Absolute and relative blindsight.

Consciousness and Cognition xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Consciousness and Cognition journal homepage: www.elsevier.com/locate/concog

Absolute and relative blindsight Tarryn Balsdon 1, Paul Azzopardi ⇑ Department of Experimental Psychology, University of Oxford, Oxford OX1 3UD, UK

a r t i c l e

i n f o

Article history: Received 7 January 2014 Revised 11 September 2014 Accepted 12 September 2014 Available online xxxx Keywords: Blindsight Conscious awareness Signal detection Metaconstrast masking Type 2 blindsight Confidence ratings Low vision

a b s t r a c t The concept of relative blindsight, referring to a difference in conscious awareness between conditions otherwise matched for performance, was introduced by Lau and Passingham (2006) as a way of identifying the neural correlates of consciousness (NCC) in fMRI experiments. By analogy, absolute blindsight refers to a difference between performance and awareness regardless of whether it is possible to match performance across conditions. Here, we address the question of whether relative and absolute blindsight in normal observers can be accounted for by response bias. In our replication of Lau and Passingham’s experiment, the relative blindsight effect was abolished when performance was assessed by means of a bias-free 2AFC task or when the criterion for awareness was varied. Furthermore, there was no evidence of either relative or absolute blindsight when both performance and awareness were assessed with bias-free measures derived from confidence ratings using signal detection theory. This suggests that both relative and absolute blindsight in normal observers amount to no more than variations in response bias in the assessment of performance and awareness. Consideration of the properties of psychometric functions reveals a number of ways in which relative and absolute blindsight could arise trivially and elucidates a basis for the distinction between Type 1 and Type 2 blindsight. Ó 2014 Published by Elsevier Inc.

1. Introduction In the study of conscious awareness, the possibility of identifying a neural correlate of consciousness (NCC) is regarded as one of the more achievable objectives and therefore has considerable appeal. The NCC refers to ‘the minimal set of neural events and mechanisms jointly sufficient for a specific conscious percept’ (Koch, 2004). In theory, there is a simple test to identify a NCC: Compare neural activity across conditions that differ only in the observer’s conscious awareness. This would throw light on important questions, such as, how much processing is needed for awareness, whether it depends on global or local processes, and whether there are different forms of awareness. In order to compare conditions differing only in awareness, it is necessary to dissociate performance from awareness. There are many neuropsychological conditions, such as blindsight, amnesia, achromatopsia, propospagnosia, anosognosia, and neglect, that are associated with this kind of dissociation. In each case, patients exhibit some residual ability to perform tasks in which they are ordinarily impaired. For example, amnesic patients are able to perform well above chance in forcedchoice recognition (e.g. Voss, Baym, & Paller, 2008) despite claiming to have no memory for the stimuli being tested. In such cases, subtractive designs could be used to identify the NCC. ⇑ Corresponding author at: Department of Experimental Psychology, University of Oxford, South Parks Rd., Oxford OX1 3UD, UK. Fax: +44 1865 310447. 1

E-mail address: [email protected] (P. Azzopardi). Current address: School of Social Sciences and Psychology, University of Western Sydney, Locked Bag 1797, Penrith 2751, New South Wales, Australia.

http://dx.doi.org/10.1016/j.concog.2014.09.010 1053-8100/Ó 2014 Published by Elsevier Inc.

Please cite this article in press as: Balsdon, T., & Azzopardi, P. Absolute and relative blindsight. Consciousness and Cognition (2014), http:// dx.doi.org/10.1016/j.concog.2014.09.010

2

T. Balsdon, P. Azzopardi / Consciousness and Cognition xxx (2014) xxx–xxx

Of all such examples, blindsight seems to be the most clear-cut. Blindsight has been defined as ‘visual capacity in a field defect in the absence of acknowledged awareness’ (Weiskrantz, 1986), reflecting the fact that patients who are clinically blind are capable of detecting and discriminating stimuli presented in their field defects when forced choice procedures are used. Because the extent of the field defect is often restricted to part of the visual field (e.g. hemianopia and quadrantanopia), it is possible to make straightforward, within-subject comparisons of conditions with and without awareness. On the other hand, the rarity of blindsight patients makes them less convenient in the study of the NCC. Many researchers have tried to emulate blindsight in normal observers by manipulating the detectability of visual stimuli using paradigms such as backward masking, but unfortunately it has proven difficult to produce dissociations between performance and awareness that are not entirely due to differences in response bias over forced-choice and yes–no tests (see Heeks & Azzopardi, this issue, for example). An interesting approach was devised by Lau and Passingham (2006), who showed a relative difference in the level of conscious awareness of normal observers between two conditions in which the performance was otherwise matched, which they referred to as ‘relative blindsight’. By implication, ‘absolute blindsight’ refers to a straightforward dissociation between performance and awareness regardless of whether performance is matched across conditions. Lau and Passingham used a metacontrast masking procedure in which performance followed a U-shaped function, i.e. decreasing and then increasing as a function of the temporal asynchrony between target and mask, and took advantage of this to identify two stimulus onset asynchronies (SOAs) at which percent correct detection was matched, yet where there was a difference between them in the rate of reported awareness (Fig. 1). By comparing neural activity in these two conditions by means of fMRI, Lau and Passingham found activity in the dorsolateral prefrontal cortex (area 46) to correlate with awareness. This finding is in agreement with a previous finding from blindsight, that area 46 forms at least part of the NCC (Sahraie, Weiskrantz, Barbur, Simmons, & Williams, 1997). That this approach reveals dissociations in performance and awareness between conditions otherwise matched for performance makes it particularly attractive in the context of fMRI, where differences in performance can be associated with different baselines of activity. Despite the elegance of this approach, a number of potential issues have since been identified (Jannati & Di Lollo, 2012). First, the stimulus actually looks different at shorter SOAs than at longer SOAs due to the relatively poor temporal resolution 100

