Measuring unconscious cognition: Beyond the zero-awareness criterion

Visual masking can be employed to manipulate observers’ awareness of critical stimuli in studies of masked priming. This paper discusses two different lines of attack for establishing unconscious cognition in such experiments. Firstly, simple dissociations between direct measures (D) of visual awareness and indirect measures (I) of processing per se occur when I has some nonzero value while D is at chance level; the traditional requirement of zero awareness is necessary for this criterion only. In contrast, double dissociations occur when some experimental manipulation has opposite effects on I and D, for instance, increasing priming effects despite decreasing prime identification performance (Schmidt & Vorberg, 2006). Double dissociations require much weaker measurement assumptions than other criteria. An attractive alternative to this dissociation approach would be to use tasks that are known to violate necessary conditions of visual awareness altogether. In particular, it is argued here that priming effects in speeded pointing movements (Schmidt, Niehaus, & Nagel, 2006) occur in the absence of the recurrent processing that is often assumed to be a necessary condition for awareness (for instance, DiLollo, Enns, & Rensink, 2000; Lamme & Roelfsema, 2000). Feedforward tasks such as this might thus be used to measure the time-course of unconscious processing directly, before intracortical feedback and awareness come into play.

Paradoxically, this controversy does not so much concern the existence of unconscious processing (most researchers seem to be convinced of this) but the question how to demonstrate unconscious processing in a given experiment.
Progress in the field has been handicapped by the unquestioned assumption that in order to demonstrate unconscious processing, one has to make sure that a critical stimulus was completely outside of awareness. In this contribution, I would like to propose two alternative lines of attack for establishing unconscious processing beyond the zero-awareness criterion. The first part of the paper will deal with different types of dissociation between measures of awareness and measures of processing per se (Schmidt & Vorberg, 2006). The conclusion of this section is that even though different methods are available, the most pow- If these conditions be known, measures known to defy them could be used to measure unconscious processing directly. As an illustration, I will argue that priming effects in speeded pointing movements (Schmidt, Niehaus, & Nagel, 2006) occur in the absence of the recurrent processing that is often assumed to be a necessary condition for awareness (for instance, DiLollo, Enns, & Rensink, 2000;Lamme, 2002;Lamme & Roelfsema, 2000;Tong, 2003).

Simple dissociations and the zero-awareness criterion
To demonstrate that a critical stimulus was processed unconsciously, one usually has to produce some dissociation between different behavioral measures of performance concerning that stimulus. Traditionally, this is done by comparing two measures obtained from different tasks. 1 One measure (called the direct measure, D) is supposed to signal the observer's awareness of the critical stimulus, for instance, in a forced-choice prime discrimination task. The second measure (called indirect measure, I) is used as an indicator that the stimulus was processed at all, for instance, a priming effect in reaction times. The traditional criterion for unconscious processing has required D to equal zero, assuming that this signals the absence of any conscious processing of the critical stimulus. At the same time, I is required to be nonzero, indicating that the stimulus was nevertheless processed (Reingold & Merikle, 1988;Shanks & St. John, 1994). Historically, this zero-awareness criterion has run into difficulties because it only works if a valid conclusion can be drawn from zero performance in the direct measure to zero awareness in the observer (Reingold & Merikle, 1988, 1990, 1993Reingold, 2004).
Recently, Dirk Vorberg and I have examined the scopes and assumptions required by the zero-awareness criterion as well as alternative approaches (Schmidt & Vorberg, 2006). We start from the assumption that direct as well as indirect measures may depend on two sources of stimulus information which may be labeled "conscious" (c) and "unconscious" (u) without loss of generality: D = D (c, u), I = I(c, u), where information is defined non-negative. The dependency is supposed to be weakly monotonic, which means that if any type of information increases, the measures can only increase or remain constant (in the long run, that is, in the expected values). These are weak assumptions that must be conceded for virtually any measurement situation.
Establishing unconscious processing then consists in refuting a Null Model which states that the influence of unconscious information is zero, or equivalently, that both measures are driven by a single source of conscious information. If the null model is discarded, performance in the two tasks must be driven by at least one additional source of information.
There is one important constraint here. If D and I are to be modeled as functions of the same arguments c and u, one has to make sure that the underlying conscious and unconscious information is the same for both measures. Therefore, the direct and indirect tasks must be designed to use identical stimuli, identical responses, and identical stimulus-response mappings (Schmidt & Vorberg, 2006). In other words, D must address exactly that stimulus distinction that drives the effect in the indirect task (Reingold & Merikle, 1988). For an example of mismatch between direct and indirect tasks, consider the study by Dehaene et al. (1998). The indirect task was to indicate as quickly as possible whether a target digit was numerically smaller or larger than five, where the target digit was preceded by a masked prime digit. Response times in this task were shorter when the prime was consistent http://www.ac-psych.org Given that D-I mismatch is efficiently avoided, how can the null model of only conscious processing be disproved? The traditional way of doing this is the zero-awareness criterion, which produces what we call a simple dissociation of direct and indirect measures: zero D in the presence of nonzero I (Figure 1). If we start from this finding, we quickly see that we don't get very far without additional assumptions, because the observation that I(c, u) > 0 only implies that c > 0, u > 0, or both. Can we use the fact that D(c, u) = 0 to make sure that c = 0? Not quite, because D(c, u) = 0 does not imply c = 0 under weak monotonicity assumptions: D may simply fail to respond to changes in information, so that there could be some c that D was not able to detect. To work around this problem, we have to make the stronger assumption that D is an exhaustive measure of conscious information, that is, that D is a strictly monotonic function of c (Reingold & Merikle, 1988;see Schmidt & Vorberg, 2006, for a more general proof). This means that D is able to detect any change in c whatsoever, so that D(c, u) = 0 implies c = 0. Given this exhaustiveness assumption, we can finally use the fact that c can no longer drive the indirect effect: I(c, u) = I(0, u) > 0 implies u > 0, which says that there is nonzero unconscious information in the system. How restrictive is the exhaustiveness assumption?
It requires that no change in awareness, however small, must escape detection by D; only then can we infer the absence of awareness from zero values in the direct measure. You may compare this with your old mechanical barometer which is likely to be a weakly monotonic measure of atmospheric air pressure: The needle of the barometer tends to rise with air pressure, but it sometimes "hangs", and you have to knock against the shell to break the needle free. A strictly monotonic, exhaustive measure of air pressure would be an infinitely sensitive barometer, one that never hangs. Strict monotonicity is violated by conditions as trivial and inescapable as random error in the direct measure. The exhaustiveness assumption is thus a strong requirement that should not be taken for granted. If the exhaustiveness assumption is wrong, it can always be argued that it was conscious process-

