The notion of implicit bias has taken root in North America and influential politicians like Hillary Clinton or FBI director James Comey used the idea to understand persistent racism and prejudice in the United States (Greenwald, 2015).
The main idea of implicit bias is that most White Americans have negative associations about Blacks that influence their behaviors without their awareness. This explains why even Americans who hold egalitarian values and do not want to discriminate end up discriminating against Black Americans.
The idea of implicit bias emerged in experimental social psychology in the 1980s. Until then most academic psychologists dismissed Freudian ideas of unconscious processes. However, research in cognitive psychology with computerized tasks suggested that some behaviors may be directly guided by unconscious processes that cannot be controlled by our conscious and may even influence behavior without our awareness (Greenwald, 1992).
Some examples of these unconscious processes are physiological processes (breathing), highly automated behaviors (driving while talking to a friend), and basic cognitive processes (e.g., color perception). These processes differ from cognitive tasks like adding 2 + 3 + 5 or deciding what take out food to order tonight. There is no controversy about this distinction. The controversial and novel suggestion was that prejudice could work like color perception. We automatically notice skin color and our unconscious guides our actions based on this information. Eventually the term implicit bias was coined to refer to automatic prejudice.
To provide evidence for implicit bias, experimental social psychologists adopted experiments from cognitive psychology to study prejudice. For example, one procedure is to present racial stimuli on a computer screen very quickly and immediately replace them with some neutral stimulus to prevent participants from actually seeing the stimulus. This method is called subliminal (below-threshold of awareness) priming.
Some highly cited studies suggested that subliminal priming influences behaviour without awareness (Bargh et al., 1996; Devine, 1989). However, in the past decade it has become apparent that these results are not credible (Schimmack, 2020). The reason is that social psychologists did not use the scientific method properly. Instead of using experiments to examine whether an effect exists, they only looked for evidence that shows an effect. Studies that failed to show the expected effects of subliminal priming were simply not reported. As a result, even incredible subliminal priming studies that reversed the order of cause and effect were successful (Bem, 2011). In the 2010s, some courageous researchers started publish replication failures (Doyen et al., 2012). They were attacked for doing so because it was a well-known secrete among experimental social psychologists that many studies fail, but you were not supposed to tell anybody about it. In short, the evidence that started the implicit revolution (Greenwald & Banaji, 2017) is invalid and casts a shadow over the whole notion of prejudice without awareness.
Measuring Implicit Bias
In the 1990s, experimental psychologists started developing methods to measure individuals’ implicit biases. The most prominent method is the Implicit Association Test (IAT, Greenwald et al., 1998) that has produced a large literature with thousands of studies that used the IAT to measure attitudes towards the self (self-esteem), exercise, political candidates, etc. etc. However, the most important literature with the IAT are studies of implicit bias. In these studies, White Americans tend to show a clear preference for Whites over Black Americans. This preference can also be shown with self-ratings. However, a notable group of participants shows much stronger preferences for Whites with the IAT than in their self-ratings. This finding has been used to claim that some White Americans are more prejudice than their are aware off.
One problem with the IAT and other measures of implicit bias is that they are not very good. That is, an individual’s test score is much more strongly influenced by measurement error than by their implicit bias. One way to demonstrate this is to examine the reliability of IAT scores. A good measure should produce similar results when it is used twice (e.g., two Covid-19 tests should be both positive or negative, not one positive and one negative). Reliability can be assessed by examining the correlation of two IATs. A correlation of r = .5 would imply that there is a 75% chance for somebody to score above average on both tests and a 25% chance to get conflicting results (i.e., above and below average).
Experimental social psychologists rarely examines reliability because most of their studies are cross-sectional ( a single experimental session lasting from 10 minutes to 1 hour). However, a few studies with repeated measurements provide some information. Short intervals are preferable to avoid any real changes in implicit bias. Bar-Anan and Nosek (2014) reported a retest-correlation of r = .4, for tests taken within a few hours. Lai et al. (2016) conducted the largest study with several hundred participants for tests taken within a few days. The retest correlations ranged from .22 to .30. Even two similar, but not identical, race IATs in the same session produce low correlations, r ~ .2 (Cunningham et al., 2001). More extensive psychometric analysis further suggest that some of the variance in implicit bias measures is systematic measurement error that influences one type of measure, but not other measures (Schimmack, 2019). Longitudinal studies over several years further show that the reliable variance in IATs is highly stable over time (Onyeador et al., 2020).