80

%

60

40

% Correct % Seen

RELATIVE BLINDSIGHT

20

0 -50

-33

-17

0

17

33

50

67

83

100

117

133

SOA (ms)

(a) Awareness

Performance 100

100

P = 0.007 *

P = 0.213 ns 80

% Seen

% Correct

80

60

40

60

40

20

20

0

0 33

104

33

104

SOA (ms)

SOA (ms)

(b)

(c)

Fig. 1. Relative blindsight (replotted from Lau and Passingham (2006)). (a) Percent scores for performance (% Correct) and awareness (% Seen) as a function of stimulus onset asynchrony (SOA) in a detection task with metacontrast masking. Relative blindsight is demonstrated between two SOAs in which performance is the same, but awareness is different. (b) There is no significant difference in performance between the shorter and the longer SOA. (c) There is a significant difference in awareness between the shorter and longer SOA.



3

of the visual system. (In their terminology, due to Kahnemann, 1968, the stimuli at the two critical SOAs have different ‘criterion contents). Secondly, without explicitly ensuring that the stimulus presentations were time-locked to the screen refresh signal of the display, there is a possibility that small variations in the time taken to draw the stimulus, relative to the start of the trial, could have had a disproportionate effect on shorter SOAs, increasing the observers’ uncertainty over whether a stimulus had been presented or not. Such technical problems should be relatively straightforward to overcome. A more pressing concern is the question of what should be considered appropriate measures of performance and awareness. Lau and Passingham reported comparing percent correct 2AFC scores for performance (‘Was the target a square or a diamond?’) and percent correct yes–no scores for awareness (‘Did you see the target?’). There are two issues to be aware of here. The first is that when measured with percent correct scores, dissociations between performance and awareness can be accounted for trivially by response bias, which can affect yes–no scores but not 2AFC scores (see Azzopardi & Cowey, 1998). The second issue, and more pertinent to the interpretation of Lau and Passingham’s experiment, is that the test that was used to assess performance was in fact not a 2AFC, in which both target and distractor were presented in separate intervals and the observer was required to indicated the interval in which the target was presented (‘In which interval, first or second, was the square presented?’), but rather a yes–no question in which either the target or the distractor was presented and the observer was required to identify which one had been presented (‘Was it a square (yes) or a diamond (no)?’). The distinction is crucially important because 2AFC scores are inherently unbiased (any bias towards an interval is not reflected as bias towards one stimulus or another) whereas yes–no scores are prone to bias. This raises the possibility that at the critical SOAs, chosen by Lau and Passingham on the basis of having identical percent yes–no detection scores, the stimuli were not equally detectable but merely appeared to be so due to different amounts of response bias at the two SOAs. Jannati and di Lollo (2012) demonstrated that response bias in the yes–no question used to measure awareness also could be a factor contributing to relative blindsight. It is up to the observer to set their own criterion for what constitutes a ‘seen’ stimulus, but the criterion can be influenced by the instructions given to observers. Lau and Passingham instructed observers to respond ‘seen’ when the stimulus was ‘‘actually consciously perceived’’, but Jannati and Di Lollo found that changing the instructions to respond ‘seen’ ‘‘if you think you saw the target’’ abolished the dissociation altogether. In this paper we describe an experiment designed to test whether ‘relative blindsight’ can be induced when response bias in the assessment of performance and awareness is eliminated using bias-free measures of performance derived from signaldetection theory (Green & Swets, 1966). 2. Methods Participants were required to detect geometric visual targets during yes–no and forced-choice tasks using a metacontrast masking paradigm in a replication of Lau and Passingham’s task. The experiment was approved by the University of Oxford’s Central University Research Ethics Committee. 2.1. Participants 13 student volunteers aged 18–22 took part in the experiment. 3 were excluded for not finishing the experiment, and 3 were excluded as they reported being able to perform the task on the basis of afterimages. 2.2. Stimuli The stimuli were based on those used by Lau and Passingham (2006). The targets were white squares (2° 2°) or diamonds (the same squares rotated by 45°) masked by a large disk (12° diameter) with a shape cut out of the centre corresponding to the shape of the two targets superimposed. The stimuli were produced by means of a ViSaGe stimulus generator (Cambridge Research Systems, Rochester) using custom software and displayed on a gamma-corrected 2400 CRT monitor (Mitsubishi Diamond pro 20708B) with 12 bits of greylevel resolution per gun, and a refresh rate of 100 Hz enabling a minimum stimulus presentation of 10 ms. The apparatus was calibrated by means of a photodiode and an oscilloscope to ensure that the stimuli and mask were presented at the specified times and for the specified durations. Preliminary tests revealed that at the presentation times reported by Lau and Passingham (target duration 33 ms, SOAs of 33 ms and 104 ms, and mask duration of 50 ms) the target was either clearly visible or produced visible afterimages. The stimulus duration was therefore reduced to 10 ms, the mask duration increased to 70 ms, and the contrast of the targets reduced to 0.05 in order to reduce visibility and minimise the possibility of afterimages. The difficulty of the task was controlled by varying the SOA between 20 ms and 110 ms in steps of 10 ms. 2.3. Procedure Prior to the commencement of the experiment, the method of constant stimuli was used in conjunction with a 2AFC task to plot bias-free psychometric functions from which the SOAs corresponding to 4 levels of difficulty (50%, 65%, 80% and 90% correct) were determined for each participant. This was also treated as practice. The main experiment consisted of two


