OUP user menu

Impaired filtering of behaviourally irrelevant visual information in dyslexia

Neil W. Roach, John H. Hogben
DOI: http://dx.doi.org/10.1093/brain/awl353 771-785 First published online: 19 January 2007

Summary

A recent proposal suggests that dyslexic individuals suffer from attentional deficiencies, which impair the ability to selectively process incoming visual information. To investigate this possibility, we employed a spatial cueing procedure in conjunction with a single fixation visual search task measuring thresholds for discriminating the orientation of a target stimulus. Replicating preliminary findings in an earlier report, we found evidence of a striking dissociation between dyslexic participants’ performance in cued and uncued conditions. Whereas uncued search results were equivalent for dyslexic and normal adult readers, the majority of dyslexic individuals failed to display a comparable benefit when the location of the target was indicated by the appearance of a brief peripheral pre-cue. Using receiver operating characteristic curve analysis, we further demonstrate that the effectiveness of the cueing task at discriminating between dyslexic and normal readers surpasses that of a range of other psychophysical tasks typically used in dyslexia research. Moreover, we find that the discriminative accuracy of the task is at least on par with measures of verbal short-term memory (a core component of phonological processing), which ranks as one of the most widely accepted areas of difficulty in dyslexia. Potential mechanisms underlying the cueing effect are outlined, and the plausibility of each considered within a signal detection theory framework of visual search. It is argued that performance benefits obtained by normal readers in cued conditions most likely reflect the prioritization of target information during decision making, and could feasibly be subserved by top-down biasing effects on pooling processes in extrastriate cortex.

  • dyslexia
  • vision
  • visual attention
  • attention deficit
  • visual cortex

Introduction

The notion that abnormal sensory processing might contribute to the reading difficulties experienced by people with dyslexia has fuelled debate for many years. A variety of deficits have been linked to the disorder, including abnormal binocular stability and vergence control (e.g. Stein and Fowler, 1980, 1981), faulty eye movements (e.g. Pavlidis, 1981a, b), abnormalities of the magnocellular retinocortical and/or dorsal extrastriate pathways of the visual system (e.g. Stein and Walsh, 1997; Stein, 2001) and specific impairments affecting the processing of auditory temporal sequences and spectral information (e.g. Tallal, 1980; McAnally and Stein, 1996). However, strong evidence for any of these proposed sensory deficits has proved elusive. Particularly problematic for the field is that poor performance is typically confined to a small minority of dyslexic individuals, and patterns of difficulties across different tasks are often inconsistent and fail to point to a common underlying cause. As a result, most extant aetiological theories of dyslexia focus almost exclusively on the role of linguistic skills in the reading process, such as the ability to understand the phonological structure of language, and to retain and retrieve speech sounds (Wagner and Torgesen, 1987; Wolf and Bowers, 1999).

Recently, it has been suggested that dyslexic individuals might exhibit deficiencies not in encoding sensory information per se, but rather in selecting such information in an effective manner (e.g. Facoetti and Molteni, 2001; Vidyasagar, 2004). The need for selectivity stems from the notion that processing and reporting sensory information is subject to some form of capacity limitation. In the case of the visual system, it is clear that although the retinae continuously encode vast amounts of information, only a subset is able to reach visual awareness at any one time. To facilitate effective interaction with the environment, it is essential that the most behaviourally relevant aspects of the visual scene can be selected or prioritized ahead of less relevant information. This selective process is typically termed ‘visual attention’ and undoubtedly plays a critical role in the reading process.

In a recent study (Roach and Hogben, 2004), we examined visual attention in dyslexic adults using spatial cueing in conjunction with a single fixation visual search task. Participants were required to discriminate the orientation of a tilted target stimulus, presented in one of 16 possible locations equidistant from fixation. The target was presented either alone, or along with a variable number of vertical distractors. On some trials a peripheral pre-cue was presented, which always indicated the location of the target. Both the cue and the search array were presented rapidly, precluding any eye movements during stimulus presentation. Normal readers’ uncued search performance was characterized by a strong dependence on the number of elements in the stimulus array, with orientation discrimination thresholds increasing sharply as more distractors were added. Cueing the location of the target dramatically reduced the detrimental effect of the distractors on task performance. Given that restricting search to a single fixation effectively equates the basic sensory representation of the stimulus array for cued and uncued conditions, we can be confident that this cueing advantage reflects the operation of selective visual attention. While the cueing effects shown by normal readers were both large and robust, a different picture emerged when the task was administered to a small group of dyslexic adults. Performance in the uncued condition was in no way different from normal readers, suggesting that they were not impaired at making orientation discrimination judgements. However, unlike normal readers, cueing the location of the target produced little or no effect on performance.

This experimental approach seems to offer a number of advantages over many other paradigms for investigating potential attentional impairments in dyslexia. First, the task offers more rigorous control over non-attentional factors that might affect performance. Second, our preliminary results point to a difference between dyslexic and normal readers of substantial magnitude. Third, the robustness of the effects permits analysis at the level of the individual observer. Finally, dyslexic participants’ difficulties appear to be specific to the cued condition. The fact that dyslexic individuals perform just as well as controls on uncued search is particularly advantageous, as it effectively rules out the possibility that dyslexics have more general problems meeting task demands (e.g. Roach et al., 2004). In the present study, we sought to establish the replicability of the spatial cueing deficit in a large sample of dyslexic adults. In addition, we aimed to address the following questions: (i) how effective is the cueing task at discriminating between dyslexic and normal readers compared to other non-reading measures? (ii) What form of mechanism underlies the success or failure of attentional selection in the cueing effect?

