Study Participants

The study was approved by the Cambridgeshire 3 National Health Service research ethics committee. An early-psychosis group (N = 31) was recruited, consisting of individuals with first-episode psychotic illness (N = 14) or with at-risk mental states (ARMS; N = 17) from the Cambridge early intervention service in psychosis (CAMEO). Inclusion criteria were as follows: age 16–35 years and current psychotic symptoms meeting either ARMS or first episode of psychosis (FEP) criteria according to the Comprehensive Assessment of At-Risk Mental States (CAARMS; Morrison et al., 2012; Yung et al., 2005). Patients with FEP were required to meet ICD-10 criteria for a schizophrenia spectrum disorder (F20, F22, F23, F25, F28, F29) or affective psychosis (F30.2, F31.2, F32.3). Healthy volunteers (N = 31) without a history of psychiatric illness or brain injury were recruited as control subjects. None of the participants had drug or alcohol dependence. Healthy volunteers had not reported any personal or family history of neurological or psychotic illness and were matched with regard to age, gender, handedness, level of education, and maternal level of education. None of the patients with ARMS were taking antipsychotic medication, and four patients with FEP were on antipsychotic medication at the time of testing. All of the experiments were completed with the participants’ written informed consent.

Behavioral JTC Task

This was a novel task (Figure 1), based on previously published tasks (Garety et al., 1991; Huq et al., 1988) of reasoning bias in psychosis, but amended in the light of decision-making theory, according to which the amount of evidence sought is inversely proportional to the costs of information sampling. These costs include the high subjective cost of uncertainty and the cost to self-esteem or other factors (Moutoussis et al., 2011). Participants were told that there are two lakes, each containing black and gold fish in two different ratios (60:40). The ratios were explicitly stated and displayed on the introductory slide. A series of fish was drawn from one of the lakes; all the previously “caught” fish were visible to reduce the working memory load. The participants were informed that fish were being “caught” randomly from either of the two lakes and then allowed to “swim away.” We used a pseudo-randomized order for each trial, which was the same for all participants. The lake from which the fish were drawn was also pseudo-randomized.

Participants could ask for a maximum of 20 fish to be shown. After each fish shown, they indicated whether the fish came from Lake G (mainly gold) or Lake B (mainly black) or asked to see another fish. The trial terminated when the subject chose the lake. There were four blocks, each with the 10 trials of the predetermined sequences to increase reliability. Block 1 was similar to the classical beads task and was included to provide a reference point. The only difference was that feedback (“correct” or “incorrect”) was provided after each trial. In Block 2, a win was assigned to a correct decision (100 points) and a loss (−100 points) to an incorrect decision. In Block 3, the cost of each extra fish after the first one was introduced (−5 points) was subtracted from the possible win or loss of 100 points for making a correct or incorrect decision, respectively. Block 4 was similar to Block 3, but the information sampling cost was incrementally increased: The first fish would cost 0 points, the second −5 points, the third −10 points, and so on. Thus a higher number of fish sampled led to more lost points. Subjects performed the task at their own pace. Whether Lake G or Lake B was correct was randomized. The task consisted of four blocks. Within each block were 10 trials of predetermined sequences of fish to increase reliability. All of the fish that were “caught” during one trial were visible on the screen to minimize the working memory load. Block order was not randomized because the task increased in complexity.

The main outcome variable was the number of fish sampled (DTD). Secondary outcomes were the accuracy of the decision, calculated according to Bayes’s theorem, based on the probability of the chosen lake given the color and number of fish seen (Everitt & Skrondal, 2010), and the dichotomous JTC variable, which is defined as making a decision after two or fewer pieces of information.

