices

Full size image

Spatial skill variability and navigational style

The SI measure was intended to capture individual differences in navigational style. In our previous study, SI was positively correlated with SPT score (r = +.37), suggesting that flexibility in perspective taking was related to the use of shortcuts. In the present data, this correlation was in the same direction, but weak and not significant, r = +.29, p = .17 (Fig. 3a). This did not differ from the previous correlation, z = 0.34, p = .37, but it is a numerically weaker effect size. One concern may be the stability of SPT as a measure of perspective-based transformations. Previous work had suggested that although SPT generally captures perspective-based transformations, it may be susceptible to mental rotation strategies associated with object-based transformations (Kozhevnikov & Hegarty, 2001). To evaluate this concern, we correlated SPT with the MRT scores that we obtained. This correlation was surprisingly high, r = +.47, p = .02, suggesting that we may have had a disproportionate number of participants using a mental rotation strategy instead of perspective-based transformations, relative to previous studies. Mental rotation represents a type of flexibility that has been associated with some aspects of spatial behavior (Fields & Shelton, 2006) and brain activation (Shelton & Gabrieli, 2004). However, it is distinct from the flexibility associated with taking novel perspectives in space (e.g., Zacks & Michelon, 2005). As such, it is unclear how it should play into a relationship with SI as a measure of navigational style.

Fig. 3
figure 3

Correlations between solution index in the DSP and additional behavioral measures for (a) spatial perspective taking (SPT) and (b) Questionnaire on Spatial Representation (QSR) self-reported preference, as indexed by the Survey–Route score

Full size image

A more direct approach to navigational style has been self-report measures such as the QSR. Correlations between the QSR measures and SI were stronger than any other measures. First, the SI had a moderate positive relationship with QSRSurvey score, r = +.39, p = .06, and a negative relationship with QSRRoute score, r = –.47, p = .02, and these correlations were significantly different from each other, t(21) = 3.32, p = .003. However, the QSRSurvey and QSRRoute scores were not correlated with each other, r = –.07, p = .74, suggesting that there is not a single trade-off between use of a route or survey strategy, and that a participant’s preference for one strategy is not clearly related to his or her preference for the other. As such, a participants’ relative bias for one over the other might be more informative. Indeed, the correlation between the QSRS-R and SI was among the strongest correlations that we observed in the present study, r = .58, p = .003 (Fig. 3b); this correlation is the only one among this set that would remain significant under a conservative Bonferroni correction (α pc = .012). These results suggest that self-report measures and the empirical SI measure are capturing at least some of the same variability in navigational style.

Functional neuroimaging analysis

Relative activation in the hippocampus and caudate during initial encoding was associated with subsequent solution use in our previous study (Marchette et al. 2011). In the present study, we were interested in how activation in the hippocampus and caudate during subsequent navigation might inform our understanding of how the networks associated with these regions give rise to specific behaviors.

ROI comparison

Given the focus on ROI analyses, we first explored how the size and location of the ROIs from the previous study compared to those observed during initial encoding versus control and subsequent navigation versus control in the present study. The center of mass, range, and volume of the right and left caudate and right and left hippocampus are shown in Table 2. As is shown in Fig. 4 and Table 2, there was substantial overlap in all four regions across the three methods used to derive them (previous study, initial encoding, and subsequent navigation). The percentage of overlap was calculated within each region for each pair of corresponding ROIs by obtaining the volume of overlapping voxels relative to the average size of each ROI (Nieto-Castanon, Ghosh, Tourville, & Guenther, 2003). In all cases, the regions of activation were larger on the right than on the left in both the striatal and hippocampal regions. However, we had no specific hypotheses with respect to laterality for the key questions. We verified that the results were not driven exclusively by left or right activation, but we present only the combined results for the sake of brevity.

Table 2 Characterization and comparison of the regions of interest (ROIs)
Full size table
Fig. 4
figure 4

ROIs for the right and left caudate in the top panel for right and left hippocampus in the bottom panel showing the regions of overlap (white outline; orange in the online version) based on the three different methods of extraction: (1) initial encoding in the Marchette et al. (2011) study (yellow in online version), (2) initial encoding of present study (red in online version), and (3) subsequent navigation phase of the present study (green in the online version)

Full size image

Initial encoding phase (replication of Marchette et al. 2011

We first established that our paradigm was replicable with respect to both whole-brain activity and predicting navigational style from the relative activation of the hippocampus and caudate during encoding. The results of the whole brain analysis (encoding tour vs. encoding control) revealed significant activity in a number of regions, including bilateral posterior parietal, retrosplenial, and inferior temporal cortices, as well as right hippocampus (see Fig. 5a). These regions, representing a navigational network (Maguire et al. 1998; Shelton & Gabrieli, 2002), are consistent with those observed in our previous whole-brain findings.

