The application of a rodent-based Morris water maze (MWM) protocol to an investigation of age-related differences in human spatial learning

Abstract

The current study applied a rodent-based Morris water maze (MWM) protocol to an investigation of search performance differences between young and older adult humans. To investigate whether similar age-related decline in search performance could be seen in humans based on the rodent-based protocol, we implemented a virtual MWM (vMWM) that has characteristics similar to those of the MWM used in previous studies of spatial learning in mice. Through the use of a proximity to platform measure, robust differences were found between healthy young and older adults in search performance. After dividing older adults into good and poor performers based on a median-split of their corrected cumulative proximity (CCProx) values, the age effects in place learning were found to be largely related to search performance differences between the young and poor-performing older adults. When compared to the young, poor-performing older adults exhibited significantly higher proximity values in 83% of 24 place trials and overall in the probe trials that assessed spatial learning in the absence of the hidden platform. In contrast, good-performing older adults exhibited patterns of search performance that were comparable to that of the younger adults in most place and probe trials. Taken together, our findings suggest that the low search accuracy in poor-performing older adults stemmed from potential differences in strategy selection, differences in assumptions or expectations of task demands, as well as possible underlying functional and/or structural changes in the brain regions involved in vMWM search performance.

The Morris water maze (MWM), originally developed by Morris (1981), is a task widely used by psychologists and neuroscientists for assessing spatial learning and memory in both rodents (e.g., Morris, 1984; Morris, Garrud, Rawlins, & O’Keefe, 1982; Sutherland, Whishaw, & Kolb, 1983) and humans (e.g., Astur, Ortiz, & Sutherland, 1998; Driscoll et al., 2003; Moffat & Resnick, 2002). In the assessment of MWM performance among human subjects, virtual simulations of the MWM have been widely employed, and previous studies have generally shown that advanced age was associated with poorer performance in finding the hidden platform over multiple learning trials. The testing protocols/procedures of these studies commonly assessed participants’ ability to reach the static location of a hidden pole (Antonova et al., 2011) or a hidden platform (Driscoll et al., 2003; Moffat & Resnick, 2002; Korthauer et al., 2016; Daugherty et al., 2015) from different starting positions over six to 28 place trials. Probe trials, assessing memory of the location of the hidden platform when the platform was removed, were only conducted at the end of place trials (Driscoll et al., 2003; Moffat & Resnick, 2002). Previous studies have assessed MWM performance by computing path-length (i.e., distance travelled to reach the hidden platform) and escape latency (i.e., time taken to reach the hidden platform) [see, e.g., Antonova et al., 2011; Daugherty & Raz, 2017; Driscoll et al., 2003; Korthauer et al., 2016; Moffat & Resnick, 2002; Moffat et al., 2007; Yuan, Daugherty, & Raz, 2014], as well as a newer measure of fractal dimensionality (Daugherty et al., 2015; Daugherty & Raz, 2017).

In order to enhance translational research involving both humans and rodents, the current study applied an experimental protocol adapted from the MWM protocols that were previously employed by Magnusson and colleagues to assess search performance in male mice (see Magnusson, 1998; Magnusson, Scruggs, Zhao, & Hammersmark, 2007; Zhao et al., 2009) to an investigation of age-related differences in spatial learning in human subjects. Critically, in view that the MWM is relevant for assessing age-related declines in spatial learning and memory formation in both humans and mice, we designed a virtual Morris water maze (vMWM) — comprising of simple geometric cues, multiple place trials, and interleaving probe trials — that emulated the procedure of Magnusson et al.’s rodent-based protocols (Magnusson, 1998; Magnusson et al., 2007; Zhao et al., 2009) with reference to the number, types and duration of trials, the spatial layout of cues, data-recording procedures, and analytical methods (see Methods, for details).

Compatible with the previous data-recording approach taken by Gallagher, Burwell, and Burchinal (1993), and Magnusson and colleagues (Das & Magnusson, 2011; Magnusson, 1997, 1998; Magnusson et al., 2007; Zhao et al., 2009), we applied the proximity measure in our vMWM task to evaluate spatial learning. Prior to this study, proximity had only been used to evaluate the MWM performance of rodents (i.e., rats and mice) (see, e.g., Das & Magnusson, 2011; Gallagher et al., 1993; Magnusson, 1997, 1998; Magnusson et al., 2007; Walsh, Booth, & Poe, 2011; Zhao et al., 2009). Based on our knowledge, this is the first study that assessed the MWM performance of human subjects in terms of proximity. Pertinently, we chose proximity over other measures as our primary dependent variable of analysis because this measure has been theorized to offer several unique advantages over conventional measures (e.g., path-length, escape latency), such as: (i) a record of the spatial distribution of the search relative to the escape platform, which cannot be conveyed by path-length per se; (ii) a fine-grained analysis of search behavior (i.e., the proximity measure can be averaged over the duration of a search or trial, or analyzed in smaller segments of a trial); and (iii) a greater sensitivity to individual differences in spatial learning among aged animals (Gallagher et al., 1993). The use of proximity was also supported by a Monte-Carlo simulation study by Maei, Zaslavsky, Teixeira, & Frankland (2009) that demonstrated it to be more sensitive at detecting group-based MWM performance differences than other commonly used measures (e.g. time in quadrant, crossings of the location of the hidden platform).

Overall, this study is the first to assess differences in search performance between young and older adults in a vMWM task that was adapted as precisely as possible from previous MWM protocols that utilized the proximity measure (Gallagher et al., 1993; Magnusson, 1998; Magnusson, et al., 2007; Zhao et al., 2009). Through the use of this protocol, we aimed to establish a parallel task to investigate whether age-related decline in search accuracy (represented primarily by the proximity measure), previously found in aged mice (Das & Magnusson, 2011; Gallagher et al., 1993; Magnusson, 1997, 1998; Magnusson et al., 2007; Zhao et al., 2009), would similarly apply to aged humans.