In short, ample evidence suggests that most of the variance in implicit bias measures is measurement error. This has important implications for research with these measures that tries to change implicit bias or use implicit bias measures to predict behaviors. However, experimental social psychologists have ignored these implications when they implicitly assumed that their measures are perfectly valid.
The Numbers do not add up
Some simple math shows the problems for experimental social psychologists to study implicit bias. The main method to study implicit bias is to conduct experiments where participants are randomly assigned to two or more groups. Each group receives a different treatment and then the effects on an implicit bias measure and actual behaviors are observed. For illustrative purposes, I assume that manipulations actually have a moderate effect size of half a standard deviation (d = .5) on implicit bias. However, because only a small proportion of the variance in the implicit bias measures is valid (here the assumption is a generous .5^2 = 25%), the effect that an experimental social psychologist could observe is only .25 standard deviations. That is, measurement error cuts the actual effect size in half. The effect on an actual behavior is even smaller because the link between attitudes and a single behavior is also small, d = .5 * .3 = .15. Thus, even under favorable conditions, experimental social psychologists can only expect to observe small effect sizes.
A good scientist would plan studies to be able to reliably detect these small effect sizes. Cohen (1988) provided guidelines for scientists how to plan sample sizes that make it possible to detect these small effects. A so-called power analysis shows that N = 500 participants are needed to detect an effect size of d = .25 and 1,400 participants are needed to detected an effect size of d = .15 for behavior.
However, experimental social psychologists tend to conduct studies with much smaller sample, often fewer than 100 participants. With N = 100, they would have only a 25% chance to reliably (with a p-value below .05) detect an effect and the observed effect size would be severely inflated because the significant result can only be significant with an inflated effect size estimate. Thus, we would expect many non-significant results in the implicit bias literature. However, we do not see these results because experimental social psychologists did not report their failures.
Implicit Bias Intervention Studies
For 20 years, experimental social psychologists have reported studies that seemed to change implicit bias (Dasgupta & Greenwald, 2001; Kawakami, Dovidio, Moll, Hermsen, Russin, 2000). The most influential article was Dasgupta and Greenwald’s (2001) article with nearly 700 citations. As this article spanned an entire literature, it is worthwhile to take a closer look at it.
There were two studies, but only Study 1 focused on implicit race bias. The sample size was N = 48. These 48 participants were divided into three groups, leaving n = 18 per group. Aside from a control group, one group was shown positive example of Blacks and negative examples of Whites and another group was shown the reverse. To get a significant result for the extreme comparison of the opposing groups, we have a study with 36 participants. To have an 80% chance to get a significant result for this contrast, an observed difference of d = .96 is needed. Taking measurement error into account this requires a change in implicit bias by 2 standard deviations. Otherwise, a non-significant result is likely and the study is risky.
Surprisingly, the authors did find a very strong effect size for their manipulation, d = 1.29. They even found a significant difference with the control group, d = .58.
As shown in Figure 1, Panel A, results revealed that exposure to pro-Black exemplars had a substantial effect on automatic racial associations (or the IAT effect).5 The magnitude of the automatic White preference effect was significantly smaller immediately after exposure to pro-Black exemplars (IAT effect = 78 ms; d = 0.58) compared with nonracial exemplars (IAT effect = 174
ms; d = 1.15), F(1, 31) = 6.79, p = .01; or pro-White exemplars (IAT effect = 176 ms; d = 1.29), F(1, 31) = 5.23, p = .029. IAT effects in control and pro-White conditions were statistically comparable
(F < 11)
Dasgupta and Greenwald not only wanted to show an immediate effect. They also wanted to show that this effect can last at least for a short time. Thus, they repeated the measurement a second day. The problem is that they now need to show two significant results, when they have a relatively low chance to show even one. The risk of failure therefore increased considerably, but they were successful again.
Panel B of Figure 1 illustrates the response latency data 24 hr after exemplar exposure. Compared with the control condition, the magnitude of the IAT effect in the pro-Black condition remained significantly diminished 1 day after encountering admired Black and disliked White images (IAT effects = 126 ms vs. 51 ms, respectively; ds = 0.98 vs. 0.38, respectively), F(1, 31) = 4.16, p = .05. Similarly, compared with the pro-White condition, the IAT effect in the pro-Black exemplar condition remained substantially smaller as well (IAT effects = 107 vs. 51 ms, respectively;
ds = 1.06 vs. 0.38, respectively), F(1, 31) = 3.67, p = .065.