Partially Observable Markov Process Decision-Making

We consider a belief-based model of decision-making, formally a partially observable Markovian decision process, to model behavior in this task. The process is Markovian because we can concisely formulate the state in which the people find themselves upon observing n_d draws, so that the state contains all information that can be extracted from observations thus far. As beads are drawn from one jar only at each trial, this can be simply defined as the number of g fishes seen so far and the total number seen: s = [n_d, n_g]. The agent is interested to infer upon the true state of the world, which is a B or G lake. This is not directly observable but partially observable. The agent maintains a belief component of their state, P(G|s). The Markov property still holds: future beliefs are independent of past beliefs given the current state (Figure 2A). As we will see, belief-state transitions can be calculated just by considering the evidence so far. Given the current belief state, the probability of the possible unfolding of the task into the future can be estimated and hence the expected returns for each possible future decision (Figure 2B). The value of the available choices can thus be estimated: choosing the B lake D_B, the G lake D_G, or sampling another piece of information D_S. The agent chooses accordingly and either terminates the trial or gathers a new datum and repeats the process (Figure 2).

Figure 2.  

State space schematic. A) Markovian transitions in this task. Top: belief (probabilistic) component of states; middle: observable part of the state (data/feedback). Down arrows: actions (sample, declare). Bottom: true state. For example, let the cost of sampling be very high. Then b₀ may be “equiprobable lakes,” Action 1 “sample,” s₁ “B,” b₁ “60% B,” Action 2 “declare B,” and s₁ “Wrong.” B) In this example, sampling cost is very low. A person has drawn 15 fishes, 7 of them g, hence the position of 15 on the x-axis and +1 on the y-axis as there is a +1 excess of black fish so far. The visible states corresponding to all possible future draws are shown. Looking ahead (example: gray arrow), the agent finds the “sampling” action more valuable in that the current preference for the B lake is likely to be strengthened at very low cost.

We now formally specify the model. Let P(G|n_d, n_g) be the probability of the lake being gold after drawing n_d fish and seeing n_g gold fish. Using Bayes’s theorem, and assuming that a gold lake and a black lake are equally probable,

$P\left.G|n_d,n_g\right)=\frac{P\left.n_d,n_g|G\right)P(G)}{P\left.n_d,n_g|G\right)P(G)+P\left.n_d,n_g|B\right)P(B)}=\frac{P\left.n_d,n_g|G\right)}{P\left.n_d,n_g|G\right)+P\left.n_d,n_g|B\right)},$

$P\left.n_d,n_g|G\right)=\left.\begin{array}{c}{n}_g\\ n_d\end{array}\right)P{\left.g|G\right)}^{n_g}{\left.1-P\left.g|G\right)\right]}^{n_d-n_g},$

$P\left.n_d,n_g|B\right)=\left.\begin{array}{c}{n}_g\\ n_d\end{array}\right)P{\left.g|B\right)}^{n_g}{\left.1-P\left.g|B\right)\right]}^{n_d-n_g},$

((1))

$P\left.G{n}_d,n_g\right)=\frac{P{(gG)}^{n_g}{\left.1-P(gG)\right]}^{n_d-n_g}}{P{(gG)}^{n_g}{\left.1-P(gG)\right]}^{n_d-n_g}+P{(gB)}^{n_g}{\left.1-P(gB)\right]}^{n_d-n_g}}.$

We then need to calculate the value of each action. For the “declare” choices, the action value is the expected reward for a correct answer minus the expected cost for a wrong answer. For example, if Q(D_G; n_d, n_g) is the action value of declaring the lake gold, after drawing n_d fish and seeing n_g gold fish,

((2))

$Q\left.D_G;n_d,n_g\right)=R_{C}P\left.G{n}_d,n_g\right)-C_{W}P\left.B{n}_d,n_g\right),$

where R_C is the reward for declaring the color of the correct lake and C_W is the cost of declaring the color of the wrong lake. In our task, R_C = C_W, so

$Q\left.D_G;n_d,n_g\right)=R_C\left.P\left.G|n_d,n_g\right)-P\left.B|n_d,n_g\right)\right]=R_C\left.2P\left.G|n_d,n_g\right)-1\right].$