Data patterns and assumptions necessary to interpret a simple dissociation as evidence for nonzero unconscious information.
An arrow from information source a to measure B indicates that B is some function of a. S-shaped inset symbols denote that weak monotonicity is assumed for that function. Abbreviations as explained in main text. a) Data pattern required for a simple dissociation. Direct and indirect measures are plotted in an opposition space in effect size units. Evidence for a simple dissociation is given by data points lying on the stippled vertical line such that I > 0 while D = 0. b) A simple dissociation gives evidence for nonzero unconscious information if it can be assumed that D is an exhaustive function of c and that I is a weakly monotonic function of u. c) Alternatively, a simple dissociation gives evidence for nonzero unconscious information if I is an exclusive measure of u. Adapted from Schmidt and Vorberg (2006). http://www.ac-psych.org

Thomas Schmidt
ing alone that influenced both D and I, but that I was sensitive enough to detect it while D was not (Reingold & Merikle, 1988).
There are some other difficulties with simple dissociations that are more on the practical side. One often stated problem is how to show statistically that D is not different from zero, because this involves "proving the null hypothesis", which is a commonplace problem in scientific research. Actually, the solution to this is straightforward and requires establishing binding criteria for effect, size, power, or confidence limits in the direct measure (Murphy & Myors, 1998). 2 However, given the conservativeness of applied statistics, this is unlikely to happen soon. Another practical problem is that finding stimulus conditions that will yield chance performance in the direct task is difficult, and largely a matter of good luck.
There is an alternative set of assumptions that abolishes the need for a direct measure altogether ( Fig. 1c). This is when the indirect measure can be assumed to be an exclusive measure of unconscious information, that is, a weakly monotonic function of u that is unaffected by c. In this case I(c, u) = I(u) > 0 implies u > 0 directly. 3 Tentative evidence for exclusive measures of unconscious processing is discussed later in this paper.

