Mild cognitive impairment (MCI) has been conceptualized as an intermediate stage between normal cognitive function and dementia (Winblad et al., Reference Winblad, Palmer, Kivipelto, Jelic, Fratiglioni, Wahlund, Nordberg, Bäckman, Albert, Almkvist, Arai, Basun, Blennow, de Leon, DeCarli, Erkinjuntti, Giacobini, Graff, Hardy, Jack, Jorm, Ritchie, van Duijn, Visser and Petersen2004). People with MCI experience objective and sometimes subjective declines in cognition, to a greater extent than would be expected for their age and education level, while maintaining independent functioning and not meeting clinical criteria for dementia or other neurological disorders (Albert et al., Reference Albert, DeKosky, Dickson, Dubois, Feldman, Fox, Gamst, Holtzman, Jagust, Petersen, Snyder, Carrillo, Thies and Phelps2011; Winblad et al., Reference Winblad, Palmer, Kivipelto, Jelic, Fratiglioni, Wahlund, Nordberg, Bäckman, Albert, Almkvist, Arai, Basun, Blennow, de Leon, DeCarli, Erkinjuntti, Giacobini, Graff, Hardy, Jack, Jorm, Ritchie, van Duijn, Visser and Petersen2004). Although MCI diagnostic criteria require that activities of daily living are essentially unimpaired, functional impairments can be observed in more complex tasks, including bill paying (Griffith et al., Reference Griffith, Belue, Sicola, Krzywanski, Zamrini, Harrell and Marson2003), driving (Anstey et al., Reference Anstey, Eramudugolla, Chopra, Price, Wood and Bondi2017), and decision-making in health and financial contexts (Griffith et al., Reference Griffith, Okonkwo, den Hollander, Belue, Copeland, Harrell, Brockington, Clark and Marson2010; Martin et al., Reference Martin, Gerstenecker, Triebel, Falola, McPherson, Cutter and Marson2019; Okonkwo et al., Reference Okonkwo, Griffith, Copeland, Belue, Lanza, Zamrini, Harrell, Brockington, Clark, Raman and Marson2008). A diagnosis of MCI confers an increased risk of conversion to dementia, estimated in the range of 5–10% annually (Mitchell & Shiri-Feshki, Reference Mitchell and Shiri-Feshki2009), or 25–65% over a five-year period (Darmanthé, Tabatabaei-Jafari, & Cherbuin, Reference Darmanthé, Tabatabaei-Jafari and Cherbuin2021).
The cognitive impairments observed in MCI often involve memory (amnestic MCI, aMCI) but can also involve other cognitive domains such as executive functions, language, or visuo-construction (Albert et al., Reference Albert, DeKosky, Dickson, Dubois, Feldman, Fox, Gamst, Holtzman, Jagust, Petersen, Snyder, Carrillo, Thies and Phelps2011). Diagnostic classifications based on the type and number of impaired domains on neuropsychological tests yield subtypes of single- and multi-domain amnestic and non-amnestic MCI, which are associated with different patterns of neuropathology (Csukly et al., Reference Csukly, Sirály, Fodor, Horváth, Salacz, Hidasi, Csibri, Rudas and Szabó2016). Amnestic MCI is associated with higher rates of Alzheimer’s disease (AD) type dementia (Reinvang, Grambaite, & Espeseth, Reference Reinvang, Grambaite and Espeseth2012), while the non-amnestic (naMCI) subtypes are more predictive of other (non-AD) forms of dementia (Allain, Etcharry-Bouyx, & Verny, Reference Allain, Etcharry-Bouyx and Verny2013; Wadley et al., Reference Wadley, Crowe, Marsiske, Cook, Unverzagt, Rosenberg and Rexroth2007). Impairment in multiple domains is associated with higher rates of progression to dementia (Jung et al., Reference Jung, Park, Jang, Cho, Kim, Kim, Kim, Na, Seo and Kim2020).
Secondary to memory impairment, impairments in attentional and executive functions are thought to be the most important and commonly affected cognitive domains in aMCI (Reinvang et al., Reference Reinvang, Grambaite and Espeseth2012) and are associated with higher rates of progression to dementia (Belleville, Fouquet, Hudon, Hervé Tchala Vignon, & Croteau, Reference Belleville, Fouquet, Hudon, Hervé Tchala Vignon and Croteau2017). Executive functions encompass a range of cognitive and functional abilities, including planning, working memory, attentional and inhibitory control, and feedback processing (Chan, Shum, Toulopoulou, & Chen, Reference Chan, Shum, Toulopoulou and Chen2008). In studies which have investigated executive functions by MCI subtypes, people with multi-domain aMCI have shown the clearest impairments (Klekociuk & Summers, Reference Klekociuk and Summers2014; Pereiro, Juncos-Rabadan, & Facal, Reference Pereiro, Juncos-Rabadan and Facal2014). Brandt et al. (Reference Brandt, Aretouli, Neijstrom, Samek, Manning, Albert and Bandeen-Roche2009) used empirically derived components from 18 executive function tests to define three subdomains of executive function, which they labeled “planning/problem-solving,” “working memory,” and “judgement.” In their study, the planning/problem-solving and working memory domains reliably discriminated between those with and without MCI, with impairments observed for all four MCI subtypes, and strongest impairments among those with multi-domain MCI. Given the established role of executive functions in deliberative decision-making processes (Schiebener & Brand, Reference Schiebener and Brand2015a; Schiebener et al., Reference Schiebener, Wegmann, Gathmann, Laier, Pawlikowski and Brand2014), this suggests that there may be impairment in decision-making among those with MCI, in particular among those with impairments in executive function domains.
