Navigating the narrative: An eye-tracking study of readers' strategies when Reading comic page layouts

In multimedia stimuli (e.g., comics), the reader must follow a narrative in which text and image both contribute information, and artists may use more irregular layouts which must still be followed correctly. While previous work has found that the exter-nal structure (outlines) of panels is a major contributor to navigation decisions in comics, other studies have shown that panel content can affect reading order. The present studies use eye-tracking to investigate these contributions further. In Experiment 1, the reading behaviors on six layout variations were compared. The influence of the external structure was replicated, but an effect of text location was also found for one layout type. Experiment 2 focused on variations of this particular layout, manipulating the location of text within critical panels. Panel content was a consistent effect for all variations. While most navigation decisions are made using the external structure, content becomes key when resolving ambiguous layouts.


| INTRODUCTION
The order in which viewers choose to look at the components of an image is not particularly strict. Unlike text, where readers predominantly move from one word to another in a single reading direction, viewing order of image regions depends on several factors, such as task setting, the viewers' own interest in particular image regions, and low-level features such as brightness or line orientation that may grab attention.
In multimedia stimuli, such as comics, images play an important role, contributing much information alongside text (or, in the case of textless comics, all the information). Perhaps due to the role of images in this medium, initial theories of comic reading suggested that the viewing process would not be particularly structured, as images communicated "largely without rules" (Postema, 2013). Tracking the eyes of someone reading a comic would, it was thought, provide a rather erratic pattern, as the reader gathered information holistically from the page (Barber, 2002). However, this interpretation is debated, with a growing number of studies showing that, when reading comics, readers follow the narrative sequence of the panels, in a fashion that is similar to reading text (Foulsham et al., 2016;Kirtley et al., 2018).
Indeed, disruptions to the panel order, either by mixing up or deleting certain items have clear effects on viewing behavior (Foulsham et al., 2016;Cohn, 2013), slowing reading speed and increasing the amount of regressive saccades and re-fixations on the disordered items. This work has shown that, when reading comics, readers are following the sequential narrative provided by the text and images.
As with text, comics mostly follow the typical reading order of the culture in which they were written: Western comics tend to be written to read left-to-right, Japanese manga are written right-to-left (Bongco, 2000;Duncan, 2000). The standard reading order is to follow the line of panels from one side of the page to the other, and upon reaching the end, drop down by one line and return to the start to begin again, also known as the Z-path (Cohn, 2013). However, within comics there is much more scope for varying panel layouts beyond a simple grid shape. Artists and writers may choose to use a variety of panel layouts, such as basing the pages around a circle or more irregular shapes or superimposing smaller images on a larger background to draw attention to them. However the panels are laid out, they should still serve the reader, and allow them to navigate successfully through the sequence. Mistakes in this aspect of comic creation can result in readers failing to follow the narrative, or losing interest in the story itself (McCloud, 1993).
The first investigations into comic page navigation were conducted by Cohn and colleagues (Cohn, 2013;Cohn & Campbell, 2015). In these studies, the focus was on the contribution of the panels' spatial arrangement, rather than their content. Cohn (2013) suggested layout alone could be a strong contributor to navigation, noting that it was possible to re-arrange comic strips into different layouts without affecting the information within the panels. To this end, participants were presented with blank comic pages, showing only the outlines of empty panels. Participants were required to number the panels in the order they thought they would read them if the panels contained information.
Five manipulations of layout in particular were examined, all created by changing the spatial relationships between the panels in question (shown in Figure 1). The baseline type was the typical Grid pattern (Figure 1a), the arrangement which could be considered the closest to the organization of rows of text, and the manipulations included increasing the distance between panels (Separation, Figure 1b) or reducing it so that panels overlapped (Overlap, Figure 1c). The Staggered pattern (Figure 1d) is an adjustment of the Grid pattern where either the vertical or horizontal gutters are no longer continuous. All these layout types are used in published comics today (Cohn, 2013). The final manipulation (Blockage, Figure 1e) is a more extreme version of the Staggered layout, where one panel effectively blocks the horizontal gutter of the neighboring panels. The Blockage layout has also been mentioned in other works, with comic theorists and creators pointing out that this is one of the more potentially misleading layouts that can be used (e.g., McCloud, 2006;Abel & Madden, 2008), due to uncertainty as to whether the reader should read across or down.