Method

Participants

Forty-two male adults (21 young and 21 older adults) participated in this study. We recruited males only in this initial study because the protocol we adopted was derived from previous MWM studies that involved male mice only (Magnusson, 1998; Magnusson, et al., 2007; Zhao et al., 2009) and because sex differences in human vMWM performance have been documented by many previous studies (e.g., Astur et al., 1998; Daugherty et al., 2015; Driscoll et al., 2003; Nowak, Diamond, Land, & Moffat, 2014; Sandstrom, Kaufman, & Huettel, 1998; Yuan et al., 2014). The young adults (aged 18 to 30) were recruited from the psychology research volunteer pool at Georgia Institute of Technology and from the academic community in Atlanta, GA, USA. The older adults (aged 60 to 79) were recruited using newspaper advertisements and notices distributed at older adult community centers in the greater Atlanta area. All participants were selected to be right-handed based on initial screening with the Edinburgh handedness questionnaire (Oldfield, 1971). Based on self-reported medical history, all participants were assessed to be in good health (i.e., no current history of coronary heart disease, high blood pressure, stroke, diabetes, and dementia) and free of medications that could affect their cognitive performance. All participants scored ≥ 27/30 on the Mini-Mental State Exam (MMSE) (Folstein, Folstein, & McHugh, 1975), a range that has been shown to offer high classification sensitivity and specificity for ruling out dementia (O’Bryant et al., 2008). They also scored below the cutoff score of 16 on the self-report Center for Epidemiologic Studies Depression scale (CES-D) scale (Radloff, 1977), which indicated that no participants had depressive symptoms that were severe enough to affect their daily functioning. Vision assessments further showed that all participants had good visual acuity (20/40 vision and better), normal color vision (Ishihara Color Plates Test), and intact contrast sensitivity (Mars Letter Contrast Sensitivity Test). Informed consent for participation in this study was approved by the Georgia Institute of Technology Institutional Review Board. Each participant gave written consent prior to participation. Table 1 presents the demographic details of the two age groups.

Table 1.

Descriptive statistics of the demographic and pretest variables of young and older adults

	Young (n = 21)			Old (n = 21)			Difference (Old – Young)
	M (SD)	Min.	Max.	M (SD)	Min.	Max.	M	95% CI
Agea	22.29 (3.13)	18	30	65.71 (4.51)	60	79	43.43	[41.01, 45.85]
Education	14.86 (1.31)	12	17	15.43 (2.18)	12	18	0.57	[− 0.56, 1.70]
MMSE	29.62 (0.59)	28	30	29.19 (0.87)	28	30	−0.43	[− 0.89, 0.04]
CES-D score	5.38 (3.64)	0	14	4.71 (4.22)	0	13	−0.67	[− 3.13, 1.79]
Ishihara color-blindness	13.52 (2.18)	4	14	14.00 (0.00)	14	14	0.48	[− 0.52, 1.47]
Contrast sensitivity	1.83 (0.07)	1.72	2.00	1.72 (0.08)	1.48	1.80	−0.10	[− 0.15, − 0.06]
Speed test (sec)a	56.57 (4.73)	48	73	63.90 (15.30)	50	120	7.33	[0.24, 14.43]
No. of attempts in Practice vMWM	2.86 (1.06)	1	5	3.57 (1.29)	2	6	0.71	[− 0.02, 1.45]

Note. Data presented in M (± SD). Education: Education of participant in years. MMSE: Mini-Mental State Examination. CES-D: Center for Epidemiologic Studies Depression scale.

^aMeasures indicative of significant age group differences (p < .05).

Procedure

Joystick control and vMWM practice

Before starting the experiment proper, the participants underwent practice with controlling a joystick for virtual movements. Specifically, the participants sat in front of a 21-inch flat panel LCD monitor at approximately a distance of 30 cm (at eye-level) and practiced using a Logitech^® Attack™ 3 joystick to explore a virtual arena [designed using Unity^® Pro 5.0., Unity Technologies, Inc., San Francisco, CA, USA].

They were instructed to guide themselves to four different objects (car, bench, tree, chair) situated at the four corners. After demonstrating their ability to reach the different target objects, the participants performed a joystick control speed test, in which they were exposed to another virtual environment — a virtual maze — and instructed to traverse a long winding passageway until they reached a flagged finishing point [designed using Unreal Engine^® v3.0 (Epic games, Inc., Cary, NC, USA)]. The maximum time allowed was two minutes. Both young (M = 57 secs) and older (M = 64 secs) adults completed the task in around one minute, demonstrating adequate proficiency at using the joystick (see Table 1).

Next, the participants practiced searching for a hidden platform in a round virtual pool (diameter = 16 virtual units) [Note: 1 virtual unit = 1 m in real-world metrics]. The platform was a round disk (diameter = 3.5 virtual units) submerged beneath the water surface and fixed in the center of the southeast quadrant of the pool. There were four object cues (soccer ball, cube, cylinder, tree) positioned equidistant from each other on the deck of the pool. The participants were instructed to search for the hidden platform over repeated trials. They began each trial by facing the deck of the northwest quadrant of the pool and had to turn around before beginning their search, in line with the protocol of Magnusson et al. (2007). The hidden platform surfaced within 0.5 seconds after the participant entered the platform area. The practice trials repeated 10 seconds later, starting from the same entry point. The participants were encouraged to practice as many trials as possible until they felt confident of handling the joystick and could find the hidden platform with ease. The control and test vMWMs were designed using SketchUp^® Pro 2014 (Trimble Navigation Ltd., Sunnyvale, CA, USA) and Unity^® Pro 5.0.2 (Unity Technologies, Inc., San Francisco, CA, USA).

Experimental testing

The vMWM used in the formal testing of search performance among participants comprised a larger round-shaped pool (diameter = 30 virtual units) situated at the center of a rectangular quadrangle [66 virtual units (length) × 46 virtual units (breadth) × 14 virtual units (height)] bounded by white walls with thickness of 1.0 virtual unit. The dimensions of the virtual pool and rectangular floor were scaled in proportion to the real-world spatial dimensions of the MWM used by Magnusson et al. (2007) [pool diameter: 116.84 cm; rectangular floor: 282 cm (length) × 203 cm (breadth)]. The corners of the quadrangle were curved to ensure that none of them would stand out as potential geometric cues for locating the hidden platform (see Figure 1A). Akin to the practice vMWM, there were four object cues positioned high up on the walls and scaled to the same size relative to each other (see Figure 1A). The types of cues and their spatial positions represented a virtual version of a real-world MWM used previously by Magnusson et al. (2007). The hidden platform had identical spatial (i.e., diameter, location coordinates) and dynamic attributes (i.e., latency and speed of elevation) as the platform installed in the practice pool. It covered approximately 1.4% of the total surface area of the test pool and occupied a fixed location near the middle of the southeast quadrant of the virtual pool across trials.

