501

British Journal of Social Psychology (2011), 50, 501–518  C 2010 The British Psychological Society

The British Psychological Society www.wileyonlinelibrary.com

Prime and prejudice: Co-occurrence in the culture as a source of automatic stereotype priming Paul Verhaeghen1 ∗ , Shelley N. Aikman2 and Ana E. Van Gulick3 1

Georgia Institute of Technology, Atlanta, USA Gainesville State College, Georgia, USA 3 Vanderbilt University, Nashville, Tennessee, USA 2

It has been argued that stereotype priming (response times are faster for stereotypical word pairs, such as black-poor, than for non-stereotypical word pairs, such as blackbalmy) is partially a function of biases in the belief system inherent in the culture. In three priming experiments, we provide direct evidence for this position, showing that stereotype priming effects associated with race, gender, and age can be very well explained through objectively measured associative co-occurrence of prime and target in the culture: (a) once objective associative strength between word pairs is taken into account, stereotype priming effects disappear; (b) the relationship between response time and associative strength is identical for social primes and non-social primes. The correlation between associative-value-controlled stereotype priming and self-report measures of racism, sexism, and ageism is near zero. The racist/sexist/ageist in all of us appears to be (at least partially) a reflection of the surrounding culture.

Everyone, or so the song goes (Lopez & Marx, 2003), is a little bit racist. Or sexist. Or ageist. Scientific evidence for this claim comes from, among other things, studies that investigate stereotype priming. In such experiments, participants are first presented with a prime, typically a word denoting a social category, such as black, female, or old, and then perform some latency task on a target, a second word presented immediately after the prime. Reaction times (RT) are typically found to be faster for stereotypical prime-target pairings (such as black-poor, female-caring, and old-forgetful) than for non-stereotypical prime-target pairings (such as black-forgetful, female-poor, and oldcaring) – the stereotype priming effect. RT differences in priming effects are thought to result from activation spreading through an individual’s semantic network; faster RTs denote that the prime and the target are closely related in the individual’s mind. The assumption in stereotype priming is that the RT difference between stereotypical prime-target pairs and non-stereotypical prime-target pairs is a measure of an individual’s

∗ Correspondence should be addressed to Dr Paul Verhaeghen, School of Psychology, Georgia Institute of Technology, 654 Cherry Street, Atlanta, GA 30332, USA (e-mail: [email protected]). DOI:10.1348/014466610X524254

502

Paul Verhaeghen et al.

‘unconscious’ or ‘automatic’ or ‘implicit’ level of prejudice towards the group denoted by the prime. Implicit prejudice is quite dissociable from explicit prejudice as measured, for instance, by questionnaires. A recent meta-analysis estimates the correlation between the implicit association task (IAT), one implicit measure of prejudice, and explicit measures to be .21 (Greenwald, Poehlman, Uhlmann, & Banaji, 2009). The modest size of the correlation is typically explained by assuming that implicit measures tap an individual’s ‘true’ level of racism, whereas self-report data may not, due to self-presentation (e.g., Dovidio, Kawakami, Johnson, Johnson, & Howard, 1997; Fazio, Jackson, Dunton, & Williams, 1995), or by assuming that each stems from a different phase of social processing: implicit measures tap a reflexive gut reaction, whereas explicit processes tap a later, more reflective process of social evaluation (Gawronski & Bodenhausen, 2006). That does not imply that implicit stereotyping is innocuous; its effects might be all the more insidious for (generally) not being accessible by conscious thought (e.g., Asendorpf, Banse, & M¨ ucke, 2002; Devine, 1989; Dovidio et al., 1997; Egloff & Schmukle, 2002; Fazio, 1990; Fazio et al., 1995; Wilson, Lindsey, & Schooler, 2000). The Greenwald et al. meta-analysis obtained a correlation between implicit measures of prejudice and prejudiced behaviour (such as voting preference, non-verbal aspects of interactions, including touching, eye-contact, perceived friendliness, hostility, and physical distance chosen) of .27; explicit measures yielded a correlation with prejudiced behaviour of .33. Both types of measures possessed incremental validity, further lending credence to the argument that they tap different systems or stages of social processing. What is the origin of the implicit prejudice observable in the stereotype priming effect? Recent work in the field (reviewed by Rudman, 2004) suggests at least four routes: early experiences, affective experiences, cognitive consistency principles, and cultural biases. The former three are idiosyncratic – every individual’s developmental and learning history is unique – whereas the latter, cultural milieu, is much less so. Direct evidence for the three idiosyncratic sources is relatively easy to obtain – one can correlate early experiences with implicit attitudes in adulthood (e.g., Rudman & Goodwin, 2004; Rudman & Heppen, 2003), one can manipulate affective experiences experimentally and measure its effect on implicit attitudes (e.g., Blair, 2002), one can measure (or manipulate) cognitive consistency and note its effect on implicit attitudes (e.g., Greenwald et al., 2002). Measuring the effects of culture on implicit attitudes is less self-evident. Cultural biases have been inferred from the frequent (but not ubiquitous; see Rudman, 2004) finding that minority/disadvantaged groups show priming for stereotypes concerning their own group (e.g., Greenwald & Krieger, 2006; Nosek, Banaji, & Greenwald, 2002) – a finding that is hard to explain through any idiosyncratic mechanism. Devine, in her seminal 1989 paper, argues that cultural influence can also be inferred from an individual’s mere knowledge of the content of stereotypes. Our goal in the present set of experiments was to observe the impact of cultural biases on implicit stereotypes more directly. We were inspired by recent developments in the (cognitive) priming literature. The traditional view on priming, which we will label the semantic view (see Lucas, 2000), holds that two items in memory get connected through meaning, typically defined as an overlapping set of features. Thus, the prime-target pair dog–cat yields fast RTs because dogs and cats share many features (smallish carnivorous mammals kept as pets). The other view on priming has been labelled the associative cooccurrence view (e.g., Balota & Paul, 1996). This framework states that two items will

