Introduction
Depression is characterized by impaired executive function (Snyder, Reference Snyder2013), difficulty sustaining attention (Biringer et al., Reference Biringer, Mykletun, Sundet, Kroken, Stordal and Lund2007), and memory problems—including trouble recalling details from encoding (MacQueen et al., Reference MacQueen, Campbell, McEwen, Macdonald, Amano, Joffe, Nahmias and Young2003) and loss of the positive memory bias typically seen in healthy adults (Burt et al., Reference Burt, Zembar and Niederehe1995; Dillon et al., Reference Dillon, Dobbins and Pizzagalli2013). Indeed, multiple meta-analyses document broad cognitive deficits in depressed adults (Burt et al., Reference Burt, Zembar and Niederehe1995; Zakzanis et al., Reference Zakzanis, Leach and Kaplan1998; Snyder, Reference Snyder2013), with problems related to executive function and attention persisting in remission (Douglas and Porter, Reference Douglas and Porter2009; Rock et al., Reference Rock, Roiser, Riedel and Blackwell2014). Each cognitive problem may involve separate pathophysiologies, but some processes may be common to most of the tests on which depressed adults show impairment. If so, then a negative effect of depression on those processes would help explain the broad range of cognitive deficits observed. This work aimed to study the impact of depression on one such process—namely, decision-making.
Decision-making is an appealing candidate because most tests of attention, executive function, and memory involve choosing among alternatives (Shadlen and Kiani, Reference Shadlen and Kiani2013). Thus, a negative effect of depression on decision-making would lead to the broad impairments that have been seen. Another reason to focus on decision-making is that computational models can parse it into component processes, providing an opportunity to pinpoint the specific mechanisms affected by depression. In particular, the drift-diffusion model (DDM) has been used for over 40 years to decompose decision-making during recognition memory tests (Ratcliff, Reference Ratcliff1978), lexical decision tasks (Ratcliff et al., Reference Ratcliff and Smith2004), purchasing games (Krajbich et al., Reference Krajbich, Lu, Camerer and Rangel2012), and many other paradigms (for review, see Ratcliff and McKoon, Reference Ratcliff and McKoon2008). The DDM and similar models have also been used to study brain systems that support decision-making in humans (Frank et al., Reference Frank, Gagne, Nyhus, Masters, Wiecki, Cavanagh and Badre2015) and non-human primates (Gold and Shadlen, Reference Gold and Shadlen2007). Finally, prior work has emphasized that by applying the DDM in clinical contexts, it may be possible to uncover deficits in patients that cannot be detected with traditional analysis of response times (RT) and accuracy (White et al., Reference White, Ratcliff, Vasey and McKoon2010). For these reasons, we elected to use a Bayesian variant of the DDM called the Hierarchical Drift Diffusion Model (HDDM; Wiecki et al., Reference Wiecki, Sofer and Frank2013) to study decision-making in depression.
The HDDM is shown in Fig. 1a. Briefly, it conceptualizes decision-making as a process of evidence accumulation. When a participant views two options and must choose between them, the HDDM assumes that the participant sets up boundaries specifying the amount of evidence needed to select one alternative over the other. Next, the participant draws a sample of the evidence for each option, computes the difference between the samples, and then increments a decision variable towards whichever boundary is favored by the difference score. This process is performed repeatedly until the evidence crosses one of the boundaries, at which point the corresponding response is rendered.
Fig. 1. (a) The Hierarchical Drift Diffusion Model (HDDM). The HDDM represents decisions as a process of evidence accumulation towards response boundaries separated by a decision threshold. The speed of evidence accumulation is referred to as the drift rate. The drift process moves left to right over time, from a starting point that can be midway between the boundaries or shifted towards one to an extent captured by the starting bias. The time needed for perception and response execution is captured by the non-decision time. In applying the model to PRT data, we mapped the upper and lower boundaries to ‘rich’ and ‘lean’, responses, respectively. The HDDM is a Bayesian extension of the original DDM (Ratcliff and McKoon, Reference Ratcliff and McKoon2008) that provides enhanced parameter estimation for studies with between-group designs. (b) The probabilistic reward task (PRT). On each trial, participants must indicate whether a short (11.5 mm) or long (13.0 mm) mouth was shown. Correct identifications of one length (the ‘rich’ stimulus) are rewarded three times more frequently than correct identifications of the other length (the ‘lean’ stimulus).