Nobody cared about p-values that are strictly not significant (p = .05, p = .068), but these days these p-values are considered red flags that may suggest the use of questionable research practices to find significance. Another sign of questionable practices is when multiple tests are all successful because each test produces a new opportunity for failure. Thus, the fact that everything always works in experimental social psychology is a sign of widespread abuse of the scientific method (Sterling, 1959; Schimmack, 2012).
Study 2 did not examine racial bias, but it is relevant because it presents more statistical tests. If they also show the desired results, we have additional evidence that QRPs were used. Study 2 examined prejudice towards old people. Notably, the reported study did not have a control group as in Study 1, thus there is only a comparison of manipulations with favorable old people versus favorable young people. Study 2 also did not bother to examine whether the changes last for a day, or at least there were no results reported if this was examined. Thus, there is only one statistical test and that was significant with p = .03.
As illustrated in Figure 2, exposure to pro-elderly exemplars yielded a substantially smaller automatic age bias effect (IAT effect = 182 ms, d = 1.23) than exposure to pro-young exemplars
(IAT effect = 336 ms, d = 1.75), F ( 1 , 24) = 5.13, p = .03.
Over the past decade, meta-scientists have developed new tools to examine the presence of questionable practices even in small sets of studies. One test examines the variability of p-values as a function of sampling error (TIVA). After converting p-values into z-scores, we would expect a variance of 1, but the variance is only 0.05. This outcome has only a probability of 1 out of 180 times to occur by chance. Even if we are conservative and make this 1 out of 100, Dasgupta and Greenwald were extremely lucky to get significant results in all of their critical tests. We can also examine the power of their studies given the reported test statistics. The average observed power is 56%, yet they had 100% successes. This suggests that QRPs were used to inflate the success rate. This test is extremely conservative because mean observed power is also inflated by the use of QRPs. A simple correction is to subtract the inflation (100% – 56% = 44%) from the observed mean power. This yields a corrected replicability index of 56% – 44% = 12%. A replicability index of 21% is obtained when there is actually no effect.
In short, power analyses and bias tests suggest that Dasgupta and Greenwald’s article contains no empirical evidence that simple experimental manipulations can produce lasting changes in implicit bias. Yet, this article suggested to other experimental social psychologists that changing IAT scores is relatively easy and worthwhile. This generated a large literature with hundreds of studies. Next we are going to examine what we can learn from 20 years of research with over 40,000 participants.
A Z-Curve Analysis of Implicit Bias Intervention Studies
Psychologists often use meta-analyses to make sense of a literature. The implicit bias literature is no exception (Forscher et al., 2019; Kurdi et al., 2019). The problem with traditional meta-analyses is that they are uninformative. Their main purpose is to claim that an effect exists and to provide an average effect size estimate that nobody cares about. Take the meta-analysis by Forscher et al. (2019) as an example. After finding as many published and unpublished studies as possible, the results are converted into effect size estimates to end up with the conclusion that
“implicit measures can be changed, but effects are often relatively weak (|ds| < .30).
What do we do with this information. After all, Dasgupta and Greenwald (2001) reported an effect size of d > 1. Does this mean, they had a more powerful manipulation or does this mean their results were inflated by QRPs?
Traditional meta-analysis suffers from two problems. First, unlike medical meta-analysis where manipulations represent a treatment with the same drug, social psychologists use very different manipulations to change implicit bias ranging from living with a Black roommate for a semester to subliminal presentation of stimuli on a computer screen. Not surprisingly there is evidence of heterogeneity, that is, effect sizes vary, making any conclusions about the average effect size meaningless. What we really want to know is which manipulations reliably can produce the largest changes in implicit attitudes.
The next problem of this meta -analysis is that it did not differentiate between IATs. Implicit measures of attitudes towards alcohol or consumer products were treated the same as implicit bias. Thus, the average results may not hold for implicit bias.
The biggest problem is that meta-analysis in psychology do not take publication bias into account. Either they do not even examine it or, as in this case, they find evidence for publication bias, but don’t correct conclusions accordingly.
“we found that procedures that directly or indirectly targeted associations, depleted mental resources, or induced goals all changed implicit measures relative to neutral procedures” (p. 541).