Discriminative accuracy

While most studies investigating perceptual functioning in dyslexia tend to focus exclusively on group-based comparisons, it is becoming increasingly clear that analysis of differences at the level of the individual is necessary. Tasks that are able to successfully discriminate between dyslexic and control individuals provide the best opportunity for guiding us towards a coherent understanding of reading disabilities. Our preliminary results suggest that the spatial cueing task may be quite efficient in this respect. However, in order to properly gauge its discriminative accuracy, it is necessary that we be able to compare the task with other non-reading measures. Here we provide two forms of comparison. First, we contrast the cueing task to other measures of perceptual functioning typically used in the field: flicker contrast sensitivity (e.g. Martin and Lovegrove, 1987, 1988; Evans et al., 1994), global dot motion detection (e.g. Cornelissen et al., 1995; Hansen et al., 2001), auditory frequency discrimination (e.g. McAnally and Stein, 1996; Ahissar et al., 2000) and auditory frequency modulation detection (e.g. Witton et al., 1998, 2002). Since support for perceptual deficits in dyslexia remains equivocal, we also chose to compare the cueing task with measures of verbal short-term memory. Problems relating to the retention of verbally presented material are, perhaps, the most commonly reported difficulties in dyslexia and are widely taken as evidence for a core phonological deficit (Snowling, 2000).

Mechanism of attentional selection

The facilitatory effects of attention in cueing tasks have been attributed to a variety of causes. Some authors have suggested that attentional selection operates at early stages of visual processing, either by affecting the relative speed at which attended and unattended stimuli are processed (e.g. Hikosaka et al., 1993; Carrasco and McElree, 2001), or by modulating the quality of processing inside and/or outside the focus of attention (e.g. Lu and Dosher, 2000; Carrasco et al., 2002). Others maintain that attention does not have any effect on visual processing per se, but instead influences the way in which decisions are formed based on the available information (e.g. Shaw, 1980; Shiu and Pashler, 1994, 1995; Eckstein et al., 2002, 2004). The difference between these forms of explanation is analogous to the traditional distinction made between early selection theories of attention, epitomised by Broadbent's (1958) seminal filter model, and late selection theories such as those put forward by Deutsch and Deutsch (1963) and Duncan (1980).

In order to consider what type of mechanism underlies the success and or failure of attentional filtering in the cueing effect, it is essential first to have a working concept of how distractors reduce target discriminability. Here we favour the application of signal detection theory (SDT, Swets, 1964; Green and Swets, 1966) to visual search. SDT is the standard theoretical model used to predict simple psychophysical thresholds, but can also be applied to more complex multi-element search displays. According to this approach, each stimulus element is independently analysed by the visual system, regardless of set size. Importantly though, it is assumed that the resulting internal representations of stimulus information are inherently noisy. For an orientation search task, this means that the perceived tilts of the target and distractors will vary somewhat from trial to trial, even if their physical values are kept constant. As a result, on any given trial, there is a possibility that the perceived tilt of one or more vertical distractors will be as large, or larger than that of the tilted target. In SDT models, there is a distinct decision stage in which a simple rule is implemented, such as choosing the largest response or summing the response across all locations. According to this approach, search performance declines as distractors are added to the display simply because each additional distractor increases the amount of decisional uncertainty. Models based on SDT have been used to successfully predict set size effects for tasks measuring accuracy and threshold values for targets defined by a number of dimensions, including orientation, luminance, colour, size and speed (e.g. Palmer et al., 1993, 2000; Verghese and Stone, 1995; Morgan et al., 1998; Baldassi and Burr, 2000; Verghese, 2001; Baldassi and Verghese, 2002). For present purposes, a particular advantage of this approach is that SDT models are neutral with respect to how attentional selection operates, allowing unbiased consideration of both early and late mechanisms.

Material and methods

Participants

Criteria for inclusion in the dyslexic group required both (i) a lifelong history of specific reading difficulties and/or prior diagnosis and (ii) impaired phonological decoding, defined as a standard score of below 85 on the Phonemic Decoding subtest of the Test of Word Reading Efficiency (TOWRE, Torgesen et al., 1999). Control participants were required to have no history of reading difficulties and a TOWRE Phonemic Decoding standard score greater or equal to 94. Additionally, all participants were required to have a performance IQ of at least 85 [Kaufman Brief Intelligence Test (KBIT); Kaufman and Kaufman, 1990], English as a first and primary language and normal or corrected to normal visual acuity. Adults (112) were recruited from poster and local press advertising, and pre-tested on each of the selection measures. Of these, 72 individuals met the selection criteria for one or the other groups resulting in a sample of 35 controls and 37 dyslexic participants. Note, results of the psychometric tests and additional psychophysical tasks for some participants have previously been published as part of the study reported by Heath et al. (2006).

Standardized psychometric tests

Participants were measured on a series of standardized tests in order to provide profiles of ability levels in reading, spelling, and phonological short-term memory. All were administered and scored according to test instructions.

Reading

Regular word reading was indexed using the Sight Word Reading subtest of the TOWRE (Form A). This is a speeded test, requiring participants to read out loud as many words as possible from a graded list of 104 real words in a fixed 45-s interval. Phonological decoding was assessed using both speeded and unspeeded tests of non-word reading: the Phonemic Decoding subtest of the TOWRE (63 graded non-words, 45 s duration) and the Word Attack subtest (form G) of the Woodcock Reading Mastery Tests (45 graded non-words, untimed, Woodcock, 1987).

Spelling

Spelling ability was measured using the Spelling subtest (Tan form) from the 3rd edition of the Wide Range Achievement Tests (WRAT-3, Wilkinson, 1993).

Verbal short-term memory

Three measures requiring the short-term retention of verbally presented information were administered: the Nonword Repetition and Memory for Digits subtests of the Comprehensive Test of Phonological Processing (CTOPP, Wagner et al., 1999) and the Memory for Sentences subtest from the 4th edition of the Stanford Binet Intelligence Scale (Thorndike et al., 1986). These tests required participants to listen to and repeat graded sequences of multisyllablic non-words, lists of single digit numbers and sentences respectively.