When labeling the empty panels, readers showed a strong tendency to follow the Z path in the Grid layout. While slight deviations were found, participants were also more likely to follow this path for Separation and Staggered layouts. For the Overlap layout, however, readers were more likely to follow the direction indicated by the overlapping panels: thus, if the arrangement indicated a Z-path pattern, readers would follow this, while if the overlap suggested a Blockage pattern, readers would use this path. Finally, when labeling the Blockage panel, readers were significantly more likely to break from the standard Z-path-as is necessary to follow the narrative in this layout-and follow a vertical pathway (from A to B, then across to C).
This tendency was enhanced for readers who scored highly on measures of comic expertise: participants who were less familiar with comics and did not read as many were more likely to follow the Z-path (A-C) even on the blockage layouts, where this path is not what was intended by the artist in order to follow the narrative.
Based on these findings, Cohn (2013) proposed a set of rules for detailing how the external compositional structure (ECS) of comic panels influenced readers' navigational choices in selecting which panel to move to next from their current position. These rules suggest that readers will try to maintain a grouping between panels, preferring smooth paths to broken ones, something that the borders (both internal and external) of the panels indicate. Readers will also usually attempt to follow a Z-path, moving to the right, and then returning to the left a line down, unless a downwards movement is indicated by the borders (as in the Blockage layout). Less expert readers, Cohn notes, may stick to the Z-path whatever the panel arrangement, due to their lack of familiarity with the more varied comic layouts. As a last catch-all, readers will finally examine any panel which has not yet been viewed, to ensure no details have been lost. Figure 2 provides an example of the proposed navigation decisions in Grid and Blocked layouts. In the Grid pattern, the reader begins in Panel A, and, since both inner and outer borders are continuous, they prefer to move to the right, into Panel B. Once in B, there are no more rightwards panels, so they follow the Z-path to move into C, before repeating the same process to move from C to D. In the F I G U R E 1 Examples of the five layout types used by Cohn (2013). Letters indicate reading order. (a) Grid; (b) Separation; (c) Overlap; (d) Staggered; (e) Overlap F I G U R E 2 Correct reading order for Grid layouts (L) and Blocked layouts (R) Blocked layout, the vertical inner border is continuous, while the horizontal inner border is blocked by Panel C. Thus, rightward movement from A to C is blocked, and so the reader moves from A to B, violating the Z-path, but following the narrative.
While Cohn's (2013) findings suggest a major role for the layout alone in guiding navigation, there is evidence that the content of the panel also makes some contribution. Text as a feature of comic panels is a strong attractor of attention: Laubrock et al. (2018) found that text within comic panels drew significantly more fixations than images, despite the fact that text regions typically occupied a much smaller area of a panel than images. Text also appears to be identified as a target for future fixations: Laubrock et al. (2018) found readers used preview information more for text than images; on a similar note, Kirtley et al. (2018) found that the presence of text within a panel increased the likelihood of readers visiting that panel, while textless panels were more likely to be skipped in first pass reading. The position of text within an upcoming panel can also influence the incidences of fixation within the region (Kirtley et al., 2018;Ishii et al., 2004). However, there is currently no work examining how content might contribute to readers' navigation, in conjunction with the panel layout.
The overall aim of the current studies, therefore, is to explore the contributions of panel content and panel layout on the behavior of participants reading short comic stories. By adding content to the layout types suggested by Cohn (2013) we create a situation in which there are right and wrong ways to follow the story. By using comics created for the purpose of this study, we were able to control for the narrative and visual content of panels across the different types of layouts studied here. Furthermore, by using eye tracking to record the gaze and viewing behavior of the participants, we can obtain more details on the participants' decision making during the reading process. Rather than only seeing the final outcome of participants' direction choices, the eye-tracking allows us to trace any cases in which participants make errors which they then correct, thus providing a clearer understanding of the processes participants go through to make their decisions about the order in which to read the panels. It is predicted that readers will show differences in how they read the same content across the different layout types, with measures such as regressions (returns to panels that are earlier in the narrative sequence) and skips indicating where readers make more errors in the reading order. Second, the content of the layouts, in particular the presence of text within the panels, will contribute to how readers choose to navigate the layouts presented. These findings will add to the understanding of page navigation in sequential art and expand upon Cohn's (2013) initial proposal.