4


blocks of trials per observer: one yes–no task, and one 2AFC. Each block consisted of 320 trials, 80 for each level of difficulty randomly interleaved. In the yes–no task, one of the stimuli (either a square or a diamond, selected at random) was presented with equal probability, followed after the specified SOA by the mask and then a tone indicating the end of the trial. Observers were instructed to respond ‘yes’ if they saw a square, and ‘no’ if they did not. In the 2AFC task, each stimulus was presented sequentially, i.e. square – SOA – mask – tone, followed by diamond – SOA – mask – tone (or vice versa) in randomized order, and observers were instructed to respond ‘1’ if the square was presented in the first interval, or ‘2’ if it was presented in the second interval. In both tasks, observers were also required to rate their confidence in their response on a scale of 1–4, where ‘1’ indicated ‘very low confidence’/‘do not know’ and ‘4’ indicated ‘high confidence’/‘certainty’. (Intermediate ratings were not defined, leaving observers free to set their own criteria for these categories. The implications of this are discussed in Section 4.3.) Responses were given verbally and typed into a computer by the experimenter who sat behind a screen out of sight of the observer. Eye movements were monitored via closed-circuit TV, and a trial was aborted (with replacement) if fixation was not maintained for its duration. 2.4. Analysis Percent correct scores were transformed using the arcsine transformation as appropriate for proportional data for analysis with parametric statistical tests. Paired sample t-tests were used for pair-wise comparisons. Where the data did not meet all of the distributional assumptions pertaining to ANOVA, non-parametric rank-sum tests (Meddis, 1984) were performed on the untransformed scores. 3. Results 3.1. Replication of relative blindsight with yes–no percent correct scores The scores were first analysed in the manner of Lau and Passingham (2006) by comparing performance (percent correct) in the yes–no task with awareness (percent seen) in the yes–no task. In their experiment, Lau and Passingham derived percent seen from the proportion of responses ‘seen’ rather than ‘guessed’, whereas in this experiment, ‘seen’ (awareness) was defined as responses with confidence ratings greater than 2, and ‘performance’ as percent correct scores regardless of confidence ratings. The data are plotted in Fig. 2a. The graph shows a clear dissociation between performance and awareness at every level of difficulty, analogous to ‘absolute blindsight’. A 2-way factorial analysis of variance using rank sum tests, with level of difficulty and task (performance or awareness) as treatments and observers as blocks, revealed a significant effect of level of difficulty (X2 = 12.75, d.f. = 1, p < .001) and a significant effect of task (X2 = 29.91, d.f. = 1, p < .001), but no significant interaction (X2 = 0.683, d.f. = 5, ns). It is also the case that no significant difference in performance between the third and fourth levels of difficulty (t = 1.196, d.f. = 6, p = .277; Paired sample t-test on arcsin transformed scores; Fig. 2b) was accompanied by a significant difference in percent ‘seen’ between the third and fourth levels (t = 2.674, p = .037; Fig. 2c), conforming to Lau and Passingham’s criteria for ‘relative’ blindsight. The results of Jannati and Di Lollo (2012) were also replicated: When a more liberal definition of ‘seen’ was used (responses with a confidence rating greater than 1), there was no significant difference in percent ‘seen’ between the third and fourth levels of difficulty (t = 2.262, p = .064, ns), showing that ‘relative blindsight’ is susceptible to response bias. 3.2. Absolute blindsight and response bias Confidence ratings reflect different levels of response bias (see, e.g. Macmillan & Creelman, 2005). It is therefore instructive to compare the effects of setting different thresholds of confidence on the level of reported awareness. In Fig. 3 are plotted percent correct yes–no scores corresponding to four possible definitions of ‘seeing’ derived from the four confidence ratings used in this experiment, that is, from the most liberal to the most conservative, all yes–no responses, all yes–no responses with a confidence level >1, all yes–no responses with a confidence level >2, all yes–no responses with a confidence level >3. Evidently, the degree of ‘absolute blindsight’ depends very much on the confidence rating taken as the threshold for awareness. 3.3. Comparison of yes–no and 2AFC percent correct responses The diagnosis of ‘absolute blindsight’ effectively boils down to a difference in percent correct scores between yes–no and 2AFC tests (Azzopardi & Cowey, 1998). The corresponding test for ‘relative blindsight’ would be a difference in percent correct yes–no scores between two conditions in which there was no corresponding difference in 2AFC scores. Lau and Passingham (2006) did not actually show this because their test of performance was not a 2AFC test after all. A comparison of percent correct yes–no scores with percent correct 2AFC scores from our experiment is shown in Fig. 4. A 2-way factorial analysis of variance using rank sum tests, with level of difficulty and test (yes–no or 2AFC) as treatments and


5


Comparison of YN detection and awareness yn (all confidence ratings) yn (confidence ratings > 2)

100

80

%

60

40

20

0

RELATIVE BLINDSIGHT 1

2

3

4

Level of difficulty

(a) Awareness

Performance 100

80

80

% Correct

% Correct

P = 0.277 ns 100

60 40

P = 0.037 * 60 40 20

20

0

0 3

4

Level of difficulty

(b)

3

4

Level of difficulty

(c)

Fig. 2. Replication of relative blindsight using by comparing yes–no performance with yes–no awareness. (a) Mean percent correct scores of performance (yes–no scores) and awareness (yes–no scores with confidence rating >2). Relative blindsight is seen between levels 3 and 4, where there is no pair-wise significant difference in performance scores between levels (b) but there is a significant difference in awareness scores between them. Bars indicate 95% confidence limits.