Prime and prejudice

503

become associated with another when they both co-occur in the natural environment. Dog primes cat because cat and dog are often encountered in identical contexts, both in the real world and in the world of media. Evidence for the co-occurrence position comes from mediated priming (e.g., lion primes stripes, although there is no feature overlap between the two entities; Balota & Lorch, 1986; Chwilla & Kolk, 2002) and from priming effects obtained with ambiguous words, such as organ, as targets (Balota & Paul, 1996). The possibility that priming might reflect associative co-occurrence rather than semantic proximity opens a potential methodological route to study culture’s impact on stereotype priming. If we take the co-occurrence view of (cognitive) priming seriously, we can surmise that stereotype priming as observed in an individual might at least partially reflect the semantic network that exists in the culture (see Arkes & Tetlock, 2004; Devine, 1989, for a related discussion). It is feasible, for instance, that stereotypical pairings are associated with a higher degree of co-occurrence in the cultural discourse. Perhaps, over a lifetime, we do see, hear, or read more instances where the word black is coupled with the word poor, or the word female with the descriptive weak, or the word old with the term fragile than instances where each of these words is coupled with, say, the word courage. Stereotype priming effects should then at least partially reflect not what is often termed an individual’s ‘personal’ (e.g., Gawronski, Peters, & LeBel, 2008) biases, but rather how well the individual has internalized the co-occurrences presented to her by her surrounding culture. The associative co-occurrence model for stereotype priming is testable. Its basic tenet – the weakest statement the model can make – is that stereotype priming effects should covary with the associative strength between prime and target. A stronger version of the model might state that stereotype priming is driven by associative cooccurrence to the same extent as standard forms of semantic priming are, at least at the group level. In other words, associative co-occurrence should have identical effects for stereotypical pairings and non-stereotypical pairings – co-occurrence in the culture builds co-occurrence in the mind, without distinguishing between the types of representations that are being correlated. In that case, we would expect the following hypotheses to hold true: (a) once objectively measured population-based associative strength between prime and target is equated, stereotype priming effects should be identical to standard lexical priming effects and (b) the assumed relationship between associative strength and the priming effect (as tested by latency by associative-value curves) should be identical for primes that are traditionally associated with stereotype priming and for primes that fall outside this realm. The stronger version is quite parsimonious (it treats stereotype priming like any other form of priming); therefore, this is the version we will test. We note that the outcome is not self-evident: it has been argued that stereotypes serve as a short-cut for social cognitive operations (e.g., Sherman, Conrey, & Groom, 2004; Wigboldus, Sherman, Franzese, & van Knippenberg, 2004) to allow humans to navigate easily through the many complexities of the social world by side-stepping intricate decision processes, leading individuals to make faster judgments about the social than the non-social world. In our experiments, then, we will investigate priming effects using both social primes able to elicit stereotypes (viz., black, white, female, male, old, young) and non-social primes (viz., night, day, summer, winter, cat, dog). The social primes will be coupled with both stereotypical targets and non-stereotypical targets of both positive and negative valence; the non-social primes will be coupled with targets of different associative strengths. Associative strength will be assessed objectively using a semantic similarity

504

Paul Verhaeghen et al.

data base compiled from a large variety of text sources assumed to reflect the lifetime body of literature read by a typical college student, our population of choice. Participants will be asked to complete one of three types of tasks, representative of the stereotype (and standard) priming literature – lexical decision (is the target a word?), evaluative judgments (does the target denote something positive or something negative?), and relatedness judgments (are prime and target related?). The three tasks can be considered as lying on a continuum of task difficulty (lexical decision implies access to the lexicon; evaluative judgment implies access to the lexicon and retrieval of valence information; relatedness judgment requires access to the lexicon for both stimuli, as well as judgment of the associative tie between them – our response times for these tasks, monotonic with assumed difficulty, bear out this expectation). One might also argue that this continuum represents decreasing degrees of implicitness of the task: subjects who are motivated to represent themselves in a non-prejudiced way might be more tempted to do so in the evaluative task than in the lexical decision task, and even more so in the associative judgment task. If we do find the same pattern of results across these tasks, this would represent strong conceptual replication. In our three experiments, we hope to reproduce the stereotype priming effect, that is, we hope to find that within social targets RTs will be faster for stereotypical pairs. At the same time, we hope to obtain evidence for the importance of associative cooccurrence, as explicated in the previous paragraph. To test whether the stereotype priming effects obtained here are indeed implicit, we will examine the correlations between these effects and scores on standard measures of explicit stereotypes, the Age Group Evaluation and Description (AGED) Inventory, The Modern Racism Scale, and the Modern Sexism Scale.

Method Our design effectively consists of three between-subject experiments, differing only in the tasks used: (a) lexical decision on the target; (b) evaluative judgments on the target; and (c) relatedness judgments on the prime-target pairing. Prime-target pairs and general aspects of the procedure are identical across experiments.

Participants Participants were undergraduate students recruited from an introductory psychology course. They received course credit in return for their participation. For the lexical decision experiment, we tested 34 participants: 50% female, 79.4% self-reported White/Caucasian ethnicity, ranging in age from 18 to 42 years (M = 20.7). For the evaluative judgment experiment, 35 participants signed up: 45.7% female, 82.9% selfreported White/Caucasian ethnicity, ranging in age from 18 to 21 years (M = 18.7). For the relatedness judgment experiment, 35 participants signed up: 51.4% female, 74.3% self-reported White/Caucasian ethnicity, ranging in age from 18 to 49 years (M = 20.2). We imposed no restrictions on the participants’ gender, age, or race/ethnicity, for two reasons: first, we did not want to sensitize participants to the theme of the study, and, second, we wanted our sample to be as representative of the typical student population as possible, given that the Touchstone Applied Science Associates (TASA) text corpus (see the Stimuli section) is aimed at this population.

