Results for the composite task are typically presented in terms of accuracy (d’) and reaction time, with accuracy receiving more attention in the literature. We computed d’ in the same way for each combination of alignment and congruency, by calculating the normal-transformed hit rate (the proportion of times each participant said “same” when the ground truth was “same”) and subtracting from it the normal-transformed false alarm rate (the proportion of times each participant said “same” when the ground truth was “changed”). Although participants were cautioned to respond faster if their response took longer than 2500 ms, we did not exclude these trials from analysis. Such trials represented only 2.3% of the overall dataset, but five subjects out of our entire set of 57 subjects had greater than 10% of their trials fall into this category. If these trials are excluded, we find similar results to those reported below with the exception of the experts-only congruency by alignment interaction in Experiment 1.
Experiment 1: upright fingerprints
Sensitivity (d’)
The two top panels of Fig. 2 show the d’ value for each of the four conditions (two levels of alignment crossed with two levels of congruency) separated into two graphs by expertise. The standard marker of holistic processing is a relatively large difference between the congruent and incongruent d’ values when the halves are aligned, and a reduced difference when the halves are misaligned (Richler & Gauthier, 2014). In other words, the incongruent, task-irrelevant halves should interfere with performance more when the halves are aligned than when they are misaligned.
A repeated-measures, mixed-factor 2 (congruency: congruent, incongruent) × 2 (alignment: aligned, misaligned) × 2 (expertise: expert, novice) analysis of variance (ANOVA) was performed for experiment 1. The ANOVA revealed a strong main effect of expertise with F
1,27 = 7.98, MSE = 1.042, p < 0.01, and η
p
2 = 0.228, confirming that experts are more sensitive to the task overall. There was a main effect of congruency (F
1,27 = 15.21, MSE = 0.308, p < 0.01, η
p
2 = 0.360), and this effect was not different between experts and novices (F
1,27 = 0.091, MSE = 0.308, p = 0.765, η
p
2 < 0.01) and a main effect of alignment (F
1,27 = 38.93, MSE = 0.195, p < 0.01, η
p
2 = 0.590), which was also not different across experts and novices (F
1,27 = 0.295, MSE = 0.195, p = 0.591, η
p
2 = 0.011). These main effects of congruency and alignment, where d’ is generally lower in all incongruent trials and in all trials where the test image is misaligned, are consistent with other work involving the composite task (Chua et al., 2015; Richler, Bukach, & Gauthier, 2009; Richler, Mack, Palmeri, & Gauthier, 2011). There was a trend-level interaction between congruency and alignment overall (F
1,27 = 3.76, MSE = 0.110, p = 0.063, η
p
2 = 0.122), but this slight interaction was not significantly different across expertise (F
1,27 = 0.572, MSE = 0.110, p = 0.456, η
p
2 = 0.021). A significant interaction between congruency and alignment signifies holistic processing in the composite task and one priori hypothesis was that this interaction would vary across latent fingerprint expertise. Separate two-way ANOVAs were done to unpack the expert and novice groups, which revealed a small but significant interaction between congruency and alignment for experts (F
1,13 = 5.000, MSE = 0.077, p = 0.0435, η
p
2 = 0.278).Footnote 1 We found no significant interaction between congruency and alignment for novices (F
1,14 = 0.564, MSE = 0.080, p = 0.465, η
p
2 = 0.039). However, given a lack of significant three-way interaction between congruency, alignment, and expertise, the significant two-way interaction between congruency and alignment for experts and the lack of thereof for novices should not be viewed as compelling evidence for differences in holistic processing due to expertise in upright fingerprints.
Response bias (criteria)
The two top panels of Fig. 3 show the response bias (criterion) values for Experiment 1 in both experts and novices. Criterion was calculated by normal-transforming the result of 1 minus the false alarm rate for each condition. The predictions for the holistic model are less clear for the criteria values, in part because they depend on complex factors such as probability matching (the tendency for participants to want to give an equal rate of “same” and “different” responses over the course of the experiment), which can interact with shifting underlying distributions as demonstrated by differences in d’. However, they do document whether participants tend to adopt more liberal or conservative response strategies for different levels of congruency or alignment. We performed a similar three-way ANOVA on criteria and only found a main effect of expertise with F
1,27 = 8.44, MSE = 0.478, p < 0.01, and η
p
2 = 0.238. An additional ANOVA on the log beta values revealed no main effect for expertise however, showing that experts were more conservative primarily because they also had higher d’ values and adjusted their criteria accordingly.