Cued visual search

Visual stimuli were preloaded onto the framestore section of a Cambridge Research Systems VSG 2/3 card and displayed on a gamma corrected Sony Multiscan 20SE monitor refreshed at 100 Hz. The dimensions of the display were 328 × 242 mm2, resulting in a pixel size of ∼1.1 arcmin when displayed at 1024 × 768 resolution at a distance of 100 cm. Each element in the search array was a Gabor patch, comprising a simple sinusoidal grating convolved by a Gaussian envelope. Stimuli were presented against a mean grey background with a luminance of 23.5 cd/m2. Participants were required to identify the orientation of a target Gabor patch that was tilted either clockwise or counter-clockwise from vertical. The target was presented on every trial for 100 ms, either alone or accompanied by 1, 3, 7 or 15 vertical distractor Gabors (i.e. set sizes of 1, 2, 4, 8 and 16). Carrier spatial frequency (2 c/degree), envelope SD (0.25°) and contrast (50% Michelson) parameters were identical for target and distractor stimuli. The target could appear at one of 16 positions located 5° from a central fixation cross (width/height = 11 arcmin). As shown in Fig. 1A, distractor stimuli were positioned regularly around this circular configuration maintaining a constant interstimulus separation for each set size condition (∼1.91 ° with 16 stimuli). In cued conditions, a small circular black dot (diameter = 11 arcmin, luminance < 1 cd/m2) was presented at 4° along the imaginary line linking the fixation cross to target location. The cue was presented for 20 ms immediately prior to the array of Gabors (ISI = 0). Each combination of set size and cue condition was tested in a block of 80 trials. In each block, the tilt of the target stimulus was controlled using an adaptive PEST routine (Taylor and Creelman, 1967) set to converge on 75% correct performance and thresholds were estimated by taking the mean of all stimulus levels following the third reversal point. The order of conditions was randomized for each participant.

Fig. 1

(A) Examples of search arrays for set sizes of 4 and 16. In each case, all distractors are vertically oriented while the target is tilted 45° clockwise. To illustrate the spatial relationship between the cue and the target, the two frames are shown superimposed upon one another in the right-hand panel. (B) Mean orientation discrimination thresholds for each group, plotted as a function of set size. Note, both the ordinate and abscissa are logarithmically scaled. Unfilled symbols denote performance in cued conditions whereas filled symbols represent uncued performance. Dotted lines show the best linear fits to the log-transformed data. Error bars indicate the 95% confidence interval around the mean.

Additional psychophysical tasks

Participants were also measured on two additional visual (flicker contrast sensitivity, global dot motion) and two auditory (frequency discrimination, frequency modulation detection) psychophysical tasks. Stimuli for the two visual tasks were produced using the same apparatus as described earlier. Auditory stimuli were digitally generated and controlled using a Tucker-Davis Technologies (TDT) System 2 and presented to participants in a sound attenuated room via Sennheiser HD 265 headphones. For each task thresholds were estimated using a similar PEST routine to that described earlier.

Contrast sensitivity

Contrast sensitivity was measured for detection of a large flickering Gaussian blob (SD = 3.15°). Flicker was produced by modulating the luminance contrast of the blob sinusoidally at 10 Hz. A two-interval forced choice design was employed, whereby participants were required to select which of two 1000 ms temporal intervals (marked by brief tones) contained the flickering stimulus. The non-target interval contained a uniform field with the same mean luminance as the target (20 cd/m2). The depth of contrast modulation of the stimulus was manipulated over the course of a 60 trial block.

Global dot motion

Ability to extract a global motion signal was estimated using 20-frame global dot motion sequences. The initial frame in each sequence consisted of 100 circular dots (diameter = 6.6 arcmin; luminance = 47.7 cd/m2), randomly distributed on a dark background (<1 cd/m2; dot density = 0.42 dots/deg2). On each frame transition dots were displaced by 11.4 arcmin: signal dots were shifted in a coherent direction (upwards or downwards); noise dots were shifted in random directions. Signal dots were reselected on each frame transition to minimize local motion cues. Additionally, a spatial proximity rule prevented dots from overlapping and reduced the possibility of false dot pairings across frames. Each stimulus frame was presented for 30 ms, resulting in a dot speed of 6.33°/s. Blocks of 80 trials were used, in which the percentage of signal dots was adaptively manipulated. Participants were required to indicate the direction of the coherent signal, which was randomly selected as up or down on each trial.

Frequency discrimination

Sensitivity to differences in pure tone frequency was measured against a 1000 Hz standard. An AXB design was used in which three 100-ms tones were presented sequentially (85 dB SPL, 10 ms rise/fall, ISI = 300 ms). The frequency of the first tone (A) was the fixed standard, while the frequency of the third tone (B) was the standard plus a variable increment. On each trial, the participant was required to indicate whether the middle tone (X) was the same frequency as either the first tone or third tone. The frequency increment was adaptively controlled over the course of 70 trials.

Frequency modulation detection

Detection of auditory frequency modulation was measured using a 2-interval forced choice design. Pairs of 500 ms tones (81 db SPL, 10 ms rise/fall, ISI = 500 ms) were presented: a pure 500 Hz standard and a test stimulus undergoing sinusoidal frequency modulation at 2 Hz around a 500 Hz mean. Participants were required to indicate which of the two tones sounded ‘warbled’. The depth of frequency modulation was varied during a 70 trial block to threshold.

Results

A comparison of demographic and psychometric scores between the two groups is presented in Table 1. As would be expected, on average the dyslexic group was significantly poorer than controls on all measures of reading and spelling. The groups did not differ significantly in chronological age, but there was a small but statistically significant difference in KBIT performance IQ scores. Significant group differences were found for each of the three verbal short-term memory measures, reflecting poorer mean performance in the dyslexic group compared with controls.