Stimulus comics
Six comics were created for the purposes of this study. These comics portrayed a range of art styles, genres and use of color or black and white, and ranged in length from 3-8 pages. Table 1 summarizes further information about the six comics, including the breakdown of panel content. A panel was considered to contain text if it presented words in speech or sound effects. Character-focused panels were those in which the focus was on a human character central to the story.
Within each comic, a set of four panels was selected to be manipulated into the following chosen layouts (examples of these layouts for one of the comics is shown in Figure 3). Table 2 summarizes the panel content for the four critical panels in each comic. It is important to note that the critical panels did not occupy the entire page of any comic. Rather they were a subset of panels on one of the pages and their position within the page-and indeed whether  Grid. The basic layout: the panels were arranged in a 2 Â 2 square.
Correct reading order required the reader to use the standard Z path to move left to right and then down the line.
Overlap Grid. The same grid layout, with panels overlapping so that Panel 1 (P1) overlapped P2 and P3 overlapped P4.
Staggered. The size of the panels in the grid layout was adapted, so that there was no longer a clear grid between the panels. Panels were manipulated horizontally for all items. Correct reading order required the reader to use the standard Z path to move left to right and then down to the line below.
Overlap Staggered. The same staggered layout was used, with panels arranged so that Panel 1 overlapped P3 and P2 overlapped P4.
Blocked. One of the four panels was removed (care was taken to ensure that the narrative was not disrupted by this removal, and no text was removed from the story). The third panel was increased in size, with the content adjusted to fit it. Correct reading order for such panels was the typical blocked pattern of moving down from 1 to 2, then across to 3.
Overlap Blocked. Panel A of the blocked layout was extended so it overlapped with Panel C. Figure 3

Eye-Tracking
Participants' eye movements were monitored and recorded on the EyeLink 1000+ eye-tracker, set to the participants' dominant eye. Eye dominance was determined using a variant of the Miles test (Miles, 1928). A 9-point calibration grid, followed by a 9-point validation grid was used to fit and test the spatial accuracy of the eyetracker at the start of the experiment. If the validation procedure showed a mean spatial accuracy worse than 0.5 , or a maximum spatial accuracy worse than 1 degree, calibration and set up were repeated. Saccades were detected using the default cognitive setting for the SR Research algorithm, whereby saccades are detected if both the velocity threshold of 30 s À1 and acceleration threshold of 8000 s À2 were exceeded for two consecutive samples.

Expertise questionnaire
Participants' familiarity with comics as a medium was assessed using the Visual Language Fluency Index (Cohn, 2014). This questionnaire asks participants to rate, on a scale of 1-7, the frequency with which they read different types of comics, both currently and during childhood (aged 16 or younger). They are also asked to rate their proficiency in drawing both currently and during childhood, and how fluent they consider themselves to be in comic reading. The VLFI score ranges from 1 to 52.5, with higher numbers indicating greater comic fluency.