It is not clear whether this conclusion holds after taking publication bias into account. Meta-scientists have developed better tools to examine and correct for the influence of questionable research practices that inflate effect sizes (QRP, John et al., 2012). A simulation study found that z-curve is superior to several alternative methods (Brunner & Schimmack, 2020). Thus, I conducted a z-curve analysis of the literature on implicit bias interventions.
The meta-analysis by Forscher et al. (2019) was very helpful to find studies until 2014. I also looked for newer studies that cited Dasgupta and Greenwald (2001), the seminal study in this field. I did not bother to get data from unpublished studies or dissertations. The reason is that these sources are only included in traditional meta-analysis to give the illusion that all studies were included and that there is no bias. However, original researchers who used QRPs are not going to share their failed studies. Z-curve can correct bias for the published studies and does not require cooperation from original researchers to correct the scientific record.
I found 214 studies with 49,1145 participants (data). Figure 1 shows the z-curve. A z-curve is a histogram of the reported test-statistics converted into z-scores. Each z-score reflects the strength of evidence (effect size over sampling error) against the null-hypothesis in each study. As the direction of the effect is irrelevant, all z-scores are positive.
The first notable finding is that the peak of the distribution is at z = 1.96, which corresponds to a two-sided p-value of .05. The second finding is the sharp drop from the peak to values below 1.96. The third observation is that the peak of the distribution has a density of 1.1, which is much larger than the peak density of a standard normal distribution (~ .4). All of these results together make it clear that non-significant results are missing. To quantify the amount of bias due to the use of QRPs, we can compare the observed discovery rate (the percentage of significant results) with the expected discovery rate based on the z-curve model (the grey curve is the predicted distribution without QRPs). The literature contains 74% significant results, when we would expect only 8% significant results.
Thus, there is strong evidence that QRPs undermine the credibility of this literature. Especially, p-values like those reported by Dasgupta and Greenwald (2001) are often a sign of studies with low power that required QRPs to produce a p-value less than .05 (see values below x-axis, 12% for z-scores 2 to 2.5). However, there is also clear evidence of heterogeneity. Studies with z-scores greater than 4 are expected to replicate with 90% or more (again values below x-axis) and 6 studies are not shown because their z-scores even exceeded the maximum value of 6 on the x-axis. To give a context, particle physicists use a z-score of 5 to claim major discoveries. Thus, a few studies produced credible evidence, while the bulk of studies used QRPs to achieve statistical significance in studies with low power.
There are two remarkable articles in this literature that deserve closer attention (Lai et al., 2014, 2016). Before I examine these two articles in more detail, I also conducted a z-curve analysis of the literature without these two articles to examine the credibility of typical articles in this literature.
The z-curve plot for traditional articles in this literature looks even worse. The expected discovery rate of 7% is just above the discovery rate of 5% that is expected from studies without any effect simply because the alpha criterion of .05 allows for 5% false positive discoveries. Moreover, the 95% confidence interval of the expected replication rate does include 5%, which means we cannot rule out that all of the published studies with significant results are false positives. This is also reflected in the maximum False Discovery Rate, 73%, but the upper limit of the 95% confidence interval includes 100%.
While there may be two or three studies with credible evidence, 154 studies with nearly 20,000 participants have produced no scientific information about implicit bias. In short, like several other areas of research in experimental social psychology, implicit bias research is junk science and the seminal study by Dasgupta and Greenwald is no excpetion.
Exception No 1: Lai et al. (2014)
The IAT is a popular measure of implicit bias in part because the developers of the IAT created an online site where visitors can get feedback on their (invalid) IAT scores, including the race IAT. This website is called Project Implicit. Some also volunteer to be participants in studies with the IAT. This makes it possible to get large samples. Lai et al. (2014) used Project Implicit to conduct 50 studies with 18 different interventions. Each study had several hundred participants, which allows for higher power to get significant results and more precise effect size estimates. The next figure shows the z-curve for these 50 studies.
Visual inspection of the histogram does not show the previous steep cliff around z = 1.96. In addition, the replication rate for significant studies is high and the lower limit of the 95%CI is still 65%. Thus, even if some minor QRPs may have produced a little bump around 1.96, this article provides credible evidence that IAT scores can be changed with some manipulations. However, it also shows that several manipulations produce hardly any effects.