Figure 1.

Overviews of the test (A) and control (B) virtual Morris water mazes (vMWM) designed for the experimental testing of participants’ search and visuomotor performance respectively. In both environments, the object cues were placed high up on the walls (11 virtual units from ground level). The four quadrants of the virtual pool were represented by the lines of division in the cubic enclosure (invisible to the participants) (A). The annulus in the southeast quadrant (invisible to the participants) marks the region that circumscribed the stationary hidden platform (A). In the control vMWM, the same pool used in the test environment was placed at the center of a circular arena. Four identical red diamonds were positioned equidistant from each other on the walls of the arena (B). A yellow halo facilitated the identification of the visible platform from afar (B). The location of the visible platform, together with the halo, changed from trial to trial in the control environment.

The participants performed 27 trials in the vMWM: 24 place trials (60 secs maximum per trial) interleaved by three probe trials (no hidden platform; 30 secs maximum per trial). Each probe trial was presented at: (i) the start of the task (before the first place trial), (ii) after the twelfth place trial, and (iii) after the twenty-fourth (final) place trial. The first probe trial served as an exploratory and familiarization phase, whereas the second and third probe trials served to investigate whether the participants searched closer to the hidden platform after experiencing 12 and 24 place trials respectively. Importantly, we administered 24 place trials because findings from previous research by Magnusson and colleagues (Magnusson, 1998; Magnusson, 2001; Magnusson et al., 2007) demonstrated that 24 place trials were adequate for detecting search performance differences between young and older subjects and in detecting slower performance improvement in aged individuals (unpublished observation).

The entry points were systematically varied between the northwest, northeast, and southwest quadrants of the pool (i.e., areas that did not contain the hidden platform). Figure 2A–C shows the progression in sample place and probe trials. At the start of the experiment, we informed the participants that there was no platform in the first trial (i.e., the first probe trial) and that they had to explore as much as they could in order to familiarize themselves with new vMWM. They were then instructed to search for the hidden platform in the subsequent 26 trials. To emulate the testing conditions of the rodent-based protocols as closely as possible (Magnusson, 1998; Magnusson et al., 2007, Zhao et al., 2009), the participants were not informed that the hidden platform was stationary across the place trials or about the presence and purpose of the two successive probe trials.

Figure 2.

Scenes from sample place (A – C), probe (A – B), and control (D – F) trials. The participants began each trial situated at the outer rim of the pool facing the deck (A, C). The maximum trial duration was 60 secs in a place trial and 30 secs in both probe (no hidden platform) and control trials. In a place trial, the hidden platform surfaced within 500 msecs after a participant crossed the boundary of the invisible circular zone containing it. The in-game display dimmed to signal contact with the platform in all place and control trials. Upon finding the platform (in the place and control trials) or exceeding the time limit (in any trial), all translational movements ceased. Thereafter, a participant could only make rotational movements about the vertical axis for 5 secs before the commencement of another trial.

Furthermore, a control vMWM was designed for assessing and controlling for visuomotor performance between the age groups (see Figure 1B). The participants performed six trials (30 secs maximum per trial) in this control environment after completing the 27 trials in the test environment, adding to a total of 33 trials. The round shape of the virtual arena and the uniformity of objects surrounding the virtual pool ensured that no geometric or object features could be used as cues for pinpointing the escape platform’s location. The escape platform was the same as that used in the test pool. However, it was kept visible over the top of the virtual water surface, and was programmed to vary in location and distance from the pool deck between trials. Specifically, the visible platform’s location varied between the four quadrants and the center of the pool across all control trials. The entry points were systematically varied between the northern, southern, eastern, and western ends. The combined variation in platform location and entry points was done to assess participants’ visuomotor ability to reach different target locations (spanning across the major sectors of the pool) from different entry points. They were instructed to move as efficiently as possible to the visible platform in every control trial. Figure 2D–F shows the progression in a sample control trial.

Data-Recording during Experimental Testing

A measure of cumulative proximity was computed by adding the participant’s initial Euclidean distance to the center of the platform (when he was at the entry point) to the participant’s subsequent Euclidean distances toward the center of the platform during virtual movements. The follow-up distances from the platform were computed once every 200 msecs until the hidden platform was found in the place trials or when the time-limit of 60 secs was reached. These distances were added to the initial Euclidean distance to provide the fully integrated cumulative proximity value in each trial. The same recording procedure was applied to the control trials until the visible platform was reached or when the time-limit of 30 secs was reached. In the probe trials, in which no platform was present, this recording procedure was applied to the integration of Euclidean distances over the entire trial duration of 30 secs in relation to the central coordinate of the circular region that encompassed the hidden platform. The cumulative proximity values attained by participants in the place and control trials were corrected by deducting the cumulative proximity of an “ideal” path. The cumulative proximity of this “ideal” path pertained to integrating the Euclidean distances from the entry point to the platform’s center over an “ideal” time taken to reach the platform. This “ideal” time was computed based on dividing the initial Euclidean distance to the platform’s center by the average speed of continuous virtual movements (i.e., movements that excluded momentary pauses) throughout a trial.

Overall, the corrected cumulative proximity (CCProx) measure accommodated for changes in entry points and their distances away from the center of the platform across multiple place or control trials, and over varying search times across such trials. By deducting the “ideal” cumulative proximity that considered the average speed of each participant, the CCProx measure also deterred average speed differences (derived from different search times over the same path-length) from confounding the interpretation of search accuracy. In other words, the CCProx measure was essentially the search error in a place or control trial after correcting for a participant’s “ideal” cumulative proximity (Gallagher et al., 1993). We applied the CCProx measure in the analyses of performance in the place and control trials and an average proximity measure in the analysis of performance in the probe trials. The latter measure was a time-weighted proximity measure derived from dividing the raw cumulative proximity value over the fixed duration of each probe trial (30 secs) (see, e.g., Magnusson, 1997, 1998, for instances of its previous use). Overall, higher proximity values reflected poorer search accuracy and lower proximity values reflected higher search accuracy.