| Procedure
Participants were asked to read the full stories, while their eye movements were tracked. Stories were presented as in sets of two pages per screen, in order to mimic natural comic reading. Readers were able to "turn" to a new pair of pages using the left and right arrow keys on the keyboard. An information screen was presented at the end of each story, allowing the reader to either turn back to previous pages, or move on to a new story. Reading was self-paced, and the study typically took less than 30 min to complete. Each participant saw each of the six layout types, counterbalanced across the stories. After reading the stories, participants were given the VLFI to complete on a laptop computer.

| Analysis
To analyze the various possible influences on reading behaviors in comics, linear mixed models (LMM) were run for analyses involving continuous variables (viewing time, saccade progression), while Values were generated using the lmerTest library (Kuznetsova et al., 2015). When an interaction was significant, we ran follow-up models to explore it. Data plots were created using the ggplot2 package (Wickham, 2016).
The main concern of the experiment was how the reading behav- Three key factors were considered as possible influences on these measures of reading behavior: the layout in which a panel was set (the six versions); the readers' prior experience of comic reading (as measured by their VLFI score), and the panel content. For these studies, panel content concentrated on the presence, amount and location of text within the critical panels. While earlier work has indicated that the presence of character is also important, the comics used in these studies had very little variation in character presence, with most panels focusing on main characters. By using LMMs, it could be determined whether effects of these measures arose independently of each other.
In the models that follow, any predictor with only two levels was coded using deviation coding. The predictor for layout type had six levels, and was coded using simple coding, with the Grid type as the reference level. This was chosen as the simplest of the layout types with the clearest reading order.
Where interactions were found between layout type and other factors, these were broken down by comparing each of the six layout types to a level of the other predictor. In all models, the score for each participant on the VLFI was included as a fixed effect, in order to consider whether any aspects of reading order were influenced by prior experience. Other potentially confounding factors were included as fixed effects where necessary and described in the following sections.

Model structures can be viewed in Appendix A, and full results tables
for Experiment 1 can be seen in Appendix B.

| Viewing time
We investigated whether differences in layout might affect how long readers took to explore the individual panels within the critical panel groups: layouts which readers found more complex might be expected to increase the participants' viewing time per panel compared to simple layout types (Model 1.1).  indicate re-reading of previously inspected panels or returns to panels that were originally skipped.

First pass skipping
For the current study, we looked only at first pass skips (where a panel is not read in sequence, but the reader returns to it later), because full skips (where a panel was not fixated at all) were very rare, occurring in only 0.8% of instances while the critical panels were read.
Closer examination of the skipping patterns showed that navigation of the panels was more erratic than would be seen for reading text alone, resulting in a high proportion of skips (43.7%). However, these skips could be divided into two distinct types: brief "lookaheads," where the reader would make a single fixation to a panel later in the sequence, but then return to the current position to continue reading, and longer deviations from the order, where the participant would continue to fixate on the later panel, exploring the image and reading the text out of sequence. As the focus of this study was readers tendency to be misled when reading, we focus only on this second category of skips, which occurred at a lower frequency (13.1%). 2 For two of the comic layouts, the critical layout began at the top of the left-hand page. Since pages were displayed to participants after a central fixation, readers would begin at the center and progress to the first panel on the left-hand page (Panel 1 of the critical layout these comics). When moving from the center to the top-left of the page, participants would often make a single fixation elsewhere on the page before moving to Panel 1. These early fixations were not counted as skips. However, if the participant made more than one fixation on later panels before reaching Panel 1, it was considered that they were focusing more on the content, and they were recorded as having read out of sequence, and therefore skipped Panel 1.
We looked initially at the proportion of skipping in all panels in the critical layout set, before looking at the panels on an individual basis. Panels were numbered 1-4, in accordance with the correct reading order. 3 As well as the layout in which the panel was set, we considered the influence of text content: specifically, the number of words present. Full tables for the models run for each panel can be viewed in Appendix C, here, we consider only the significant findings.  Taking the skipping and regression together, it appears that readers had more difficulty in inspecting panels in the intended order with layout types that included blocked panels. Initially, it seems they missed out the second panel and went directly to the third, before realizing their mistake and regressing back to the second. Our previous analyses showed consistently that the readers behaved similarly when viewing the Grid, Overlap Grid, Staggered and Overlap Staggered layouts. However, Blocked and Overlap Block layout differed from these four layouts. In order to maximize our ability to detect any differences in these analyses, the layout types were re-classified into three groups. These were the Grid-Based layouts (Grid, Overlap Grid, Staggered, Overlap Staggered), the Blocked layout and the Overlap Block layout, with the fixed effect of layout being entered into models with these three levels.
As previous work has shown readers will often direct saccades to text in upcoming panels (e.g., Kirtley et al., 2018;Laubrock et al., 2018), the location of text in Panels 1 and 2 was used as a fixed effect in the LMMs that follow in order to assess whether read- interact (see Figure 4b). were, therefore, influenced by the content of Panels 2 and 3, specifically the location of the text in these panels.