Moreover, it is possible that the little bump around 1.96 is a chance finding. This can be examined by fitting z-curve to all values, including no-significant ones. Now the estimated discovery rate perfectly matches the observed discovery rate, suggesting that no QRPs were used.
In short, a single study with well-powered studies that honestly reported results provided more informative results than a literature with hundreds of underpowered studies that used QRPs to publish significant results. This just shows how powerful real science can be, while at the same time exposing the flaws of the way most experimental social psychologists to this day conduct their research.
Do Successful Changes of IAT scores Reveal Changes in Implicit Bias?
If we think about measures as perfect representations of constructs, any change in a measure implies that we changed the construct. However, Figure 1 showed that we need to distinguish measures and constructs. This brings up a new question. Did Lai et al. successfully change implicit biases or did they merely change IAT scores without changing attitudes.
This question can be difficult to answer. One way to examine this would be to see whether the manipulation also influenced behaviour. In the Figure a change of actual implicit bias would also produce a change in behavior, whereas the direct effect on the measure (red path) would not imply a change in behavior. However, as we saw studies with actual behaviors require even larger samples than used in the Project Implicit studies. So, this information is not available.
This brings us to the second exceptional study, which was also conducted by Lai and colleagues (2016). It is essentially a replication and extension of their first study. Focussing on the successful intervention in Lai et al. (2014), the authors examined whether the immediate effects would persist for a few days. First, the authors successfully replicated the immediate effects. More important, they failed to find significant effects a few days later, despite high power to do so. Even participants who were trained to fake the IAT did not bother to fake the IAT again the second time. Thus, even successful interventions that change IAT scores do not seem to change implicit biases measured with the IAT.
Don’t just trust me. Even Greenwald himself has declared that there are no proven ways to change implicit bias, although he fails to explain how he obtained strong effects in his seminal study.
“Importantly, there are no such situational interventions that have been established to have durable effects on IAT measures (Lai et al., 2016)” (Rae and Greenwald, 2017).
“None of the eight effective interventions produced an effect that persisted after a delay of one or a few days.This lack of persistence was not previously known because more than 90% of prior intervention studies had considered changes only within a single experimental session (Lai et al. 2013).” (Greenwald and Lai, 2020).
In short, 20 years of research that started with strong and persistent effects in Dasgupta and Greenwald’s seminal article has produced no useful information how to change implicit bias, despite hundreds of articles that claimed to change implicit bias successfully.
Where do we go from here?
Based on the famous saying “insanity is doing the same thing over and over again and expecting different results” we have to declare experimental social psychologists insane. For decades they have tried to make a contribution to the understanding of prejudice by bringing White students at White universities into labs run by mostly White professors, expose them to some stimuli and measured prejudice right afterwards. The only things that changed is that social psychologists now do even shorter studies with larger samples over the Internet. Should anybody expect that a brief manipulation can have profound effects? The only people who think this could work are social psychologists who have been deluded by inflated effect sizes in p-hacked studies that even subliminal manipulations can have profound effects on prejudice. Meanwhile, racisms remains a troubling reality in the United States as the summer in 2020 made clear.
It is time to use research funding wisely and not to waste it on experimental social psychology that is more concerned with publications and citations than with affecting real change. Resources need to be invested in longitudinal studies, studies with children, studies at work places with real outcome measures. Right now, this research does not attract funding because researchers who pump out five quick, p-hacked experiments get more publications, funding, and positions than researchers who do one well-designed longitudinal study that may fail to show a statistically significant result. Junk is drowning out good science. Maybe a new administration that actually cares about racial justice will allocate research money more wisely. Meanwhile, experimental social psychologists need to rethink their research practices and wonder what their real priorities are. As a group, they can either continue to do meaningless research or step up. However, they can no longer deceive themselves or others that their past research made a real contribution. Denial is not an answer, unless they want to take a place next to Trump in history. Publishing only studies that work was a big mistake. It is time to own up to it.
Onyeador, I. N., Wittlin, N. M., Burke, S. E., Dovidio, J. F., Perry, S. P., Hardeman, R. R., … van Ryn, M. (2020). The Value of Interracial Contact for Reducing Anti-Black Bias Among Non-Black Physicians: A Cognitive Habits and Growth Evaluation (CHANGE) Study Report. Psychological Science, 31(1), 18–30. https://doi.org/10.1177/0956797619879139