Decision-making performance has been assessed behaviorally using tasks which vary in terms of task parameters (e.g., risk/reward contingencies), availability of information, and optimal strategies. In “ambiguous” decision-making tasks, no information is provided about task parameters or advantageous choices, and participants must learn these associations from experience (Bechara, Damasio, Tranel, & Damasio, Reference Bechara, Damasio, Tranel and Damasio1997). In “explicit risk” decision-making tasks information is provided about the task parameters, enabling participants to deduce and implement optimal strategies (Schiebener & Brand, Reference Schiebener and Brand2015a). These two types of tasks broadly map onto the conceptual heuristic of two separable “decision-making systems”; the fast, implicit and relatively effortless “impulsive” system and the slow, explicit and controlled “deliberative” system (Liebherr, Schiebener, Averbeck, & Brand, Reference Liebherr, Schiebener, Averbeck and Brand2017; Tversky & Kahneman, Reference Tversky and Kahneman1986). Explicit risk decision-making tasks are thought to require processing within the “deliberative” system, with performance drawing on executive functions, logical reasoning, and feedback processing (Brand, Laier, Pawlikowski, & Markowitsch, Reference Brand, Laier, Pawlikowski and Markowitsch2009; Schiebener & Brand, Reference Schiebener and Brand2015a).
The Game of Dice Task (GDT) is an important test of decision-making under explicit risk conditions, and has been used extensively, in both clinical and nonclinical samples (Brand et al., Reference Brand, Fujiwara, Borsutzky, Kalbe, Kessler and Markowitsch2005). In the GDT participants make a series of gambles on the number of a rolled dice, attempting to maximize their capital. Participants can select between one and four numbers on each trial, thus enabling a conservative (e.g., four number choices have a moderate success probability along with smaller gains and losses) or risky strategy (e.g., single numbers have low success probability and larger gains and losses). Performance is typically measured by quantifying the “net score” (quantity of “safe” three or four number choices minus quantity of “risky” one or two number choices), single number choices (quantity of “riskiest” single number choices), and the number of strategy changes (changes between “safe” and “risky” response strategies on consecutive trials). The availability of task parameter information enables participants to deduce that a conservative strategy will be more profitable over time. Studies have suggested that executive functions (Brand & Schiebener, Reference Brand and Schiebener2013; Schiebener et al., Reference Schiebener, Wegmann, Gathmann, Laier, Pawlikowski and Brand2014), logical reasoning (Schiebener & Brand, Reference Schiebener and Brand2015b), numerical processing (Brand, Schiebener, Pertl, & Delazer, Reference Brand, Schiebener, Pertl and Delazer2014), working memory resources (Starcke, Pawlikowski, Wolf, Altstotter-Gleich, & Brand, Reference Starcke, Pawlikowski, Wolf, Altstotter-Gleich and Brand2011), and episodic learning (Sinclair, Eramudugolla, Brady, Cherbuin, & Anstey, Reference Sinclair, Eramudugolla, Brady, Cherbuin and Anstey2021) underpin successful performance on the GDT. It is therefore not surprising that impaired GDT performance has been observed among clinical populations in which cognition is impaired, including people with Korsakoff’s dementia (Brand et al., Reference Brand, Fujiwara, Borsutzky, Kalbe, Kessler and Markowitsch2005), AD (Delazer, Sinz, Zamarian, & Benke, Reference Delazer, Sinz, Zamarian and Benke2007; Sun et al., Reference Sun, Xie, Wang, Zhang, Tian, Wang, Yu and Wang2020), and Parkinson’s disease (Euteneuer et al., Reference Euteneuer, Schaefer, Stuermer, Boucsein, Timmermann, Barbe, Ebersbach, Otto, Kessler and Kalbe2009).
Existing studies of explicit risk decision-making among people with MCI have indicated subtle patterns of impairment, which are typically observable when decision tasks are ambiguous, complex, or draw heavily on affected domains of cognition (Pertl, Benke, Zamarian, & Delazer, Reference Pertl, Benke, Zamarian and Delazer2017). Zamarian et al. (Reference Zamarian, Weiss and Delazer2011) found that people with MCI showed subtle impairments on a modified version of the revised Probability Assisted Gambling task (PAG-R), an explicit risk task. Studies employing the GDT have shown a mixed pattern of results among people with MCI (Fernandes, Macedo, Barbosa, & Marques-Teixeira, Reference Fernandes, Macedo, Barbosa and Marques-Teixeira2021). Jacus et al. (Reference Jacus, Fau, Raffard and Gély-Nargeot2013) found that people with MCI showed lower net scores on the GDT compared to older adults without cognitive impairment, along with an increased number of highest-risk “single number” choices. Sun et al. (Reference Sun, Xie, Wang, Zhang, Tian, Wang, Yu and Wang2020) also showed this increased number of single number choices among people with MCI, but no differences on the overall net score. Pertl et al. (Reference Pertl, Benke, Zamarian and Delazer2015) found that people with MCI were unimpaired on basic decision tasks, but showed suboptimal decision-making on the modified “Game of Dice Task-Double” (GDT-D), which places increased demands on numerical and probability processing abilities relative to the standard GDT. These observational studies were based on smaller samples recruited in clinical settings, limiting the ability to control for potentially confounding variables between MCI and control cases. Furthermore, none of these studies were able to report data by MCI subtype, and it is not as yet known whether GDT performance varies by MCI subtype.