observers as blocks, revealed a significant effect of level of difficulty (X2 = 38.96 v = 1, p < .0001) but no significant effect of test (X2 = 0.368, v = 1, ns), and no significant interaction (X2 = 1.818, v = 5, ns). Thus there is no evidence for ‘absolute blindsight’ consistent with the yes–no scores being based on the most liberal criterion (confidence ratings of 1 or more). (Note, as shown in Fig. 3, there would have been had more a conservative definition of awareness been used.) There was no pair-wise comparison of 2AFC scores among the set of levels of difficulty in the experiment that was not significantly different and therefore there was no pair of levels at which it was possible to show ‘relative blindsight’ over the range of the SOAs tested. The fact that the ‘relative blindsight’ effect revealed by comparing yes–no performance with yes–no awareness in Levels 3 and 4 (Fig. 2) is no longer evident when bias-free 2AFC scores are used as the measure of performance suggests that the effect must have been due to response bias associated with the yes–no measure of performance. 3.4. Signal detection analysis If blindsight can be attributed to response bias then the effect should disappear when bias-free measures of both performance and awareness are used. In order to test this, observers’ confidence ratings were used to construct ROCs (graphs of hit rate against false alarm rate) using a maximum-likelihood algorithm (Dorfman & Alf, 1969) from which the statistic da was obtained. da is equivalent to d0 , corrected for the possibility of unequal variances in the underlying signal and noise Please cite this article in press as: Balsdon, T., & Azzopardi, P. Absolute and relative blindsight. Consciousness and Cognition (2014), http:// dx.doi.org/10.1016/j.concog.2014.09.010

6


100

All responses

80

Confidence > 1

%

60 Confidence > 2 40 Confidence > 3

20 0 -20 1

2

3

4

Level of difficulty Fig. 3. Showing how varying confidence can affect the extent of absolute blindsight. The curves represent the proportion of stimuli ‘seen’ according to which confidence rating is taken as the threshold for seeing. The extent of blindsight, denoted by the difference between performance (All responses) and awareness (Confidence >1, or 2, or 3), clearly depends on which confidence rating is taken as the threshold for awareness.

2AFC % Correct YN % Seen

100

%

80

60

40

20

0 1

2

3

4

Level of difficulty Fig. 4. Comparison of 2afc performance with yes–no awareness, using the most liberal definition of awareness (confidence rating 1). Bars indicate 95% confidence limits.

2AFC yes-no

4

da

3

2

1

1

2

3

4

Level of difficulty Fig. 5. Signal detection analysis. (a) Mean da scores of performance (2AFC) and awareness (yes–no) derived from ROCs (not shown). There are no pairwise significant differences in (b) 2AFC scores (c) yes–no scores between levels 3 and 4. Bars indicate 95% confidence limits.

p distributions (Green & Swets, 1966; Simpson & Fitter, 1973). A factor of 1/ 2 correction was also applied to the 2AFC values as required to compensate for the fact that the forced choice task allows two opportunities to detect the stimulus (Macmillan & Creelman, 2005). The data are presented in Fig. 5. A 2-way factorial analysis of variance using rank sum tests, with level of



7

difficulty and test (yes–no or 2AFC) as treatments and observers as blocks, revealed a highly significant effect of level of difficulty (X2 = 32.038, v = 1, p < .0001) but no significant of effect of test (X2 = 0.107, v = 1, p = .371) and no significant interaction (X2 = 5.47, v = 5, ns). Therefore, across the whole experiment, there is no evidence for a dissociation between 2AFC and yes–no detection using a bias free measure: in other words no evidence for absolute blindsight. Individual psychometric functions are plotted in Fig. 6. Pair-wise comparisons of da values for 2AFC and yes–no detection for every combination of observer and level of difficulty, incorporating a Bonferroni correction, only yielded one instance out of 28 comparisons in which da(2AFC) was significantly greater than da(yes–no) at an experiment-wise level of p < .05. 4. Discussion 4.1. Relative blindsight We were able to replicate Lau and Passingham’s (2006) results showing relative blindsight by comparing percent correct scores in a yes–no test of performance with percent seen scores in a yes–no test of awareness. Given that percent correct scores in yes–no tests can vary with the observer’s response bias, it could be argued that relative blindsight is nothing more than the manifestation of a difference in response bias between two conditions. This is supported by the fact that varying the threshold for the definition of ‘seeing’ can abolish the dissociation (Jannati & Di Lollo, 2012, and the present paper). Lau and Passingham reported using a 2AFC task to measure the performance against which ‘seeing’ could be compared. The advantage of 2AFC tests is that they do not depend on the observer setting a response criterion: the observer simply indicates the interval in which the signal was stronger, meaning that the decision threshold is inherently unbiased, and therefore percent correct 2AFC scores are not affected by response bias. But, in fact, Lau and Passingham measured performance using a yes–no task: Rather than presenting each alternative in separate intervals and asking the observer to identify the interval in which a particular stimulus was selected (i.e. a bias-free, 2AFC test), they presented one or the other stimulus and instructed the observers to indicate which was presented. This leaves open the possibility that the percent correct scores used to identify two different conditions may nevertheless have reflected unequal performance in the two conditions if the response bias associated with the two conditions differed. In that case any difference in fMRI activation between the two conditions would have been correlated with a difference in performance (masked by response bias), which is precisely what the concept of ‘relative blindsight’ was invented to avoid. This interpretation is supported by the fact that the ‘relative blindsight’ that we showed between levels 3 and 4 using a yes–no measure of performance (Fig. 2), evaporated when unbiased 2AFC measure of performance were used instead. This suggests that ‘relative blindsight’ reflected response bias associated with the yes–no measure of performance. Another potential problem that affects the assessment of relative blindsight is that comparing percent correct scores obtained from unbiased 2AFC tests with percent correct scores from potentially biased yes–no tests leaves open the possibility that the dissociation comes down entirely to a difference in response bias between the two tasks. This is exactly the problem faced when trying to interpret blindsight. To circumvent it Azzopardi and Cowey (1997, 1998) compared 2AFC detection and yes–no detection in a blindsighted patient using bias free measures derived from signal detection theory (Green & Swets, 1966), and found d0 (yes–no) was significantly lower than d0 (2AFC), implying that not all of the dissociation between performance and awareness in blindsight can be accounted for by differences in response bias. In other words, there could be more to blindsight than response bias (Azzopardi & Cowey, 2001). When d0 was used to assess performance and awareness in the present experiment, there was no significant difference between 2AFC performance and yes–no awareness, and no evidence for relative blindsight. These results tell us two important things: First, that dissociations between performance and awareness as measured with percent correct scores can be explained entirely by the difference in the response bias in yes–no and 2AFC tasks; and second, because of that, the dissociations found here are different from the dissociations found in blindsight. Thus, the assumption that studying normal observers with absolute- or relative- blindsight will reveal the same things about the NCC as studying patients with blindsight is not supported. 4.2. Relative blindsight and response bias The role of detectability and response bias in relative blindsight can be illustrated with reference to pairs of psychometric functions. Psychometric functions are graphs of performance versus stimulus intensity which typically have a sigmoid shape reflecting the way in which performance usually increases as the strength of the stimulus increases (Fig. 7). The position of the mid-point of the curve is called the point of subjective equality (PSE) and is taken to correspond to the absolute threshold. The slope of the curve relates to the difference threshold (the just noticeable difference or JND). The psychometric functions of any two stimuli may vary in their PSE (absolute threshold) or in their slope (difference threshold), or both. The PSE (the position of the curve in relation to the x-axis) is affected both by the detectability of the stimulus and by the observer’s response bias (unless a bias-free measure is used). For example, when scoring a yes–no task with a percentage score which is susceptible to response bias, a conservative response bias would push the psychometric function to the right (the PSE would be higher) relative to unbiased responding (Fig. 7a), whereas a liberal response bias would push the curve to the left (the PSE would be lower).