For example, several non-human primate studies have used a task in which monkeys must decide if a swirling dot pattern is moving mainly to the left or right (Gold and Shadlen, Reference Gold and Shadlen2007). Monkeys solve this task by repeatedly drawing samples of evidence from motion-sensitive neurons that encode movement to the left v. right, respectively, until the evidence crosses a threshold (boundary) in favor of one direction; at this point, the monkeys respond. The speed of evidence accumulation is called the drift rate. Drift rate sets a limit on the speed and accuracy of decision-making, and it can be estimated by applying the HDDM to RT and accuracy data. Importantly, the HDDM also includes a non-decision time parameter that captures the time needed to perceive stimuli and execute a response once a decision has been made. Thus, the model can distinguish between factors that affect the speed of evidence accumulation, quantified by drift rate, v. those that affect sensorimotor aspects of behavior, quantified by non-decision time.
As a first step towards testing the hypothesis that depression affects decision-making, we applied the HDDM to behavioral data collected from 281 unmedicated adults with Major Depressive Disorder (MDD) and 61 healthy controls as they performed a probabilistic reward task (PRT; Pizzagalli et al., Reference Pizzagalli, Jahn and O'Shea2005). As described below, on each trial the participants viewed a schematic face onto which a mouth was briefly flashed. The task was to indicate whether the mouth was long or short. Correct responses elicited monetary rewards, and correct identifications of one mouth length (the ‘rich’ stimulus) were rewarded three times more often than correct identifications of the other length (the ‘lean’ stimulus). This asymmetric reinforcement rate was used to induce a response bias. Many studies have found that depressed adults develop a weaker response bias than controls (Pizzagalli et al., Reference Pizzagalli, Jahn and O'Shea2005, Reference Pizzagalli, Iosifescu, Hallett, Ratner and Fava2008; Vrieze et al., Reference Vrieze, Pizzagalli, Demyttenaere, Hompes, Sienaert, de Boer, Schmidt and Claes2013; Liu et al., Reference Liu, Roiser, Wang, Zhu, Huang, Neumann, Shum, Cheung and Chan2016), and we report on this below. However, our main goal was to conduct a fine-grained analysis of the impact of MDD on decision-making.
We expected that MDD would be associated with slow evidence accumulation. To test this account, we first compared RT distributions across the groups. Prior research (Ratcliff and McKoon, Reference Ratcliff and McKoon2008) indicates that differences in drift rate have a small effect on fast RTs but a large effect on slow RTs, such that the RT distribution should be more positively skewed in the group with the slower drift rate (because the impact of the drift rate difference is magnified as RT increases). Therefore, we predicted that the RT distribution would be more positively skewed in the MDD group. Next, we fit the HDDM to the data. Our second prediction was that the model would reveal slower drift rates in the MDD group. We also examined the non-decision time parameter to determine whether MDD affected sensorimotor processes.
We also report on decision threshold (Fig. 1a). Wider thresholds correspond to greater distance between the boundaries, which means more evidence must accumulate for the participant to respond. Manipulations that increase response caution—such as prioritizing accuracy over speed (Ratcliff and Rouder, Reference Ratcliff and Rouder1998)—result in slower RTs and wider thresholds. We did not have an a priori hypothesis about this parameter, but because depressed adults often respond more slowly than controls in this task (Pizzagalli et al., Reference Pizzagalli, Iosifescu, Hallett, Ratner and Fava2008), wider thresholds in MDD might be expected.
The drift process can also begin from a position shifted towards either boundary to an extent captured by the starting point bias. White and Poldrack (Reference White and Poldrack2014) found that a response bias in behavior typically maps onto a starting point bias in the DDM: the drift process begins closer to one boundary such that even a little evidence in favor of that option will elicit a response. Given the asymmetric reinforcement rate, we expected the starting point to be biased towards the more frequently rewarded ‘rich’ boundary. If so, then fast responses should predominantly be rich responses because if the starting point is close to the rich boundary, then the accumulator will need to travel just a short distance to reach it. To our knowledge, no PRT study has examined whether the response bias is stronger for fast v. slow RTs, as these considerations predict. Finally, we examined the split-half reliability of our measures to assess internal consistency (Levinson et al., Reference Levinson, Speed, Infantolino and Hajcak2017; Luking et al., Reference Luking, Nelson, Infantolino, Sauder and Hajcak2017).