In the current study we aimed to better understand the impact of MCI on decision-making, through a cross-sectional investigation of decision-making performance on the GDT among older adults with and without MCI. Using a population-based sample recruited as part of a large longitudinal study enabled matching of people with MCI and cognitively unimpaired control cases on relevant confounding variables, as well as analysis of MCI subtypes. Based on the known contribution of executive function abilities to performance on the GDT (Schiebener et al., Reference Schiebener, Wegmann, Gathmann, Laier, Pawlikowski and Brand2014), it is reasonable to propose that GDT performance may be more impacted among those with MCI subtypes in which executive functions are affected, in particular those with multi-domain aMCI. We hypothesized (H1) that compared to those without cognitive impairment, older adults with MCI would show evidence of decision-making impairment on the GDT, and (H2) that the level of impairment would be most apparent among those with multi-domain MCI. We also hypothesized (H3) that decision-making performance would be associated with measures of executive functions.
Methods
The Personality and Total Health (PATH) Through Life study is a population-based longitudinal cohort study, which recruited participants residing in the Australian cities of Canberra and Queanbeyan and aged within narrow age cohorts (20–24, 40–44, and 60–64 years) at wave 1 (1999–2002) via random sampling from the electoral roll (Anstey et al., Reference Anstey, Christensen, Butterworth, Easteal, Mackinnon, Jacomb, Maxwell, Rodgers, Windsor, Cherbuin and Jorm2012). Electoral roll enrolment is compulsory in Australia. The current study uses data from the older adult cohort (N = 2551 participants aged 60–66 years, with 58.3% response rate at wave 1), with a focus on outcome measures from the wave 4 data collection (2014–2015), in which 2048 participants were invited to respond, with data collected from 1644 participants aged 72–78 years (Anstey et al., Reference Anstey, Butterworth, Christensen, Easteal, Cherbuin, Leach, Burns, Kiely, Mortby, Eramudugolla and Gad2021). The current study is reported in line with the Strengthening Reporting of OBservational studies in Epidemiology (STROBE) checklist and guidelines (von Elm et al., Reference von Elm, Altman, Egger, Pocock, Gøtzsche, Vandenbroucke and Vandenbroucke2007).
Participants
Of the 1644 older participants surveyed at wave 4, 1287 completed all GDT trials and were eligible for inclusion in the study. We defined a subgroup of participants meeting International Working Group (IWG) criteria for MCI (Winblad et al., Reference Winblad, Palmer, Kivipelto, Jelic, Fratiglioni, Wahlund, Nordberg, Bäckman, Albert, Almkvist, Arai, Basun, Blennow, de Leon, DeCarli, Erkinjuntti, Giacobini, Graff, Hardy, Jack, Jorm, Ritchie, van Duijn, Visser and Petersen2004), using a validated diagnostic algorithm, which has been described previously in greater detail (Eramudugolla et al., Reference Eramudugolla, Mortby, Sachdev, Meslin, Kumar and Anstey2017). The diagnostic algorithm utilized a combination of neuropsychological assessments, participant and informant survey responses, and participant medical history information for existing clinical diagnoses. Table 1 shows how DSM-IV criteria were operationalized using data collected in the PATH study, to identify participants with suspected cognitive disorders, and the subgroup with suspected MCI. We note that at the time of applying this algorithm, wave 4 GDT net scores were included in the battery of measures to screen for a cognitive disorder (Eramudugolla et al., Reference Eramudugolla, Mortby, Sachdev, Meslin, Kumar and Anstey2017). While this raises a possibility of circularity, the GDT was just one of five measures used as part of the executive functions domain (17 tests across all domains). Participants with combined z scores for the measures on a domain of ≥ −2.0 and ≤ −1.0 SD below the gender- and education-standardized age group cohort mean were identified as having “objective impairment” on a domain. Other algorithmic criteria included reports of subjective cognitive changes, either by the participant on the Memory Assessment Complaint Questionnaire (Crook, Feher, & Larrabee, Reference Crook, Feher and Larrabee1992) or by an informant on the Informant Questionnaire on Cognitive Decline in the Elderly (Jorm, Reference Jorm1994). Participants identified by this diagnostic algorithm were also reviewed clinically, with full case file review by a research neurologist, along with a psychiatrist for complex cases, to confirm the diagnosis (Eramudugolla et al., Reference Eramudugolla, Mortby, Sachdev, Meslin, Kumar and Anstey2017). Among those identified as having MCI, subtyping was undertaken using standard criteria (Winblad et al., Reference Winblad, Palmer, Kivipelto, Jelic, Fratiglioni, Wahlund, Nordberg, Bäckman, Albert, Almkvist, Arai, Basun, Blennow, de Leon, DeCarli, Erkinjuntti, Giacobini, Graff, Hardy, Jack, Jorm, Ritchie, van Duijn, Visser and Petersen2004). Those with evidence of impairment on the memory domain were classified as amnestic (aMCI), others without evidence of memory impairment were classed as non-amnestic (naMCI). For both aMCI and naMCI cases, those with evidence of impairment on two or more domains (either executive, language or visuo-spatial) were classed as multi-domain cases (aMCI-multi or naMCI-multi), while those without evidence of more than a single impaired domain were classed as single-domain (aMCI-single or naMCI-single). The process for classifying MCI and MCI subtypes was undertaken and published prior to the commencement of the current study (Eramudugolla et al., Reference Eramudugolla, Mortby, Sachdev, Meslin, Kumar and Anstey2017).