Similarly,

$Q\left.D_B;n_d,n_g\right)=R_C\left.P\left.B|n_d,n_g\right)-P\left.G|n_d,n_g\right)\right]=R_C\left.1-2P\left.G|n_d,n_g\right)\right].$

The action value of sampling again is the expectation over the value of the next state minus the cost of sampling C_S(n_d). If the value of the (possible) next state is V(s′), the action value for sampling again is the sum over the new possible states, weighted by their probabilities. The latter depends, in turn, on the identity of the true underlying lake, L:

((3a))

$Q\left.D_S;s\right)=-C_S+\sum _{L\in \{G,B\}}P\left.L|s\right)\sum _{s^{\prime }}V\left.s^{\prime }\right)P\left.s^{\prime }|L\right).$

The possible outcomes for sampling again, and getting a black fish (going from (n_d, n_g) to (n_d + 1, n_g)) and getting a gold fish (going from (n_d, n_g) to (n_d + 1, n_g + 1)), are

$Q\left.D_S;n_d,n_g\right)=-C_S\left.n_d\right)+P\left.G|n_d,n_g\right)\times \left.V\left.n_d+1,n_g+1\right)P\left.g|G\right)+V\left.n_d+1,n_g\right)P\left.b|G\right)\right\}$

((3b))

$+P\left.B{n}_d,n_g\right)\times \left.V\left.n_d+1,n_g+1\right)P(gB)+V\left.n_d+1,n_g\right)P(bB)\right\}.$

Agents will tend to prefer actions with the greatest value. An ideal, reward-maximizing agent will always choose the action with the maximum value and will thus endow the corresponding state with this value. Denoting q as the vector of action values,

((4a))

$q\text{(s)}=\left.Q(D_G;n_d,n_g),Q\left.D_B;n_d,n_g\right),Q(D_S;n_d,n_g)\right]\Rightarrow V(n_d,n_g)=max(q(s)).$

Real agents will choose probabilistically as a function of action values, so

((4b))

$V(s)=\sum _{a\in \left.D_{S,G,B}\right\}}{q}_{s}p(a|q(s)).$

Agents cannot fill in the action values for sampling starting from their current state, as the next state value is not known. However, they can fill in all values by backward inference. At the very end, n_d = 20, sampling is not an option, and the action values can be calculated directly from Equation 2 and the state value from Equation 4a. Once all possible state values for n_d = 20 have been calculated, Equations 2, 3a, 3b, 4a, and 4b are used to calculate action and state values for n_d = 19 and downward to n_d = 1.

We now turn to the ideal, deterministically maximizing agent against which we can compare human performance. When given the same sequence of fish as human participants, on average, the ideal Bayesian agent samples 20, 20, 3.5, and 1 fish in each of the four blocks, respectively, achieving total winnings of 1,070 points (Table 3). For no cost (Blocks 1 and 2), an ideal Bayesian agent samples all the fish and has p = 0.835 of being correct (100% if we had infinite fish). For constant cost (5 points, Block 3), it samples until the difference between black and gold fish (Nb − Ng) is 2, with p = 0.692 probability of being correct. For increasing cost (Block 4), it guesses after the first fish, so there is 0.6 probability of being correct. To model real agents, the probabilistic action choice in Equation 4b took the softmax form:

((5))

$lnp(a|q(s))=q_a/T+z,$

with z a normalizing constant that is the same for all the actions at a specific state and T a decision temperature parameter, described later (Moutoussis et al., 2011).

Thus two parameters shape a participant’s behavior:

• •   CS, the cost of sampling, that is, the subjective cost of each additional piece of information compared to the final reward. It may be greater or smaller than the costs imposed by the experimenter but is here taken to be constant (in a given block). High values of CS mean that decisions are made early.
• •   T, the noise parameter. As this value increases, the probability of the participant following the ideal behavior (for their given value of CS) decreases and their actions become more uniformly random.