View this table:
Table 1

Group statistics for demographic and psychometric variables

Control Mean (SD)Dyslexic Mean (SD)
Age (years)32.91 (8.36)35.43 (9.77)t (70) = −1.17, P > 0.05
KBIT matrices111.97 (8.19)105.32 (8.99)t (70) = 3.28, P < 0.05
TOWRE phonemic decoding107.40 (8.78)72.89 (8.76)t (70) = 16.69, P < 0.05
TOWRE sight word efficiency99.63 (10.89)81.78 (9.40)t (70) = 7.46, P < 0.05
WRMT word attack111.60 (7.84)88.46 (9.05)t (70) = 11.56, P < 0.05
WRAT-3 spelling112.31 (7.36)87.41 (14.90)t (70) = 8.91, P < 0.05
CTOPP memory for digits (scaled score)12.31(2.61)8.62 (3.31)t (70) = 5.24, P < 0.05
CTOPP non-word repetition (scaled score)6.91 (1.68)5.00 (1.72)t (69) = 4.74, P < 0.05
Binet memory for sentences (scaled score)52.69 (7.52)42.81 (8.66)t (70) = 5.15, P < 0.05
  • Unless otherwise stated, numbers represent standard scores (M = 100, SD = 15). Note, CTOPP non-word repetition data for one control participant were lost as the result of an equipment malfunction.

Cued visual search

Mean group data for the spatial cueing task is shown in Fig. 1B. Performance in the uncued conditions was similar for the two groups. Orientation discrimination thresholds show a strong positive relationship with set size, approximating a steep linear function when plotted on log–log axes. In the control group spatial cueing produced a substantial reduction in the set size effect, producing a linear fit with a relatively shallow gradient. However, the effect of cueing was far less pronounced in the dyslexic group.

To test whether this pattern of results was statistically significant, separate linear regressions were run on the log-transformed thresholds for each participant, and the gradients entered into an analysis of variance (ANOVA). Data for two participants (one from each group) were excluded from the analysis because of the poorness of their linear fits. A 2-way mixed ANOVA was employed, with one within subject factor (cue condition) and one between subject factor. A significant interaction between group and cue condition was found [F(1,67) = 5.61, P < 0.05], reflecting the presence of a significant simple effect of group in the cued condition [MControl = 0.37, MDyslexic = 0.47, F(1,67) = 4.88, P < 0.05] but not in the uncued condition [MControl = 0.62, MDyslexic = 0.60, F(1,67) = 0.25, P > 0.05].

While group based comparisons of this form provide a useful starting point for data analysis, they tell us little about the extent to which each task discriminates between individuals. To explore this further we concentrated on the largest set size, as group differences are most striking in this condition. Figure 2A plots individual cued and uncued thresholds for set size 16, which have been log-transformed and standardized against the control distribution. As might be expected given the similarity of the group means, the uncued condition failed to discriminate between dyslexic and control individuals. Far greater separation is evident in the cued condition, though there is some overlap between the scores of individuals from each group.

Fig. 2

(A) Scatter plot of cued and uncued search performance for the largest set size (16). Thresholds are expressed as standard (z) scores relative to the distribution of log-transformed thresholds in the control group. The dotted line indicates 1 SD above the mean for the cued condition. (B) Receiver operating characteristic (ROC) curve for cued and uncued conditions, showing the trade-off between sensitivity and specificity for a range of cut-off points.

A common practice is to define deficient performance as falling more than one SD above the control mean. Using this criterion, 23 of the 37 dyslexic participants (62%) and 2 of the 35 control participants (6%) could be classified as displaying deficient performance in the cued, set size 16 condition. The task appears to exhibit reasonable levels of both sensitivity and specificity—a good proportion of dyslexic individuals are identified as poor performers, whereas very few controls are. Rather than setting a single arbitrary cut-off point, a better approach is to evaluate the trade-off between sensitivity and specificity for a range of cut-offs. Figure 2B shows the complete receiver operating characteristic (ROC) for cued and uncued conditions—hit rate refers to the proportion of dyslexic individuals identified; false alarm rate refers to the proportion of control individuals identified. The smooth curves were obtained by fitting binormal ROC functions to the data sets using a maximum likelihood technique. By calculating the area under each fitted curve, we are able to obtain single measures of classification accuracy that can be directly compared across tasks (e.g. Swets, 1988). In the case of the cued condition, an area of 0.84 implies that if pairs of dyslexic and control individuals were randomly drawn from the sample, the dyslexic individual would have a higher threshold 84% of the time. In contrast, applying the equivalent procedure to uncued thresholds would discriminate individuals with only 58% accuracy.

Group comparisons on additional psychophysical tasks

Descriptive statistics for each of the four visual and auditory tasks are shown in Table 2. For all tasks, the mean threshold for the dyslexic group was higher than for controls. As distributions of thresholds for each task were positively skewed, all scores were log-transformed prior to the calculation of any parametric inferential statistics. This produced reasonably symmetrical and unimodal distributions for both groups. Simple t-tests performed on the log-transformed thresholds revealed significant group differences for global dot motion, frequency discrimination and frequency modulation, but not for contrast sensitivity.

View this table:
Table 2

Summary statistics and group comparisons between control and dyslexic groups on measures of contrast sensitivity, global dot motion, frequency discrimination and frequency modulation

Control geometric mean (geometric SD)Dyslexic geometric mean (geometric SD)
Contrast sensitivity (threshold Michelson contrast)5.05 × 10−3 (1.57)5.80 × 10−3 (1.76)t (70) = 1.14, P > 0.05
Global dot motion (% signal dots)7.37 (1.53)10.52 (1.54)t (70) = 3.46, P < 0.05
Frequency discrimination (Hz)5.97 (1.88)11.28 (2.84)t (70) = 3.13, P < 0.05
Freqency modulation detection (Hz)2.69 (1.75)3.84 (2.08)t (70) = 2.26, P < 0.05
  • As t-tests were conducted on log transformed thresholds, geometric means and SD are shown for each group. Note, since a geometric SD represents a factor rather than a quantity, the specific units for each measure do not apply.