Fig. 5
figure 5

fMRI results and relationship to navigational performance. (a) Whole-brain differences between initial encoding and encoding control; (b) correlation between the normed ratio of hippocampal/caudate activation during initial encoding and the solution index; (c) whole-brain differences between subsequent navigation and navigational control; and (d) correlation between the normed ratio of hippocampal/caudate activation during subsequent navigation and the solution index

Full size image

Next, we computed normed ratios of hippocampal and caudate activity for encoding versus control in the ROIs from the previous study, as described above (NREncod). As is shown in Fig. 5b, the correlation between SI and NREncod was significant, r = +.41, p = .05. This value is consistent with that in the previous study (r = +.44), from which it did not differ significantly, z = 0.11, p = .45, suggesting that again the relative activation in the putative place and response regions during encoding was predictive of the solutions used during subsequent navigation. These results were replicated using the ROIs based on initial encoding versus control and subsequent navigation versus control in the present study (see the supplemental materials).

Subsequent navigation

The critical addition of the present study was the investigation of activation during subsequent navigation. We first established that the navigational brain regions were active during the memory-based navigation to targets. The results revealed widespread activation throughout the brain for navigation trials relative to navigational control trials, including the navigation network regions and, specifically, the hippocampus and caudate (Fig. 5c).

To address the initial question of whether the bias observed during encoding is preserved during subsequent navigation, we calculated the normed ratio of activation in the hippocampus and caudate on all navigational trials versus control using the predefined ROIs (NRNavig) and compared it to NREncod. Indeed, the correlation between NREncod and NRNavig was significantly positive, r = +.42, p = .04. As is shown in Fig. 6, there may be a slight shift in the entire group toward greater caudate activation during subsequent navigation, but overall, individuals remained in the same relative positions on the normed ratio continuum from initial encoding to subsequent navigation. Again, these results were replicated when we calculated these same normed ratios using the ROIs from the present study (see the supplemental materials).

Fig. 6
figure 6

Correlation between the normed ratios of hippocampal/caudate activation at initial encoding and subsequent retrieval

Full size image

To address the critical question of how the putative place and response regions participate during subsequent navigation to give rise to solution use, we examined the overall relationship between SI and NRNavig. As we expected from the relationship between NREncod and NRNavig, we found a positive correlation, r = +.47, p = .02, reflecting greater activation in the hippocampus with more use of shortcuts, and greater activation in the caudate with more use of familiar paths (Fig. 5d). This correlation did not differ in magnitude from the relationship of SI to activation during initial encoding, t(21) = 0.28, p = .78, and these results were replicated using the ROIs from the present study (see the supplemental materials).

In the introduction, we highlighted two ways that a positive correlation might emerge during subsequent navigation. First, it could be due to trial-by-trial variability in the networks associated with the hippocampus and caudate. In this case, we expected to find a difference in the relative activations of the hippocampus and caudate as a function of which type of solution was used on a trial: more hippocampal activation for shortcut trials, and more caudate activation for familiar-path trials. Alternatively, an individual might have an overall bias that affects the availability of different solution types in a more probabilistic fashion, such that the relative activations of the hippocampus and caudate are largely stable across different solution types within an individual. To test these alternatives, we separated out trials that were completed using shortcuts and trials completed using familiar paths. For these analyses, we only used individuals who had at least two of each trial type, resulting in 18 participants. Using whole-brain analysis, we compared each trial type to the navigational control, revealing that similar overall networks were activated (Fig. 7a & c). We then did the direct comparison of shortcut and familiar-path trials and found no differences in brain activation, even with more liberal thresholds applied, suggesting that any differences in the trial types were not coming through as strong regional differences. Finally, we repeated these comparisons using the time points attributable to the first 6 s of the trial, when planning was likely to occur, and we still failed to observe any differences between the trials on which familiar paths or shortcuts were used.