Rating Scales and Questionnaires

The participants underwent a general psychiatric interview and assessment (Tables 1 and 2) using the CAARMS (Morrison et al., 2012; Yung et al., 2005), the Positive and Negative Symptom Scale (PANSS; Kay, Fiszbein, & Opler, 1987), the Scale for the Assessment of Negative Symptoms (SANS; Andreasen, 1989), and the Global Assessment of Functioning (GAF; Hall, 1995). The Beck Depression Inventory (BDI; Beck, Steer, & Brown, 1996) was used to assess depressive symptoms during the last 2 weeks. IQ was estimated using the Culture Fair Intelligence Test (Cattell & Cattell, 1973). Schizotypy was measured with the 21-item Peters et al. Delusions Inventory (PDI-21; Peters, Joseph, Day, & Garety, 2004).

Statistical Analysis of the Behavioral Data

The effect of manipulations of wins, losses, and costs on the DTD and accuracy was assessed by repeated-measures analysis of variance (ANOVA) using SPSS 23. Although DTD was not normally distributed, repeated-measures ANOVA is robust to violations of normality and was therefore an appropriate test to run. IQ and depressive symptoms scores were initially included as covariates and dropped from the subsequent analysis where nonsignificant. We report two-tailed p-values, which were significant at p < 0.05. When the assumption of sphericity was violated, we applied Greenhouse–Geisser corrections. We also examined whether DTD was correlated with the severity of psychotic symptoms in the patient group (CAARMS positive symptoms) and with schizotypy scores on the PDI in controls using Spearman’s correlation coefficients. The intraclass correlation coefficients (ICCs) were used to estimate the consistency of decision-making. We calculated the ICCs of the mean number of choices in each block separately for patients and controls.

For completeness and to help relate our study to prior literature (or for future meta-analyses), we report data on the dichotomous variable JTC, defined as making a decision after receiving one or two pieces of information, and we compare the FEP and ARMS patient groups separately to controls on Block 1 DTD and estimated model parameters.

Demographical Characteristics of the Participants

In Tables 1 and 2, the sociodemographic and cognitive parameters as well as clinical measures of the participants are presented. Although healthy volunteers had higher IQ compared to the psychosis group, the difference was not significant, and the groups were matched with regard to level of maternal education and number of years of education. The control and patient groups did not differ with regard to gender or age. There were significantly more smokers in the patient group compared to the control group, and some subjects of the patient group used more recreational drugs. For all of the other measures (e.g., alcohol or cannabis), there were no significant group differences. As expected, there are significant differences between the healthy volunteers and the patients in all subscales of CAARMS (Table 2), on the self-reported depression questionnaire (BDI), and in self-reported schizotypy scores (PDI). In the patient group only, we performed additional clinical assessments. The mean (±SD) score for PANSS positive symptoms was 13.68 (±3.99); for PANSS negative symptoms 9.87 (±4.88); for SANS 0.35 (±0.75), and for GAF 55.00 (±18.54).

Group Differences in the Number of DTD and Points

Inspection of Figure 3A reveals that in all of the blocks, the controls took more DTDs than the patients. Mauchly’s test indicated that the assumption of sphericity was not violated, W(5) = 0.211, p < 0.001. On mixed-model ANOVA, there were significant main effects of block, F(3) = 94.49, p < 0.001, and of group, F(1) = 5.99, p = 0.017, on the number of DTD. The interaction between the group and the block was also significant, F(3) = 4.32, p = 0.006. This indicates that, depending on the group, block change had different effects on DTD. Group differences were statistically significant in the first two blocks (Block 1, p = 0.007; Block 2, p = 0.028), whereas in Blocks 3 and 4, the group differences were increasingly attenuated, as the cost of decision-making became increasingly high (Block 3, p = 0.059; Block 4, p = 0.419). Control behavior was more similar to the ideal Bayesian agent than patients on the first three blocks but not Block 4 (Table 3). Group differences in the probability of being correct were very similar to the results in the number of DTD (Figure 3B).