Response times
The two top panels of Fig. 4 show the mean reaction times for all trials in each of the eight conditions for Experiment 1. Holistic processing is sometimes observed in response times, but not always (Curby et al., 2013; Wong et al., 2009). Examining response times also demonstrates when participants may have made speed/accuracy tradeoffs. The reaction times showed a main effect of alignment (F
1,27 = 17.102, MSE = 14911, p < 0.01, η
p
2 = 0.388) and a trend-level effect of expertise (F
1,27 = 3.003, MSE = 295695, p = 0.094, η
p
2 = 0.100). There were no significant interactions between any variables (all p values > 0.15). The lack of interactions suggests that participants did not operate in vastly different speed/accuracy tradeoff regimes.
Summary
Overall, there is weak evidence for holistic processing as an element of expertise matching upright fingerprints. Although the interaction between congruency and alignment for d’ may appear stronger for experts, there was no significant difference when compared with the novice interaction. As expected, experts are better at the task in general, but this is difficult to attribute to holistic processing. These results contrast with other domains such as musical notation (Wong & Gauthier, 2010), cars (Curby et al., 2013), faces (Richler & Gauthier, 2014), and novel objects (Gauthier & Tarr, 2002), which demonstrated clear evidence for holistic processing.
Experiment 2: inverted fingerprints
As a further test of the hypothesis that fingerprint examiners would demonstrate holistic processing for fingerprints, we repeated Experiment 1 with a new group of experts and novices, but instead used inverted stimuli. Stimulus inversion is commonly used to disrupt holistic processing, as seen in the familiar Thatcher Illusion (Thompson, 1980). However, even with faces, the effects of inversion remain somewhat complicated, with some authors arguing that inversion does not qualitatively change the nature of perceptual processing (e.g. Sekuler, Gaspar, Gold, and Bennett, 2004). Richler et al. (2011) manipulated exposure duration and inversion of faces to identify the time-course of holistic processing effects and found that holistic processing was present for inverted faces but that it emerged later compared to upright faces. With fingerprints, Thompson and Tangen (2014) manipulated orientation and found no performance difference during a 60-s side-by-side comparison of fingerprints corrupted by noise. This result seems in contrast to from the conclusion drawn by Busey and Vanderkolk (2005), who argued that fingerprint inversion produced changes in the N170 component that were consistent with those seen when faces are inverted.
Experiment 2 is designed to assess whether inversion affects the holistic processing of fingerprints by experts specifically with the composite task, so it replicated and extended Experiment 1 with new observers and inverted fingerprints. Otherwise all procedures were identical.
Sensitivity (d’)
The two bottom panels of Fig. 2 show the d’ value for each of the four conditions separated by expertise. A clear interaction is present in the expert data, where misalignment completely eliminates any interference caused by the incongruent un-attended half. The repeated-measures, mixed-factor 2 (congruency: congruent, incongruent) × 2 (alignment: aligned, misaligned) × 2 (expertise: expert, novice) ANOVA confirmed this result.
With inverted fingerprints, there was only a trend-level effect of expertise with F
1,26 = 3.424, MSE = 1.442, p = 0.076, and η
p
2 = 0.116. There was a main effect of congruency (F
1,26 = 29.586, MSE = 0.200, p < 0.01, η
p
2 = 0.532) and this effect was not different between experts and novices (F
1,26 = 0.422, MSE = 0.200, p = 0.522, η
p
2 = 0.016) and a main effect of alignment (F
1,26 = 22.673, MSE = 0.126, p < 0.01, η
p
2 = 0.466), which was also not different across expertise (F
1,26 = 0.903, MSE = 0.126, p = 0.351, η
p
2 = 0.034). These main effects of congruency and alignment are very similar to Experiment 1.