Results

Place Trial Analysis

In the analysis of place trial performance, we followed the practice of other studies (e.g., Driscoll, Hamilton, Yeo, Brooks, & Sutherland, 2005; Gallagher et al., 1993; Magnusson et al., 2007) and averaged the CCProx values consecutively through every block of four trials, giving rise to six blocks of place trials. A mixed-model ANCOVA was performed with Age Groups (young versus older adults) and Blocks of Place Trials (6) as the independent variables, CCProx as the repeated dependent variable, and the mean path-length in the control trials (i.e., mean path-length when the platform was visible) as the covariate. This covariate was regarded as a direct behavioral measure of the level of visuomotor control in using the joystick to head toward a visible virtual target and was used to control for potential age group differences in joystick control. The inclusion of this covariate was also commensurate with its use as a control measure in a recent study on longitudinal change in vMWM search performance (see Daugherty & Raz, 2017).

After controlling for the significant covariate effect of mean visible path-length, F (1, 39) = 4.73, p = .036, η² = .108, there were significant effects of age group, F (1, 39) = 16.23, p < .001, η² = .294, and trial, F (5, 35) = 3.95, p = .006, η² = .361, and a non-significant interaction effect, F (5, 195) = 1.68, p = .142, η² = .041 (sphericity assumed) [see Figure 3A]. On average, older adults spent more time exploring sectors of the pool away from the hidden platform across the place trials.

Figure 3.

Corrected cumulative proximity (CCProx) from blocks of four place trials as a function of two age groups (A). Adjusted means and SEs (± 1) are shown after controlling for the covariate effect of mean control trial path-length. There were significant performance differences between the young and older adults (*) over all blocks of place trials. Scatterplot of the CCProx values (averaged across all place trials) in the young and older adults (B). For each age group, the means and SDs (± 1) are shown by the corresponding horizontal bars in the background. The dotted horizontal line shows the median CCProx value attained by the older adults (261.77).

To gain a closer examination of the performance differences between young and older adults, we divided the older adults into two subgroups based on the median CCProx they attained (261.77 virtual units). We performed this median-split based on the theoretical notion that proximity can be used as an index of individual differences in spatial learning among aged animals (Gallagher et al., 1993). We also consulted a similar dichotomization procedure used in previous MWM studies in rodents that focused on separating older subjects into two performance groups with respect to proximity (i.e., separating older subjects who scored above and below the mean cumulative or average proximity value of the younger subjects by 2 to 3 SDs of the young’s mean, see Zamzow, Elias, Acosta, Escobedo, & Magnusson, 2016; Gallagher et al., 1993; Rowe et al., 2007; Yetimler, Ulusoy, Çelik, & Jakubowska-Doğru, 2012). In addition, we examined the distribution of the CCProx values of the older adults [skewness = 0.48 (“0” in a normal distribution); kurtosis = −0.57 (“3” in a normal distribution)]. A one-sample Komolgorov-Smirnov test of goodness-of-fit on the older adults’ CCProx values showed no significant deviation from normality, D (21) = 0.76, p (two-tailed) = .608.

Based on the median-split, we obtained 10 older adults who scored above the median CCProx (> 261.77) and 11 older adults who scored at or below the median CCProx (≤ 261.77). The older adults’ median CCProx was higher than the young’s mean CCProx by 1.43 SD of the young’s mean (see Figure 3B). The mean CCProx of the first subgroup of 10 older adults was higher than the young’s mean CCProx by 2.48 SD of the young’s mean, and the mean CCProx of the second subgroup of 11 older adults was higher than the young’s mean CCProx by 0.65 SD of the young’s mean.¹ Consequently, with reference to the young’s search performance, we dubbed the first subgroup of older adults (mean CCProx > 261.77) as “poor” performers and the second subgroup of older adults (mean CCProx ≤ 261.77) as “good” performers.² Crucially, the two subgroups of older adults did not differ significantly in terms of age or any other demographic and pre-test variables (see Table 2), indicating that their vMWM search performance was not confounded by pre-existing group differences in any of those variables.

Table 2.

Descriptive statistics of the demographic and pretest performance variables of two subgroups of older adults derived from the median-split of corrected cumulative proximity (CCProx) values

	Old: Good performers (n = 11)			Old: Poor performers (n = 10)
	M (SD)	Min.	Max.	M (SD)	Min.	Max.
Age	66.27 (3.55)	60	70	65.10 (5.10)	60	79
Education	15.27 (2.00)	12	18	15.60 (2.46)	12	18
MMSE	29.45 (0.82)	28	30	28.90 (0.88)	28	30
CES-D score	5.36 (5.18)	0	13	4.00 (2.94)	0	9
Ishihara color-blindness	14.00 (0.00)	14	14	14.00 (0.00)	14	14
Contrast sensitivity	1.72 (0.06)	1.60	1.80	1.72 (0.09)	1.48	1.80
Speed test (sec)	64.91 (18.79)	55	120	62.80 (10.35)	50	80
No. of attempts in Practice vMWM	3.36 (1.12)	2	6	3.65 (1.48)	2	6

Note. Data presented in M (± SD). Education: Education of participant in years. MMSE: Mini-Mental State Examination. CES-D: Center for Epidemiologic Studies Depression scale.

We repeated the mixed-model ANCOVA as above with the Age/Performance groups adjusted to three levels (good- and poor-performing older adults, and younger adults). After controlling for the significant covariate effect of mean visible path-length, F (1, 38) = 6.82, p = .014, η² = .150, there were significant main effects of group, F (2, 38) = 26.63, p < .001, η² = .584, and trial, F (5, 34) = 3.15, p = .019, η² = .316, and a significant group x trial interaction, F (5, 35) = 3.94, p = .006, η² = .360 (Roy’s largest root criterion) (see Figure 4A). In finding the hidden platform, post-hoc Bonferroni comparisons showed that the young adults did not exhibit lower CCProx values than the good-performing older adults overall (p = .379), and that both groups exhibited lower CCProx values than the poor-performing older adults overall (ps < .001). Table 3 shows a summary of post-hoc independent t-tests in each block of place trials with the poor-performing older adults designated as the reference group. Based on an alpha of .004 (Bonferroni-corrected), the young adults and good-performing older adults exhibited significantly lower CCProx values than the poor-performing older adults from trial block 2 to 5 (ps ≤ .002). Only the young adults exhibited significantly lower CCProx than the poor-performing older adults in trial block 6 (p < .001).