Prime and prejudice

505

Stimuli We created 240 prime-target pairs out of 12 different primes and 240 different targets (these are listed in Table 1). Six of our primes were social primes that could elicit stereotypes (i.e., black, white, female, male, old, young) and six were non-social (i.e., day, night, summer, winter, cat, dog). For each of the social primes, we used 10 stereotype targets, defined as targets that have previously been used in stereotype priming experiments. Half of these had a positive valence (e.g., black-charming), half had a negative valence (e.g., black-poor). For each of the stereotype primes, we also used 10 non-stereotype targets (these were, in fact, associates generated for the non-social primes), half of these positive (e.g., black-balmy), half negative (e.g., black-mundane). For the non-stereotype primes, six members of our laboratory generated associates; associates generated by two or more members were assumed to be of high associative strength with the prime. For each of the non-stereotype primes, we used 10 highassociate targets, 5 positive (e.g., cat-independent), 5 negative (e.g., cat-sly), and 10 low-associate targets, that is, words that were associated with the stereotype primes, 5 positive (e.g., cat-memorable) and 5 negative (e.g., cat-senile).1 Associative strength for each prime-target pair was generated by the Bound Encoding of the AGgregate Language Environment model (BEAGLE; Jones & Mewhort, 2007), a computational model to represent semantic space as a set of high-dimensional vectors. BEAGLE is distinct from other such systems, like latent semantic analysis (LSA; e.g., Landauer, Foltz, & Laham, 1998) in that it takes word order into account when computing similarity statistics. BEAGLE ‘reads’ a large-scale corpus of text data, one sentence at a time, and records both context information and order information for each word in the sentence. BEAGLE stores both types of information in a vector pattern; the associative strength between two words is measured as the cosine of the vectors representing the two words. The database is dense: dimensionality of the vector space in BEAGLE is set at 1,024, and it contains 90,000 lexical entries. The corpus of text data used is the TASA corpus (Landauer et al., 1998), designed to be approximately equivalent to what the average college-level student has read in her lifetime. We note that, as expected, within the social primes, associative strengths for stereotype prime-target pairs were significantly higher than those on non-stereotype pairs: .295 and .045, respectively, t(118 df ) = 17.51, p < .001, a confound that is likely to contaminate many studies of stereotype priming. Word pairs and their associative strength values are provided in Table 1. The 240 non-lexical stimuli used in the lexical decision task were pronounceable nonwords, downloaded from the English Lexicon Project Web Site (Balota et al., 2002; http: //elexicon.wustl.edu/default.asp), between 6 and 12 letters long, with 3–10 orthographic neighbours, a bigram frequency of 2,000 to 3,500 and a response accuracy equal to or higher than .89. Examples include basner, tarren, carrow, pingers, chirling, and drater.

Prejudice-related questionnaires The AGED Inventory (Knox, Gekoski, & Kelly, 1995) measures evaluative (i.e., goodness and positiveness) and descriptive (i.e., vitality and maturity) dimensions of age group 1 Our

concern was to closely replicate experiments on stereotype priming as reported in the literature. Ideally, additional constraints on word frequency, word length, and the number of orthographic neighbours should be imposed; this would, however, have restricted the possible set of stimuli considerably. Because our generator of associative values, the BEAGLE model, is not available as a look-up table (yet), we needed to resort to a priori selection of prime-target pairs, which were then fed through the model by its originator, Michael Jones, to access their actual associative values.

506

Paul Verhaeghen et al.

Table 1. All prime-target pairs, with the BEAGLE-derived associative strength values

Prime and prejudice Table 1. (Continued)

507

508

Paul Verhaeghen et al.

Table 1. (Continued)

perceptions. In the present study, participants were asked to rate ‘the “average” or “typical” person in their seventies’ along 28 bipolar adjective pairs (e.g., independent– dependent, timid–assertive) using seven-point scales. We used the evaluative dimension as our measure of ageism. The Modern Racism Scale (McConahay, 1986) purports to measure a more subtle form of racism (distinct from more overt racism) that has been described by Sears (1988) to encompass denial of continued discrimination against, antagonism towards demands of, and resentment about special favours for African-Americans.2 The Modern Sexism Scale (Swim, Aikin, Hall, & Hunter, 1995) purports to measure a more subtle form of sexism (i.e., negative reactions towards women), distinct from overt sexism. This scale encompasses three dimensions: denial of continued discrimination 2 One

have’.

item was omitted for being obsolete: ‘Blacks have more influence upon school desegregation plans than they ought to

Prime and prejudice

509

against, antagonism towards demands of, and resentment about special favours for women.

Procedure The lexical decision experiment consisted of 480 trials: 240 trials in which a primetarget pairing was presented mixed randomly with 240 trials in which a prime-non-word pairing was presented (primes were the same as used for the prime-target pairs). In the evaluative judgment and relatedness judgment tasks, 240 prime-target trials were presented. Presentation order of trials was randomized for each individual participant. Participants were instructed to be as fast and accurate as possible, and to keep their index fingers on the response buttons at all times. Each trial consisted of the following sequence of events. First, a fixation cross appeared at the centre of the screen for 500 ms. The prime was then presented (Courier New font, size 24, bold) for 100 ms. In the lexical decision condition, primes were presented in uppercase font; in the other two conditions primes were presented in lowercase font; the discrepancy was due to a clerical error. Next, a blank screen was presented for 200 ms, immediately followed by the target, presented in a lowercase font (Courier New font, size 24, bold). The target remained onscreen until the participant logged her response by pressing either the ‘z’ or ‘/’ key. A card was placed behind the keys, identifying what each key signified: for the lexical decision task, the keys signified ‘nonword’ and ‘word’, respectively; for evaluative judgment, ‘negative’ and ‘positive’, respectively; and for relatedness judgments ‘unrelated’ and ‘related’, respectively. After a 700 ms delay, the next trial was presented. Upon conclusion of the RT task, participants filled out the prejudice-related questionnaires. Finally, they were handed a debriefing statement that explained the goals of the study and given the opportunity to ask questions.

