The own-race bias in face processing is a well-known effect that refers to the fact that people generally find it easier to identify faces of people of their own race. Although the general effect has been known for decades, the source of the bias is not well understood. There are a number of broad explanations for why it might occur; perhaps the most popular explanation is that faces of people of one’s own race are processed differently.
The popular holistic processing account suggests that our extensive experience with faces of our own race, and their distinguishing features, changes the way we process them. This is the way things work with non-face stimuli: experience with cars, for instance, enables us to ignore small feature differences and focus on the broad differences between, say, different car makes and models. Without car expertise, one would have to focus on much finer details because one would not know the ways in which features generally group together in cars. The holistic processing account suggests that faces of people of races other than our own are more difficult because we process them at a finer feature level due to having less experience discriminating them.
One major problem with the holistic processing account is vagueness. What does “holistic processing” mean, exactly? What are its necessary and sufficient characteristics? The problem of the vagueness of the explanation, along with the inconsistency of the previous literature in finding evidence for holistic processing, led Cheng-Ta Yang, Mario Fifič, Ting-Yun Chang, and Daniel Little to take a different tack in a new article in Psychonomic Bulletin & Review.
The researchers used a mathematical technique called Systems Factorial Technology (SFT) to examine how own- and other-race faces were processed among a sample of Taiwanese students. Systems Factorial Technology was applied to data from an experiment in which participants were asked to make judgments about the features of faces. The own-race faces were “Asian”, and the other-race faces were “Caucasian”.
Yang and colleagues conclude that indeed, Asian faces were processed differently by students in their sample than Caucasian faces, but the difference is not holistic versus feature based, but is rather serial versus parallel: features of the face of a person of the same race are processed simultaneously (in parallel), while features of the face of a person of another race are processed one after another (serially). Moreover, their participants tended to use more features to distinguish Asian faces than they did to distinguish Caucasian faces.
Systems Factorial Technology
To understand how Yang and colleagues came to their conclusion, we need to understand how SFT works. Suppose we were designing a novel proofreading company, and we hire three people: two proofreaders and one person (“Judy”) to compile the results of the two proofreaders and make a decision. Proofreader A (“Anna”) will read the book for grammar, and proofreader B (“Bill”) will read the book for plot mistakes. If any problems are found, the book is sent to the editing department.
How can we arrange the way Anna, Bill, and Judy work? One broad distinction we can make is between “serial” and “parallel”. Suppose Anna reads the book, passes the book with her assessment to Bill, who then reads the book, then Bill passes his judgement to Judy, who looks at Anna and Bill’s assessments. Judy sends the book to editing if either Anna or Bill flagged it. This process, shown in the top of the figure below, is called a serial, exhaustive process. It is serial because the book is passed serially from Anna to Bill; it is exhaustive because the process does not terminate early when a problem is found. If both processes take 10 minutes, the whole process takes 20 minutes.
We could also arrange the system in a parallel, exhaustive fashion. Suppose we made two copies of the book, and Anna and Bill read them in separate rooms. They then pass their assessment to Judy, who either approves it or sends it to editing. Suppose that instead of taking 10 minutes, Anna and Bill take 20 minutes to perform their assessment; then the whole process takes the same 20 minutes that we saw when each part of the serial task took half has long, because they can work simultaneously. So just by the raw amount of time, we can’t tell the difference between a serial and a parallel process.
Now suppose there are five kinds of books: books with no problems, and every combination of either glaring or subtle grammar or plot problems. When a problem is subtle, it takes five minutes longer for the proofreader to find it. What happens to the time it takes for Judy to make a decision?
In the serial arrangement, when both problems are glaring, it takes 20 minutes, as before. When only one problem is glaring, it takes 25 minutes. When both grammar and plot problems are subtle, it takes 30 minutes. The effect of the harder tasks are additive, because the tasks are performed in a chain.
In the parallel exhaustive arrangement, when both problems are glaring it again takes 20 minutes, as before. When only one problem is glaring, it again takes 25 minutes, because Judy has to wait for Bill’s assessment (the processing is exhaustive). But when both problems are subtle, it would still only take 25 minutes, instead of the 30 minutes that it would take the serial process. The times are sub-additive, because making both problems hard increases the time less than twice the time increase that would result from the same manipulation in a serial process.
Now suppose that Anna or Bill could tell Judy there was a problem immediately on finding one, and that Judy could send it immediately on. The proofreading process would not be exhaustive anymore; it would be self-terminating. The serial process is still additive, but the parallel process is no longer sub-additive. With two glaring problems, the parallel, self-terminating process again takes 20 minutes. With just one glaring problem, the proofreader with the glaring problem can tell Judy early; hence, the whole process takes 20 minutes again. But with two hard problems, it takes Judy 25 minutes to hear from both Anna and Bill. Thus, the times are super-additive: adding a second hard task increases the time taken by much more than a single one.
SFT draws on these insights about the consequences of serial and parallel processes to “reverse-engineer” the processing arrangement from observed response times.
Specifically, if we didn’t know how the proofreaders’ workflow was arranged, we could infer it by giving them books and seeing how long Judy took to make a decision. This is called the mean interaction contrast (MIC) in SFT, because it uses interactions in the average time taken to infer something about the processing. Another aspect of SFT that I have not outlined here is the survivor interaction contrast (SIC) which uses the whole response time distribution to make more subtle distinctions between processing types.
For more information about SFT, check out this 2014 article in the Psychonomic Society Journal Behavior Research Methods, where Joseph Haupt and colleagues present a package for analyzing data using SFT in R (called, appropriately, sft).
A face-processing experiment
Of course Yang and colleague’s experiment was not about books and proofreading; it was about faces. The same basic task structure applied, however, as shown in the figure below.x
Participants were asked to tell whether the eye-to-eye separation or nose-to-mouth separation was different from that of a reference face. If either were different from the reference, participants had to respond with one name (e.g., “Amy”); if both were the same as the reference (that is, the face was the reference face) then participants had to respond with another name (e.g., “Mary”). Like with our books, there are five different possible faces. The four on the left of the figure above correspond to combinations of easy and hard discriminations for our proofreaders. As with our books, it is response times to these four faces that yield tell-tale signs of the underlying processing.
Yang and colleagues found that processing times in the four types of Caucasian faces were additive, suggesting a serial process. Moreover, they also suggest that their participants processed the nose-to-mouth separation first, terminating the process if that allowed a judgment. The processing was thus inferred to be serial and self-terminating.
In contrast, the own-race faces appeared to be processed in a parallel self-terminating fashion. Using the survivor interaction contrast, the authors also inferred the processes of assessing the two target features, nose-to-mouth and eye separation, operate independent of one another.
Taken together, the results suggest that the “holistic” account — at least the strongest holistic accounts — are inconsistent with these data. As Yang and colleagues point out, the independence of the processing pathways for features of own-race faces may be consistent with a weaker form of holism, but strong holistic processing is more closely identifiable with communication across the processing pathways: so-called “co-activity”. No evidence of co-activity was found.
Yang and colleagues’ work shows the value of considering psychological theories in a more mathematically-grounded way. While “holistic processing” is hard to pin down, the more specific concepts in the Systems Factorial Technology approach allow for a finer-grained exploration of psychological processes. Surely this is something psychological science could use more of.
Psychonomics article focused on in this post:
Yang, Fifić, Chang, & Little (2017). Systems Factorial Technology provides new insights on the other-race effect. Psychonomic Bulletin & Review. doi: 10.3758/s13423-017-1305-9.