Analyzing the points won/lost in Blocks 2–4 (Figure 4), we identified four outliers in each group that significantly exceeded the ±2 standard deviation threshold. After excluding these subjects, the mixed-model ANOVA revealed a significant effect for block, F(2) = 93.73, p < 0.001, a marginally significant interaction between Block ⋅ Group, F(2) = 2.52, p = 0.086, and a significant groups effect, F(1) = 4.14, p = 0.047. Patients won significantly fewer points in Block 2, F(2) = 4.65, p = 0.035, but did not differ from controls in Blocks 3 and 4, both p > 0.3. Percentage and count of people displaying a JTC reasoning style are presented in Table 4.

Intraclass Correlations of DTD and SD of the Mean DTD

ICCs of the mean number of DTD within each block were calculated separately for patients and controls. Within all blocks in both groups, the correlations were very high, indicating that behavior was consistent within each block (ICC values: Block 1, patients 0.965, controls 0.943; Block 2, patients 0.982, controls 0.973; Block 3, patients 0.972, controls 0.976; Block 4, patients 0.973, controls 0.978).

Correlation of Symptoms and IQ With DTD

We calculated two-tailed Spearman’s correlations with positive psychotic symptoms in patients and with PDI scores in controls. In the patient group, we used a summary score of the three CAARMS subscales that quantify positive psychotic symptoms, namely, Unusual Thought Content, Non-bizarre Ideas (mainly persecutory ideas), and Perceptual Aberrations. An additional advantage of the summary measure is that it provides one measure to reduce the number of correlations that need to be performed. We also ran correlation BDI scores because the groups differed on this measure. In the group with psychotic symptoms, the correlations with the overall CAARMS score were significant in the first three blocks (Block 1, ρ = −0.515, p = 0.003; Block 2, ρ = −0.489, p = 0.005; Block 3, ρ = −0.491, p = 0.005). There were no correlations in the psychotic symptoms group between DTD and BDI score, for all p > 0.1.

In controls, we found significant correlations of DTD in Block 1 with the Distress, ρ = −394, p = 0.035, and Preoccupation, ρ = −0.462, p = 0.012, subscales of the PDI. DTD in Block 2 correlated with the Preoccupation subscale of PDI, ρ = −0.387, p = 0.038. There were no significant correlations with BDI scores (e.g., for Block 1, ρ = −0.023, p = 0.905).

In the patients, we furthermore found a positive correlation between IQ and the first three blocks (Block 1, ρ = −0.481, p = 0.006; Block 2, ρ = −0.362, p = 0.045; Block 3, ρ = −0.364, p = 0.044). There was no such correlation in the controls.

Group Differences in the Probability of Being Correct (Accuracy)

Mauchly’s test indicated that the assumption of sphericity was not violated, W(5) = 0.667, p < 0.001. There was a significant main effect of block on the probability of being correct, F(3) = 53.502, p < 0.001, and a significant effect of group, F(1) = 5.514, p = 0.022. There was a significant interaction between the group and the block, F(1) = 2.791, p = 0.042, indicating that as the cost of decision-making increased, the group differences became attenuated.

Additional Analyses Having Excluded Participants on Medication or to Make Groups More Closely Matched on IQ

To demonstrate that the findings were unaffected by antipsychotic medication, we conducted repeat analyses having excluded the four patients taking antipsychotic medication. The results were similar to in the full sample. As before, on mixed-model ANOVA, there were significant main effects of block, F(3) = 88.59, p < 0.001, and of group, F(1) = 5.21, p = 0.026, on the number of DTD. The interaction between the group and the block was also significant, F(3) = 4.54, p = 0.004. Group differences were statistically significant in the first two blocks (Block 1, p = 0.013; Block 2, p = 0.023) but reduced in Blocks 3 and 4 (Block 3, p = 0.098; Block 4, p = 0.55).