Fig. 7
figure 7

fMRI results broken down by solution type during subsequent navigation. (a) Whole-brain differences between shortcut trials and navigational control; (b) correlation between the normed ratio of hippocampal/caudate activation during shortcut trials and the solution index; (c) whole-brain differences between familiar-path trials and navigational control; and (d) correlation between the normed ratio of hippocampal/caudate activation during familiar-path trials and the solution index

Full size image

As a stronger test of regional modulation, we extracted the percentages of signal change (β weights) in the predefined ROIs of the hippocampus and caudate separately as a function of whether shortcuts or familiar paths were used. We subjected these data to a repeated measures analysis of variance (ANOVA) with Region (hippocampus/caudate) and Solution Type (shortcut/familiar path) as factors.Footnote 3 As is shown in Fig. 8, the regional effect was the largest observed (η G 2 = .06), but it did not reach significance, F(1, 17) = 3.09, p = .10. Moreover, differences in these two regions could have been due to differences in vasculature, as they are not scaled for size or location. Critically, we found no difference in activation based on the solution selected, F(1, 17) = 0.004, p = .95, and no interaction with region, F(1, 17) = 0.21, p = .65. These data provide clear support for no overall differences in activation as a function of the solution that was selected.

Fig. 8
figure 8

Percentages of signal change as a function of brain region and type of solution used during subsequent navigation. Error bars reflect ±1 standard error of the mean, to indicate between-participant variability

Full size image

Finally, to assess more directly whether the relative activations were preserved at the level of individuals, we calculated the normed ratio of hippocampal and caudate activation in the predefined ROIs separately as a function of whether shortcuts (NRSC) or familiar paths (NRFam) were used. Although not statistically significant, both NRSC and NRFam had positive correlations of similar magnitude to those observed when the normed ratios were taken over initial encoding and subsequent navigation in aggregate, r = +.36, p = .14, and r = +.41, p = .09, for trials using NRSC and NRFam, respectively (Fig. 7b and d). Critically, these correlations did not differ from each other, t(15) = 0.20, p = .84, suggesting that the relationship between solution use and relative activation in the hippocampus and caudate did not depend on the particular solution selected. Notably, the correlation between NRSC and NRFam was also positive, r = +.44, p = .07, supporting the claim that the relative activation was largely stable across trial types (Fig. 9).

Fig. 9
figure 9

Correlation between the normed ratios of hippocampal/caudate activation for shortcut trials and familiar-path trials during subsequent navigation

Full size image

Given that the lack of differentiation for trial types is essential to making the case for this stability of activation, it is important to consider whether the ROIs from encoding (and from a different set of participants) might be attenuating or masking differences. Here, the replication of results using the ROIs from the present study can address potential bias and substantiate the similarity across trial types. First, we repeated the ANOVAs using region and trial type, and again found no significant effects or interactions with trial type, all Fs < 1. The correlations with SI are shown in Fig. 10a–d, and the correlations between the normed ratios for each trial type are shown in Fig. 10e and f. Again, the critical result is the similarity in the strengths and directions of correlations across the two trial types, adding support to the claim that these systems are not showing changes in relative activation as a function of the solutions employed on any given trial.

Fig. 10
figure 10

Correlation results of the normed ratios of hippocampal/caudate activation (NRs), broken down by solution type during subsequent navigation. (a) Correlation between NR during shortcut trials and the solution index (SI), for regions of interest (ROIs) derived from initial encoding (IE); (b) correlation between NR during shortcut trials and SI, for ROIs derived from subsequent navigation (SN); (c) correlation between NR during familiar-path trials and SI for IE ROIs; (d) correlation between NR during familiar-path trials and SI for SN ROIs; (e) correlation between NRs for shortcut and familiar-path trials computed from IE ROIs; and (f) correlation between NRs for shortcut and familiar-path trials computed from SN ROIs

Full size image

Predicting SI from behavior and brain

Looking broadly at the results, it is clear that performance on our navigation task is associated with self-reported navigational preferences (QSRS-R), as well as the relative activation of the hippocampus and caudate during both initial encoding (NREncod) and subsequent navigation (NRNavig). The correlations among these measures are summarized in Table 3. In addition to the correlations reported above, we found that the QSRS-R score was positively correlated with both normed ratios (NREncod and NRNavig), suggesting that self-report, brain activation, and the empirically based SI all share some variability. We then used regression analyses to discern the relative contributions of brain activation and self-reported preferences to predicting navigational performance.