Figure 4.

Corrected cumulative proximity values from the place and control trials (A, C) and average proximity values from the probe trials (B) as a function of age/performance groups and blocks of four trials or individual trials. In the analyses of place and probe trials, adjusted means and SEs (± 1) are shown after controlling for the covariate effect of mean control trial path-length. There were significant performance differences [p < .004 (Bonferroni-corrected)] between the young adults and the poor-performing adults (*) and between the two subgroups of older adults (#) within individual blocks of place trials (A) [block 2 to 5]. The young adults outperformed the poor-performing older adults in the probe trials overall (B), as well as both older subgroups in the first control trial (C).

Table 3.

Post-hoc between-groups contrasts across six place trial blocks

Measure	Place Trial Block	Young > Old: Poor performers			Old: Good performers > Old: Poor performers
		M (SE) Difference	t (29)	p-value	M (SE) Difference	t (19)	p-value
Corrected Cumulative Proximity (CCProx) [in virtual units]	1	− 102.85 (57.24)	− 1.80	.082	− 58.63 (67.96)	− 0.86	.401
	2	− 217.48 (51.28)	− 4.24	< .001	− 202.58 (57.75)	− 3.51	.002
	3	− 269.15 (51.09)	− 5.27	< .001	− 246.20 (58.72)	− 4.19	< .001
	4	− 299.26 (50.20)	− 5.96	< .001	− 243.68 (56.00)	− 4.35	< .001
	5	− 370.44 (58.80)	− 6.30	< .001	− 296.20 (58.58)	− 5.06	.001
	6	− 275.79 (58.63)	− 4.70	< .001	− 167.90 (80.76)	− 2.08	.051
Path-length (in virtual units)	1	− 10.12 (10.36)	− 0.98	.335	− 4.36 (10.83)	− 0.40	.694
	2	− 39.70 (11.78)	− 3.37	.002	− 36.97 (13.43)	− 2.75	.013
	3	− 42.36 (10.98)	− 3.86	< .001	− 41.66 (13.14)	− 3.17	.005
	4	− 47.36 (9.73)	− 4.87	< .001	− 41.32 (10.80)	− 3.82	.001
	5	− 52.58 (10.46)	− 5.02	< .001	− 42.30 (11.44)	− 3.70	.002
	6	− 35.89 (91.19)	− 3.91	< .001	− 16.18 (12.66)	− 1.28	.216
Escape Latency (secs)	1	− 9.01 (4.51)	− 2.00	.055	− 4.65 (5.17)	− 0.90	.379
	2	− 19.32 (4.37)	− 4.42	< .001	− 15.99 (4.98)	− 3.21	.005
	3	− 20.94 (4.17)	− 5.02	< .001	− 18.56 (4.89)	− 3.80	.001
	4	− 22.98 (3.87)	− 5.93	< .001	− 17.93 (4.35)	− 4.13	< .001
	5	− 26.51 (4.35)	− 6.09	< .001	− 19.23 (4.47)	− 4.30	< .001
	6	− 19.88 (4.18)	− 4.75	< .001	− 8.05 (5.76)	− 1.40	.178

Note. Post-hoc t-tests were performed on the corrected M (SE) values adjusted for the covariate effect of mean visible pathlength (as shown in Figures 4A and 5).

With regard to the trial effect in each group, the young adults exhibited steady declines in CCProx values across the trial blocks, F (5, 15) = 19.65, p < .001, η² = .868, signifying improvements in search accuracy. The good-performing older adults appeared to behave like the younger adults in exhibiting relatively lower CCProx values in the trial blocks that succeeded the first block, but the overall trial/learning effect was only marginally significant, F (5, 5) = 4.65, p = .058, η² = .823. As for the poor-performing older adults, they maintained a relatively high level of CCProx values across the trial blocks, signifying negligible improvements in search accuracy, F (5, 4) = 0.14, p = .974, η= .147.

To ascertain that these findings, derived from the median-split of older adults’ CCProx values, were not reflective of a superficial performance-based effect that occurred irrespective of age, we divided the young adults into two subgroups based on a median-split of their CCProx values (median = 93.03), with 11 good-performing young adults scoring below the median and 10 poor-performing young adults scoring above the median, and then repeated the mixed-model ANCOVA as above with the Age/Performance groups updated to four levels [good- and poor-performing older adults versus good- and poor-performing younger adults]. The overall pattern of results was similar to that found with three groups. There were significant main effects of group, F (3, 37) = 21.74, p < .001, η² = .638, and trial, F (5, 33) = 4.57, p = .003, η² = .409, and a significant group x trial interaction, F (5, 35) = 4.07, p = .005, η² = .368 (Roy’s largest root criterion).

Focusing on the two subgroups of young adults, post-hoc Bonferroni comparisons of their overall performance showed that they did not differ significantly between each other (p = .142), nor with the good-performing older adults (p = .060 for the comparison involving young good-performers; p = 1.00 for the comparison involving young poor performers). Both younger subgroups differed significantly from the poor-performing older adults (ps < .001). It is worth noting that the two younger subgroups did not differ significantly between each other despite the median-split because their CCProx values were largely clustered around the median/mean and exhibited minimal variation (see Figure 3B). Overall, in view that the median split of the young adults did not yield any meaningful between-subjects findings that were discrepant from the pre-existing findings (no median split of the young adults), we interpreted our results with the young adults maintained as one group and conducted all subsequent analyses without any division of the young adults.

Additional analyses of path-length and escape latency

With due consideration that path-length and escape latency are traditional and popular metrics of MWM search performance, we analyzed our findings further in terms of both types of measures. We repeated the mixed-model ANCOVA as above with the Age/Performance groups re-adjusted to three levels (good- and poor-performing older adults, and younger adults) and with the dependent variable entered as path-length (in virtual units) and escape latency (in seconds) in succession.