| Discussion
The influence of content is worth considering. The strongest effects were seen in viewing time on panels, where the amount of text increased viewing time. This is in line with previous work on comics (e.g., Laubrock et al., 2018;Kirtley et al., 2018) and other researchers looking at different types of mixed media (e.g., Rayner et al., 2001;Wang and Pomplun, 2011). Text is of high semantic importance when understanding a narrative, and so its presence in a panel means readers are more inclined to spend time on the region to take in the information.
In skipping measures, content also plays a (smaller) role, alongside the layout. The more words a panel contained, the less likely readers were to skip over it-although this was only seen when considering panels overall-at the level of the individual panel, the effect of words did not emerge. Thus, while layout had the stronger effect on navigation, the word content of the panels played a small but significant role in how readers selected the panels they moved to.
Expertise was not a strong influence on readers' tendency to skip panels. Previous work by Cohn (e.g., Cohn, 2014;Cohn and Campbell, 2015) has found that expertise and familiarity does play a role in comic navigation, with more expert readers better able to follow the Blocked layout versions, and so it might be expected that this would also be seen in skipping as a measure of navigation. In text reading, readers with higher spelling ability have been found to show greater skipping of words (Slattery and Yates, 2018) thus, it is possible that this behavior in comic reading has a similar, ability-linked reason behind it. However, unlike individual words, individual comic panels contain more information, from both the image and text, and thus skipping may not reflect quicker processing, but rather lead to poorer comprehension. Furthermore, the range of expertise scores within this sample is small, so further work with a greater range of scores is needed to explore this aspect more.
Overall, it appears that three aspects arise as influences on the navigation process, one very strong and two apparent modifiers of reading order in other areas. The strongest aspect is horizontal versus vertical progression of the narrative, while overlap between panels and the text location are the modifiers. These are discussed below, in relation to Cohn's (2013) findings.

| Horizontal versus vertical progression
Our participants showed a strong tendency to move through the panels horizontally, maintaining the Z-path. They did this even when the Z-path was incorrect for following the narrative, as in the Blocked and Overlap Blocked layouts. Cohn (2013) showed a similar tendency for participants to follow the Z-path, but that when confronted with a Blocked layout, his readers were more likely to deviate from the Z-path for this particular version. This difference appeared to be mediated by the readers' expertise and familiarity with comics, and this is likely to be the reason for the difference between our study and Cohn's. While Cohn's (2013) participants showed a range of scores on the VLFI, including those who were frequent comic readers, the participants in our study had an average score of 7.27 (SD = 4.03), indicating most readers had little or no familiarity with comics. The Blocked layout is one that is largely unique to comics, and so for this layout, they persisted in using the Z-path, despite the fact that this meant they often had to correct themselves.