8

T. Balsdon, P. Azzopardi / Consciousness and Cognition xxx (2014) xxx–xxx 8

10

PT03

6

8

4

6

da

da

PT02

2

4

0

2

2-

0 -2

-4 0

20

40

60

80

0

100

20

40

60

80

100

120

20

40

60

80

100

120

5

2.0

PT04

PT06 4

1.5

3

1.0

da

da

2 0.5

1 0.0 0 -0.5

-1

-1.0

-2 0

20

40

60

80

100

0

18

20 PT08

PT07

16 14

15

12 10

8

da

da

10

6

5

4 2

0

0 -2

-5

-4 0

20

40

60

80

100

120

20

30

40

50

60

70

80

90

100

14

PT10

12 10

da

8 6

yes-no 2AFC

4 2 0 -2 10

20

30

40

50

60

70

80

90

t (msec; but see Method) Fig. 6. Psychometric functions of da plotted against SOA for individual observers. The levels of difficulty (SOA) were calibrated individually. Bars indicate 95% confidence limits.


Performance


9

0.5

T1

T2

Stimulus strength

Performance

(a)

0.75

0.5

JND1 JND2 Stimulus strength

(b) Fig. 7. Psychometric functions. (a) A pair of psychometric functions which have different PSEs (absolute thresholds) and identical slopes (difference thresholds). The PSE is determined by the salience of the stimulus and, if assessed using a bias-prone measure, by the observer’s response bias, in which case a conservative response bias would displace the curve to the right (higher absolute threshold) and a liberal response bias would displace the curve to the left (lower absolute threshold). (b) A pair of psychometric functions with identical PSEs (absolute thresholds) and dissimilar slopes (difference thresholds). The absolute and difference thresholds are independent.

If a bias-free measure of performance were used, then two equally detectable stimuli (having the same PSE) with the same difference thresholds (slope), would have identical psychometric functions, and it would not be possible to identify any combination of stimulus intensities, I1 and I2, that would result in relative blindsight (Fig. 8a). It is therefore a requirement that the two stimuli must differ in some respect, either their PSE or their slope (or both), before relative blindsight can be produced. One way in which a difference could arise between the PSEs of a pair of stimuli which had identical absolute and difference thresholds (as determined with a bias-free measure), is if the detectability of at least one of them was assessed with a measure of performance that was susceptible to bias, so that the curve associated with it was shifted to the right. In that case, it should be possible to identify a combination of stimulus intensities, I1 and I2, that would produce ‘relative blindsight’ (Fig. 8b). Lau and Passingham (2006) argued that ‘relative blindsight’ in their experiment could not possibly be accounted for by difference in the response criteria between the two conditions (different SOAs), as the trials were intermixed and an observer would be unlikely to know which of two different criteria to apply in a given trial if the stimuli were equally detectable. The results of the current experiment suggest that the stimuli they presented were not in fact equally detectable, but even if they had been, their argument would only apply if the stimuli had both the same absolute thresholds and the same difference thresholds. If the absolute thresholds were the same, and the difference thresholds differed, then irrespective of response bias it should be possible to identify a combination of stimulus intensities, I1 and I2, that would produce ‘relative blindsight’ (Fig. 8c) which, in the context of identifying the NCC, is trivial. Thus, relative blindsight could arise trivially in a number of ways: a difference in the objective PSEs between two conditions (which would look like relative blindsight if the difference was masked by response bias in the preliminary stage of identifying equally salient conditions); a difference in response bias between conditions with the same objective PSEs;