Beyond zero awareness I: Double dissociations
One interesting way to circumvent the exhaustiveness or exclusiveness assumptions is to let awareness vary over experimental conditions. It may then be possible to establish a double dissociation, which consists of Our concept of double dissociations is analogous to the widely used methodology in neuropsychology and medicine (Shallice, 1988;Sternberg, 2001).
Double dissociations have surprising features (see Schmidt & Vorberg, 2006, for details). Firstly, they re-quire D to be nonconstant: They cannot be obtained in the complete absence of awareness but require variation of awareness over a range of experimental conditions, so that there must be nonzero awareness for the prime under at least some conditions. Secondly, double dissociations require weaker assumptions than simple dissociations: There is no need for an exhaustiveness or an exclusiveness assumption, and we can even drop the assumption of weak monotonicity for all functions of u. Adopting the barometer metaphor from the last section, not only is the direct measure allowed to "hang" with respect to conscious information, but neither direct nor indirect measures have to be monotonically related to unconscious information at all. Because of this, c and u are allowed to produce arbitrary interactive effects on D and I like, for instance, when c and u are mutually inhibitory (Snodgrass, Bernat, & Shevrin, 2004;see Schmidt & Vorberg, 2006, for proof). The surprising outcome is thus that unconscious stimuli are not required for demonstrating unconscious processing.
In experiments by Schwarzbach (2003, 2004), participants performed speeded keypress responses to the direction of an arrow-shaped masking stimulus that was preceded by an arrow-shaped prime. The mask had a dual purpose here, acting as the target of the response and at the same time reducing visibility of the prime by metacontrast, a form of visual backward masking (Breitmeyer & Öğmen, 2006;Francis, 1997). As the stimulus-onset asynchrony (SOA) between prime and mask increased, priming effects also increased, such that primes pointing into the same direction as the mask shortened response times, while primes pointing into the opposite direction prolonged them. Strikingly, this priming effect was independent of visual awareness of the prime. We determined this by using stimulus conditions that produced different time-courses of metacontrast masking.
When a 17-ms prime was followed by a 140-ms mask, primes were virtually invisible, and participants were unable to perform better than chance when asked to discriminate the pointing direction of the prime (in over 3,000 trials per participant). These findings provide strong evidence for a simple dissociation as traditionhttp://www.ac-psych.org ally required. In a second experiment, however, we compared all four pairings of short-duration (14 ms) and long-duration (42 ms) primes and masks, yielding very different types of masking functions. When 14-ms primes were combined with 42-ms masks, prime identification performance was low and slightly increased with SOA; performance was better when mask duration was reduced to 14 ms. When a 42-ms prime was paired with a 14-ms mask, prime identification performance was nearly perfect. However, a 42-ms prime combined with a 14-ms mask yielded an effect called type-B masking (Breitmeyer & Öğmen, 2006), where prime identification performance markedly decreases with the primemask SOA, then increases again. Still, the priming effect increased monotonically, producing a strong double dissociation between priming and prime identification performance. These data defy the claim that direct and indirect measures tend to convey similar amounts of information about the critical stimulus (Franz, 2006): Priming increased linearly with SOA no matter whether the prime was completely visible, completely invisible, or whether visibility increased or decreased with SOA.
Clearly, this data pattern reveals a relationship that would never have been found by simple dissociation: Response priming is independent of prime identification

Beyond zero awareness II: Violating necessary conditions for awareness
Dissociations at the task level are able to provide only indirect evidence for a distinction of underlying processes. An exciting alternative would be to work out the necessary conditions for visual awareness and then try to find behavioral tasks that violate those conditions.
In other words, we could search for indirect measures that are exclusive measures of unconscious processing.
In a metaanalysis of 48 studies investigating the response latencies of various cortical areas to a sudden visual stimulus, Lamme and Roelfsema (2000) showed that the stimulus creates a wave of activation by primes and masks before recurrent processes set in to gradually wipe out the prime signal before it enters visual awareness. The feedforward processes associated with priming should therefore be independent of the recurrent processes leading to visual awareness and backward masking, which is nicely compatible with the double-dissociation findings by Vorberg et al. (2003Vorberg et al. ( , 2004 and Mattler (2003) as well as the abundant evidence from simple dissociations.
To convincingly link response priming to feedforward processing, one has to show that visuomotor activation is not only transmitted rapidly, but that the dynamics of this transmission are consistent with a feedforward process. Evidence for rapid visuomotor transmission in masked priming comes from the study of primed pointing responses (Schmidt, 2002; see also Brenner & Smeets, 2004). In that study, participants were   Schmidt (2002) and Schmidt et al. (2006). b) Stimulus conditions were varied by crossing two levels of color contrast (for primes and masks together) with two types of masks ("strong" metacontrast masks or "weak" pseudomasks Results clearly showed that responses were con- can be very rapid (Bullier, 2001;Girard, Hupé, & Bullier, 2001), and there are many possibilities for signals processed in parallel visual streams to cross or overtake each other well before the first overt signs of motor activation (Merigan & Maunsell, 1993). It is therefore worthwhile to step back a bit and focus at the input-output dynamics of the system as a whole instead of claiming purely feedforward processing in all its subcomponents. To do this, we introduced the concept of a rapid chase (Schmidt et al., 2006). This concept applies to visuomotor tasks where sequential visual stimuli run for control of the same speeded motor response -for instance, when a pointing response is initiated by one stimulus and then altered in mid-flight by an immediately following stimulus (Brenner & Smeets, 2004;Schmidt, 2002). By our definition, two successive visuomotor signals are said to be engaged in a rapid chase if