Results Prior to all analyses, we removed RTs lying outside three interquartile ranges from the mean from the data set. This procedure removed 8.2% of the lexical decision data, 6.6% of the evaluative judgment data, and 5.0% of the relatedness judgment data. Given that the latter two tasks do not yield ‘correct’ answers – they reflect personal judgments – all non-outlying RTs, regardless of the participant’s answer, were included in all analyses. We report our findings in three sections. First, we report on the influence of associative strength on stereotypical prime-target response times compared to response times for non-stereotypical pairings – the main focus of interest. Second, we investigate the relationship between response times and associative strength within stereotype primes and non-stereotype primes. Third, we report on the relationship between stereotype priming effects and our more explicit measures of prejudice.

Stereotype priming versus non-stereotype priming To mimic the results of a typical stereotype priming experiment, we first contrasted, within the social prime-target pairs, average RT for the 60 stereotypical pairs with average

Paul Verhaeghen et al.

A 1100 Response time (ms)

1050

Stereotype vs. non-stereotype social prime pairs, associative value not controlled

B 7 Response time ln (ms)

510

1000 950 900 850 800 750 700 650 Lexical decision

1050

Evaluative judgment

6.8 6.7 6.6 6.5

Relatedness judgment

Stereotype social pairs vs. all others, overlapping associative values

Lexical decision

D

1000 950

Response time (ms)

Response time (ms)

6.9

6.4

600

C

Stereotype vs. non-stereotype social prime pairs, associative value controlled in ANCOVA

900 850 800 750 700 650 600 Lexical decision

Evaluative judgment

Relatedness judgment

1100 1050 1000 950 900 850 800 750 700 650 600

Evaluative judgment

Relatedness judgment

Racist/sexist/ageist pairs vs. all others, overlapping associative values

Lexical decision

Evaluative judgment

Relatedness judgment

Figure 1. Average response time for stereotype (white bars) and non-stereotype (hatched bars) prime-target pairs for the three experiments (lexical decision, evaluative judgment, and relatedness judgments, N = 35 for each), using three different definitions. Panel A: 60 traditional stereotype primetarget pairs contrasted with 60 non-stereotype prime-target pairs. Panel B: The same 60 traditional stereotype prime-target pairs contrasted with the same 60 non-stereotype prime-target pairs, marginal means of ln(response times) as derived from an analysis of covariance with ln(associative strength) as the covariate. Panel C: 54 stereotype prime-target pairs contrasted with 59 non-stereotype primetarget pairs with overlapping associative strength values as measured by BEAGLE, see text. Panel D: 11 negatively valenced stereotypical prime-target pairs contrasted with 102 other prime-target pairs with overlapping associative strength values as measured by BEAGLE, see text. Only in Panel A is the difference in latency between stereotypical and non-stereotypical prime-target pairs significant, indicating that control for semantic co-occurrence makes stereotype priming (statistically) disappear. Error bars represent standard error of the mean.

RT for the 60 non-stereotypical pairs. The results are presented in Figure 1, Panel A. All three t tests, one for each task, were highly significant, t(118 df ) = 5.37, 3.66, and 2.70; p < .001, .001, and .005; the average difference between stereotypical pairs and nonstereotypical pairs, that is, the stereotype priming effect, over the three experiments, was 102 ms (108 ms for lexical decision, 142 ms for evaluative judgments, and 56 ms for relatedness judgments). In a second set of analyses, we implemented three separate univariate ANCOVA (analysis of covariance) models, one for each task. In a first model, we analysed response time for stereotypical pairs versus non-stereotypical pairs, using associative strength as