Performance

10

P2

P2

P1

P1

0.5

0.5

T I1

I2

T1 T2 I1

(a) P2

P2

P1

P1

Performance

I2

(b)

0.5

0.5

T

I1

I2

T1 T2

I1

Stimulus intensity

Stimulus intensity

(c)

(d)

I2

Fig. 8. Relative blindsight can arise from differences in the PSEs and slopes of the psychometric functions associated with a pair of stimuli or with the same stimulus assessed in two different conditions. (a) In this example, the two stimuli or conditions, S1 (solid curve) and S2 (dotted curve), have psychometric functions with identical PSEs and slopes, in which case there is no possible pair of stimulus intensities, I1 and I2, common to S1 and S2 that is compatible with relative blindsight. (b) S1 and S2 have psychometric functions with the same slope but different PSEs, T1 and T2. In this case it is possible to identify pairs of stimulus intensities, e.g. I1 and I2, common to S1 and S2 that give rise to relative blindsight. Note that the difference between PSEs could be due either to a difference between the salience of the stimuli as measured with a bias-free metric, or to S2 being associated with a more conservative response bias than S1 when salience is assessed with a bias-prone metric. (c) S1 and S2 have psychometric functions with identical PSEs and different slopes. In this case, it is possible to identify pairs of stimulus intensities common to both S1 and S2 that trivially give rise to relative blindsight, even if bias-free measures of performance are used. (d) S1 and S2 have psychometric functions with different PSEs and different slopes. In this case it is possible to identify pairs of intensities common to both conditions that give rise to relative blindsight that is due to the combination of difference in slope (as in c) and difference in response bias (as in b). A similar analysis applies to the distinction between Type 1 and Type 2 blindsight.

and a difference in difference thresholds between the two conditions (regardless of whether or not they have the same PSEs). All these possibilities would have to be ruled out before relative blindsight can be established. 4.3. Confidence ratings and response bias Our results highlight the fact that confidence ratings can be used in different ways to produce different results. First, we used the different degrees of confidence as thresholds for awareness, i.e. all yes–no responses, all yes–no responses with a confidence level >1, all yes–no responses with a confidence level >2, all yes–no responses with a confidence level >3, and showed that the degree of absolute blindsight depended on the threshold level of confidence taken as the definition of awareness (Fig. 3). It is easy to see why this should be the case. Fig. 9a depicts a psychometric function associated with a stimulus assessed with a bias-free measure (e.g. 2AFC), accompanied by a series of psychometric functions associated with successively more conservative (i.e. higher) confidence rating thresholds (from left to right) assessed with a bias-prone metric (e.g. proportion of yes–no responses). Taking the most liberal definition of awareness, i.e. a confidence rating of 1 or more, ensures that the psychometric function is relatively close to the bias-free psychometric function, whereas taking the most conservative definition of awareness, i.e. confidence rating of 4, ensures that the psychometric function is shifted as far to the right as possible. As a result, for a stimulus of fixed intensity (e.g. I1 in Fig. 9a), there is a striking difference in performance between the two conditions that could be taken as evidence of blindsight. The second way in which confidence ratings were used takes advantage of the fact that the observers’ choice of confidence rating on a particular trial reflects their response criterion (or response bias) on that trial, to produce ROCs (plots of hit rate versus false alarm rate over a range of response criteria) from which a bias-free measure (d0 ) can be derived using signal detection theory (Fig. 9b). Using the same raw data, this method showed there was no dissociation between performance measured as d0 (2AFC) and awareness measured as d0 (yes/no) (Figs. 5 and 6). Thus, the same set of confidence rating data can be used to show the presence or absence of blindsight depending on how they are analysed. Please cite this article in press as: Balsdon, T., & Azzopardi, P. Absolute and relative blindsight. Consciousness and Cognition (2014), http:// dx.doi.org/10.1016/j.concog.2014.09.010


11

Performance

P0 P1 P2 0.5 P3 P4 OT ST1 ST2 ST3 ST4 I1

Stimulus intensity

(a) 1.0

Hit rate

P0 P1 0.5

P2 P3 P4

0.0 0.0

0.5

1.0

False alarm rate

(b) Fig. 9. Blindsight can arise from response bias associated with confidence ratings. (a) The solid curve depicts the psychometric function associated with a stimulus having a PSE and slope assessed with a bias free metric. (OT = objective threshold). The dotted curves depict the psychometric functions associated with successively more conservative (i.e. lower) confidence rating thresholds (from left to right) assessed with a bias-prone metric, analogous to the data presented in Fig. 3. (ST1–ST4 = subjective thresholds.) Almost any stimulus intensity (e.g. I1) will yield a dissociation between performance (P0) and awareness (P1–P4). The PSEs of the psychometric functions associated with the different confidence ratings, and hence the corresponding measures of performance, will be influenced by the instructions given to the observer and by the way the observer interprets them. (b) An ROC curve used to compute a bias-free measure of performance from confidence ratings assigned to a stimulus of fixed intensity (e.g. I1 in Fig. 9a). In this hypothetical example, the performance levels corresponding to the different confidence ratings are plotted as hit rate versus false alarm rate pairs which fall on a curve corresponding to d0 = 1.0. Response bias varies with position along the curve from conservative (bottom left) to liberal (top right), but signal detection theory provides an unbiased measure of performance, d0 , that is constant regardless of the inclination of the observer to favour more or less confident ratings.