With path-length set as the dependent variable, after controlling for the significant covariate effect of mean visible path-length, F (1, 38) = 9.68, p = .004, η² = .203, there were significant main effects of group, F (2, 38) = 14.06, p < .001, η² = .425, and trial, F (5,34) = 4.37, p = .004, η² = .391, and a significant group x trial interaction, F (5, 35) = 4.22, p = .004, η² = .376 (Roy’s largest root criterion) in blocks of place trials [see Figure 5A]. In finding the hidden platform, post-hoc Bonferroni comparisons showed that the young adults and good-performing older adults exhibited comparable path-lengths overall (p = .862), and that both groups traveled shorter path-lengths than the poor-performing older adults overall (ps ≤ .001). Table 3 shows a summary of post-hoc independent t-tests in each block of place trials with the poor-performing older adults designated as the reference group. Based on an alpha of .004 (Bonferroni-corrected), the young adults exhibited significantly shorter path-lengths than the poor-performing older adults from trial block 2 to 6 (ps ≤ .002). The good-performing older adults exhibited significantly shorter path-lengths than the poor-performing older adults in trial blocks 4 and 5 (ps ≤ .002), and marginally significant shorter path-lengths than the poor-performing older adults in trial blocks 2 and 3 (ps ≤. 013).

Figure 5.

Path-length (A) and escape latency (B) from the place trials as a function of age/performance groups and blocks of four place trials. Adjusted means and SEs (± 1) are shown after controlling for the covariate effect of mean control trial path-length (A, B). In each block, significant performance differences [p < .004 (Bonferroni-corrected)] between the young adults and the poor-performing older adults were indicated by the * signs while significant performance differences between the two older subgroups were indicated by the # signs.

An examination of the trial effect in each group showed that the young adults exhibited a steady shortening of path-lengths across the trial blocks, F (5, 15) = 13.69, p < .001, η² = .820. The good-performing older adults exhibited shorter path-lengths in the blocks of place trials that succeeded the first block, with an overall trial/learning effect that trended toward but did not reach significance, F (5, 5) = 3.33, p = .107, η² = .769. In contrast, the poor-performing older adults exhibited relatively long path-lengths across all trial blocks, (5, 4) = 0.138, p = .974, η² = .147.

With escape latency set as the dependent variable, there were a significant main effect of group, F (2, 38) = 21.06, p < .001, η² = .526, a significant group x trial interaction, F (5, 35) = 4.00, p = .006, η² = .363 (Roy’s largest root criterion), and a trial effect that was marginally significant, F (5, 34) = 2.31, p = .066, η² = .254, in blocks of place trials (see Figure 5B). In finding the hidden platform, post-hoc Bonferroni comparisons showed that the young adults did not spend less travel time than the good-performing older adults overall (p = .178), but that both groups spent less travel time than the poor-performing older adults overall (ps ≤ .001). Table 3 shows a summary of post-hoc independent t-tests in each block of place trials with the poor-performing older adults designated as the reference group. Based on an alpha of .004 (Bonferroni-corrected), the young adults exhibited significantly less travel time than the poor-performing older adults from trial block 2 to 6 (ps < .001). The good-performing older adults exhibited significantly less travel time than the poor-performing older adults from trial block 3 to 5 (ps ≤ .001). The former group exhibited less travel time than the latter group in trial block 2 with a difference that was marginally significant (p = .005).

An examination of the trial effect in each group showed that the young adults exhibited gradual declines in travel time across the trials, F (5, 15) = 13.69, p < .001, η² = .820. The good-performing older adults spent less travel time in the blocks of place trials that succeeded the first block, with an overall trial/learning effect that was marginally significant, F (5, 5) = 7.33, p = .024, η² = .880 (based on a Bonferroni-corrected alpha of .017), approximating the trial effect attained with CCProx. In contrast, the poor-performing older adults spent a relatively large and even amount of travel time across the trials, F (5, 4) = 0.348, p = .861, η² = .303.

Correlations between vMWM measures

In general, the findings above were similar across vMWM measures, although escape latency appeared to offer a more similar pattern of results to the findings based on CCProx than did path-length. To further assess the magnitude of the association between these measures, we conducted correlations between vMWM outcome variables after averaging across all place trials. The correlation between CCProx and escape latency was [r (42) = .95, p < .001] and was marginally higher than the correlation between CCProx and path-length [r (42) = .87, p < .001] and between path-length and escape latency [r (42) = .80, p < .001].

Probe Trial Analysis

A mixed-model ANCOVA was performed with Age/Performance Groups (3) (same classification as that in the place trial analysis, with one young group and two older subgroups) and Probe Trials (3) as the independent variables, average proximity as the repeated dependent variable, and mean visible path-length as the covariate. There were significant effects of group, F (2, 38) = 3.83., p < .031, η² = .168, and trial, F (2, 37) = 30.82, p < .001, η² = .625. The group x trial interaction did not reach significance, F (2, 38) = 1.09, p = .346, η= .054 (Roy’s largest root criterion) (see Figure 4B). Post-hoc Bonferroni comparisons showed that the poor-performing older adults exhibited significantly higher average proximity values than the younger adults overall (p = .024), but not significantly higher average proximity values than the good-performing older adults overall (p = .264). As in the place trial analyses, the young adults exhibited comparable performance as the good-performing older adults overall (p = .569). With regard to the trial/learning effect, the average proximity values in the first probe trial were significantly higher than those in the second (p < .001) and third probe trials (p < .001) across all the participants.

Control Trial Analysis

A mixed-model ANOVA was performed with Age/Performance Groups (3) (same classification as that in the place trial analysis, with one young group and two older subgroups) and Control Trials (6) as the independent variables, CCProx as the repeated dependent variable. There were significant effects of group, F (2, 39) = 8.22, p < .001, η² = .331, and trial, F (5, 35) = 37.99, p < .001, η² = .844, and a significant group x trial effect, F (5, 36) = 2.53, p = .047, η² = .260 (Roy’s largest root criterion). Post-hoc Tukey’s HSD test showed that the younger adults exhibited better visible platform search performance than both good- (p = .017) and poor-performing older adults overall (p = .005). More specifically, older adults (both good and poor performers) exhibited higher CCProx values than the younger adults in the first trial, F (2, 39) = 7.30, p = .002, η² = .272, but not in the remaining trials, based on a Bonferroni-corrected alpha of .008 (see Figure 4C).