Comparison of discriminative accuracy of measures

In order to compare the accuracy of the psychophysical and verbal short-term memory measures at discriminating between dyslexic individuals and normal readers, separate ROC curves were constructed for each (see Supplementary material). ROC curves were also constructed for each of the reading and spelling measures (other than that used to select the groups), so as to provide an indication of the upper bound of higest discriminative accuracy. A comparison of the area under ROC curves derived from each measure is shown in Fig. 3. As would be expected given their close association with the group selection variable, the reading and spelling measures produced the highest discriminative accuracy. Of the remaining measures, the spatial cueing task displayed the discriminative accuracy, closely followed by the verbal short-term memory measures. The ROC areas for the additional psychophysical tasks were all noticeably smaller. A method outlined by Hanley and McNeil (1983) was employed to test the statistical significance of these differences (see Supplementary material). The ROC area for the cueing task was found to be significantly smaller than for the WRMT Word Attack subtest and the WRAT-3 Spelling subtest, but significantly larger than for contrast sensitivity, global dot motion, frequency discrimination and frequency modulation. Differences between the cueing task and the TOWRE Sight Word Reading subtest and each of the verbal memory measures were not statistically significant.

Fig. 3

Comparison of classification accuracy of the cueing task relative to visual, auditory, memory, reading and spelling tasks. Asterisks are used to denote tasks for which the area under the ROC curve was significantly different to the cueing task (P < 0.05, 1 tailed). Error bars are ±1 standard error of the area statistic. C16, cued search, set size 16; CS, contrast sensitivity; GDM, global dot motion; FD, frequency discrimination; FM, frequency modulation; Digit, CTOPP memory for digits; Nonword, CTOPP non-word repetition; Sentences, Binet memory for sentences; SWE, TOWRE sight word efficiency; spelling, WRAT-3 spelling; Word attack, WRMT word attack.

Modelling the role of attention in visual search

A simple SDT model of uncued search performance

We assumed that information at each stimulus location is analysed by two independent orientation detectors with preferences for clockwise and counter-clockwise tilt from vertical. Each detector was broadly tuned to its preferred orientation. The response of each detector was modelled by a Gaussian function of the form Embedded Image where M is the maximum response of the detector (arbitrarily fixed at 10), φ is the stimulus orientation, φP is the preferred orientation of the detector (−30° or 30°) and σ is the SD of the tuning function (fixed at 18°). The coarse sampling and broad tuning of these orientation selective mechanisms is consistent with known physiological (e.g. Hubel and Wiesel, 1968; Blakemore and Campbell, 1969) and psychophysical (e.g. Campbell and Kulikowski, 1966; Thomas and Gille, 1979) evidence. Orientation tuning functions for the pair of detectors are shown on the left hand side of Fig. 4A. The values given by the tuning functions represent the mean response values for a particular orientation. However, to generate an individual response to a stimulus, the value given by the tuning function is perturbed by noise. The internal noise associated with each detector was assumed to be independent and normally distributed with a constant SD (σ). The resulting probability density functions for responses to a stimulus oriented 10° counter-clockwise from vertical are shown on the right hand side of Fig. 4A. Similar patterns of variability tend to occur when the responses of single neurons are repeatedly measured (e.g. Bradley et al., 1987). We chose to implement the most widely used decision process—a maximum-of-outputs rule. This is a simple winner-takes-all scheme where the model monitors responses from the full set of clockwise and counter-clockwise detectors and forms a decision based on which detector produces the maximum response. It should be noted, however, that similar results to those presented here can be obtained by using simple linear summation or a range of other non-linear combination strategies.

Fig. 4

(A) Generation of internal representations of stimulus orientation in the simple SDT model. On the left are tuning functions for two detectors with preferences for clockwise and counter-clockwise orientations. The tuning functions specify the mean response of each type of detector to stimuli with a range of orientations. Individual responses are assumed to be subject to internal noise. The right-hand portion of the figure shows probability distributions for the response of each detector to a stimulus tilted 10° counter-clockwise. The amount of noise is equivalent for each detector and is specified by the SD (σ) of the probability distributions. (B) Simulated orientation discrimination thresholds for a range of internal noise values and set sizes. Data points and error bars represent the mean and SD of 100 simulated threshold estimates respectively. (C) Comparison of uncued search performance for the control group of normal readers (replotted from Fig. 1B) and for a single experienced observer (N.W.R.).

Figure 4B shows the relationship between simulated orientation discrimination thresholds and set size for a range of internal noise values. A distinct set size effect is present for all noise levels—thresholds rise as more distractors are added to the display, approximating a linear function when plotted on log–log axes. As can be seen by comparing the gradients of each fit, the magnitude of this effect is very consistent across noise conditions. This reflects the fact that set size effects index the proportional change in accumulative noise as the number of stimuli is increased, a quantity that remains invariant to changes in the actual quantity of noise contributed by each individual stimulus. Increasing internal noise does, however, produce a substantial decline in the absolute level of performance, as seen by the vertical displacement of the linear functions. Similar results can be obtained experimentally by adding external noise to each element in the stimulus array (Baldassi and Burr, 2000). For comparison, empirical data for the group of normal readers is replotted along with that of a highly practised single observer (N.W.R.) in Fig. 4C. The pattern of results is similar to the predictions of the simple SDT model. In particular, the vertical separation between the search functions of the control group and N.W.R. is reminiscent of a shift in the amount of noise associated with each detector response. Indeed, this interpretation is consistent with the view expressed in the literature on perceptual learning that extended practice on a task can lead to reductions in the internal noise associated with relevant detectors (e.g. McLaren et al., 1989; Dosher and Lu, 1998).

Early selection: signal enhancement

In recent years, a vast amount of research has been carried out, which documents the neural correlates of attentional selection. A number of single-cell recording studies have compared the responses of neurons when attention is covertly directed to a stimulus falling within the neuron's receptive field with when attention is directed to a stimulus outside of the receptive field. Enhanced firing rates in attended conditions have been reported in a number of cortical areas including V1, V2, V4, MT and LIP (e.g. Motter, 1993; Colby et al., 1996; Treue and Maunsell, 1996, 1999; McAdams and Maunsell, 1999). Findings consistent with signal enhancement in the human cortex have also been reported using fMRI and PET (e.g. Heinze et al., 1994; Brefczynski and DeYoe, 1999; Ghandi et al., 1999). Support for enhancement occurring at the behavioural level has been drawn from studies showing that using spatial cues to direct attention to a stimulus can increase contrast sensitivity (e.g. Carrasco et al., 2000) and spatial resolution (e.g. Yeshurun and Carrasco, 1998, 1999).

