Sequential dynamics of prediction : from shifting to biasing of decision thresholds
Assuming the fundamental and pervasive cognitive operation of prediction—even about objectively unpredictable events—the role of prediction in a trial-by-trial sequential effect on choice speed was studied: the adverse between-choice (ABC) sequential effect. As identified in lexical decision, this effect involves faster responses on word and nonword trials after a word trial. This effect is distinct to more generally observed sequential effectsthat, in terms of the parameters of evidence-accumulation models (EAMs) of choice performance, can be explained by biasing thresholds to local choice probability. In contrast, the ABC effect can be explained by prospective threshold shifting: Rather than biasing thresholds, the total evidence required across choices(response caution) is shifted up or down according to the choice made on the previous trial. Lower response caution after a word trial thereby leads to faster word and nonword responses on the next trial. Alternatively, the ABC effect can be explained by sequential variations in accumulation rates, with the rates becoming positively or negatively correlated between trials depending on prior stimulus difficulty. The relative roles of these mechanisms in producing the ABC effect remained to be empirically determined, while the effect potentially offered insight into how prediction dynamically influences serial decisions—by prospective threshold shifts, not only by threshold biasing.
A review of studies reporting the ABC effect identified two common procedural constraints in earlier studies: short time-on-task and long response-stimulus interval (RSI). Explaining the ABC effect by threshold shifting under these conditions implicated the relation between response caution and speed/accuracy trade-off (SATO)—that the effect is a by-product of strategic learning of a SATO policy in early trials that is contingent on the temporal opportunity to volitionally shift thresholds between trials. Explaining the ABC effect by autocorrelation of accumulation rates, in contrast, should be favoured by short RSI without being constrained by time-on-task, and without such strategic implications. Replicating the effect, manipulating these conditions, and comparing EAMs of the data, were undertaken to test these hypotheses and examine these implications.
In replication attempts using data from Mega Lexicon Projects (MLPs), the ABC effect was only observed in those MLPs with RSIs of at least 1700 ms, and relatively few pre-formal practice trials, and then only within their first 500 formal trials. Fitting the data to EAMs of lexical decision, model comparison favoured a choice-dynamic model with one-back target effects on threshold shifting and biasing, and on non-decision time, with the shifting effect constrained by RSI and time-on-task in the same way as seen for response time. This supported the explanation of the ABC effect by early threshold shifting, with a likely role in strategic development of SATO policy. Furthermore, in later trials (up to 20,000), threshold biasing with a cost-benefit pattern on choice speed became dominant, although both shifting and biasing occurred for initial word trials, suggesting parallel development of SATO policy and cost-benefit prediction. Reinforcement learning models further revealed that threshold biasing was informed by choice (base) probabilities rather than repetition probabilities, further implicating their production by response rather than stimulus processes.
A prospective study of the effects of time-on-task and RSI on threshold shifting was then made in a new experiment. The role of speed/accuracy-stress—which was variously constrained in the MLPs— was also examined, as was the difficulty of word and nonword decisions (per word frequency, and nonword base frequency). Uniformity of target and transition sequencing was controlled to account for an issue of target non-randomness that could be statistically identified within a subset of the MLPs and that could qualify interpretation of their sequential effects. The experiment involved 88 participants run through 6 × 160-trial blocks of a web-based lexical decision task, with three consecutive blocks per the within-subject speed- and accuracy-stress manipulation, and with between-subject RSIs of 1s and 2s. Results showed pervasive repetition effects for both choices— except for an ABC effect in the first block under the long 2s RSI, and only substantially under accuracy stress. This confirmed that the ABC effect depends on the temporal opportunity for volitional threshold control, in relatively unpracticed trials.
A further experiment examined what it is about the one-back trial that cues threshold shifting. The review, archival analyses, and the former experiment all pointed to a limited role of within-choice difficulty, and so of shifting per one-back ease of evaluation. A response-based explanation was, alternatively, rationalised, assuming a role for task-wide response and error-management biases between words and nonwords. The issue was tested by changing the lexical decision task by instruction into one of error detection where nonwords were assigned the “yes” response. It was hypothesised that this could flip the sign of the ABC effect between choices—so that caution is downshifted after a nonword response. This was tested in a web-based experiment involving three 160-trial blocks administered to 22 participants. Even as nonword responses remained slower than word responses, the ABC effect was no longer evident except that, within the per-participant scatter of effects across 40-trial sub-blocks, the early operation of a nonword repetition effect was detected in the presence of a small word alternation effect. A response-related, strategic basis to the ABC effect was therefore indicated in its sensitivity to task instruction independently of stimulus difficulty.
In summary, this research identified how prospective changes in decision thresholds produce sequential effects on performance in lexical decision. These effects were firstly dominated by dynamic shifts of overall decision thresholds based on the previous choice in a way that supports learning to optimize SATO. The initial ABC effect that resulted from threshold shifting quickly gave way to threshold biases according to local choice probability as a cost-benefit pattern involving higher-order sequences came to dominate performance. The change with time-on-task in the form of sequential effects can thereby signify the evolution of SATO policy, and a transition from more controlled prediction of overall evidence requirements to more automatic prediction of relative evidence requirements. These observations are likely to generalise to any task that, like lexical decision, involves a between-choice response bias and difference in evaluative difficulty.
History
Sub-type
- PhD Thesis