Table 1. Diagnostic algorithm stages and alignment with international working group criteria for mild cognitive impairment
Note
a Cognitive measures used to define objective impairment included: Symbol Digit Modalities Test, Trail Making Test Part A, Reaction Time Test (complex attention), Digit Span Backwards, Trail Making Test Part B, Stroop Color Word Test, Zoo Map Test, Game of Dice Test (executive function), California Verbal Learning Test, Benton Visual Retention Test (Administration B) (learning and memory), Letter Fluency, Boston Naming Test-15 item, Spot the Word Test (language), Purdue Pegboard, Ideomotor Apraxia Test (perceptual motor) and Reading the Mind in the Eyes.
b Measures available to define decline over time included: California Verbal Learning Test immediate and delayed recall, Digit Span Backwards, Symbol Digit Modalities Test, Purdue Pegboard, Controlled Oral Word Association Test, Trail Making Test Part B, Simple Reaction Time, Complex Reaction Time.
IADL = Instrumental Activities of Daily Living; MAC-Q = Memory Assessment Complaint Questionnaire; IQCODE = Informant Questionnaire on Cognitive Decline in the Elderly.
Of the 1287 participants who completed the GDT, a total of 224 met the criteria for a cognitive disorder and 116 of these met IWG criteria for MCI. Among the participants with MCI, those with a self-reported history of stroke (n = 7), Parkinson’s disease (n = 3), or missing data on these comorbidity flags (n = 3) were excluded, leaving a total of 103 participants with MCI for further analysis (see Figure 1). Of these participants, 38 were categorized as aMCI-single, 31 as aMCI-multi, 26 as naMCI-single, and 7 as naMCI-multi. Due to the small number of participants in the naMCI-single and naMCI-multi groups, these were collapsed into an overall naMCI subtype, with 33 participants. One participant with MCI had missing data relating to MCI subtype, and was excluded from analyses at the subtype level. Of the 103 included participants categorized as having MCI at wave 4, 15 (14.6%) had previously been identified as meeting IWG criteria for MCI at wave 3 (see Table 2).
Figure 1. Participant flowchart showing exclusions (grey arrows and boxes) at each stage of recruitment and data processing. Gender breakdown is shown for major groups, percentages indicate the proportion of females in each wave cohort and the analytic sample.
Note: aMCI = amnestic mild cognitive impairment; naMCI = non-amnestic cognitive impairment.
Table 2. Participant demographic characteristics and cognitive measures by diagnostic grouping
Note. Test statistic and effect size results refer to independent paired samples tests or bivariate associations between cognitively unimpaired (n = 142) and mild cognitive impairment (n = 103) groups. For the cognitive measures higher scores reflect higher levels of performance, except for the Trail Making Test Parts A and B and Stroop Color-Word Interference ratio score.
W = Wilcoxon rank sum test, χ2 = chi-square test, V = Cramer’s V effect size coefficient, r 2 = effect size coefficient (r 2 < 0.3 = small effect, 0.3 < r 2 < 0.5 = medium effect), SD = standard deviation, CVLT = California Verbal Learning Test.
The comparison group was a sample of older adults without MCI, drawn from the same cohort and wave of the PATH study, and matched on age, sex, years of education, and household income. The MatchIt package (version 4.3.4) was used to perform two-to-one, nearest-neighbor, log-odds propensity score matching with replacement. The comparison group was drawn from the 1000 eligible participants with complete data on the GDT and relevant matching covariates, not meeting the criteria for MCI (or any other neurological condition or cognitive disorder) and with no self-reported history of stroke. This yielded a matched sample of 142 participants without MCI. Love Plots and Balance Plots were inspected to ensure acceptable covariate balance in the matched sample. As expected, Mini-Mental State Examination (MMSE) scores (Folstein, Folstein, & McHugh, Reference Folstein, Folstein and McHugh1975) were lower among participants with MCI than those without MCI (27.8 vs 29.2, W = 10,648, p < .001).
Game of Dice Task
The GDT is a computerized decision-making task (Brand et al., Reference Brand, Fujiwara, Borsutzky, Kalbe, Kessler and Markowitsch2005), in which participants commence with a hypothetical balance of $1000 AUD, and are asked to try to maximize their earnings across 18 trials. On each trial, participants choose a combination of one, two, three, or four numbers, aiming to match what appears to be a random dice roll (but is actually prespecified by software code and identical for all participants). The amount bet on each trial is specified by the combination of numbers chosen (one number = $1000 AUD bet, 16.7% chance to win; two numbers = $500 AUD bet, 33.3% chance to win; three numbers = $200 AUD bet, 50% chance to win; four numbers = $100 AUD bet, 66.7% chance to win) and is constant across the task. The gambles associated with each of the combinations are displayed on the screen, along with the participant’s current balance. If the dice roll matches one of the numbers chosen by the participant on that trial, they win the gambled amount, which is added to their displayed balance. If the dice roll does not match any of the numbers chosen by the participant on that trial, they lose the gambled amount, which is subtracted from their displayed balance. Participants can continue to play even if their balance falls into a negative amount. The rewards and losses in the current task were hypothetical and participants were not compensated based on their performance or results. Based on the task reward structure and probabilities of success (which are constant across the task and always explicitly available to participants) a conservative strategy is the most optimal on this task (Brand, Heinze, Labudda, & Markowitsch, Reference Brand, Heinze, Labudda and Markowitsch2008). A choice of three or four numbers is classified as “low-risk,” while a choice of one or two numbers is classified as “high-risk.” Following previous studies (Brand et al., Reference Brand, Fujiwara, Borsutzky, Kalbe, Kessler and Markowitsch2005; Pertl et al., Reference Pertl, Benke, Zamarian and Delazer2015), performance on the GDT was assessed using the following measures:
-
1) Net score: the number of high-risk options (one or two numbers) subtracted from the number of low-risk options (three or four numbers) across the 18 trials, yielding a score between plus or minus 18, with higher scores indicating more advantageous decision-making.