Simulating the effect of signal enhancement on search performance

To simulate the effect of signal enhancement in our simple SDT model, we assumed that responses of both clockwise and counter-clockwise detectors were amplified by a multiplicative factor at the attended (target) location. Support for this assumption can be found in the results of McAdams and Maunsell (1999), who measured the attention related changes in response rate for a population of orientation selective V4 neurons. Figure 5A shows mean normalized tuning curves obtained by McAdams and Maunsell in conditions where monkeys attended either to a grating presented within the neuron's receptive field or to a location in the opposite hemifield. Attention appears to increase the gain of the neuronal response without affecting the shape of the tuning function. Indeed, the upper curve fit can be almost perfectly obtained by multiplying the lower curve by a constant factor (∼1.22).

Fig. 5

(A) Mean normalized responses of 197 V4 neurons to an oriented grating. Filled symbols indicate responses when attention was directed to the target grating, while unfilled symbols indicate responses when attention was directed to another stimulus outside the neuron's receptive field (data replotted from McAdams and Maunsell, 1999, Fig. 4). (B) Simulated outcomes for the effect of signal enhancement on search performance. The magnitude of the set size effects (indexed by the log–log gradient of the search function) is plotted against amount of signal enhancement for different levels of internal noise.

Implementing this type of multiplicative increase, orientation discrimination thresholds were simulated for a range of enhancement levels (0–50%), set sizes (1, 2, 4, 8 and 16) and internal noise levels (σ = 0.5, 1.0 and 2.0°). Results are summarized in Fig. 5B. Increasing the response of detectors at the target location had the effect of reducing the magnitude of the set size effect. However, this result was heavily dependent on the amount of internal noise associated with each detector response. At the lowest noise level (σ = 0.5°), increasing the gain at attended locations rapidly removed the influence of the distractors, such that no set size effect was evident with a 30% signal enhancement. With noise σ = 1°, 50% enhancement was required to achieve the same result while with noise σ = 2°, a considerable set size effect persisted even at the highest enhancement level simulated.

Plausibility of signal enhancement

The results of our simulations indicate that, notionally, early enhancement of this type could produce similar benefits to search performance to that observed in normal readers when the target location is cued—in both cases we observe a systematic reduction of the set size effect. If we assume for a moment that signal enhancement is the primary mechanism underlying the cueing effect, we can also generate more subtle predictions that can be compared with the empirical dataset.

As mentioned earlier, differences in absolute uncued performance between the control group of normal readers and highly practised observers may reflect differences in the amount of internal noise associated with detector responses. If this were true, then the model would predict that an equivalent degree of signal enhancement could result in a large reduction of the set size effect for the experienced observers (for whom internal noise is relatively low), but only partially reduce the set size effect for control participants (with higher internal noise). Indeed, this seems to be the case. For NWR cueing seems to totally remove any influence of the distractors, resulting in an almost horizontal linear fit (see Supplementary data). Comparable results in experienced observers have also been reported previously by Morgan et al. (1998) and Baldassi and Burr (2000). In contrast, cueing only partially reduced the set size effect in the control group in the present study, leaving a noticeable residual effect (Fig. 1B).

A clear characteristic of signal enhancement is that it acts to amplify the signal provided by the target, while not affecting the variability of responses. Given that this increases the basic signal to noise ratio at the target location, an obvious prediction is that cueing should improve performance even when no distractors are present. As can be seen in Fig. 1B, a small but statistically non-significant cueing effect at set size 1 is present for the control group. However, drawing a firm conclusion about the plausibility of the signal enhancement account based on this prediction is difficult. The predicted improvements at set size 1 are small and one might question the power of our experimental design to detect a significant effect. More importantly though, the prediction rests upon the questionable assumption that attention is not directed to the target stimulus on uncued trials. When the target is presented in isolation, it seems highly likely that attention would be allocated to the single stimulus. As a result, one might argue that no cueing effect should be expected at set size 1.

A bigger problem for the signal enhancement account stems from predictions it makes about the effect of introducing tilted distractors. Figure 6 compares experimental results and model predictions for a cued, set size 16 search in which the orientation of each distractor stimulus was subjected to random ‘jitter’ on each trial. The amount of jitter was controlled by drawing distractor orientations from a noisy Gaussian probability distribution with a mean of 90 (vertical) and a variable SD (σorientation). Separate simulations were run for each of four observers—internal noise values were derived from orientation discrimination thresholds for targets presented in isolation (σNWR = 0.51, σBSW = 0.53, σPVM = 0.58, σCAR = 0.85) and signal enhancement levels were approximated from cued, set size 16 performance in the absence of any distractor orientation jitter (19, 11, 9 and 12% for N.W.R., B.S.W., P.V.M. and C.A.R., respectively). As shown by the triangular symbols in Fig. 6, increasing the amount of distractor orientation noise resulted in a progressive increase in simulated orientation discrimination thresholds. The basis of this effect is fairly straightforward. In the model, decisions are based on the maximum response from the set of oriented detectors. On average a tilted distractor will produce a larger response in one of the detectors than a stationary one, and this difference will increase with the amount of tilt (peaking at 30° either side of vertical). Thus, the more tilted the distractors are, the more chance there is that the decision will ultimately be decided by a completely random detector response generated at a distractor location. Enhancement of the target response can counteract this effect to a point—notice how simulated thresholds are fairly resilient to minor perturbations of distractor orientation (e.g. σorientation = 1°). However, under early selection of this type, unattended distractors are never entirely excluded from the decision process.

Fig. 6

Comparison of simulated and empirical data showing the effect of distractor orientation jitter on cued performance with a set size of 16. Thresholds for discriminating the orientation of the cued target stimulus are shown as a function of the amount of distractor orientation noise. Divergent predictions are obtained via simulations of signal enhancement (triangular symbols) and uncertainty reduction (square symbols) models of attentional selection (see text for details).