Table 3 Correlation coefficients (with uncorrected probabilities) supporting the relationships among brain activation, navigational performance, and self-reported preferences
Full size table

First, we examined the relative contributions of NREncod and NRNavig to predicting SI. The simultaneous regression model was significant, F(2, 21) = 3.92, p = .04, and accounted for about 20 % of the variance in SI (AdjR 2 = .202). We observed that NRNavig had a slightly larger contribution (β = .36) and unique contribution (sr 2 = .11) to the variance in SI than did NREncod (β = .26 and sr 2 = .06). However, neither regressor was statistically significant, suggesting that the model was reflecting the combined contributions. When we added QSRS-R score to the model, the model was again significant, F(3, 20) = 4.66, p = .013, AdjR 2 = .323, but only QSRS-R was a significant regressor, t(20) = 2.18, p = .04. More specifically, the beta weights and unique contributions were diminished for both NREncod and NRNavig, but this reduction was more pronounced for NREncod (see Table 4), suggesting that the explanatory power of the QSRS-R was overlapping more with activation during encoding than during subsequent navigation.

Table 4 Summary of simultaneous regressions using brain activation alone (Model 1) or with QSR Survey–Route score (Model 2) to predict the solution index during subsequent navigation
Full size table

Discussion

The overarching aim of the present study was to deepen the exploration of the neurobiologically based learning-mechanisms approach to individual differences in navigational styles (see Shelton et al. 2013). In our previous work, we observed that the relative activation of the hippocampus and caudate predicted the solutions used in subsequent navigation (Marchette et al. 2011). In the present study, we replicated this finding, and further observed that the balance of activation during initial encoding was maintained during subsequent navigation in two ways. First, the balance of activation in the hippocampus and caudate was correlated from initial encoding to subsequent navigation, suggesting that a given individual had similar balances of activation during both phases of the experiment. Second, the balance of activation during subsequent navigation was correlated with the solution index observed. That the individual differences were present during both initial encoding and subsequent navigation in the same form suggest that the link between the learning mechanisms and navigational performance is a reflection of both how a space is engaged during initial encoding to learn the environmental information and during the dynamic situation of navigating to specific goal locations from memory.

The interpretation of the normed ratio of hippocampal and caudate activation as representing a stable bias is also supported by the consistent relationship among these brain activation measures, behavioral performance, and self-reported spatial learning preferences. The QSR, which has been used frequently in the literature on individual differences, is presumed to reflect a stable approach to navigation by asking a variety of different questions about how one prefers to learn and navigate in environments (e.g., Denis, Pazzaglia, Cornoldi, & Bertolo, 1999; Garden, Cornoldi, & Logie, 2002; Meneghetti, Pazzaglia, & De Beni, 2011). The observation that this presumably stable trait was correlated with the brain activation at least as strongly as the actual behavioral performance on the task suggests that our measures may also reflect some stable underlying trait(s) that participate in the emergence of navigational style. These links offer compelling support for the mechanistic approach, suggesting that activation of underlying learning mechanisms can be associated with actual behavior and explicitly described strategies. In addition, there is still unexplained variance among these measures, which opens the door for exploring how one gets from a bias in the engagement of underlying learning mechanisms to these complex preferences for navigational information and strategies.

One of the more remarkable findings of the present study was that the overall balance of activation in the hippocampus and caudate during subsequent navigation did not depend on the particular solution that one was engaging on a given trial. That is, when we broke down the data according to the type of solution that was used, shortcut or familiar path, we found that the regions associated with the putative place and response networks, the hippocampus and caudate, showed a consistent balance of activation, regardless of trial type. Moreover, the normed ratios of activation were correlated with each other—individuals with a lower normed ratio overall were lower on both trials types, whereas those with higher normed ratios were higher on both trial types—and had the same correlation with the measure of overall performance (SI). This suggests that the activation in the putative place and response networks was not an index of which solution was being selected at any given time, but reflected some overall bias in the system that was stable from initial encoding to subsequent navigation.