Prime and prejudice

511

the covariate. All three F tests for the main effect of type of pair were now reduced to non-significance, F (1, 117) = 1.56, 3.38, and 1.63. This analysis, however, is suboptimal: ANCOVA presupposes a linear relationship between dependent variable and covariate. In the present case, this assumption is incorrect: as described in the next section, the relationship between associative strength and response time is better described by a power function. The net result is that the traditional linear analysis overestimates the effect of pair type. Therefore, in a second model, we analysed the main effect of pair type on the natural logarithm of response time with the natural logarithm of associative strength as a covariate; this transformation provides a better approximation of the data (see also Figure 2). Prior to the logarithmic transformation, associative strength was multiplied by 1,000, and then 100 was added; this linear rescaling was done to avoid negative numbers for the logarithmic transformation. The data for this model are represented in Figure 1, Panel B. As expected, F values for the main effect of type of pair were even more reduced for this model, F (1, 117) = 0.27, 1.10, and 1.30, respectively – once associative strength was partialed out, there was no longer a significant difference in response times for stereotypical pairs and non-stereotypical pairs. Our statistical control analysis thus yielded non-significant results. It is, however, also possible to control for associative value more directly by comparing stereotypical pairs with non-stereotypical pairs with overlapping values for associative strength. In a third analysis, therefore, designed to control for the fact that stereotypical primetarget pairs tend to have higher associative strengths than other prime-target pairs, we directly contrasted RTs for stereotype prime-target pairs (e.g., female-weak, malestrong) with non-stereotype prime-target pairs, both social and non-social (e.g., femalelucky, dog-smelly). We restricted the analysis to prime-target pairs with associative strengths that yielded large overlap between the two categories. We chose associative values between .15 and .50, contrasting 54 stereotypical prime-target pairs with 59 non-stereotype pairs (mean associative value for this sample of pairs was 0.30, SD = 0.08, for the stereotypical pairs and 0.29, SD = 0.09, for the non-stereotypical pairs). Results are depicted in Figure 1, Panel C; t tests revealed no significant differences in RT between the two categories for any of the three tasks, t(111 df ) = 1.30, 1.33, and −0.39; the average difference between stereotypical pairs and non-stereotypical pairs, over the three experiments, was 22 ms (32 ms for lexical decision, 42 ms for evaluative judgments, and −8 ms for semantic relatedness judgments). (We note that this analysis could not be performed on social primes only – the associative strengths for the nonstereotype prime-target pairs within social primes were generally low, leaving only 2 social prime-target pairs in the non-stereotypical category.) It could be argued that the previous analysis is not stringent enough, and that what needs to be examined are negative stereotypes for the disadvantaged groups only – truly racist, sexist, and ageist stereotypes. Therefore, in an final analysis, we directly contrasted RTs for negatively valenced stereotypical prime-target pairs for the three primes black, female, and old with all other RTs (i.e., we contrasted RTs for stimulus pairs of the type female-weak with RTs for stimulus pairs of the type female-gentle [nonnegative stereotype], female-lucky [non-stereotype], male-strong [prime not typically associated with negative stereotype], and dog-smelly or dog-striped [non-stereotype prime]). As in the previous set of analyses, the analyses were restricted to prime-target pairs with associative strengths between .15 and .50, contrasting 11 negatively valenced stereotypical prime-target pairs with 102 other pairs (mean associate values were 0.29 and 0.30, respectively, with SD = 0.09 and 0.08, respectively). Results are depicted in Figure 1, Panel D; t tests revealed no significant differences in RT between the

512

Paul Verhaeghen et al. A

Lexical decision task 1100

Non-stereotype primes Stereotype primes Power (non-stereotype primes) Power (stereotype primes)

1050

Response time (ms)

1000 950 900 850 800 750 700 650 600 0

100

200

300

400

500

600

700

800

700

800

700

800

Associative strength

B

Evaluative judgment task 1300

Response time (ms)

1200 1100 1000 900 800 700 600 0

100

200

300

400

500

600

Associative strength

C

Relatedness judgment task 1300

Response time (ms)

1200 1100 1000 900 800 700 0

100

200

300

400

500

600

Associative strength

Figure 2. The relationship between objectively measured semantic co-occurrence (associative strength in a large text corpus, as computed by BEAGLE, see text) and response time for social and non-social primes, for lexical decision (panel A), evaluative judgment (panel B), and relatedness judgment (panel C), along with freely estimated power functions for each type of prime. Statistical analysis (see text) shows that the two power functions in each display coincide, indicating that the co-occurrence–latency relationship is identical for social and non-social primes.

Prime and prejudice

513

two categories for any of the three tasks, t(111 df ) = 0.67, 0.20, and −1.08; the average difference between negative-stereotype pairs and other pairs, over the three experiments, was −1 ms (16 ms for lexical decision, 8 ms for evaluative judgments, and −28 ms for relatedness judgments).

The relationship between RT and associative strength in stereotype primes and non-stereotype primes Figure 2 plots mean RT, averaged across all individuals, for each of the 240 prime-target pairs as a function of prime-target associative strength as derived from BEAGLE, for each of the three tasks separately (for these plots and the associated regression analyses, associative strength was first multiplied by 1,000, and then 100 was added; this linear rescaling was done to facilitate regression analyses and avoid negative numbers for the logarithmic transformation discussed below). The plots suggest a non-linear, negatively accelerating relationship; we decided to fit a power function to the data, mainly because of its mathematical tractability. Overlaid on the plots are the freely estimated best-fitting power functions for each of the prime types, namely social (i.e., black, white, female, male, old and young) and non-social (i.e., day, night, summer, winter, cat, dog). To test for significant differences between the curves for stereotype (i.e., social) primes and non-stereotype (i.e., non-social) primes for each of the panels, we first linearized the data by performing a logarithmic transformation on both associative strength and RT, and then conducted three hierarchical linear regressions, one for each task. In a first step, we entered the logarithm of the associative strength to predict the logarithm of RT; in a second step, we entered a dummy variable, coding for type of prime (to test for intercept differences between the two curves) as well as the interaction between this dummy variable and the logarithm of the associative strength (to test for rate differences between the two curves). In neither of the three tasks did the second step lead to a significant increase in R2 ; for lexical decision, R2 = .000, F(2, 236) = 0.03; for evaluative judgment, R2 = .009, F(2, 236) = 1.16; for relatedness judgment, R2 = .008, F(2, 236) = 1.04. This indicates that the two curves for each plot coincide statistically. Calculating back to power functions in raw units, the three lines are: for lexical decision, RT = 1,222 (associative strength)−.083 , non-linear R = .40; for evaluative judgment, RT = 1,417 (associative strength)−.072 , non-linear R = .30; and for relatedness judgment, RT = 1,1,239 (associative strength)−.038 , non-linear R = .24. To check whether the shape of the RT by associative strength function influenced the results, we repeated these analyses using linear hierarchical regression on the raw, untransformed data. The conclusion remained the same: there is no significant difference between the lines for stereotype prime-target pairs and non-stereotype prime target pairs. Put simply: the relationship between response time and associative strength is exactly the same for stereotype prime-target pairs and non-stereotype prime-target pairs across all three tasks. This strongly suggests there is nothing ‘special’ about stereotype pairs over and beyond their higher associative value in the general culture.