Method
Participants
Study 1
Data were collected from 296 adults with MDD and 40 healthy participants in the multi-site ‘Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care’ (EMBARC) study. EMBARC was a randomized, placebo-controlled trial of sertraline. The goal was to identify predictors of treatment response, thus a variety of measures—including the PRT—was administered before randomization to drug or placebo. Only pre-treatment data were included in this analysis. Depressed participants were outpatients between 18 and 65 who met DSM-IV criteria for MDD, as assessed by the Structured Clinical Interview for DSM-IV Disorders (SCID; First et al., Reference First, Spitzer, Gibbon and Williams2002). The SCID was administered by graduate-level clinicians. To limit heterogeneity, depressed participants had to report early onset (before age 30), chronicity (current episode >2 years duration), or recurrence (two or more episodes). Data were collected at Columbia University Medical Center, Massachusetts General Hospital (MGH)/McLean Hospital, University of Texas Southwestern Medical Center, and the University of Michigan. Participants consented to a protocol approved by local Institutional Review Boards. See Trivedi et al. (Reference Trivedi, McGrath, Fava, Parsey, Kurian, Phillips, Oquendo, Bruder, Pizzagalli, Toups and Cooper2016) for details.
Study 2
To determine if the results from Study 1 could be replicated, we reanalyzed data from 23 unmedicated adults with MDD and 25 healthy controls, previously published in Pizzagalli et al. (Reference Pizzagalli, Iosifescu, Hallett, Ratner and Fava2008). The depressed participants were recruited from treatment studies at MGH, whereas controls came from the community. Depressed participants met DSM-IV criteria for current MDD based on the SCID, which was administered by trained psychiatrists. All participants consented to a protocol approved by the Harvard University and Partners Healthcare IRBs. See the original publication for details.
Self-report
Participants provided demographic information and completed the Mood and Anxiety Symptoms Questionnaire (MASQ: Watson et al., Reference Watson, Weber, Assenheimer, Clark, Strauss and McCormick1995); Studies 1 and 2 used 30 and 62 item versions, respectively. Both include scales for anhedonic depression (MASQ-AD) and anxious arousal (MASQ-AA). The shorter version includes a ‘general distress’ (MASQ-GD) scale, whereas the longer one has scales for general distress due to depression (MASQ-GDD) and anxiety (MASQ-GDA). All participants except the Study 2 controls were administered the 17-item Hamilton Rating Scale for Depression (HRSD; Hamilton, Reference Hamilton1960), which was scored by the diagnostic interviewer. Both studies included additional questionnaires not considered here.
Task
E-Prime version 1.1 (Psychology Software Tools, Sharpsburg PA) was used to present the PRT, which is shown in Fig. 1b. On each PRT trial, participants view a schematic face for 500 ms. A mouth is then shown for 100 ms, and the task is to indicate by button press whether the mouth was long (13.0 mm) or short (11.5 mm). Correct identifications of one length (the ‘rich’ stimulus) are rewarded three times more often than correct identifications of the other length (the ‘lean’ stimulus). In Study 1, participants completed two 100-trial blocks in which they pressed ‘c’ or ‘m’ on a keyboard to report seeing the short or long mouth, which served as the rich and lean stimuli, respectively. The PRT was programmed to deliver 40 20-cent rewards per block—30 rich rewards v. 10 lean rewards—although there was some variability due to differences in behavior. No feedback was presented on non-reward trials. Study 2 was very similar, except that participants completed three 100-trial blocks, rewards were worth 5 cents, and the rich and lean keys were counterbalanced across subjects. See Pizzagalli et al. (Reference Pizzagalli, Jahn and O'Shea2005) for additional task details.
Analyses
Quality Control (QC)
Trials were excluded for extreme RTs (<150 ms, >2500 ms), or if the remaining (log transformed) RT exceeded the participant's mean ± 3s.d. Participants' datasets were excluded if, in any block, there were more than 20 RT outliers, fewer than 24 rich or 7 lean rewards, a rich-to-lean reward ratio lower than 2.5, or lower than 40% correct accuracy. In Study 1, 258 depressed adults and 36 controls passed the QC criteria. Study 2 data are from participants who passed these QC checks.
Quantile-probability plots
We used quantile-probability plots to determine whether the RT distribution was more positively skewed in the MDD group. To generate these plots, we binned responses by RT quantile. The quantiles used, from fastest to slowest, were 0.1, 0.3, 0.5, 0.7, 0.9, and 0.995. Each quantile served as the RT ceiling for its bin, with the previous quantile as the floor. We used the 0.005 quantile as the floor for the 0.1 quantile, so that the sizes of the fastest RT bin (0.100–0.005 = 0.095) and slowest RT bin (0.995–0.900 = 0.095) would be identical. For each bin, we plotted the percent correct and incorrect on the right and left sides of the x-axis, respectively, with the mean correct/incorrect RT plotted on the y-axis.