In addition to providing support for a stable bias in learning systems, the lack of differentiation between the trial types raises further questions about how an individual selects a particular solution on any given trial. Not only did we fail to see modulation for the different solutions in the regions associated with the putative place- and response-learning systems, we also failed to see overall brain differences on the basis of the type of solution that was selected. Clearly, at some point in the process a path is selected, which allows the participant to begin navigating; that is, a path must be selected to proceed at all. Moreover, given that we did see differences across individuals associated with the types of paths used, we expect that some sources of neural activity must differentiate the solutions. Our results suggest that wherever this selection is occurring, it is not manifesting as large regional differences in activation, and must, instead, be occurring at some within-region level. Multivariate techniques such as multivoxel pattern analysis (MVPA) may be able to address this question more carefully; however, the present study was not powered for MVPA, nor do we have clear hypotheses about where one would expect to find these effects. This lack of specificity on the solution selection process does not alter or detract from the main conclusions of the study, but it highlights another potential area for future exploratory investigations.

The working model from our results is that an individual has a stable balance of engagement in the place and response systems that can be observed in the average activation during navigational activities (in our case, initial encoding and subsequent navigation). This balance of activation provides the framework and intrinsic limits for how an individual engages and processes the information in the environment. The ability to use different solutions will then be dependent on what information and processes are available at the time of retrieval, as well as what information that solution requires. In an ideal scenario, we would be able to predict on which trials one would be likely to use a shortcut or a familiar path, on the basis of the information that is available. This might be accomplished using a careful spatial analysis of the environment to understand which locations offer certain kinds of information. However, we would also need some understanding of the information or features to which the participant paid attention, since it is not a given that two participants with the same balance of activation would pay attention to all of the same information in the environment. Indeed, when we did an item analysis on these and previous data from the DSP, we did not find that certain trials were more likely to induce the use of one solution or the other across all individuals, or even among individuals with similar behavioral profiles. This lack of item-specific response suggests that the information gleaned by a given individual is probably influenced by a wide range of task-related, personal, and motivational factors. Future studies using eyetracking may weigh in on this issue, but understanding how an individual will approach a specific trial will likely involve a complicated set of interactions.

Our proposed model has a number of characteristics that distinguish it from other approaches. First, our model places an emphasis on learning; that is, in addition to considering what type of information people might prefer to use when navigating, we seek to consider what information is initially available during learning and how it might interact with the relative engagement of the different learning mechanisms. Although the present data are not sufficient to elaborate on this interaction, emphasizing learning offers a novel way of thinking about how people might attend to the initial environment in different ways that lead to differences in the availability of solutions. In addition, this learning-based approach might offer links to individual differences in learning in other domains. For example, we might ask how an individual who seems biased to more heavily engage the habit-based response-learning system would approach learning a new language or practicing math skills. For this individual, offering repetitive contexts in which he or she experiences the new information might have an advantage over effortful attempts to memorize new words. There is a long history of using rodent place and response learning as analogues to human explicit and implicit memory (e.g., Packard & Knowlton, 2002; Poldrack et al. 2001; Reber, Knowlton, & Squire, 1996; Squire, 2004), so extending this analysis to learning styles more broadly has substantial translational potential for clinical and educational applications.

A second important characteristic of our proposed model is that navigational or learning style is not rooted in the exclusive use of a single solution. An important conclusion from the DSP is that individuals might use all familiar paths or all shortcuts, but we more often see individuals who shift between these solutions, and their styles are then tied to the characteristic balance of this solution use. This bias or balance of systems suggests that it would be inappropriate to try to classify individuals into distinct categories, and instead emphasizes the need to think in terms of the likelihood of using one solution over another. This has important implications for how an individual might react to navigational challenges, such as being faced with a detour or needing to retrace one’s steps. Rather than predicting whether one will or will not succeed on a given challenge, we can offer a richer language of how more or less likely one is to succeed. This more probabilistic approach to navigational styles also motivates an appreciation of the flexibility of human learning and navigational processes. That is, any given individual may have multiple resources upon which to draw flexibly, depending on his or her particular biases interacting with the availability of information in the environment.

Taken together, our results suggest a new model for thinking about navigation and memory in the context of a cognitive system that includes parallel mechanisms for successful learning. The reactivation of a memory mechanism reflects the availability of a pool of stored knowledge, rather than the construction of a specific solution for a specific situation. But learning is not fate: These discrete pools of knowledge do not determine what an individual will do, but merely gate the solutions that will be available on the basis of other idiosyncratic decision criteria (e.g., trading off route certainty for route efficiency, or balancing the desire to explore with timely arrival at a destination). This novel approach offers an exciting framework for understanding both the complex processes that occur during learning and how the knowledge acquired might be differentially accessed and utilized in the face of behavioral challenges.