In contrast to the predictions of the signal enhancement hypothesis, experimental data suggests that distractor orientation noise has minimal effect on cued search performance. As shown by the circular symbols in Fig. 6, introducing tilted distractors produced small increases in orientation discrimination thresholds for each observer, though not nearly of the same magnitude of that seen in the simulated data. Contrary to the signal enhancement hypothesis, directing attention to the cued location seems to allow fairly efficient exclusion of distractor information. Thus, while signal enhancement accounts nicely for certain aspects of the experimental dataset, it appears unlikely that it is the primary mechanism of attentional selection underlying the cueing effect.

Late selection: reducing decisional uncertainty

How might attentional selection occur at the decisional stage? Next we consider the possibility that attention acts primarily by restricting the amount of information that is incorporated into the final decision. In the absence of a cue, the location of the target stimulus is uncertain, and an observer must monitor information relating to each and every stimulus in the search display. However, as cueing acts to remove any uncertainty about which element is the target, the most efficient strategy would be to form a decision based solely on information derived from the cued location. This would effectively negate the effect of the noisy distractors, making the task equivalent to when the target is presented in isolation. Late selection of this type occurring at the decisional stage has been referred to in the literature as a process of uncertainty reduction (e.g. Palmer, 1994; Morgan et al., 1998) or noise reduction (e.g. Shiu and Pashler, 1994). We will adopt the former term here.

Plausibility of uncertainty reduction

Proponents of uncertainty reduction as the major mechanism for attentional selection point to a number of forms of evidence. First, cueing often seems to affect performance only when there are distracting stimuli or positional uncertainty, consistent with the view that attention does not alter the quality of processing of the cued stimulus. This is certainly also the case with the cueing effect considered in this article—cueing acts primarily to reduce the magnitude of the set size effect and typically there is little or no cueing effect at set size 1. Second, the magnitude of performance gains is often consistent with the statistical advantage conferred by using cue information to reduce decisional noise. As mentioned earlier, the 100% valid cue used in the present study should act to remove all uncertainty about the target location. Accordingly, uncertainty reduction predicts that performance ought to be completely independent from set size in cued conditions. Critically, in contrast to the predictions of the signal enhancement hypothesis, the fact that distractors are excluded from the decision process altogether means that cued search performance should be unaffected by jittering distractor orientation. As shown in Fig. 6, the predictions of uncertainty reduction (square symbols) provide a far more convincing approximation of the empirical data than the signal enhancement account.

These features of uncertainty reduction make it a more feasible explanation of the cueing effect described in the present study than the signal enhancement. This being said, not all facets of the data set are entirely consistent with a strong form of uncertainty reduction. As noted previously, data for the control group shows evidence of a residual set size effect in the cued condition. How can this finding be reconciled with the total elimination of distractor influence predicted by uncertainty reduction? One possibility is that relatively inexperienced participants may have occasional difficulties detecting the brief cue. Assuming that detection failures occur irrespective of set size, and that on such trials observers resort to a standard uncued strategy, it seems reasonable to predict that a residual set size effect would eventuate. An alternative possibility might be that while the cue completely removes any positional uncertainty regarding the target, the mechanism that allows distractor information to be filtered out of the decision process is not always absolute. This would be consistent with the suggestion by a number of researchers that information outside the focus of attention continues to be processed to a certain degree (e.g. Eriksen, 1990; Carrasco and Yeshurun, 1998). If the efficiency of distractor exclusion improves as a function of practice on a task, one might expect to find a residual set size effect in inexperienced observers, but little or no set size effect in highly practised individuals.

Discussion

When the location of a task-relevant stimulus is pre-cued, covert attentional mechanisms allow normal readers to effectively filter out irrelevant visual information, thereby maximizing performance. The present study has demonstrated, however, that many dyslexic individuals seem unable to obtain the same benefit from this form of attentional filtering. The deficit does not appear to be a general feature of all dyslexic individuals. However, we have shown that performance on the cueing paradigm is a remarkably accurate discriminator of dyslexic and normal readers. Indeed for our sample, discriminative accuracy for this task was found to be significantly higher than for a range of other psychophysical tasks commonly used in dyslexia and was on par with measures of a core component of phonological processing (verbal short-term memory). These findings indicate that attentional deficits are particularly prevalent in dyslexia and warrant more detailed consideration.

To begin to investigate the nature of mechanisms underlying successful attentional filtering, we outlined biologically plausible early and late selection accounts and implemented them within a simple signal detection model of visual search. Based on a comparison of simulated and empirical data, we propose that cueing effects in normal readers most likely reflect late attentional selection, in which information pertaining to the target stimulus is prioritized during decision making. According to this view, cueing benefits performance primarily because it allows observers to reduce the amount of noise or uncertainty associated with a particular decision. In the research literature, uncertainty reduction is most often considered in an abstract manner, with little discussion of the neural mechanisms that allow distractor information to be filtered out of the decision making process. However, these considerations are of particular importance here. Given that uncertainty reduction seems the most plausible explanation of the cueing effect in normal observers, the implication is that the cueing deficit observed in dyslexic individuals most likely constitutes a failure of the uncertainty reduction process.

Potential neural mechanisms for uncertainty reduction

Visual processing in the cerebral cortex exhibits hierarchical organization. Neurons in early areas such as V1 show tuning for simple stimulus features and are characterized by relatively small, retinotopically organized receptive fields. Moving along the cortical pathways, projections from lower areas converge onto cells with larger receptive fields. This results in a progressive change in selectivity towards more complex stimulus features and an accompanying trend towards greater invariance with respect to stimulus size and position. An implication of this pooling process is that multiple stimuli will necessarily compete for neural representation at higher visual areas (e.g. Desimone and Duncan, 1995). This competition can be demonstrated by recording a neuron's responses when multiple stimuli are presented within its receptive field. If a pair of stimuli is presented, one of which is well matched to the tuning characteristics of the neuron and the other is not, firing rates tend to be intermediate between the responses to the two stimuli presented in isolation. Results of this kind have been reported in a variety of extrastriate areas (V2, V4, MT, IT) and suggest that individual stimuli are not processed independently, but rather interact in a mutually suppressive manner (e.g. Miller et al., 1993; Recanzone et al., 1997).