In the present experiment the observers were given minimal instructions as to what meaning should be assigned to each confidence rating, other than that 1 indicated ‘very low confidence’/‘do not know’ and ‘4’ indicated ‘high confidence’/‘certainty’, and there were no controls as to whether the observers actually adopted these criteria or substituted them with their own. Nor was there any attempt to influence how observers assigned responses to the intermediate categories. In other words the observers were given as much freedom as possible, consistent with experimenters’ need to convey what the task entailed, to adopt their own criteria for the different ratings. When analysed by calculating the proportion of confidence ratings above and below a specified confidence rating threshold, the scores will depend on the criteria adopted by the observer as well as on the salience of the stimulus. Such scores will also be affected by another source of bias, which is the observers’ inclination to assign one rating rather than another irrespective of the salience of the stimulus and regardless of how the categories are defined. Methods that attempt to standardise the definitions of the categories, such as the Perceptual Awareness Scale devised by Ramsøy and Overgaard (2004), may minimise observers’ response bias due to the placement of category boundaries but they do not prevent response bias due to the observer’s preference of some ratings over others, and as a result, different labels and different observers will produce variations in scores that are independent of their ability to detect the stimuli. The important thing to note about signal detection analysis, in contrast, is that not only does it produce a measure of performance, d0 , that is impervious to variations in response bias arising for any reason, including those discussed here, but in fact the correct way to determine d0 relies on plotting ROCs by measuring performance associated with a range of different response biases, and taking advantage of the range of response criteria associated with confidence ratings to do so is far more efficient than, say, varying the proportions of targets and distractors across multiple blocks of trials (Macmillan & Creelman, 2005). Please cite this article in press as: Balsdon, T., & Azzopardi, P. Absolute and relative blindsight. Consciousness and Cognition (2014), http:// dx.doi.org/10.1016/j.concog.2014.09.010

12


4.4. Relative blindsight and the NCC If ‘relative blindsight’, at least by Lau and Passingham’s measures, can be explained by response bias, then the question remains, what does Lau and Passingham’s (2006) finding of the correlation between activity in area 46 and awareness really mean? Considering the dissociation they measured was likely not in awareness, but in response bias, it is possible that activity in the dorsolateral prefrontal cortex correlates with response bias. This is supported by Azzopardi and Cowey’s (2002) finding of a correlation between the prefrontal cortex and response bias when response bias was manipulated by requiring observers to guess on trials when a target was not detected. 4.5. Type 1 and Type 2 blindsight Despite his original definition of blindsight as ‘visual capacity in a field defect in the absence of acknowledged awareness’, Weiskrantz subsequently recognised that there were certain conditions under which patients were able to report some kind of awareness in their field defects, and therefore drew a distinction between Type 1 blindsight (visual capacity in the absence of acknowledged awareness) and Type 2 blindsight (visual capacity accompanied by some form of awareness) (Weiskrantz, Barbur, & Sahraie, 1995). Sahraie et al. (1997) took advantage of this distinction to compare brain activity associated with Type 1 blindsight (where patients report absolutely no ‘sensation, feeling, or experience of the visual event’) and Type 2 blindsight (where patients report some kind of awareness) and found neural activity in prefrontal cortical area 46 that correlated with awareness. However the measures used were the same as Lau and Passingham, namely subjective reports of ‘awareness’ and % correct in a yes–no task, which does not rule out the possibility that the difference in brain activation between Type 1 and Type 2 blindsight is also a matter of response bias (Azzopardi & Cowey, 2002). There has been considerable debate as to whether the distinction between Type 1 and Type 2 blindsight is meaningful, which rest on doubts about whether any blindsighted patient could be fully unaware of the stimuli which they are capable of discriminating within their field defects (Azzopardi & Cowey, 1998; Zeki & ffytche, 1998; plus articles in this issue). A consideration of the psychometric functions associated with ‘unaware’ and ‘aware’ conditions, analogous to that which applies to relative blindsight (Fig. 8), suggests a possible resolution. Blindsight – a dissociation between forced-choice performance and yes–no awareness – arises because the PSE of the psychometric function associated with yes–no awareness is shifted to the right of the PSE of that associated with forced-choice performance (e.g. Fig. 8b compared to Fig. 8a, or Fig. 8d compared to Fig. 8c, where yes–no awareness is depicted by the dotted line). The shift could be due to a difference in the detectability of the target, evident when bias-free measures of performance and awareness are used, (such as a difference between d0 (2AFC) and d0 (yes–no)), or to a difference in response bias, evident when bias-prone measures are used (such as a difference between bias-free percent correct 2AFC and bias-prone percent correct yes–no scores), or to a combination of both. As long as there is some difference between the PSEs or slopes of the psychometric functions associated with measures of performance and measures of awareness, it should in theory always be possible to show Type 1 and Type 2 blindsight with certain pairs of stimuli as illustrated in Fig. 8c and d. It is worth remembering that ‘Type 1’ and ‘Type 2’ are merely labels attached to different patterns of behaviour that do not necessarily map independently onto two distinct mechanisms. The important question is how the difference between Type I and Type 2 blindsight arises. In a two dimensional space defined by variation in sensitivity and variation in response bias, blindsight could be due to a difference in sensitivity, or response bias, or both. This was investigated explicitly by Azzopardi and Cowey (1997, 1998), who compared 2AFC and yes–no responding to static and moving targets in a blindsighted patient, and found that d0 (2AFC) was significantly greater that d0 (yes–no) for static targets, but not for moving targets, implying that not all of the dissociation between performance and awareness with static targets could be accounted for by response bias. In other words, any distinction between Type 1 and Type 2 blindsight in the patient’s responses to moving targets could only arise from response bias, but any distinction between Type 1 and Type 2 blindsight in his responses to static targets could be due to differences in sensitivity as well as bias. Azzopardi and Cowey (2001) subsequently showed that the reduction in yes–no sensitivity to static targets relative to 2AFC sensitivity, contributing to Type 1 blindsight, could be accounted for by the patient’s inability to maintain a stable response criterion during yes–no responding. 4.6. Conclusion We have shown, in a replication of Lau and Passingham’s experiment, that it is not possible to induce either relative blindsight or absolute blindsight in normal observers once response bias in the assessment of performance and awareness was eliminated by using bias-free measures (d0 (2AFC) and d0 (yes–no)). Consideration of the possible relations between psychometric functions representing pairs of conditions suggests several ways in which either type of blindsight could arise trivially in normal observers that bear no relation to impaired awareness, and also suggests a basis for distinguishing the mechanisms underlying Type 1 and Type 2 blindsight. The lesson is that in order to diagnose blindsight, and to understand how it arises, it is essential to compare the psychometric functions associated with the conditions being contrasted. References Azzopardi, P., & Cowey, A. (1997). Is blindsight like normal, near-threshold vision? Proceedings of the National Academy of Sciences of the United States of America.



