. 2020 Feb 10;3:59. doi: 10.1038/s42003-020-0786-7

Table 1.

Descriptive model comparison.

	Experiment	Comparison method	Model
	Experiment	Comparison method	Fixed bias (DDM drift rate change)	DDM start point change	DDM, start point and drift rate change	Unbounded, start point change
Prior-dependent	Discrimination	AICc	82	20	24	59
	Discrimination	BIC	85	23	26	61
	Discrimination MTurk	AICc	504	301	303	417
	Discrimination MTurk	BIC	637	432	437	551
	Detection	AICc	83	35	24	54
	Detection	BIC	90	42	31	61
Context-dependent	TI until-response	AICc	131	116	77	64
	TI until-response	BIC	140	125	86	72
	TI 200 ms	AICc	124	125	128	84
	TI 200 ms	BIC	133	134	137	92
	TI MTurk until-response	AICc	580	846	495	489
	TI MTurk until-response	BIC	742	1008	659	653
	TI MTurk 200 ms	AICc	589	793	562	519
	TI MTurk 200 ms	BIC	732	936	707	664
	TI MTurk mix	AICc	5083	6928	5006	4673
	TI MTurk mix	BIC	7406	9250	7333	7000
	TAE fixation	AICc	154	158	118	117
	TAE fixation	BIC	168	172	132	130
	TAE fixation Pinchuk et al.¹⁶	AICc	159	104	76	87
	TAE fixation Pinchuk et al.¹⁶	BIC	170	115	87	98
	TAE periphery	AICc	160	166	132	134
	TAE periphery	BIC	178	185	150	152
	TAE periphery non-retinotopic	AICc	148	122	116	128
	TAE periphery non-retinotopic	BIC	166	140	134	146

Shown are Akaike information criterion (AICc) and Bayesian information criterion (BIC) values which indicate how well a given model accounts for the measured behavioral bias compared to other models, when taking into account the number of fitted model parameters. Within a row, lower values indicate better model performance, where the standard rule-of-thumb is that a difference of at least ten can be interpreted as a “very strong” evidence in favor for the winning model⁴². Correspondingly, models having a difference of approximately ten from the winning model are in bold typeface. (Between rows, differences mostly reflect differences in the amount of data.) In all models, a single parameter was fit per observer, corresponding to the magnitude of the individual bias. In addition, a single group-level parameter was fit in two of the models: the “DDM, start point and drift rate change” model with the group level parameter indicating the percentage of the overall bias that is due to a starting point rather than a drift rate change; and the unbounded model where the group level parameter was the nondecision time, namely, t₀. Results show, for both prior-dependent and context-dependent bias, the validity of a model which assumes a change in the starting point of the process, as indicated by reduced AICc and BIC values relative to the “Fixed bias” null-hypothesis model.