-
2) Single number choices: the number of trials participants chose a single number option.
-
3) Strategy changes: the number of times participants changed between high-risk and low-risk options on consecutive trials.
Cognitive measures
Previous studies have implicated the role of executive functions (Brand et al., Reference Brand, Fujiwara, Borsutzky, Kalbe, Kessler and Markowitsch2005; Schiebener & Brand, Reference Schiebener and Brand2015a), logical and numerical processing (Pertl, Zamarian, & Delazer, Reference Pertl, Zamarian and Delazer2017), and working memory (Brand & Schiebener, Reference Brand and Schiebener2013; Starcke et al., Reference Starcke, Pawlikowski, Wolf, Altstotter-Gleich and Brand2011) in behavioral tasks measuring decision-making under explicit risk, suggesting that higher-order cognitive and fluid processing abilities are of particular importance. The following measures were selected from the PATH wave 4 cognitive assessment battery, in order to i) describe levels of cognitive functioning among the participant groups, and ii) determine the cognitive abilities associated with decision-making performance for participants with MCI. The Symbol Digit Modalities Test (SDMT, Smith, Reference Smith1982) was used as a measure of attention and perceptual processing speed, scored by the number of successfully completed symbols in 90 s (maximum score = 110). The Trail Making Test (TMT, Reitan & Wolfson, Reference Reitan and Wolfson1995) was used as a measure of psychomotor speed (Part A); and cognitive flexibility and set shifting (Part B). Part A and Part B completion times were used rather than a difference score, as the direct measure of Part B completion time has been shown to be more strongly predictive of GDT performance (Schiebener et al., Reference Schiebener, Wegmann, Gathmann, Laier, Pawlikowski and Brand2014). Following the methods described by Heaton et al. (Correia et al., Reference Correia, Ahern, Rabinowitz, Farrer, Watts, Salloway, Malloy and Deoni2015; Heaton, Miller, Taylor, & Grant, Reference Heaton, Miller, Taylor and Grant2004), a prorated score was calculated for the (n = 9) participants who did not complete all 25 circles within the maximum allowable time of 300 s, for both Part A and Part B. The Stroop Color-Word Interference Test (Spreen & Strauss, Reference Spreen and Strauss1998; Stroop, Reference Stroop1935) was used as a measure of inhibitory control, by calculating an interference ratio score (color-word interference task time divided by color-dot naming task time), with higher scores reflecting an increased difficulty in inhibiting automatic responses. The first trial of the Zoo Map task (number of correct places visited minus number of errors, with broad instructions only) from the Behavioral Assessment of Dysexecutive Syndrome (Wilson, Alderman, Burgess, Emslie, & Evans, Reference Wilson, Alderman, Burgess, Emslie and Evans1996) was used as a measure of planning and goal-directed behavior (Oosterman, Wijers, & Kessels, Reference Oosterman, Wijers and Kessels2013). The Controlled Oral Word Association Test (COWAT, Benton, Hamsher, & Sivan, Reference Benton, Hamsher and Sivan1983) was used as a measure of phonemic verbal fluency, scored by the sum of the number of “a” words and “f” words spontaneously produced in separate trials of 60 s duration (only two of the three stimulus letters were used in the PATH study). The Digit Span Backwards Test from the Wechsler Memory Scale (Wechsler, Reference Wechsler1945, Reference Wechsler1997) was used as a measure of verbal working memory. Immediate and delayed recall scores from the first list of the California Verbal Learning Test (CVLT, Delis et al., Reference Delis, Massman, Kaplan, McKee, Kramer and Gettman1991) were used to assess episodic verbal learning and episodic verbal memory, respectively. The MMSE (Folstein et al., Reference Folstein, Folstein and McHugh1975) was used to describe global levels of cognitive function across participant groups.
Data analysis
The data were analyzed using R (version 12.6.3) in the R Studio environment (version 1.3.1093). Continuous measures were inspected for normality and homogeneity of variance. Graphical inspection of GDT performance measures indicated deviations from normality, which were confirmed statistically. The GDT net scores deviated from normality and were negatively skewed (Shapiro–Wilk = .93, p < .001; skewness = −.45, p < .001; kurtosis = 2.16, p < .001). Single number choices (Shapiro–Wilk = .76, p < .001; skewness = 1.74, p < .001; kurtosis = 2.98, p < .001) and the number of strategy changes (Shapiro–Wilk = .93, p < .001; skewness = .31, p < .001; kurtosis = −.97, p < .001) were positively skewed and zero-inflated. Analysis used nonparametric tests, including Wilcoxon rank-sum tests and Kruskal–Wallis tests for between group differences with Benjamini–Hochberg correction (Benjamini & Hochberg, Reference Benjamini and Hochberg1995) to control the false discovery rate for pairwise comparisons. Associations between continuous measures were assessed using Spearman’s ρ rank order correlations.