| Overlap
When panels overlap with one another, this can have an effect on the direction the reader chooses-although only in some situations. Cohn (2013) found that readers would use the overlap to guide their reading, although he noted that readers could be led into rejecting the Z-path if the overlap suggested moving horizontally. Our own readers preferred to follow the overlap only when it indicated the Z-path, as seen in the Overlap Block layout. However, vertical overlap, such as in the Overlap Staggered condition, appeared to have no such effect on readers' navigation decisions. This is perhaps another indication of the effect of our readers' low expertise, leading to their preference of a familiar path, whatever the context.

| Placement of text
The final influence appears to come from the content of the panels being read-specifically how the text was placed in key panels. Experiment 1 suggests that location was not a strong influence for all panel types, but is important in the Blocked layout, the most ambiguous type. In a version with no strong overlap cues, our less experienced readers seemed equally likely to follow the Z-path or not; however, when the presence and location of text is closer in one panel, it seems to increase the likelihood that they will choose that panel as their next step. While this is the first demonstration of panel content contributing to panel navigation  However, it is important to note that while Experiment 1 showed a likely pattern of this behavior, it was only marginally significant. This is likely due to the relatively small amount of data for these conditions. We therefore followed up this finding in a second study, which focused on Blocked layouts alone, as these are the most sensitive to the influence of panel content, in order to experimentally test the influence of text position within panels on reading order.
In Experiment 2, new layouts were created, all variations of the Blocked arrangement. Both the overlap of panels and text location were manipulated, allowing us to determine how much these aspects contributed to the navigation behavior. Panels could either have no overlap (the standard Blocked layout), an overlap between Panels 1 and 3 (the Overlap Block, emphasizing a horizontal path), or an overlap between Panels 1 and 2 (a new layout, emphasizing a vertical path). Text in Panels 2 and 3 was also manipulated, so that it was either clearly closer to or further from Panel 1 for each layout type, resulting in six new critical layouts.
Given the findings of Experiment 1 and previous work by Cohn and colleagues, we predicted different reading patterns for each of the layout variations. For layouts in which Panels 1 and 3 overlap horizontally, we expected to see readers continue to follow the Z-path, regardless of text location, as the overlap encourages the horizontal reading path, which seems to take precedence over text. For layouts in which Panels 1 and 2 overlapped vertically, we expected to find that readers would also continue to prefer the incorrect horizontal path: however, this may be affected by text location. When text in Panel 2 is closer to Panel 1, this may encourage readers to break out of the Z-path and follow the correct vertical route. Finally, for the Blocked layout, with no overlaps, we predict that the reading direction will be influenced by text location, affecting readers' tendency to follow the z-path, with readers moving from Panel 1 to the panel which contains the closest text. For all layouts, it is also possible that readers' expertise will mediate this behavior.

| Participants
Thirty-three participants (11 males, M age = 25.58, SD = 9.89) were recruited from the Psychology undergraduate and postgraduate population at the University of Aberdeen. Three participants were removed from the following analyses due to errors in recording. In return for their participation, they received either course credits, or £3 as reimbursement. All participants had normal or corrected-to-normal vision.
The average VLFI score was found to be 10.67 (SD = 6.62), with the majority of participants being novice readers. The study was approved by the local ethics committee (reference PEC/3625/2017/4) and was in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki).

Stimulus comics
The six comics used in Experiment 1 were again employed for Experiment 2. Six versions of each story were produced, based on three variations of the Blocked layout, each with two versions of text location in Panels 2 and 3. In half the versions, the text in Panels 2 and 3 was moved, so that the Panel 2 text was closer to Panel 1. For the other half, the Panel 3 text was closest to Panel 1. Figure 5 shows the six variations in layout type and text location. The three layout types were as follows: Blocked. This was identical to the Blocked manipulation used in

| Procedure
The study's procedure was identical to that of Experiment 1.