The progressive pooling of stimulus information in extrastriate visual areas bears a strong resemblance to the way in which information is integrated before making a decision in SDT models of visual search. The simple SDT model, which employs a maximum of outputs rule across orientated filters at each location, is undoubtedly a simplistic imitation of the true neural process. In reality initial orientation coding is likely to occur via the responses of populations of neurons with different preferences, rather than a single pair of oriented filters. Additionally, integration of information across space is likely to occur progressively across multiple brain areas, and be subject to additional influences not implemented in the model, such as lateral inhibition. This being said, the physiological and anatomical data are certainly supportive of the general type of scheme we have implemented. In uncued search, local coding of stimulus orientation in early cortical areas is progressively pooled across the search display, allowing responses to be formed based on the output of a smaller number of cells in higher areas.

According to the uncertainty reduction account, informative spatial cueing allows an observer to avoid the pooling process and form a decision based primarily on information derived from the target location. Again, we can look to single cell recordings in extrastriate cortex for a neural correlate of this function. When attention is directed to one of a number of stimuli falling with in neuron's receptive field, the resulting response is often biased towards that which occurs when that stimulus is presented in isolation (e.g. Moran and Desimone, 1985; Chelazzi et al., 1993, 2001; Reynolds et al., 1999). Attention seems to minimize, or filter out the effect of the competing information, perhaps by causing the receptive field to shrink around the attended location (see Reynolds and Desimone, 1999; Kastner, 2004). In addition, fMRI results suggest that similar mechanisms also operate in human extrastriate cortex, particularly at intermediate levels such as V4 and temporo-occipital area (TEO) (e.g. Kastner et al., 1998; Pinsk et al., 2004).

Similarity with acquired difficulties following extrastriate lesions

Perhaps the most insightful findings for the present discussion come from studies documenting the breakdown of attentional selection at a behavioural level following lesions to extrastriate visual areas. De Weerd et al. (1999) trained two monkeys to discriminate the orientation of parafoveally presented sinusoidal gratings that were surrounded by distractor stimuli. While spatially restricted lesions of V4 and TEO had no substantial effect on discrimination thresholds when the grating was presented in isolation, performance deteriorated rapidly in affected regions as the distractor contrast was increased. These results provide an interesting comparison to the spatial cueing deficit shown by dyslexic individuals in this study. In both cases, the observed deficit is not in orientation discrimination per se, but rather in the inability to inhibit or suppress the influence of distractor stimuli on performance. An additional similarity is that distractor interference occurs even in the absence of uncertainty about the target location—in the De Weerd et al. study, the grating was always presented in the centre of the stimulus configuration; in the present study the position of the target was cued. Given these similarities it might be tempting to infer that an impairment affecting extrastriate areas such as V4 or TEO might underlie the spatial cueing deficit in dyslexia. The appeal of this argument is further strengthened by recent studies showing that losses in attentional selection following lesions in these areas also extend to humans (Gallant et al., 2000) and generalize across a range of tasks and stimuli (De Weerd et al., 2003). However, if this was the case then we would also expect to see evidence of a range of concomitant difficulties associated with loss of ventral stream function. For example, Gallant and colleagues (2000) reported a case study of a human neuropsychological patient who had a V4 lesion as the result of a right hemisphere posterior cerebral artery infarct. Replicating the methods used by De Weerd et al. (1999), they found evidence of a marked difficulty with inhibiting the influence of irrelevant distractors. In addition though, they reported a major disturbance of hue (colour) discrimination and very poor performance on a range of tasks measuring aspects of intermediate form processing (discrimination of contrast-defined borders, glass patterns and spiral non-Cartesian gratings). Other functions, such as luminance contrast sensitivity and motion direction discrimination, were not affected. While many types of visual deficiencies have been proposed to exist in dyslexia, the literature does not provide much support for a similar pattern of difficulties to that shown following V4 lesions. In fact, the precise functions that were spared in the patient reported by Gallant and colleagues are perhaps those most commonly reported to be deficient in dyslexia.

Source or site of attentional selection?

The results of the lesion studies described earlier support the notion that extrastriate visual areas permit attentional selection by biasing competitive interactions between multiple stimuli. However, it is thought that the signals causing this biasing are not generated in the visual cortex itself, but are fed back from higher-order brain areas. Much of what we know about top-down attentional signals is drawn from the results of neuroimaging studies in human observers. Pre-cueing techniques have been used to temporally separate the processes of generating the attentional signal from those involved in actually processing the attended stimulus. Results point towards a network of frontal and parietal areas as the most likely sources of attentional signals, including the superior frontal sulcus, inferior and superior parietal lobules, frontal eye fields, and supplementary eye fields (for reviews see Kastner and Ungerleider, 2000; Corbetta and Shulman, 2002; Kastner, 2004). Recently it has been further suggested that this network subserves attentional deployment supporting both signal enhancement and biased competition effects (Marois et al., 2004), but that a distinct ventral frontoparietal network lateralized to the right hemisphere might be recruited in automatic, stimulus-induced orienting (Corbetta and Shulman, 2002). It is feasible that the cueing deficit displayed by dyslexic individuals could result from a problem with generating attentional signals in frontal or parietal cortex. Alternatively, the deficit could result from incomplete or damaged connections allowing top-down feedback of these signals to extrastriate visual areas. In either case, one would expect that filtering of distractor information would be impaired, even if structures supporting the filtering process were uncompromised. Support for this possibility can be found in a recent study documenting ineffective attentional filtering in a patient with bilateral damage to posterior parietal cortex (Friedman-Hill et al., 2003).

Supplementary material

Supplementary data are available at Brain Online.

Acknowledgements

This research was supported by a grant from the Australian Research Council. N.W.R. is supported by the Wellcome Trust.

Footnotes

  • Abbreviations:
    Abbreviations:
    fMRI
    functional magnetic resonance Imaging
    PET
    positron emission tomography
    PEST
    parameter estimation by sequential testing

References

View Abstract