Correlations between stereotype priming and self-reported racism, sexism, and ageism In a final analysis, we investigated the correlations between stereotype priming for racist, sexist, and ageist pairs, and self-reported racism, sexism, and ageism, respectively. Within each individual, we construed a racism priming scale by subtracting average RT

514

Paul Verhaeghen et al.

on the five negatively valenced stereotype targets associated with the prime black (mean associative strength of 0.42) from average RT on the five negatively valenced targets on non-stereotype primes selected to yield the same average associative strength (viz., catwicked, dog-dirty, winter-dim, night-silent, night-gloomy; mean associative strength of 0.42). This score reflects how much faster (or slower) the subject responds to racist prime-target pairs compared to non-stereotypical prime-target pairs. We constructed a sexism (mean associative strength of 0.25) and ageism (mean associative strength of 0.22) scale in analogous fashion (baseline pairs had a mean associative strength of 0.25 and 0.22, respectively). No significant correlations were found between the priming measure of racism and the Modern Racism Scale, between the priming measure of sexism and the Modern Sexism Scale, and between the priming measure of ageism and the AGED inventory for any of the three tasks (r = .04, .20, and .06 for racism; .19, .17, and −.02 for sexism, and −.12, .03, and −.29 for ageism; the average correlation is .03).

Discussion Our experiments were designed to test the conjecture (e.g., Arkes & Tetlock, 2004; Banaji, 2001; Devine, 1989; Rudman, 2004) that implicit stereotypes are at least partially acquired through the umbilical cord we all share with the surrounding culture. Our measure of implicit stereotype was stereotype priming, implemented in three types of tasks – lexical decision, evaluative judgments, and relatedness judgments. Our measure of cultural influence was associative co-occurrence between the prime-target pairs as measured objectively in a large corpus of text data assumed to be representative of the accumulated reading experience of the average undergraduate. The basic tenet of a model using associative co-occurrence as the proxy for cultural influence would be that stereotype priming effects are reduced when associative co-occurrence is taken into account; the strongest version of such model would predict that associative cooccurrence would eliminate all stereotype priming effects at the group level – that is, it would expect that co-occurrence in the culture is as successful an explaining mechanism for the implicit stereotype priming effect as it is for other, non-stereotype, priming effects. Our main results are unambiguous, and support the strong version of the cooccurrence view: there is no special status for priming related either to social stimuli or to stereotypical prime-target pairs. This suggests that at least part of the alleged racist/sexist/ageist hiding inside us all is a monster not of our own making; it is built out of memes borrowed from close contact with our environment – the books, newspapers, and magazines we read, the television shows we watch, the radio programmes we tune into. Two lines of evidence support this conclusion; each of these pieces of evidence was obtained in three independent groups of participants, each performing a different task on the same set of prime-target pairs. First, in all three groups and tasks, we replicated the standard stereotypical priming effect: participants responded faster to stereotypical pairs, such as black-poor than to non-stereotypical pairs, such as black-goofy. Judged by this criterion, our participants are, indeed, racist, sexist, and ageist. When, however, associative co-occurrence in the culture was taken into account, a different picture emerged: our participants responded equally fast to either general or negative stereotype prime-target pairs as to non-stereotype prime-target

Prime and prejudice

515

pairs. It is therefore not the stereotypical nature of the prime-target relation that drives the stereotype priming effect, it is the associative strength between prime and target. Second, the relationship between semantic co-occurrence and RT was statistically identical for social prime-target pairs and non-social prime-target pairs. Thus, there is no special status for primes associated with stereotypical responses – the mechanism that links associative co-occurrence in the culture to priming effects as observed in groups of individuals is a general one that does not make a distinction between the social or non-social nature of the primes. We note that in all three experiments, the correlations between individual stereotype priming effects (with associative co-occurrence taken into account) and standard measures of explicit prejudice were non-significant and rather small, ranging from −.29 to .20, with a mean of .03. This might suggest – although more research is necessary before this conclusion can be stated with confidence – that our co-occurrence-controlled stereotype priming measure has stronger divergent validity from explicit measures than is typically observed in the literature (Greenwald et al., 2009). One side effect of our study is that it provides further support for the associative co-occurrence view on semantic priming (Balota & Paul, 1996), and extends this view to social and stereotypical priming. We want to reiterate that our proposition that implicit stereotypes at least partially reflect internalized cultural ‘knowledge’ of stereotypes is not new. Devine (1989) demonstrated that even low-prejudice individuals (as determine by self-report) show automatic stereotype activation effects, and research using the IAT (Greenwald, McGhee, & Schwarzt, 1998) has examined the role of familiarity/strength of associations in stereotype priming effects (e.g., Fazio, Sanbonmatsu, Powell, & Kardes, 1986; Rudman, Greenwald, Mellott, & Schwartz, 1999). However, to our knowledge, no study has used objectively measured population-based associative strength values to control for associative strength, examining both stereotypic and non-stereotypic primes and stereotypic and non-stereotypic targets. We also note that although our results indicate that, at the group level, stereotype priming effects can be explained as completely as can be expected by the cultural biases captured by co-occurrence in the language (i.e., to the same extent as priming with nonsocial primes can be explained), they do not at all implicate complete explainability at the level of the individual. Idiosyncratic influences, such as early experiences, affective experiences, motivational dynamics, attentional biases, and cognitive consistency, also play a role in the formation of implicit stereotypes (e.g., Rudman, 2004; Sherman et al., 2009) – our results cannot speak to the mechanisms by which individuals internalize the cultural dynamic (for recent review, see Fiedler & Walther, 2004). These experiences, needless to say, would be inter-individually different, and would therefore show up as inter-individual variability around the mean values derived here. Given that the mean stereotype priming effect is zero (or, more precisely, −1 ms) after controlling for associative co-occurrence, these individual experiences on average appear to neither strengthen nor weaken the prejudice that is present in the culture, perhaps suggesting that as different individuals move through life, what they individually and implicitly experience as evidence for or against a stereotype ultimately balances out on the group level. Of course, our results are based on a sample of college students; the results may fall out differently when investigated within groups scoring higher (or perhaps lower) on one of more of the different types of prejudice examined here. We also note that although our experiment uses objectively derived measures of co-occurrence in the language, the source of those co-occurrences cannot be traced (at least not