| Analysis
Analyses methods for Experiment 2 were similar to those for Experiment pattern readers were following. We also included the new factor of text location in the analysis, as the presence of nearby text might attract the reader's attention to an incorrect panel. Table 6 presents the average proportion of first pass skips for panels in the three layouts.
The average proportion of first pass skips in all critical layouts was 0.198, (SD = 0.40) skips occurring across all three panels. For Panel 2 (Model 6.1), the average proportion of first pass skips was higher, at 0.48 (SD = 0.50). Again, this was uniformly high for all three layout types, and did not differ significantly between these three layout types (see Table 6).   We predicted that readers, particularly those of low expertise, would continue to follow the incorrect, but more familiar horizontal path when Panels 1 and 3 overlapped with each other. When the overlap indicated the correct but unfamiliar path of 1 to 2, readers would be more likely to take this route. For the plain Blocked layout with no overlaps, readers would be directed by text location, choosing the panel, which had the closest text. However, while it was found that text location did have an influence on navigation, this effect did not interact strongly with the layout type.
It is worth noting first that our findings regarding behavior such as regressions and skipping in all versions indicate that, whatever the other aspects, this layout is a difficult one for participants. In every version, readers frequently made the mistake of skipping panel two and having to return from the third panel at the same high incidence level, which is comparable with those seen in Experiment 1. This is in line with work by Abel and Madden (2008), as well as the general anecdotal accounts of many comic creators. For inexperienced readers, like our participants (VLFI M = 10.67, SD = 6.62), it seems the novel nature of this layout is one that cannot be overcome easily without repeated exposure. Indeed, there is a strong similarity between the values of skips and regressions around Panel 2 for the layouts used in this study, and in Experiment 1, indicating consistency between how participants responded within the two studies. There is also a similar influence of content in this second study-again, the panel content exerts the strongest effect on viewing time, and also influences skipping behavior. The effect of the more specific content manipulation, changing text location, is considered in the following section.

| Influence of text location
According to the External Compositional Structure (ECS; Cohn, 2013), when presented with a Blocked layout, experts should use the inner borders to move from 1 to 2; while non-experts will ignore the borders to some extent, and move from 1 to 3, at least initially (Cohn, 2013). The overlap between 1 and 2 in the Vertical Overlap condition should encourage the correct path to be followed -Cohn and Campbell (2015), looking at vertical overlaps in Grid layouts, found that overlaps on the left hand side of a layout led to the most deviation from the Z-path, due to the strengthened continuation of the inner borders. Similarly, in the Horizontal Overlap, 1 to 3 should be preferred, due again to the continuous borders, and the grouping indicated by the panel closeness (Cohn, 2013). We support these pre-  (Cousineau, 2005;Morey, 2008), these intervals are adjusted to better represent the patterns of differences cross the experiment, however, unlike the Cousineau-Morey method, these LMEM-based intervals are suitable for crossed random effect designs.
For saccades that incorrectly targeted Panel 3 instead of Panel 2, there was also influences from the external panel structure: in the horizontally overlapping layout, where everything indicates the more familiar but incorrect path, there were more mistakes overall. This was also independent of the text placement, where again participants were more likely to be drawn to the closest text region across all layouts.
Overall, it appears that the panel content, specifically the presence and location of text, was a stronger factor in guiding reading than the findings of Experiment 1 would indicate.
The finding that the Blocked layout in particular is not significantly affected by text location is perhaps unexpected: this is the layout that has no other hints at the correct direction from the panel borders, so it might be expected that text would have the strongest effect here. It is possible that this may also be a finding which would be mediated by expertise (low expertise readers, as in this study, use text for all layouts, higher expertise readers use a combination of text and border information as cues.) The importance of text is something, which has been noted in earlier studies of comics and word-image combinations in general. Rayner et al. (2001), looking at how readers viewed multimedia magazine adverts, noted that the first goal seemed to be to get to the text first, before attending to the image for a longer period. This is further seen in Caroll et al.'s (1992) study of single panel cartoons, and the earlier work on comic reading by Kirtley et al. (2018) and Ishii et al. (2004). In the present study, 41% of first fixations into a panel were to text regions, despite the fact text occupies a smaller area of a panel than the image (on average text occupied just 14% of the comic page).
Indeed, even in real-world scenes, gaze is strongly drawn to regions of text (Wang and Pomplun, 2012). Text is highly informative, and, as Rayner et al. (2001) note, prioritizing this information may be the result of much exposure to multimedia stimuli over the years. This suggestion raises further questions with regard to expertise, however: while text is typically more informative, it is not always the case in comics (see McCloud's (1993McCloud's ( , 2006 seven word-image combinations, where text can play a much reduced role in providing information compared to image). Less expert readers who have not had this experience may be more swayed by text location than experts, who are guided by the External Compositional Structure alone. This would be in line with Cohn's previous findings (Cohn, 2013;Cohn and Campbell, 2015), but the current study cannot provide stronger evidence for this proposal. Further work which compared high and low expertise readers is required to determine if the two groups perceive text in the context of comics differently as a guide for navigation through the panels.
The pattern we predicted for Experiment 2 based on the findings in Experiment 1 were not borne out-but the suggestion from Experiment 1 was based on only a small subset of the data, in which there were naturally occurring variations in text placement within the panel.
Experiment 2 shows more clearly that, for variations of the Blocked panel layout, internal panel content is as strong a guide as the external panel structure.