Sensitivity analyses were undertaken for the primary hypothesis (H1), to determine whether the reported findings were dependent on i) exclusion of participants with MCI who had other neurological abnormalities, or ii) inclusion of participants with existing MCI prior to wave 4 (see Supplementary File 1).
Ethics approvals
The study procedures were conducted in accordance with the Declaration of Helsinki and approved by the Australian National University Human Research Ethics Committee. All participants provided written informed consent.
Results
Participant characteristics and cognitive test score summaries by diagnostic group are shown in Table 2. The propensity score matched sample of participants without cognitive impairment did not differ significantly from participants with MCI on age (75.1 vs 75.0 years, W = 7,561, p = .646), gender (45.8% vs 47.5% female χ2 (1, N = 245) = .02, p = .882), level of education (13.3 vs 13.1 years, W = 7,393, p = .883), or likelihood of reporting equal to or greater than $575 AUD per week in household income (65.5% vs 60.2%, χ2 (2, N = 245) = .744, p = .689) than participants with MCI. As expected, participants with MCI performed more poorly than participants without cognitive impairment on the cognitive measures (p’s < .001), with the exception of the Stroop Color-Word Interference ratio score (2.66 vs 2.44, W = 6,830, p = .43).
Decision-making performance scores
Decision-making performance scores by diagnostic group are shown in Table 3. GDT net scores were not significantly lower among participants with MCI compared to participants without cognitive impairment (2.56 vs 4.27, W = 8,086, p = .157). The frequency of single number choices was significantly higher among participants with MCI compared to those without cognitive impairment (3.85 vs 3.00, W = 6060, p = .02, r 2 = .15). Participants with MCI also made more strategy changes than those without cognitive impairment (5.50 vs 4.65, W = 6,224, p = .046, r 2 = .13).
Table 3. Game of Dice Task (GDT) performance measures by diagnostic grouping
Note. Test statistic results refer to independent samples tests between cognitively unimpaired (n = 142) and mild cognitive impairment (n = 103) groups, along with pairwise tests for each MCI subtype (against the cognitively unimpaired group). For the GDT net score measure (−18 minimum to 18 maximum) higher scores reflect higher levels of performance. For the frequency of single number choices (0 minimum to 18 maximum) and strategy changes (0 minimum to 17 maximum) lower scores reflect higher levels of performance.
W = Wilcoxon rank sum test, r 2 = effect size coefficient (r 2 < 0.3 = small effect), SD = standard deviation; MCI = mild cognitive impairment.
Sensitivity analyses investigated whether the reported study findings are robust to modifications in the exclusion criteria. Separate analyses tested H1 by i) including all participants categorized as having MCI at wave 4 (n = 116) or ii) further limiting the MCI sample to those with incident MCI at wave 4 (n = 88), by excluding those (n = 15) who had previously met MCI criteria at wave 3. In both cases the sensitivity analyses found the same pattern of results for the (H1) primary study outcomes for GDT net scores and the frequency of single number scores, but not for strategy changes (see Supplementary File 1).
Analysis at the subtype level compared the aMCI-single, aMCI-multi, naMCI, and cognitively unimpaired groups on each of the three GDT performance measures. For GDT net scores (see Figure 2) there was no significant effect of MCI subtype (χ2 (3, N = 244) = 4.72, p = .19). For the number of single dice choices there was a significant effect of MCI subtype (χ2 (3, N = 244) = 9.42, p = .02). Pairwise comparisons indicated that the only significant effect was between the cognitively unimpaired and aMCI-multi groups (3.00 vs 4.35, p = .02), with all other comparisons p > .29. For the number of strategy changes there was no significant effect of MCI subtype (χ2 (3, N = 244) = 6.24, p = .10). However due to the main effect of participants with MCI compared to those without cognitive impairment on this measure, pairwise comparisons were investigated, to better understand the pattern of results across MCI subtypes (see Table 3).
Figure 2. Game of Dice Task (GDT) net scores by mild cognitive impairment (MCI) subtypes, with jittered dots indicating individual participant scores within each MCI subtype. Boxplots indicate median and inter-quartile range, with notches indicating confidence intervals around the median for each group.
Spearman rank order correlation analyses were conducted to assess associations between decision-making performance and cognitive measures across the entire study sample (see Table 4). GDT net scores were significantly associated with scores on the SDMT (ρ = .13, p = .046), TMT Part B (ρ = −.21, p = .001), Zoomap Part 1 raw scores (ρ = .17, p = .009), Stroop interference scores (ρ = −.14, p = .02), and CVLT immediate recall (ρ = .13, p = .04). The number of single dice choices were significantly associated with scores on the SDMT (ρ = −.18, p = .004), TMT Part A (ρ = .15, p = .02), TMT Part B (ρ = .26, p < .001), Zoomap Part 1 raw scores (ρ = −.19, p = .003), Stroop interference scores (ρ = .14, p = .03), Digit Span Backwards (ρ = −.15, p = .02), CVLT immediate recall (ρ = −.18, p = .004), and CVLT delayed recall (ρ = −.15, p = .02). The number of strategy changes was significantly associated with scores on the TMT Part B (ρ = .22, p < .001), COWAT (ρ = −.16, p = .01), CVLT immediate recall (ρ = −.15, p = .02), and CVLT delayed recall (ρ = −.20, p = .002).