516

Paul Verhaeghen et al.

experimentally), and is unlikely to be objective. That is, associations between concepts in the language do not necessarily reflect objective co-occurrences in the world – these associations are themselves construed by the speakers of the language, thus forming a cycle of prejudiced thought begetting prejudiced speech which begets prejudiced thought and so on, without a clear beginning or end point. Our experiments capture only a very specific time slice in this cultural transmission – the influence of exposure to a broad corpus of information (acquired over the course of many years) on prejudiced behaviour as measured by priming effects within a group of individuals. Our study has other limitations as well. We did not achieve complete stimulus control: stereotypical end non-stereotypical prime-target pairs should ideally all have the same (or the same narrow range of) associative values. We were not able to construct such stimuli, because stereotypical pairs have much higher associative values – one might even argue that this is what makes them stereotypical. Likewise, the ideal experiment would keep targets constant and vary primes while keeping the associative values constant; this proved undoable, for the same reason. On a final note, we want to make it very clear that the results of this study should not be used to condone prejudice, encourage intolerance, or excuse hate crimes. While it appears reasonable to state that what the automatically activated prejudice at the level of an individual may be determined to a potentially large part by the individual’s immersion in the prevailing culture, society’s influence on its individual constituents does not absolve these individuals from their own personal responsibilities and the consequences of their actions. As a great deal of research has demonstrated, regardless of the origin of stereotype priming effects, these effects have measurable real-world implications for behaviours (e.g., Fazio & Olson, 2003; Greenwald et al., 2009). The consequences of bias and prejudice are all too real, even if part of the origin of personal prejudice may be situated within the culture.

Acknowledgements We would like to thank Michael H. Jones for providing us with the BEAGLE associative strengths for the prime-target pairs used in our experiments.

References Arkes, H. R., & Tetlock, P. E. (2004). ‘Attributions of implicit prejudice, or Would Jesse Jackson “fail” the Implicit Association Test?’. Psychological Inquiry, 15, 257–278. doi:10.1207/ s15327965pli1504 01 Asendorpf, J., Banse, R., & M¨ ucke, D. (2002). Double dissociation between implicit and explicit personality self-concept: The case of shy behavior. Journal of Personality and Social Psychology, 83(2), 380–393. doi:10.1037//0022-3514.83.2.380 Balota, D. A., Cortese, M. J., Hutchison, K. A., Neely, J. H., Nelson, D., Simpson, G. B., & Treiman, R. (2002). The English Lexicon Project: A web-based repository of descriptive and behavioral measures for 40,481 English words and nonwords. Retrieved from http://elexicon.wustl.edu/, Washington University. Balota, D. A., & Lorch, R. F. (1986). Depth of automatic spreading activation: Mediated priming effects in pronunciation but not in lexical decision. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 336–345. doi:10.1037/0278-7393.12.3.336

Prime and prejudice

517

Balota, D. A., & Paul, S. T. (1996). Summation of activation: Evidence from multiple primes that converge and diverge within semantic memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 827–845. doi:10.1037/0278-7393.22.4.827 Banaji, M. R. (2001). Implicit attitudes can be measured. In H. L. Roediger, J. S. Nairne, I. Neath, & A. Surprenant (Eds.), The nature of remembering: Essays in honor of Robert G. Crowder (pp. 117–150). Washington, DC: American Psychological Association. Blair, I. (2002). The malleability of automatic stereotypes and prejudice. Personality and Social Psychology Review, 6(3), 242–261. doi:10.1207/S15327957PSPR0603 8 Chwilla, D. J., & Kolk, H. J. (2002). Three-step priming in lexical decision. Memory and Cognition, 30, 217–225. Devine, P. G. (1989). Stereotypes and prejudice: Their automatic and controlled components. Journal of Personality and Social Psychology, 56, 5–18. doi:10.1037/0022-3514.56.1.5 Dovidio, J. F., Kawakami, K., Johnson, C., Johnson, B., & Howard, A. (1997). On the nature of prejudice: Automatic and controlled processes. Journal of Experimental Social Psychology, 33, 510–540. doi:10.1006/jesp.1997.1331 Egloff, B., & Schmukle, S. (2002). Predictive validity of an implicit association test for assessing anxiety. Journal of Personality and Social Psychology, 83(6), 1441–1455. doi:10.1037/ 0022-3514.83.6.1441 Fazio, R. (1990). A practical guide to the use of response latency in social psychological research. Research methods in personality and social psychology (pp. 74–97). Thousand Oaks, CA: Sage Publications Retrieved from PsycINFO database. Fazio, R. H., Jackson, J. R., Dunton, B. C., & Williams, C. J. (1995). Variability in automatic activation as an unobstrusive measure of racial attitudes: A bona fide pipeline? Journal of Personality and Social Psychology, 69, 1013–1027. doi:10.1037/0022-3514.69.6.1013 Fazio, R. H., & Olson, M. A. (2003). Implicit measures in social cognition research: Their meaning and uses. Annual Review of Psychology, 54, 297–327. doi:10.1146/annurev.psych.54.101601. 145225 Fazio, R. H., Sanbonmatsu, D. M., Powell, M. C., & Kardes, F. R. (1986). On the automatic activation of attitudes. Journal of Personality and Social Psychology, 50, 229–238. doi:10.1037/0022-3514.50.2.229 Fiedler, K., & Walther, E. (2004). Stereotyping as inductive hypothesis testing. New York: Psychology Press. Gawronski, B., & Bodenhausen, G. V. (2006). Associative and propositional processes in evaluation: An integrative review of implicit and explicit attitude change. Psychological Bulletin, 132, 692–731. doi:10.1037/0033-2909.132.5.692 Gawronski, B., Peters, K., & LeBel, E. (2008). What makes mental associations personal or extrapersonal? Conceptual issues in the methodological debate about implicit attitude measures. Social and Personality Psychology Compass, 2(2), 1002–1023. doi:10.1111/j.1751-9004.2008. 00085.x Greenwald, A., Banaji, M., Rudman, L., Farnham, S., Nosek, B., & Mellott, D. (2002). A unified theory of implicit attitudes, stereotypes, self-esteem, and self-concept. Psychological Review, 109(1), 3–25. doi:10.1037/0033-295X.109.1.3 Greenwald, A. G., & Krieger, L. H. (2006). Implicit bias: Scientific foundations. California Law Review, 94, 945–967. doi:10.2307/20439056 Greenwald, A. G., McGhee, D. E., & Schwarzt, J. L. K. (1998). Measuring individual differences in implicit cognition: The implicit association test. Journal of Personality and Social Psychology, 74, 1464–1480. doi:10.1037/0022-3514.74.6.1464 Greenwald, A. G., Poehlman, T. A., Uhlmann, E., & Banaji, M. R. (2009). Understanding and using the Implicit Association Test: III. Meta-analysis of predictive validity. Journal of Personality and Social Psychology, 97(1), 17–41. doi:10.1037/a0015575 Jones, M. N., & Mewhort, D. J. K. (2007). Representing word meaning and order information in a composite holographic lexicon. Psychological Review, 114, 1–37. doi:10.1037/0033-295X. 114.1.1