a) Experimental tasks and procedures employed by
(1) the response is initiated by the first stimulus, (2) the response is influenced by the second stimulus before it is completed, and (3) the response to the first stimulus is initially independent of the second stimulus.
These rapid-chase criteria say that if successive signals are in a rapid chase, they will take strictly successive control over the same motor response, such that the response will initially be controlled by the first stimulus alone. The third criterion is crucial because it demands sequential stimulus signals to exert strictly sequential response control. We adopted the pointing task used by Schmidt This finding provides independent behavioral evidence for an early phase of visuomotor processing that is at least primarily if not entirely feedforward (Lamme & Roelfsema, 2000)  It is still unclear how far the recurrent-processing hypothesis will actually carry, and for the time being, the dissociation approach is probably still the safer bet.
But measurement theory can only take us so far. In order to use dissociations for demonstrating unconscious processing, one has to work from the assumption that the "conscious/unconscious" distinction is valid in the first place. All that dissociations can teach us is that a single source of information is not sufficient to explain the data, including a single source of "conscious" information. But the insight that there must be at least two sources does not by itself imply that one of them is unconscious: There might be two dissociable types of conscious (or unconscious) information. One reviewer of this article asked whether this wouldn't render the search for double dissociations a fruitless exercise -if true, of course, this conclusion would hold for simple as well as double dissociations. However, the validity of the "conscious/unconscious" distinction is a conceptual issue that is simply beyond the scope of the measurement-theoretical arguments presented by Schmidt and Vorberg (2006). Whether or not the concept of unconscious processing will stand the test of time or be replaced by a different concept must be the outcome, not the starting point, of scientific investigation.
Dissociations at the measurement level provide tools for performing this investigation, and our analysis only shows which of these tools will work best in the widest range of situations.

Notes
1 For formal proofs and definitions, please refer to Schmidt and Vorberg (2006), especially the mathematical appendix.
2 The regression technique advocated by Greenwald and coworkers (Draine & Greenwald, 1998;Greenwald, Klinger, & Schuh, 1995) is an alternative method for demonstrating simple dissociations that has been strongly criticized on methodological and conceptual grounds (Dosher, 1998;Merikle & Reingold, 1998;Miller, 2000;Schmidt & Vorberg, 2006) and requires all the assumptions of the conventional simple-dissociation logic in addition to the statistical ones introduced by the regression methodology. Arguably, then, this method is worse than the original approach.
3 This exclusiveness assumption must not be confused with the one stated by Reingold and Merikle (1988).
These authors propose that simple dissociations can be interpreted as evidence for unconscious processing only if the direct measure is both exhaustive and exclusive for conscious information. The latter requirement would be highly problematic because unconscious processing is probably a ubiquitous precursor to conscious processing. Fortunately, from our analysis, it is sufficient that D be exhaustive for c or that I be exclusive for u -and note that these are alternative sets of assumptions that do not have to be met at the same time. Also note that the proofs given in our paper (Schmidt & Vorberg, 2006) are more general than those in Reingold and Merikle (1988) because they do not assume additivity of conscious and unconscious effects. ergy of the prime (for instance, its contrast, intensity, or duration) to control its visibility. Such manipulations should be avoided because they affect priming and visual awareness of the prime in similar ways. In particular, any disappearance of priming effects as prime energy is reduced does not constitute logically valid evidence against unconscious processing, because it cuts the necessary input to conscious and unconscious processes alike. 6 Note that cells left in the immediate wake of the feedforward sweep may pick up feedback very rapidly. The claim here is not that the entire system is feedbackfree for the duration of the feedforward sweep, but that the wavefront of the sweep travels just ahead of rapidly developing recurrent processes. 7 Reanalysis of the data reported in Schmidt (2002) confirmed these results, showing that the findings by Schmidt et al. (2006) are not acciden-tal.
http://www.ac-psych.org account allows for local recurrent activity as long as sequential signals still lead to strictly sequential motor output. Therefore, rapid chases suggest but do not logically imply feedforward processing.