Table 4. Correlations between Game of Dice Task (GDT) performance measures and cognitive measures among all study participants
Note. Correlations are expressed using Spearman’s rho coefficient. For the GDT net score measure (−18 minimum to 18 maximum) higher scores reflect higher levels of performance. For the frequency of single number choices (0 minimum to 18 maximum) and strategy changes (0 minimum to 17 maximum) lower scores reflect higher levels of performance. For the cognitive measures higher scores reflect higher levels of performance, except for the Trail Making Test Parts A and B and Stroop Color-Word Interference ratio score.
Discussion
The current study demonstrated a pattern of less advantageous decision-making performance on the GDT among participants with MCI relative to a matched sample of older adults without cognitive impairment. While impaired decision-making was observed across two of the three outcome measures (frequency of single number choices and frequency of strategy changes), the clearest impairments were seen on the frequency of single number choices measure. In a follow-up analysis of single number choices, participants classified in the aMCI-multi group were the only subtype to show impairments in decision-making performance compared with cognitively unimpaired participants.
The current study findings are consistent with Sun et al. (Reference Sun, Xie, Wang, Zhang, Tian, Wang, Yu and Wang2020) who showed that people with MCI chose single numbers more frequently, but did not differ from older adults without cognitive impairment on the use of negative feedback or overall GDT net score. Our observation of statistically significant differences in the frequency of single number choices and strategy changes on the GDT suggests a tendency toward more high-risk response patterns, along with a greater propensity for changing response patterns between trials, among older adults with MCI. Within existing theoretical models of decision-making under explicit risk conditions, it is proposed that “deliberative” and “impulsive” systems are activated concurrently, with a range of individual, contextual and decision-related factors influencing which is dominant for a specific decision (Schiebener & Brand, Reference Schiebener and Brand2015a). The presence of cognitive impairment may reduce the fluid processing resources available to activate (limited capacity) deliberative decision-making systems, resulting in greater dominance of the impulsive decision-making system. This could lead to an increased tendency to choose on the basis of somatic responses activated by prospects of large rewards (e.g., high-risk responses), or to change response strategies repeatedly in response to feedback from immediately prior trials.
Investigations by MCI subtype indicated that the only group with reliable differences in decision-making performance relative to those without cognitive impairment was the aMCI-multi group. This group showed more frequent single number choices than the cognitively unimpaired group. Executive function abilities have been consistently shown to predict performance on the GDT (Brand & Schiebener, Reference Brand and Schiebener2013; Schiebener, Zamarian, Delazer, & Brand, Reference Schiebener, Zamarian, Delazer and Brand2011) and are also found to be more impaired among those with multi-domain, as opposed to single-domain MCI (Brandt et al., Reference Brandt, Aretouli, Neijstrom, Samek, Manning, Albert and Bandeen-Roche2009; Klekociuk & Summers, Reference Klekociuk and Summers2014; Pereiro et al., Reference Pereiro, Juncos-Rabadan and Facal2014). In the current study, the aMCI-multi group also showed relatively poorer performance on a number of the included cognitive assessments, including measures of executive function abilities (TMT Part B, Zoomap test, COWAT) compared to participants without cognitive impairment or other MCI subtypes. This suggests that the aMCI-multi group included a higher proportion of participants with attentional or executive function impairments in addition to memory impairments, whose poorer executive function abilities may have contributed to their poorer performance on the GDT (Brand & Schiebener, Reference Brand and Schiebener2013; Schiebener et al., Reference Schiebener, Wegmann, Gathmann, Laier, Pawlikowski and Brand2014). Changes in frontal and striatal brain regions associated with the “dysexecutive” form of MCI (Grambaite et al., Reference Grambaite, Selnes, Reinvang, Aarsland, Hessen, Gjerstad and Fladby2011) may result in difficulties in inhibiting impulsive responses, and difficulties activating deliberative systems to integrate task information, process feedback, and select optimal responses to maximize probabilities of success (Schiebener & Brand, Reference Schiebener and Brand2015a). However, while the aMCI-multi group appears to include participants with attentional and/or executive impairments, we also note that other subtypes, in particular participants in the naMCI subtype, also showed evidence of executive function impairments. The absence of reliable impairment on the GDT in the naMCI group suggests that impairments in memory and/or learning may also play a role in performance on the GDT (Sinclair et al., Reference Sinclair, Eramudugolla, Brady, Cherbuin and Anstey2021; Starcke et al., Reference Starcke, Pawlikowski, Wolf, Altstotter-Gleich and Brand2011). Hence the observed findings may also reflect broader impacts across multiple cognitive domains. Future work might investigate whether impairment on the GDT emerges in the context of combined amnestic and dysexecutive patterns of impairment.