HDDM
Computational modeling was performed in Jupyter Notebooks (Kluyver et al., Reference Kluyver, Ragan-Kelley, Pérez, Granger, Bussonnier, Frederic, Kelley, Hamrick, Grout, Corlay, Ivanov, Loizides and Schmidt2016) and fit to trial-level RT and response data following published recommendations (Wiecki et al., Reference Wiecki, Sofer and Frank2013). Briefly, the HDDM is initialized with priors that reflect established findings in the literature, and then the Markov Chain Monte Carlo method fits the model to the data by estimating the joint posterior distribution for all parameters. All HDDM parameters were allowed to vary by group. We drew 10 000 samples from the posterior distribution, discarding the first 1000 ‘burn-in’ samples (Kruschke, Reference Kruschke2014). Trace and autocorrelation plots were inspected to assess convergence. To evaluate model quality, the estimated parameters were used to generate simulated data (posterior predictive checks). Summary statistics from the actual data fell well within 95% intervals of the simulated data, indicating a good fit. We examined the between-group overlap of the posterior distributions for all parameters, defining significance as less than 5% overlap. Because these are comparisons of Bayesian posterior distributions, we report the HDDM outcomes as q-values rather than p-values.
Signal-detection analyses
We computed the signal detection metrics response bias and discriminability using published formulas (Pizzagalli et al., Reference Pizzagalli, Jahn and O'Shea2005), analyzing them in Group x Block ANOVAs implemented in the R (R Core Team, 2018) package afex (Singmann et al., Reference Singmann, Bolker, Westfall, Aust, Højsgaard, Fox, Lawrence, Mertens and Love2016). The Greenhouse-Geisser correction was applied to all ANOVAs. To determine how signal detection measures related to HDDM parameters, we computed linear mixed models with response bias or discriminability as the dependent variable, HDDM parameters and Group as predictors, and Subject as a random effect (Bates et al., Reference Bates, Mächler, Bolker and Walker2014). Site was included as a covariate in all Study 1 analyses.
Individual differences
We computed Pearson correlations between MASQ scores and HDDM parameters in depressed adults, to determine if variability in symptoms was related to decision-making. We also used regressions to determine if there was a group difference in cumulative rewards earned, and to see if such a group difference could be explained by the HDDM parameters. Because the samples were larger in Study 1 and the methods varied somewhat across the studies, we restricted these analyses to Study 1.
Psychometrics
To assess split-half reliability, we computed response bias and discriminability separately for odd and even trials for each participant and ran the HDDM on the odd and even trials to generate two sets of model parameters per person. We then computed Pearson correlations between the odd and even results and applied the Spearman-Brown prophesy formula (2r/(1 + r)) to quantify internal consistency (Luking et al., Reference Luking, Nelson, Infantolino, Sauder and Hajcak2017).
Results
Demographics and self-report
Groups did not differ on age, education or gender (Table 1). As expected, depressed adults reported more anhedonia, higher anxiety, and greater general distress on the MASQ than did controls. Based on HRSD cut-offs (Zimmerman et al., Reference Zimmerman, Martinez, Young, Chelminski and Dalrymple2013), adults with MDD were moderately depressed.
Table 1. Demographics and self-report data
Note. p-values reflect between-group t tests except for gender and % of participants with college education in Study 2, which were evaluated by chi-square. All tests were two-sided. Effect size: Cramer's V for Gender, otherwise Cohen's d. HRSD data were not obtained from Study 2 controls. As detailed in the text, different MASQ versions were used in each study (Study 1: 30 items; Study 2: 62 items)
f, female; m, male; HRSD, Hamilton Rating Scale for Depression (17 items); MASQ, Mood and Anxiety Symptom Questionnaire (AD, anhedonic depression; AA, anxious arousal; GD, general distress; GDD, general distress, depression; GDA, general distress, anxiety).
Study 1
Quantile-probability plots
Figure 2 shows the Study 1 quantile-probability plot. This depicts the mean percentage of correct (plotted to the right) and incorrect (plotted to left) responses to the rich (circles) and lean (crosses) stimuli as a function of RT quartile (plotted on the y-axis) for the two groups. This figure supports two main conclusions. First, as expected the RT distribution was more skewed in depressed adults. Specifically, although the mean ± s.d. RT in the 0.1 quantile bin, averaged over stimulus type and response accuracy, was 5 ms slower in depressed (242.73 ± 17.22 ms) v. healthy (237.37 ± 10.87 ms) participants, t(239) = 2.23, p = 0.031, Cohen's d = 0.38, by the 0.995 quantile bin this group difference had grown to 158 ms (MDD: 841.00 ± 76.33 ms; controls: 683.07 ± 48.54 ms; t(240) = 15.34, p < 0.001, d = 2.53), a 32-fold increase (note that degrees of freedom vary in the quantile-probability analyses as not all participants contributed responses to the more extreme bins). Second, Fig. 2 reveals that response bias was constrained to the fastest 30% of responses. Notice that the circles are farther to the right than the crosses for the 0.100 and 0.300 quantiles, indicating higher accuracy for responses to the rich v. lean stimulus. This reflects response bias: when replying quickly, participants pressed ‘rich’ more than ‘lean’ and so achieved a higher proportion correct for the rich v. lean stimulus. This effect is absent for the remaining quantiles. Thus, the PRT induced a response bias but this was constrained to fast RTs.