When we excluded the three highest IQ controls and two lowest IQ patients to make groups more similar in IQ (resulting in control mean IQ 107 and patient mean 105), the results were similar: On mixed-model ANOVA, there were significant main effects of block, F(3) = 90.2, p < 0.001, and of group, F(1) = 5.25, p = 0.026, on the number of DTD. The interaction between the group and the block was also significant, F(3) = 3.50, p = 0.017. Group differences were statistically significant in the first two blocks (Block 1, p = 0.013; Block 2, p = 0.039) but reduced in Blocks 3 and 4 (Block 3, p = 0.062; Block 4, p = 0.45).

Computational Modeling Results: Analysis of Group Estimates of Model Parameters

In our computational analysis of Blocks 1–3, we found that patients with early psychosis assigned a higher cost to sampling more data than did healthy controls (Table 5). For example, in Block 1, where there was no explicit cost, the modeled mean cost of sampling in the controls was very low (1.9 ⋅ 10⁻³) compared to that of patients (1.7). The modeled variance was also higher for the patient group (13) compared to the controls (2.0 ⋅ 10⁻⁶). The estimated noise parameters were similar in both groups. For example, in Blocks 1 and 2, the respective group estimated mean noise parameters were 3.4 and 3.6 in controls and 4.2 and 2.9 in patients, respectively.

In Block 3, where an explicit cost of 5 points per fish sampled was assigned, the model shows that patients very slightly overestimated the cost of sampling compared to the assigned cost (estimated mean cost of sample for patients 5.2), whereas the controls underestimated it (estimated mean cost of sample 1.6).

In Block 4, we explicitly told participants that the cost for additional information is not fixed but increases with amount of information requested. We did not fit the computational model to Block 4 because the model does not take into account the increasing cost structure set up.

To test the null hypothesis that both groups are drawn from the same distribution of behavior parameters, we used iBIC. Table 6 shows the results of that calculation, where the null hypothesis can be rejected with strong evidence. For example, in Block 1, despite the fact that the iBIC penalizes the use of extra parameters substantially, when fitting CS_m, CS_v, T_m, and T_v separately for each group, iBIC was 3,450.2, compared to an iBIC of 3,489.7 for fitting them as one combined group. Taken together, the iBIC can be used to compare the hypotheses about the two models of interest. In our case, those are the following two: (a) Unhealthy and healthy groups have independent CS and T parameter distributions, characterized by eight numbers (mean and variance for CS and for T, for each group), and (b) all participants come from the same pool of noise and cost parameters, characterized by four numbers. The iBIC suggests that Hypothesis 1 is better explained by the separate model in Blocks 1 and 2 and by the combined model in Block 3. Model fits were not changed substantially after exclusion for medication (Table 6).

Computational Modeling Results: Analysis of Individual Participant-Level Cost and Noise Parameters

A very strong test of the hypothesis that the groups differ in the cost (or noise) parameters can be created by forcing the model to treat all participants as coming from one group, with a single group mean and variance, then using the model’s estimates of the single subject parameters to conduct a test of whether there are differences according to diagnostic group. Although we found (see earlier) that in Blocks 1 and 2, this assumption did not fit the data as well as modeling the groups separately, so the approach is overconservative, this procedure serves a purpose in subjecting the test of group differences to a stern challenge. The groups differed significantly on estimated cost parameters in this procedure (Block 1, controls mean = −0.24, median = −0.04, interquartile range = 0.07, psychosis mean = −1.1, median = −0.1, interquartile range 1.2, Mann–Whitney U = 289, p = 0.007; Block 2, controls mean = −0.8, median = −0.04, interquartile range = 0.08; psychosis mean = −2.2, median = −0.14, interquartile range = 4.9, U = 363, p = 0.038; Block 3, controls mean = −2.0, median = −0.17, interquartile range = 2.3, psychosis mean = −4.2, median = −0.48, interquartile range = 8.62, U = 338, p = 0.045) but generally were similar on estimated noise parameters (Block 1, controls mean = 2.96, SD = 2.08, patients mean = 4.14, SD = 2.86, t[60] = 1.86, p = 0.07; Block 2, controls mean = 2.27, SD = 1.61, patients mean = 2.81, SD = 1.95, t[60] = 1.2, p = 0.23; Block 3, controls mean = 4.1, SD = 0.74, patients mean = 4.31, SD = 0.41, t[60] = 1.2, p = 0.23).