13

Azzopardi, P., & Cowey, A. (1998). Blindsight and visual awareness. Consciousness and Cognition, 7, 292–311. Azzopardi, P., & Cowey, A. (2002). Cerebral activity related to guessing and attention during a visual detection task. Cortex, 38, 833–836. Azzopardi, P., & Cowey, A. (2001). Why is blindsight blind? In B. de Gelder, E. de Haan, & C. Heywood (Eds.), Varieties of unconscious processing: New findings and new models (pp. 3–19). Oxford: O.U.P. Dorfman, D. D., & Alf, E. Jr., (1969). Maximum likelihood estimation of parameters of signal detection theory and determination of confidence intervalsrating-method data. Journal of Mathematical Psychology, 6, 487–496. Green, D., & Swets, J. (1966). Signal detection theory and psychophysics. New York: Wiley. Heeks, F., & Azzopardi, P. (2014). Thresholds for detection and awareness of masked facial stimuli. Consciousness and Cognition (this issue). Jannati, A., & Di Lollo, V. (2012). Relative blindsight arises from a criterion confound in metacontrast masking: Implications for theories of consciousness. Consciousness and Cognition, 21, 307–314. Kahnemann, D. (1968). Methods, findings, and theory in studies of visual masking. Psychological Bulletin, 70, 404–425. Koch, C. (2004). The quest for consciousness: A neuroscientific approach. Colorado: Roberts & Co. Lau, H. C., & Passingham, R. E. (2006). Relative blindsight in normal observers and the neural correlate of visual consciousness. Proceedings of the National Academy of Sciences of the United States of America, 103, 18763–19768. Macmillan, N., & Creelman, D. (2005). Signal detection theory: A user’s guide (2nd ed.). Cambridge: Cambridge University Press. Meddis, R. (1984). Statistics using ranks: A unified approach. Oxford: Blackwell Publications. Ramsøy, T. Z., & Overgaard, M. (2004). Introspection and subliminal perception. Phenomenology and the Cognitive Sciences, 3(1), 1–23. Sahraie, A., Weiskrantz, L., Barbur, J. L., Simmons, A., & Williams, S. C. R. (1997). Pattern of neuronal activity associated with conscious and unconscious processing of visual signals. Proceedings of the National Academy of Sciences of the United States of America, 94, 9406–9411. Simpson, A., & Fitter, M. (1973). What is the best index of detectability? Psychological Bulletin, 80, 481–488. Voss, J. L., Baym, C. L., & Paller, K. A. (2008). Accurate forced-choice recognition without awareness of memory retrieval. Learning & Memory, 15, 454–459. Weiskrantz, L. (1986). Blindsight: A case study and its implications. Oxford: Oxford University. Weiskrantz, L., Barbur, J. L., & Sahraie, A. (1995). Parameters affecting conscious versus unconscious visual discrimination with damage to the visual cortex (V1). Proceedings of the National Academy of Sciences, 92, 6122–6126. Zeki, S., & ffytche, D. H. (1998). The Riddoch syndrome: Insights into the neurobiology of conscious vision. Brain, 121, 25–45.


Sensitivity to relative and absolute motion.

Blindsight: recent and historical controversies on the blindness of blindsight.

Relative versus absolute quantitation in disease glycomics.

Relative errors can cue absolute visuomotor mappings.

Measuring health disparities: a comparison of absolute and relative disparities.

Absolute and relative pitch: Global versus local processing of chords.

Absolute and relative quantification of RNA modifications via biosynthetic isotopomers.

The absolute disparity anomaly and the mechanism of relative disparities.

Symposium: Contraindications to tympanoplasty. I. Absolute and relative contraindications.

Absolute and Relative Socioeconomic Health Inequalities across Age Groups.

Transfer of absolute and relative predictiveness in human contingency learning.

Relative or absolute? A significant intervention for chlamydia screening with small absolute benefit.

Absolute or Relative Jogging Pace: What Makes the Difference?

Absolute or relative effects? Arm-based synthesis of trial data.

JPEN Journal Club 14. Relative vs Absolute Efficacy.

Choice deferral can arise from absolute evaluations or relative comparisons.

Income and Well-Being: Relative Income and Absolute Income Weaken Negative Emotion, but Only Relative Income Improves Positive Emotion.

U.S. internal migration and occupational attainment: Assessing absolute and relative outcomes by region and race.

Entrenched geographical and socioeconomic disparities in child mortality: trends in absolute and relative inequalities in Cambodia.

Absolute and Relative Training Load and Its Relation to Fatigue in Football.

Absolute and relative reliability of percentage of syllables stuttered and severity rating scales.

From blindsight to blindsmell: a mini review.

Affective Blindsight relies on Low Spatial Frequencies.

Relative and absolute availability of healthier food and beverage alternatives across communities in the United States.