Fig. 2. Quantile-probability plots: Study 1. Percent correct (right) and incorrect (left) for rich (circles) and lean (crosses) stimuli as a function of RT quantiles, for adults with MDD (left column) and healthy controls (right column). The six quantiles are marked on the controls' data; they are shifted upwards on the y-axis for the MDD group, with the magnitude of the shift increasing with longer response latency. The effect of stimulus type on accuracy (rich > lean) is restricted to the 0.1 and 0.3 quantiles, indicating that response bias is carried by fast RTs.
To confirm these impressions, Group x Stimulus ANOVAs were run on response accuracy for fast (RT ⩽ 0.3 quantile) and slow (RT > 0.3 quantile) responses. As expected from Fig. 2, for fast RTs the Stimulus effect was strong, F(1, 267) = 89.44, p < 0.001, η p2 = 0.25, corresponding to higher mean ± s.e. accuracy for rich (84.15 ± 1.53%) v. lean (59.95 ± 1.53%) stimuli. An effect of Group was also found, F(1, 267) = 13.62, p < 0.001, η p2 = 0.05, reflecting lower accuracy in depressed (67.11 ± 1.56%) v. healthy (76.98 ± 1.56%) participants. By contrast, for slow RTs no significant results emerged, Fs < 2.28, ps > 0.12.
HDDM
Figure 3 shows the HDDM results. The posterior distributions revealed slower drift rates (q < 0.001) and wider decision thresholds (q = 0.004) in depressed adults, but no group differences in starting bias or non-decision time, qs > 0.17.
Fig. 3. HDDM results: Study 1. Plots of posterior probabilities for HDDM parameters. Relative to results in the controls (dashed lines), drift rate was reduced and threshold width was increased in the MDD group (solid lines). There were no group differences in starting point bias or non-decision time. *q < 0.005.
Signal-detection analyses
The response bias and discriminability data are in Figure S1. Group x Block x Site ANOVAs yielded no significant results, although there was a trend (p = 0.06) for worse discriminability in depressed adults (MDD: 0.59 ± 0.24; controls: 0.67 ± 0.24). The lack of a group difference in response bias was unexpected; exploratory t tests against zero revealed reliable biases in both groups (MDD: 0.09 ± 0.16; controls: 0.11 ± 0.13; ts > 5, ps < 0.001).
Prediction of signal-detection metrics by HDDM parameters
Linear mixed models were used to predict response bias and discriminability with Group, Site, and HDDM parameters. Starting point bias was the strongest predictor of response bias (Z = 12.71, p < 0.001), and drift rate was a remarkably strong predictor of discriminability (Z = 45.56, p < 0.001); these relationships are shown in Figure S2. Response bias was also predicted by drift rate (Z = −5.17, p < 0.001), and discriminability was also predicted by decision threshold (Z = 14.81, p < 0.001) and non-decision time (Z = 2.99, p = 0.003). Group did not predict either variable (ps > 0.23).
Individual differences
In the MDD group, weak relationships emerged between MASQ-AA scores and drift rate, r = −0.15, p = 0.02, and between MASQ-GD scores and decision threshold, r = −0.17, p = 0.01. Neither of these relationships remains significant, however, when a Bonferroni-corrected alpha of 0.004 (0.5/12 comparisons) is applied.
Finally, Group predicted cumulative reward, β = −0.13, p = 0.029. Depressed adults received fewer rewards than controls did, although the difference was small (controls: 79.08 ± 1.27; MDD: 78.25 ± 2.22) as the PRT is programmed to equate reward delivery across participants. Adding the HDDM parameters improved the model, ΔR 2 = 0.34, F(4, 288) = 38.41, p < 0.001, and with these parameters included the effect of Group was no longer significant (β = −0.05, p = 0.30). Instead, drift rate (β = 0.39), non-decision time (β = 0.29), and starting point bias (β = 0.18)—but not decision threshold (β = 0.09)—emerged as predictors of cumulative reward, ps < 0.001. The fact that drift rate strongly predicted cumulative reward is sensible because individuals with high drift rates have high discriminability (Figure S2B), which allows them to respond accurately and thus efficiently harvest rewards on rich and lean trials.