| General discussion
Two experiments were conducted to examine the influence of panel content and layout on readers' navigation decisions. Experiment 1 showed that the external structure of the panels was a major factor in participants' decisions of how to move between panels, although a small role of panel content, specifically text location, was seen in certain more ambiguous layouts. Experiment 2 focused on these ambiguous Blocked layouts and text location, and found text location had an influence regardless of the layout variations for that type.
Our findings are largely in agreement with Cohn's proposals about the role of the external structures of a panel (Cohn, 2013;Cohn and Campbell, 2015). The key aspect of Cohn's theory is that, while the content of a panel is important for comprehending the narrative, reading order is not. As stated in Cohn and Campbell (2015), 'A single four-panel comic strip might be arranged horizontally, vertically or in a 2 Â 2 grid, so long as the panels are still read in the same successive order, the meaning of the sequence should not change ' (p.194). However, our inclusion of content and the findings regarding the text indicates that additions should be made to the process.
Both studies show that content has its strongest effect on the viewing time, with high-interest images (characters) and text increasing the time a reader spends on the location. However, while less prominent, content also affects how readers move through the panels.
Content, particularly that which may be informative to the readers' understanding of the story, is therefore playing a role in how readers select panels. From Experiment 2, it seems that we can go further, and identify that text location in particular is affecting the initial navigation of the layout.
Cohn's (2013) iteration of the ECS suggests two first steps when readers are deciding how to progress from the first panel of a layoutchecking the inner and outer borders to determine whether one or both are contiguous (see Figure 2). This influences the decision to move right, or down. We suggest that another aspect should be added here, specifically for ambiguous layouts, such as variations on the Blocked item. This addition would state that readers should search for the closest text source in the upcoming panels, and go to this panel if uncertain regarding the next step. From Experiment 2, it seems this ambiguity is not just related to lack of features such as overlap between panels, but any situation where there is a vertical blockage of the inner borders. It is possible that this may only be a factor for low-expertise readers, and raises questions about how affected high expertise readers are likely to be by this: for example, will they ignore text location entirely, and rely solely on the ECS, or will there be situations when text is enough to draw them in? These questions are something that should be addressed in future work.
Overall, the two experiments reported here add more detail to the current literature on comic navigation, by both confirming earlier findings, and showing a role for content. Furthermore, these findings contribute to the growing body of evidence that supports the view that comic navigation is not a random or disorganized process but a strategic one. Reading comics, just as reading text, is about using the information from the material, and the readers' own experience in order to properly navigate the narrative. Wickham, H. (2016)