To further examine whether the group differences in search performance for the visible platform were underpinned by corresponding differences in visuomotor control (i.e., in the path-length traversed) and escape latency, the mixed-model ANOVA was repeated with path-length and latency (in secs) as the dependent variables in succession. The group effect was non-significant (p = .301) with respect to path-length but was significant with respect to latency, F (2, 39) = 4.38, p = .019, η² = .183. The trial and interaction effects were also significant with respect to latency (Fs ≥ 3.82, ps ≤ .007), reproducing similar patterns of results as those attained with CCProx. There were significant group differences in the first control trial, F (2, 39) = 7.16, p = .002, η² = .268. Post-hoc Tukey’s HSD test showed that the young adults were quicker at reaching the visible platform performance than both good- (p = 0.081) and poor-performing older adults (p = 0.002) in the first control trial. No group differences in latency were present in the remaining trials based on a Bonferroni-corrected alpha of .008.

Discussion

The current study examined age-related differences in spatial learning and memory in a vMWM task adapted from previous protocols applied to mouse models (Magnusson, 1998; Magnusson et al., 2007; Zhao et al., 2009) and with proximity as the main dependent measure. The subdivision of older adults into good and poor performers based on a median-split of CCProx values showed that the age effects observed in the analyses of place and probe trials were largely attributed to the significant differences in search performance between young and poor-performing older adults. While the good-performing older adults exhibited comparable search performance as the younger adults across most place and probe trials, the poor-performing older adults exhibited significantly poorer spatial learning than the younger adults in 83% of the place trials. Critically, these group differences occurred even though the good-performing older adults were closely matched to the poor-performing adults in terms of age, education, MMSE scores and visible platform search performance. Therefore, this study showed that CCProx can be used as a sensitive marker of individual differences in spatial learning not just among rodents (Gallagher et al., 1993), but among aged humans as well.

Moreover, it is worth noting the larger CCProx values exhibited by the older adults in the control trials were related to them taking more time (but not longer path-length) in the first control trial than the young. This suggests that older adults need more time than the young to readjust to the navigational demands of a new virtual environment when exploring it for the first time. Notably, this pattern of group differences in control trial performance (with respect to CCProx) complemented the same pattern of results that were widely exhibited by C57BL/6 mice (Das et al., 2012; Magnusson, 2001; Magnusson et al., 2007; Zamzow, Elias, Shumaker, Larson, & Magnusson, 2013; Zhao et al., 2009). In conjunction, these matching patterns of results suggest that there might be cross-species age-related declines in cognitive flexibility (i.e., readapting to changing virtual environments in the current study), which have been previously found in both human and non-human primates (e.g., Boone, Ghaffarian, Lesser, Hill-Gutierrez, & Berman, 1993; Moore, Killiany, Herndon, Rosene, & Moss, 2006), and rodents (e.g., Schoenbaum, Setlow, Saddoris, & Gallagher, 2006; Zamzow et al., 2016). The negligible group differences in path-length across the control trials, together with the smaller group differences in CCProx in the first control trial than in the place trials, further suggest that age-related difficulties with visuomotor control could not fully explain the place learning difficulties experienced by the poor-performing older adults.

Interestingly, the differences in the rate of spatial learning between the young and older adults (good and poor performers combined) were similar to the patterns of age-related decline in MWM performance documented by previous MWM studies (in humans, see Driscoll et al., 2003; in mice, see Magnusson et al., 2007). The current study, however, was novel for showing that the average search performance of all older adults was mediated by the disparate search performances of two subgroups, which exhibited two distinct learning trends compared to the young. With reference to previous fMRI studies, which showed that good-performing older adults engaged in compensatory increases in prefrontal activity to achieve the same level of performance as the young (with respect to accuracy) across a wide variety of memory tasks (e.g., word recognition and recall, see Cabeza et al., 1997, 2004; Cabeza, Anderson, Houle, Mangels, & Nyberg, 2000; Madden et al., 1999; spatial location matching, see Grady et al., 1994; lexical decision, see Madden et al., 1996; letter-span encoding and recognition, see Rypma & D’Esposito, 2000), it is possible that this functional compensation might similarly apply to the good-performing older adults in the current task. Taking a comparative approach, there is complementary evidence from mice showing that there are differential effects of aging on the phosphorylation and cleavage of N-methyl-D-aspartate (NMDA) receptors in the frontal cortices of poor-versus good-performing older mice (in MWM place trials), suggesting that performance differences in aged animals may be manifested at the synaptic level (Zamzow et al., 2016).

In addition, the poor-performing older adults may have had lower functional activity in the hippocampus and/or prefrontal cortex when compared to the young or good-performing older adults. This inference corresponds with previous fMRI studies showing that healthy older adults have lower activation in the hippocampus/parahippocampal complex when performing spatial navigation tasks in virtual environments (e.g., Moffat, Elkins, & Resnick, 2006; Meulenbroek, Petersson, Voermans, Weber, & Fernandez, 2004; Antonova et al., 2009). Importantly, as cognitive mapping through the implementation of a place strategy (i.e., a spatial strategy that attends to the processing of interobject relations) has been widely touted as being hippocampal-dependent in nature (see, e.g., Bohbot, Iaria, & Petrides, 2004; Driscoll et al., 2003, 2005; Iaria, Petrides, Dagher, Pike, & Bohbot, 2003; Konishi & Bohbot, 2013; Laczó et al., 2009; Moffat & Resnick, 2002; Rodriguez, 2010; Zhong, 2013), a potential deficit in place strategy use among the poor-performing older adults could have stemmed from a reduction in hippocampal activity (Antonova et al., 2009; Moffat et al., 2006) and/or volume (Daugherty et al., 2015; Driscoll et al.,2003, Moffat, Kennedy, Rodrigue, & Raz 2007; Raz et al., 2005; Yuan et al., 2014). This possibility gains support from additional MRI findings showing that vMWM search accuracy was positively associated with both hippocampal BOLD responses (Astur et al., 2006; Folley, Astur, Jagannathan, Calhoun, & Pearlson, 2010; Shipman & Astur, 2008) and volume (Daugherty et al., 2015; Driscoll et al., 2003; Konishi & Bohbot, 2013; Korthauer et al., 2016; Moffat et al., 2007). To our knowledge, there are no studies directly linking differences in hippocampal functional activity between poor-performing older adults and younger adults to the vMWM or any other spatial navigation task.