Study 2
The next goal was to determine if the quantile-probability plot and HDDM findings from Study 1 would replicate in a dataset characterized by a group difference in response bias. To this end, we reanalyzed data published by Pizzagalli et al. (Reference Pizzagalli, Iosifescu, Hallett, Ratner and Fava2008), referred to as Study 2.
Quantile-probability plots
Figure 4 shows the quantile-probability plot. As in Study 1, RTs were drastically more skewed in the MDD group. Specifically, while the fastest RTs (0.1 quantile) were 34 ms slower in depressed (298.64 ± 21.09 ms) v. healthy (264.31 ± 17.62 ms) adults, t(34) = 5.22, p < 0.001, d = 1.77, by the 0.995 quantile bin this group difference had grown to 339 ms (|MDD: 1365.85 ± 91.57 ms; controls: 1026.81 ± 96.92 ms), t(36) = 11.08, p < 0.001, d = 3.60, a nearly 10-fold increase.
Fig. 4. Quantile-probability plots: Study 2 (Pizzagalli et al., Reference Pizzagalli, Iosifescu, Hallett, Ratner and Fava2008). Percent correct (right) and incorrect (left) for rich (circles) and lean (crosses) stimuli are shown as a function of RT quantiles, for adults with MDD (left column) and healthy controls (right column). The six quantiles are marked on the controls' data; they are shifted upwards on the y-axis for the MDD group, with the magnitude of the shift increasing with longer response latency. The effect of stimulus type on accuracy (rich > lean) is apparent in the 0.1 and 0.3 quantiles in the MDD group, but it is evident at every quantile for the controls.
The restriction of response bias to fast RTs was partially replicated. As depicted in Fig. 4, the MDD group showed an accuracy advantage for the rich stimulus in the 0.1 quantile that was reduced in the 0.3 quantile and absent thereafter. This mirrors Study 1. By contrast, in controls the rich > lean accuracy effect was visible at every quantile—notice the consistent horizontal separation between circles and crosses. Accordingly, a Group x Stimulus ANOVA on accuracy for fast RTs (⩽0.3 quantile) revealed a main effect of Stimulus, F(1, 41) = 22.40, p < 0.001, η p2 = 0.35, due to higher mean ± s.e. percent correct in response to the rich (89.98 ± 3.47%) v. lean (65.99 ± 3.47%) stimulus across both groups. By contrast, an analogous ANOVA on slower RTs (>0.3 quantile) revealed a Group x Stimulus interaction, F(1, 46) = 5.42, p = 0.02, η p2 = 0.11, due to a rich > lean effect that was significant by post-hoc Tukey test only in controls (rich: 86.21 ± 1.83%; lean: 75.05 ± 1.83%; t = 4.60, p < 0.001).
HDDM
Figure 5 shows that the HDDM again revealed slower drift rates (q = 0.037) and higher decision thresholds (q < 0.001) in depressed adults. There was no group difference in starting bias or non-decision time, qs > 0.23. These results replicate Study 1.
Fig. 5. HDDM results: Study 2 (Pizzagalli et al., Reference Pizzagalli, Iosifescu, Hallett, Ratner and Fava2008). Plots of posterior probabilities for HDDM parameters. Relative to results in the controls (dashed lines), drift rate was reduced and threshold width was increased in the MDD group (solid lines). There were no group differences in starting point bias or non-decision time. *q < 0.038.
As in Study 1, the strongest predictors of response bias and discriminability were starting point bias (Z = 3.29, p = 0.002) and drift rate (Z = 13.26, p < 0.001), respectively (Figure S3). In contrast to Study 1, response bias was also predicted by Group (Z = −2.12, p = 0.040). As in Study 1, discriminability was also predicted by decision threshold (Z = 3.58) and non-decision time (Z = 3.45), ps < 0.001.
Psychometrics
Split-half reliabilities for Studies 1 and 2 are in Figures S4 and S5. Reliability was good for response bias (SBs > 0.677) and discriminability (SBs > 0.844). For the HDDM, reliability was good for starting bias (SBs > 0.777) and outstanding for all other parameters (SBs > 0.910).