The results were similar after exclusions for medication (Block 1, cost group difference Mann– Whitney U = 248, p = 0.008; noise group difference t[60] = 1.79, p = 0.08) or to equalized IQ (Block 1, cost group difference U = 246, p = 0.01; noise group difference t[60] = 1.5, p = 0.13).

Individually estimated greater cost parameters predicted higher psychotic symptom severity in patients, ρ = 0.58, p = 0.001. There was no significant association between estimated noise parameters and psychotic symptom severity, ρ = 0.27, p = 0.14. In controls, cost was associated with PDI preoccupation, ρ = 0.41, p = 0.03, and marginally with distress, ρ = 0.35, p = 0.06, and noise was associated with overall PDI score, ρ = 0.34, p = 0.07, distress, ρ = 0.42, p = 0.02, and preoccupation, ρ = 0.45, p = 0.01.

Cost parameter estimates were highly correlated across blocks (Blocks 1 and 2, ρ = 0.9; Blocks 1 and 3, ρ = 0.6; Blocks 2 and 3, ρ = 0.7). Noise parameters were also correlated across blocks (noise on the first two blocks, ρ = 0.8; noise on Blocks 1 and 3, ρ = 0.3; noise on Blocks 2 and 3, ρ = 0.4). However, despite associations across blocks, there were significant effects of block on both cost and noise (repeated-measures ANOVA effect of block on cost, F[2, 122] = 14, p = 0.000003; effect of block on noise, F[2, 122] = 22, p = 5 ⋅ 10⁻⁹). Cost and noise parameters were related to each other (e.g., on Block 1, ρ = 0.7) and, to a lesser extent, to IQ on some blocks (e.g., IQ vs. Block 1 cost, ρ = 0.3; Block 3, ρ = 0.3; Block 1 noise, ρ = 0.2; Block 3 noise, ρ = 0).

Subgroup Analysis

On Block 1 DTD, controls (mean DTD 12.1, SD = 5.4) gathered more evidence than FEP (mean 6.7, SD = 6.0), one-tailed t(44) = 3.0, p = 0.002, and ARMS (mean 9.2, SD = 5.9), one-tailed t(47) = 1.7, p = 0.045. On Block 1, cost parameter, controls (median 0.04, interquartile range 0.07) had lower values than FEP (median 2.5, interquartile range 4.1), Mann–Whitney U = 106, one-tailed p = 0.003, and ARMS (median = 0.06, interquartile range 0.7), Mann–Whitney U = 183, one-tailed p = 0.04. On Block 1, noise parameter, controls (mean 3.0, SD = 2.1) had lower values than ARMS (mean 4.2, SD = 3.1), t(45) = 1.7, one-tailed p = 0.045, and, marginally, than FEP (mean 4.0, SD = 2.7), t(44) = 1.5, one-tailed p = 0.075).

Summary

In summary, we found that early-psychosis patients demonstrate a hasty decision-making style compared with healthy volunteers, sampling significantly less information. This decision-making style was correlated with delusion severity, consistent with the possibility that it may be a cognitive mechanism contributing to delusion formation. Our data are not consistent with the account that patients sample less information because they are in general more noisy decision makers. Rather, our data suggest that patients with psychosis sample less information before making a decision because they attribute a higher cost to information sampling. Although psychosis patients were less able to adapt to the changing demands of the task, they did alter their decision-making style in response to the changing explicit costs of information, indicating that an impulsive decision-making style is not completely fixed in psychosis. This finding is consistent with the possibility that information sampling may be a treatment target, for example, for psychotherapy (Moritz et al., 2014), and that patients with psychosis may benefit in this neuropsychological domain, as they have in other domains, from cognitive scaffolding approaches exemplified in cognitive remediation therapy (Cella & Wykes, 2017).