518

Paul Verhaeghen et al.

Knox, V. J., Gekoski, W. L., & Kelly, L. E. (1995). The age group evaluation and description (AGED) inventory: A new instrument for assessing stereotypes of and attitudes toward age groups. International Journal of Aging and Human Development, 40, 31–55. Landauer, T. K., Foltz, P. W., & Laham, D. (1998). An introduction to latent semantic analysis. Discourse Processes, 25, 259–284. doi:10.1080/01638539809545028 Lopez, R., & Marx, J. (2003). Everyone’s a little bit racist. In J. Tartaglia, S. D’Abruzzo, N. V. Belcon, J. Gelber, & A. Harada (Eds.), On Avenue Q the musical: Original Broadway cast recording [CD]. New York: Max Merchandising, LLC. Lucas, M. (2000). Semantic priming without association: A meta-analytic review. Psychonomic Bulletin and Review, 7, 618–630. McConahay, J. B. (1986). Modern racism, ambivalence, and the Modern Racism Scale. In J. F. Dovidio & S. L. Gaertner (Eds.), Prejudice, discrimination, and racism (pp. 91–125). San Diego, CA: Academic Press. Nosek, B. A., Banaji, M., & Greenwald, A. G. (2002). Harvesting implicit group attitudes and beliefs from a demonstration web site. Group Dynamics: Theory, Research, and Practice, 6, 101–115. doi:10.1037/1089-2699.6.1.101 Rudman, L. A. (2004). Sources of implicit attitudes. Current Directions in Psychological Science, 13, 80–83. doi:10.1111/j.0963-7214.2004.00279.x Rudman, L. A., & Goodwin, S. A. (2004). Gender differences in automatic in-group bias: Why do women like women more than men like men? Journal of Personality and Social Psychology, 87(4), 494–509. doi:10.1037/0022-3514.87.4.494 Rudman, L. A., Greenwald, A. G., Mellott, D. S., & Schwartz, J. L. K. (1999). Measuring the automatic components of prejudice: Flexibility and generality of the implicit association test. Social Cognition, 17, 437–465. Rudman, L. A., & Heppen, J. B. (2003). Implicit romantic fantasies and women’s interest in personal power: A glass slipper effect? Personality and Social Psychology Bulletin, 29, 1357–1370. doi:10.1177/0146167203256906 Sears, D. (1988). Symbolic racism. Eliminating racism: Profiles in controversy (pp. 53–84). New York: Plenum Press Retrieved from PsycINFO database. Sherman, J. W., Conrey, F. R., & Groom, C. J. (2004). Encoding flexibility revisited: Evidence for enhanced encoding of stereotype-inconsistent information under cognitive load. Social Cognition, 22, 214–232. doi:10.1521/soco.22.2.214.35464 Sherman, J. W., Kruschke, J. K., Sherman, S. J., Percy, E. J., Petrocelli, J. V., & Conrey, F. R. (2009). Attentional processes in stereotype formation: A common model for category accentuation and illusory correlation. Journal of Personality and Social Psychology, 96, 305–323. doi:10. 1037/a0013778 Swim, J. K., Aikin, K. J., Hall, W. S., & Hunter, B. A. (1995). Sexism and racism: Old-fashioned and modern prejudices. Journal of Personality and Social Psychology, 68, 199–214. doi:10.1037/ 0022-3514.68.2.199 Wigboldus, D. H. J., Sherman, J. W., Franzese, H. L., & van Knippenberg, A. (2004). Capacity and comprehension: Spontaneous stereotyping under cognitive load. Social Cognition, 23, 292–309. doi:10.1521/soco.22.3.292.35967 Wilson, T., Lindsey, S., & Schooler, T. (2000). A model of dual attitudes. Psychological Review, 107(1), 101–126. doi:10.1037/0033-295X.107.1.101 Received 3 September 2009; revised version received 18 June 2010