There was a significant interaction between congruency and alignment (F
1,26 = 16.204, MSE = 0.110, p < 0.01, η
p
2 = 0.384) and a trend-level interaction between congruency, alignment, and expertise (F
1,26 = 4.207, MSE = 0.110, p = 0.050, η
p
2 = 0.139) which supports our initial observation of the graphs. Recall that this three-way interaction between congruency, alignment, and expertise that is predicted by the holistic processing hypothesis (and was not significant in Experiment 1). To specify the exact interaction that varies across expertise, we again divided the three-way ANOVA into separate two-way ANOVAs by expertise. For the expert group, there was a significant interaction between congruency and alignment (F
1,12 = 23.924, MSE = 1.896, p < 0.01, η
p
2 = 0.666); for the novice group, there was not a significant interaction between congruency and alignment (F
1,14 = 1.692, MSE = 0.231, p = 0.214, η
p
2 = 0.108). With inverted fingerprints, we see that experts find it more difficult to ignore the task-irrelevant half when it is incongruent and conflicts with the response they should be making, leading to a decrease in d’ for those trials. When the halves are misaligned, however, this interference is eliminated as shown in the lower-left panel of Fig. 2. This interaction was not present in novices and our three-way ANOVA confirms this with the trend-level interaction (p = 0.050). Possible reasons of this stronger interaction for inverted prints compared to upright prints are discussed in the “General discussion”.
Response bias (criteria)
The two bottom panels of Fig. 3 show the response bias (criterion) values for Experiment 2 in the eight conditions. We performed a similar three-way ANOVA on criteria and found a main effect of congruency with F
1,26 = 9.836, MSE = 0.086, p < 0.05, and η
p
2 = 0.274. No additional main effects or interactions between any variables were found (p values all > 0.32).
Response times
The two bottom panels of Fig. 4 show the mean reaction times for all trials for each of the eight conditions for Experiment 2. The three-way ANOVA showed the main effect of alignment (F
1,26 = 31.857, MSE = 7778, p < 0.01, η
p
2 = 0.551) and an interaction between congruency and alignment (F
1,26 = 7.573, MSE = 23987, p = 0.011, η
p
2 = 0.226), but this interaction did not interact with expertise (F
1,26 = 0.018, MSE = 3167.5, p = 0.895, η
p
2 < 0.01). This interaction shows that both novices and experts were faster when the halves were both congruent and aligned. There were no other main effects or significant interactions between any variables (all p values > 0.27).
Cross-experiment analysis
We examined d-prime values once more with a repeated measure, mixed-factor 2 (congruency: congruent, incongruent) × 2 (alignment: aligned, misaligned) × 2 (expertise: expert, novice) × 2 (orientation: upright, inverted) ANOVA to look for effects and interactions across Experiments 1 and 2. There was a between-subjects main effect of expertise (F
1,53 = 10.496, MSE = 1.238, p < 0.01, η
p
2 = 0.165), but no main effect of orientation (F
1,53 = 1.008, MSE = 1.238, p = 0.320, η
p
2 = 0.019). This is consistent with Thompson and Tangen (2014), who also found no effect of inversion on the accuracies of experts or novices. There was also a main effect of congruency (F
1,53 = 41.462, MSE = 0.255, p < 0.01, η
p
2 = 0.439) and a main effect of alignment (F
1,53 = 61.022, MSE = 0.161, p < 0.01, η
p
2 = 0.535) as well as an interaction between congruency and alignment (F
1,53 = 17.891, MSE = 0.110, p < 0.01, η
p
2 = 0.252). While this is consistent with the holistic processing hypothesis, combining across both upright and inverted fingerprints, the congruency by alignment by expertise three-way interaction was only of trend-level significance (F
1,53 = 3.974, MSE = 1.110, p = 0.051, η
p
2 = 0.070), which can be interpreted as weak support for holistic processing as a function of expertise. Aside from a trend-level interaction between alignment and orientation (F
1,53 = 3.224, MSE = 0.161, p = 0.078, η
p
2 = 0.057), there were no other main effects or interactions between factors (p > 0.13).