Across this sample of older adult participants, and for all three of the GDT performance scores (net scores, single number choices and strategy changes) the cognitive assessment showing the strongest association with GDT performance was the TMT Part B completion time. This supports previous observations of associations between GDT performance and measures of executive functions, in particular the TMT Part B, as a measure that is correlated with GDT scores (Schiebener et al., Reference Schiebener, Wegmann, Gathmann, Laier, Pawlikowski and Brand2014). However, additional correlations were noted between GDT performance and measures of attention, planning, inhibitory control, working memory, and episodic learning. These are consistent with previous studies, including in populations without cognitive impairment (Brand, Labudda, & Markowitsch, Reference Brand, Labudda and Markowitsch2006; Sinclair et al., Reference Sinclair, Eramudugolla, Brady, Cherbuin and Anstey2021; Starcke et al., Reference Starcke, Pawlikowski, Wolf, Altstotter-Gleich and Brand2011). The GDT is a multicomponent decision-making task, and optimal response has been proposed to require directing attention toward the risk-reward contingencies for the task, undertaking numerical processing to identify an optimal strategy, activating responses in line with this strategy, and integrating feedback from prior trials to refine and update the strategy (Schiebener & Brand, Reference Schiebener and Brand2015a). The overall low magnitude of the correlations between GDT scores and all of the cognitive assessments suggest that the GDT may reflect a more complex range of cognitive abilities, albeit with an emphasis on executive functions (Gathmann, Brand, & Schiebener, Reference Gathmann, Brand and Schiebener2017; Schiebener et al., Reference Schiebener, Zamarian, Delazer and Brand2011).
The inconsistent effects from previous studies, along with the small effect sizes observed in the current study, suggest that MCI is associated with subtle impairments in decision-making when tasks are predictable and straightforward. Clearer impairments are observed in the context of more advanced neuropathology (e.g., dementia diagnosis; Mueller et al., Reference Mueller, Arias, Vazquez, Schiebener, Brand and Wegmann2019; Sun et al., Reference Sun, Xie, Wang, Zhang, Tian, Wang, Yu and Wang2020) or more complex tasks, such as the GDT-D (Pertl et al., Reference Pertl, Benke, Zamarian and Delazer2015) or PAG-R (Zamarian et al., Reference Zamarian, Weiss and Delazer2011). Given our findings that differences between participants with and without MCI on the GDT appear to be primarily driven by the aMCI-multi subtype (at least on single number choices), it may be that the inconsistent findings in previous studies were due to differences in the (unreported) proportions of participants with different MCI subtypes.
Strengths and limitations
The current study has some limitations, which should be considered in interpreting the results. Sample representativeness is impacted by prior sample attrition (e.g., loss to follow-up or death). Also, 357 potentially eligible participants (28 of whom met IWG criteria for MCI) were excluded as they did not complete the GDT. The propensity score matching achieved acceptable matching on age, sex, years of education, and level of household income between participants with MCI and those without MCI, however other unobserved variables may also contribute to the observed between group effects. On the other hand, the population-based, prospective recruitment methods employed in the current study enables more representative sampling and control for a broader range of covariates across multiple observation points, and is a strength. While the inclusion of wave 4 GDT net scores as part of the screening for cognitive disorders raises a possibility of circularity, this measure was just one of five assessments used in the executive functions domain, and those identified by the algorithm were also reviewed clinically to confirm the diagnostic classification (Eramudugolla et al., Reference Eramudugolla, Mortby, Sachdev, Meslin, Kumar and Anstey2017). Finally, the GDT is typically thought to measure individual decision-making under explicit risk conditions, within a financial (gambling) domain, and without the prospect of real monetary gains or losses that might be associated with real-world financial decision-making. Hence, this task may lack personal relevance to participants, thus limiting its generalizability to real-world contexts. However, previous work has suggested that performance impairments on behavioral decision-making tasks are associated with real-world decision-making impairments (Pertl, Benke, et al., Reference Pertl, Benke, Zamarian and Delazer2017), such as susceptibility to fraudulent advertising (Denburg et al., Reference Denburg, Cole, Hernandez, Yamada, Tranel, Bechara and Wallace2007).
Implications
The current findings identified impaired decision-making performance among participants with MCI, and suggest that the GDT can discriminate between participants with and without MCI. This may have implications for clinicians who are designing or selecting measures aimed at detecting early signs of cognitive impairment or discriminating between MCI subtypes, particularly in community-based samples. The current findings also demonstrate the importance of MCI subtypes in decision-making performance, suggesting that MCI subtype categories should be reported where possible.
Conclusion
The current study demonstrates impaired decision-making performance among participants with MCI, relative to older adults without cognitive impairment. To our knowledge, this study is the first to enable comparative analysis by MCI subtype, showing subtle decision-making impairments among participants with multi-domain amnestic MCI. Further research is required to understand the specific pattern of pathology associated with impaired decision-making performance on the GDT, and its relevance for providing supportive interventions to assist people with MCI in real-world decision-making contexts.
Supplementary material
To view supplementary material for this article, please visit https://doi.org/10.1017/S1355617722000765
Data availability statement
The data from the Personality and Total Health (PATH) Through Life study are not publicly available, however they can be accessed via an application to the Research Subcommittee http://pathstudy.org.au. Full analysis script will be made available upon request.
Acknowledgements
We thank the PATH participants and research staff and acknowledge funding from NHMRC Grants (No. 973302, 179839, 418039, 1002160) and Chief Investigators Tony Jorm, Helen Christensen, Bryan Rodgers, Keith Dear, Simon Easteal, Andrew Mackinnon, and Peter Butterworth.
Funding statement
This study was supported by the Australian Research Council Centre of Excellence in Population Ageing Research (project number CE170100005). Kaarin Anstey is supported by an ARC Laureate Fellowship (FL190100011).
Conflicts of interest
None.