Furthermore, the lower overall search accuracy of the poor-performing older adults might have stemmed from certain difficulties in switching between egocentric (response/procedural) and allocentric (place) strategies (Harris, Wiener, & Wolbers, 2012), or an overall predisposition toward using a landmark-based egocentric strategy that eschewed attention to interobject/allocentric spatial relations when searching for the hidden platform (cf. Bohbot et al., 2012; Wiener, de Condappa, Harris, & Wolbers, 2013). At the neural level, the potential implementation of non-spatial egocentric strategies might have corresponded to an age-related shift in the locus of functional activity from the hippocampus toward the striatum when processing navigationally relevant information (Konishi et al., 2013; Schuck, Doeller, Polk, Lindenberger, & Li, 2015a; Bohbot et al., 2012; Chersi & Burgess, 2015). This has been proposed as an adaptive mechanism that frees up hippocampal-dependent resources for facilitating repetitive or stereotyped navigational behavior (Bohbot et al., 2012) and it may be worthwhile for future studies to ascertain whether the poor-performing older adults in the vMWM would exhibit a greater engagement of the striatum than of the hippocampus. In addition, considering the recent evidence implicating the medial prefrontal cortex (mPFC) as being pertinent for encoding information about the successful implementation of an alternative learning strategy in future task performance (see Schuck et al., 2015b), it will be interesting to see if the young adults and good-performing older adults would exhibit differential levels of activation in the mPFC compared to the poor-performing older adults because of flexible adaptation of alternative learning strategies during vMWM performance.

Along with these strategy-related concerns, it is also important to note that other procedural factors could also explain the relatively poor performance of a subgroup of older adults. As our task was designed to replicate previous procedures used in mice (Magnusson, 1998; Magnusson et al., 2007; Zhao et al., 2009), we did not inform participants that the platform was stationary across trials. Therefore, it is possible that the good-performing older adults were quicker than the poor-performing older adults at realizing that the platform was stationary and engaged in more effective encoding and/or retrieval operations than the poor-performing older adults. It is possible that the poor-performing older adults assumed a moving platform or otherwise not realized that it remained in the same location across trials. However, we could not verify these possibilities because we did not query participants on their assumptions or expectations about the workings of the task and about how such assumptions/expectations might have changed during the progression of the task. Nevertheless, it seems reasonable that when facing ambiguity in vMWM performance (i.e., not being explicitly told that the platform is stationary), individual differences in the comprehension or interpretation of task demands have the potential to affect search performance. If one were to assume a stationary platform, engaging in an associative learning strategy (analogous to the place strategy) to learn the cue locations and the relationships between those locations and the stationary platform shall greatly facilitate the search for the hidden platform (see, e.g., Head & Isom, 2010; Ngo, Weisberg, Newcombe, & Olson, 2016; Zhong & Moffat, 2016, for instances of associative learning in spatial navigation). On the other hand, if one were to assume a randomly moving platform, the use of a place strategy would most likely become ineffectual at finding the hidden platform in the fastest way. Crucially, it is also worth noting that the poor-performing older group showed virtually no evidence of learning across the place trials, despite experiencing 24 trials that provided adequate opportunities to learn that the platform was stationary. This incapacity for spatial learning in the current vMWM task suggests that the poor-performing older adults possessed some form of undocumented perceptual or cognitive impairment. As aforementioned, the observed impairment may relate to individual differences among the older adults in forming assumptions/expectations of the task, and may not relate to a deficit in spatial processing or learning per se.

To our knowledge, no studies have directly evaluated the influence of different vMWM instructions on subsequent performance. Further investigations on the differential effects of instructions would not only be useful for a greater understanding of MWM spatial learning in humans, but could also broaden the interpretations of rodent-based MWM results. Since verbal instructions cannot be given to mice, the current findings suggest that MWM performance difference between healthy young and aged mice should not just emphasize spatial learning or memory deficits (as stipulated by previous studies, see Das & Magnusson, 2011; Gallagher et al., 1993; Magnusson, 1997, 1998, 2001; Magnusson et al., 2007; Zamzow et al., 2013, 2016; Zhao et al., 2009), but should also consider the possibilities of mice (of different age or performance groups) adopting different “expectations” of task demands and different search strategies.

At the same time, future studies should also consider administering relevant cognitive or neuropsychological tests for screening subjects or as covariate measures during data analysis. This is because cognitive tests that assess visuospatial working memory (e.g., Vandenberg Mental Rotation Test) and executive functioning (e.g., Raven’s Matrices Test) have been recently shown to mediate the effect of age on navigational performance in both real-world and virtual settings (see Taillade, N’Kaoua, & Sauzéon, 2016); hence, it will be important to assess and control for the potential influence of age differences in such cognitive abilities on vMWM performance. These psychometric measures can also be used in conjunction with self-report questionnaires on navigational ability and navigation strategy use [see, e.g., the Santa Barbara Sense-of-Direction Scale (Hegarty et al., 2002); the Navigation Strategy Questionnaire (Zhong & Kozhevnikov, 2016); the German Questionnaire on Spatial Strategies (Münzer, Fehringer, & Kühl, 2016)] if individual differences in inherent navigation strategy preferences were to be assessed and controlled for.

In summary, we found robust age differences in vMWM performance and described the performance of two subgroups of older adults, one of which is indistinguishable from younger adults and the other which is markedly impaired compared to both younger adults and their age-matched peers. The poor behavioral performance of the poor-performing subgroup could be related to existing brain-related differences, and/or to adopting different strategies in understanding and performing the task. Considering the evidence showing that spatial memory impairment in the form of poor MWM search performance is commonly seen among human subjects who are at risk of getting Alzheimer’s disease (AD) (e.g., patients with amnesic mild cognitive impairment (MCI), see, e.g., Hort et al., 2007; Laczó et al., 2009, 2010, 2015), we recommend further investigations into the tendency for developing MCI or AD among healthy older adults with behavioral signs of spatial memory impairment.