Skip to main content
. 2019 Aug 20;8:e43499. doi: 10.7554/eLife.43499

Figure 5. Confidence-dependent modulation of decision bounds generalizes to different experimental tasks.

(A) In Experiment 2, participants need to decide as quickly as possible which of two boxes has more dots. In Experiment 3, participants needed to decide, as fast and accurately as possible, whether the average color of the eight elements was more red or blue. (B) Subsequent RT, accuracy and their product as a model-free measure of decision bound as a function of confidence. Green crosses show fits from the DDM. (C) Subsequent decision bounds and subsequent drift rates as a function of confidence. Inset in B shows the distribution of empirical and fitted RTs.

Figure 5.

Figure 5—figure supplement 1. Behavioral results of Experiment 2.

Figure 5—figure supplement 1.

(A) Accuracy as a function of decision confidence. Confidence ratings were closely linked to choice accuracy, even on a trial-by-trial basis (logistic regression of confidence on accuracy: significant positive slopes for all 16 observers, with all ps < 0.001). Dot size reflects the percentage of trials per confidence label, separately for each participant. (B) Accuracy as a function of reaction time. Although there was a significant linear relationship between RTs and accuracy, b = 0.022, t(15) = 4.19, p<0.001, a polynomial regression confirmed the apparent inverted U-shape seen in the data, t(15) = −7.52, p<0.001.
Figure 5—figure supplement 2. Simple effects of confidence on trialn and confidence on trialn+2 on the product of subsequent RTs and accuracy as a model-free measure of decision bound for Experiment 1 (A) and Experiment 2 (B).

Figure 5—figure supplement 2.

Same conventions as in Figure 3.
Figure 5—figure supplement 3. Simple effects of confidence on trialn and confidence on trialn+2 on decision bound (A) and drift rate (B) on trialn+1 (Experiment 2).

Figure 5—figure supplement 3.

Trials with high confidence are treated as reference category.
Figure 5—figure supplement 4. Complementary approach controlling for slow fluctuations (Experiment 2).

Figure 5—figure supplement 4.

Same procedure as in Figure 3—figure supplement 3. These results remained unchanged when only including correct trials (there were not enough trials when only including errors).
Figure 5—figure supplement 5. Behavioral results of Experiment 3.

Figure 5—figure supplement 5.

(A) Mean reaction time on correct trials (top), accuracy (upper middle) estimated mean drift rate (lower middle), and confidence (bottom) as a function of coherence level. RTs on correct trials and choice accuracy scaled with SNR In both experiments (Exp 3A: RT, F(4, 3012.3)=21.60, p<0.001, error rates, X²(4)=170.74, p<0.001; Exp 3B: RT, F(3, 2173.1)=3.07, p=0.027, error rates, X²(3)=83.311, p<0.001). Correspondingly, drift rates estimated from fits of a hierarchical drift diffusion model also increased monotonically with SNR in both experiments (Exp 3A: Friedman χ2(4)=41.02, p<0.001; Exp 2B: Friedman χ2(3)=31.61, p<0.001). In these model fits, decision bound was not allowed to vary as a function of SNR, and its average was 1.33 (SD = 0.17) and 1.33 (SD = 0.20) in Experiment 3A and 3b, respectively. Similarly, non-decision time was estimated independently from SNR, and its average was 0.28 (SD = 0.08) and. 022 (SD = 0.05). Green crosses show the fits from the DDM. (B) Accuracy as a function of decision confidence. Confidence ratings were closely linked to choice accuracy, even on a trial-by-trial basis while including SNR as a covariate (logistic regression of confidence on accuracy: significant positive slopes for 21 observers, with all ps < 0.001, and non-significant positive slopes for two observers, with both ps > 0.354), showing that confidence predicts accuracy over and above SNR. Dot size reflects the percentage of trials per confidence label, separately for each participant. (C) Accuracy as a function of reaction time. There was no linear relationship between RTs and accuracy, b = −0.001, p>0.545. A polynomial regression did, however, confirm the inverted U-shape seen in the data, t(29.2) = −3.19, p=0.003. (D) Confidence as a function of coherence level, separately for corrects and errors. Confidence judgments only scaled with SNR in Experiment 3A (Exp 3A: F(4,3953.9) = 37.34, p<0.001; Exp 3B: F(3,2867.1) < 1, p=0.141), but there was a highly significant interaction between SNR and accuracy in both versions of the experiment (Exp 3A: F(4,3955.1) = 87.59, p<0.001; Exp 3B: F(3,2867.8) = 48.16, p<0.001): confidence increased with SNR for corrects (linear contrast: Exp 3A: p<0.001; Exp 3B: p<0.001) and decreased with SNR for errors (linear contrast: Exp 3A: p<0.001; Exp 3B: p<0.001).
Figure 5—figure supplement 6. Simple effects of confidence on trialn and confidence on trialn+2 on decision bound (A) and drift rate (B) on trialn+1 (Experiment 3).

Figure 5—figure supplement 6.

Trials with high confidence are treated as reference category.
Figure 5—figure supplement 7. Complementary approach controlling for slow drifts in performance (Experiment 3).

Figure 5—figure supplement 7.

Same procedure as in Figure 3—figure supplement 3. When only including correct trials, the effect of perceived errors on drift rate turned non-significant, p=0.250, apart from that the results remained unchanged (there were not enough trials when only including errors).