Discussion
This study found support for the hypothesis that decision-making deficits are present in MDD. First, RT distributions were more positively skewed in depressed v. healthy adults. Indeed, although there were differences in the fastest RTs, judging by Cohen's d the group differences for the slowest RTs were 6.66 (2.53/0.38) and 2.03 (3.60/1.77) times larger in Studies 1 and 2. As detailed in the Introduction, this suggests slow evidence accumulation in MDD. Second, the HDDM returned a result consistent with this interpretation: drift rate was lower in depressed adults. Decision thresholds were also wider in the MDD groups, potentially indicating a cautious response style. However, the effect on drift rate was practically more important as drift rate predicted cumulative reward totals, which were reduced in MDD in Study 1, while decision threshold did not. Neither study returned a group difference in non-decision time, thus the results do not appear to reflect group differences in sensorimotor processes. There were also no group differences in starting point bias, although starting points were consistently shifted away from the midpoint and towards the rich boundary, consistent with the asymmetric reinforcement rate used in the PRT. In summary, this study established a decision-making deficit in MDD across two independent samples, and localized that deficit to the evidence accumulation process, indexed by drift rate.
A goal for future work is to establish the generality of these results. They may reflect a domain-general effect of MDD that would be detectable across various tasks and thus could help explain the broad cognitive deficits seen in MDD. Additional research using different tasks is needed to investigate this possibility, but limited extant work is encouraging. For instance, a recognition memory study identified a drift rate advantage for positive v. negative material that was reduced in dysphoric students (White et al., Reference White, Ratcliff, Vasey and McKoon2009). If similar results can be obtained in MDD, that would increase confidence that depression reliably reduces the speed of evidence accumulation, perhaps especially when stimuli are positively valenced.
However, the results may depend on details of the PRT. In particular, the evidence accumulated during decision-making is influenced by stimulus properties, and drift rates are typically lower for short v. long duration presentations (Thapar et al., Reference Thapar, Ratcliff and McKoon2003). It is thus possible that the reduced drift rates in the current MDD groups reflect difficulty extracting high-quality evidence from rapidly presented stimuli. If so, then the results would not generalize to tasks with longer duration stimuli. Additional work is needed to distinguish between these possibilities.
This research also provides new insights into the PRT, which is widely used to study reward processing across different clinical groups (Barch et al., Reference Barch, Carter, Gold, Johnson, Kring, MacDonald, Pizzagalli, Ragland, Silverstein and Strauss2017) and non-human animals (Der-Avakian et al., Reference Der-Avakian, D'souza, Pizzagalli and Markou2013). First, we found that response bias was stronger for fast v. slow RTs; in Study 1, the response bias was confined to the 0.1 and 0.3 quantiles, which means that approximately 70% of responses were not biased. Similarly, in both studies starting point bias (in the HDDM) was the strongest predictor of response bias (in the PRT). This helps explain why biased responses are fast: the drift process starts close to the rich boundary such that minimal evidence needs to accumulate before a rich response is made. These findings are consistent with results from White and Poldrack (Reference White and Poldrack2014), but they are the first to demonstrate a dependency between response bias and RT in the PRT. Second, the internal consistency of response bias, discriminability, and the HDDM parameters was good to excellent. We conclude that PRT data are reliable, especially if the HDDM is used to extract estimates of underlying processes.
These results also raise several questions. First, given that a reduced response bias is often found in adults who are depressed or at risk for depression, why did a group difference in response bias emerge in Study 2 but not Study 1? The quantile-probability plots provide some insight. White and Poldrack (Reference White and Poldrack2014) distinguished between biases confined to response execution (response biases) v. stimulus processing (stimulus biases), showing that response biases affect fast but not slow RTs, while stimulus biases affect fast and slow RTs. On this analysis, both Study 1 groups and the Study 2 MDD group showed response biases: they responded ‘rich’ more than ‘lean’ only when replying quickly. By contrast, the Study 2 controls developed a stimulus bias, showing a rich > lean accuracy advantage at every quantile. The presence of biased responses for slow RTs in controls but not depressed adults explains why a group difference emerged in Study 2, and the lack of this effect explains why no group difference emerged in Study 1. Of course, this does not explain why only the Study 2 controls developed a stimulus bias, but it suggests that researchers using the PRT should distinguish between response and stimulus bias in order to determine why one v. the other is more likely to emerge.
