Mind wandering at the fingertips: automatic parsing of subjective states based on response time variability

Abstract

RESEARCH FROM THE LAST DECADE HAS SUCCESSFULLY USED TWO KINDS OF THOUGHT REPORTS IN ORDER TO ASSESS WHETHER THE MIND IS WANDERING: random thought-probes and spontaneous reports. However, none of these two methods allows any assessment of the subjective state of the participant between two reports. In this paper, we present a step by step elaboration and testing of a continuous index, based on response time variability within Sustained Attention to Response Tasks (N = 106, for a total of 10 conditions). We first show that increased response time variability predicts mind wandering. We then compute a continuous index of response time variability throughout full experiments and show that the temporal position of a probe relative to the nearest local peak of the continuous index is predictive of mind wandering. This suggests that our index carries information about the subjective state of the subject even when he or she is not probed, and opens the way for on-line tracking of mind wandering. Finally we proceed a step further and infer the internal attentional states on the basis of the variability of response times. To this end we use the Hidden Markov Model framework, which allows us to estimate the durations of on-task and off-task episodes.

Keywords: Hidden Markov Models; mind wandering; response times variability; subjective report; time-course analysis.

Figure 1

RTCV (Response time coefficient of variability) as a function of whether participants are on-task or mind wandering (“off-task”), for both the trials preceding and following a report. Error bars are standard errors of the mean.

Figure 2

Distribution of the Continuous Variability Index (CVI) across the experimental session of an arbitrary participant (time in minutes). Vertical lines represent the CVI value (in RTCV units) at the moment of the report, horizontal lines represent temporal distance from the report to the closest peak in the CVI. We predict that On-task reports (blue) have a lower CVI (shorter vertical lines), and a higher temporal distance to their closest peak (longer horizontal lines) than Off-task reports (red).

Figure 3

Markov Chain of attentional states. Illustrative time series of on-task (OT) and mind wandering (MW) states, with two pairs of complementary probability transitions to stay (ex. P_ot/ot: stay focused) in the same state or transition (ex. P_ot/mw = 1 − P_ot/ot: start mind wandering) to the other.

Figure 4

Six-Parameters Models accounting for increased variability during mind wandering. P_ot/mw: transition probability to start mind wandering when on-task, P_mw/ot: transition probability to come back on task when mind wandering, μ: mean of the distribution, σ: variance of the distribution, τ: skewness of the distribution. The critical parameter is d, “difference parameter,” applied either to σ (“variance model”) if variance increases during mind wandering, or to τ (“exponential model”) if skewness increases during mind wandering. (A) Variance Model (B) Exponential Model.

Figure 5

(A) Overall posterior distribution of the parameter d in the exponential model, pooled across all 47 participants (60000 samples per participant). Note that the prior was uniform over [0, 2]. (B) Overall posterior distribution of the two transition probabilities in the exponential model. The prior was uniform over [0, 1] (60000 samples per participant).

Figure 6

(A) Posterior distribution of hidden state across the experimental session of an arbitrary participant (time in minutes). (B) Posterior distribution of the hidden states across all participants and trials. This distribution seems bimodal, meaning that the model categorically distinguishes between the two hidden states.