Another question that remains unanswered is, what causes low drift rates in MDD? Two candidate hypotheses include reduced integrity of white matter pathways that enable evidence accumulation from distant brain regions (Madden et al., Reference Madden, Spaniol, Costello, Bucur, White, Cabeza, Davis, Dennis, Provenzale and Huettel2008), or reduction in cortical dopamine levels that can support fast drift rates (Beste et al., Reference Beste, Adelhöfer, Gohil, Passow, Roessner and Li2018). These hypotheses are informed by prior studies linking MDD to white matter abnormalities (Jiang et al., Reference Jiang, Zhao, Hu, Du, Chen, Wu, Li, Zhu, Kumar and Gong2017) and dopamine dysfunction (Treadway and Zald, Reference Treadway and Zald2011), but they await empirical test. An important future direction is thus to pair DDM-based analysis of behavior with electrophysiological and neuroimaging data collected from patient groups, to study the pathophysiology underlying deficits like the one reported here.
Finally, it is worth underscoring the fact that the HDDM detected effects of MDD on cognition that were not obvious in the traditional PRT analyses. For example, in Study 1 the MDD group showed poorer discriminability than the controls, but only at trend levels. Discriminability depends on a participant's ability to rapidly accumulate evidence in favor of each response option, captured by drift rate, but also on the criterion the participant uses to judge when there is sufficient evidence to make an accurate response, captured by decision threshold. The HDDM revealed that drift rates were markedly lower in depressed v. healthy adults, but decision thresholds were also wider. These two results appear to have roughly counterbalanced each other, leading to the weak group difference in discriminability. Nevertheless, they were clearly dissociable—for example, drift rate predicted cumulative reward totals while decision thresholds did not. This is an example of the explanatory power of computational modeling: it can identify the distinct contributions that different cognitive processes make to behavior.
In conclusion, in two PRT studies, we found evidence of decision-making deficits in unmedicated adults with MDD. These deficits were localized to the evidence accumulation process and were reflected in lower drift rate. It would be valuable to conduct additional studies to determine if these results are specific to tasks that use briefly presented stimuli, such as the PRT, or if they generalize to those that use longer stimulus durations. Moreover, pairing this approach with collection of electrophysiological or neuroimaging data could advance our understanding of pathophysiology. Finally, if we can identify a domain-general decision-making deficit in MDD, then we should gain valuable insight into the broad cognitive deficits that accompany depression.
Declaration of conflicts of interest
In the past 3 years, Dr Dillon has provided consulting services to Pfizer, Inc., for projects unrelated to this report. Over the past 3 years, Dr Pizzagalli has received consulting fees from Akili Interactive Labs, BlackThorn Therapeutics, Boehringer Ingelheim, Posit Science, and Takeda Pharmaceuticals for activities unrelated to the current research. Dr Trivedi reports the following lifetime disclosures: research support from the Agency for Healthcare Research and Quality, Cyberonics Inc., National Alliance for Research in Schizophrenia and Depression, National Institute of Mental Health, National Institute on Drug Abuse, National Institute of Diabetes and Digestive and Kidney Diseases, Johnson & Johnson, and consulting and speaker fees from Abbott Laboratories Inc., Akzo (Organon Pharmaceuticals Inc.), Allergan Sales LLC, Alkermes, AstraZeneca, Axon Advisors, Brintellix, Bristol-Myers Squibb Company, Cephalon Inc., Cerecor, Eli Lilly & Company, Evotec, Fabre Kramer Pharmaceuticals Inc., Forest Pharmaceuticals, GlaxoSmithKline, Health Research Associates, Johnson & Johnson, Lundbeck, MedAvante Medscape, Medtronic, Merck, Mitsubishi Tanabe Pharma Development America Inc., MSI Methylation Sciences Inc., Nestle Health Science-PamLab Inc., Naurex, Neuronetics, One Carbon Therapeutics Ltd., Otsuka Pharmaceuticals, Pamlab, Parke-Davis Pharmaceuticals Inc., Pfizer Inc., PgxHealth, Phoenix Marketing Solutions, Rexahn Pharmaceuticals, Ridge Diagnostics, Roche Products Ltd., Sepracor, SHIRE Development, Sierra, SK Life and Science, Sunovion, Takeda, Tal Medical/Puretech Venture, Targacept, Transcept, VantagePoint, Vivus, and Wyeth-Ayerst Laboratories. All other authors report no biomedical financial interests or potential conflicts of interest.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291719001570
Author ORCIDs
Daniel G. Dillon, 0000-0002-1977-700X.
Acknowledgements
The analysis was made possible by funding from McLean Hospital and a grant from NIMH (R00MH094438) awarded to Dr Dillon. Data collection and prior analyses were made possible by additional funding from NIMH (Pizzagalli: R01MH68376; Trivedi: U01MH092221; McGrath, Parsey, Weissman: U01MH092250). The authors are solely responsible for the content of this manuscript, which does not necessarily represent the official views of the NIH.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
