We tested the set of predictions under (1) using linear regression with the individual scales as dependent variables and the two conditions and respondent information as predictors (we elaborate on this later). Predictions (i)a-(i)d are not supported by our results. However, we found a significant difference in ratings on the masculine-feminine scale depending on the sound condition; the raised clip is rated as less feminine (p < 0.5). We return to this result in the discussion. It is worth noting, however, that respondents reacted differently to the raised and the lowered clip on this scale, even if the direction of the effect is the inverse of what was predicted. Priming did not have any effect on the ratings for the entire respondent set. Respondents who self-identified as gay (n = 28) were significantly more likely to give both clips a higher rating on the straight-gay scale if they were primed (p < 0.5).
The set of predictions under (2) cannot be straightforwardly tested. Essentially, they are assumptions on the shape of the response set, that is, the way the responses on the individual scales correlate with each other (we elaborate on this later). It is possible, however, to say whether the predictions are confirmed by the shape of the response set. Prediction (2)a is supported by our results. Higher ratings on straight-gay and masculine-feminine pattern together, and these also pattern together with higher ratings on positive-negative scales like rich-poor and handsome-ugly. This remains true irrespective of respondent background. For instance, women and men pattern together in their responses. Gay respondents are even less forgiving—their high ratings on straight-gay and masculine-feminine pattern together with higher ratings on even more positive-negative scales. This relates to the fact that these respondents are more perceptible to priming.
EFFECTS OF CONDITIONS
We analyzed the effect of condition on ratings in the following way. We looked at individual scales (like masculine-feminine) and used stepwise linear regression to determine whether the information we have about the listener (age, gender, and background; which clip he or she listened to, whether he or she was primed) affects the result. For a given scale, the dependent variable was the score on the scale, and the predictors were the condition, the age, gender, and years of education of the respondent, as well as the town in which the respondent went to high school.
We had predictions for three scales, so we report results that have a significance value smaller than 0.5, following standard practice. Since we had no hypotheses for the other scales, we would have only reported effects that are twelve times stronger (since there are twelve scales)—any effect smaller than that we attribute to chance alone.
Pitch has an effect on ratings on the masculine-feminine scale (but not on the short-tall, straight-gay scales). The direction of this effect is, however, the opposite of what we expected: the raised clip is perceived as more masculine (p < 0.5). The only relevant predictor is sound (i.e., the pitch difference). Figure 8.1 shows a density plot of the answers.
If we focus on the gay respondents (n = 28), we find that sound had no effect on their ratings. The prime, however, did—gay respondents gave higher ratings on the straight-gay scale if they were primed. The only relevant predictor is prime. Figure 8.2 shows a density plot of the answers. We must note, however, that the sample size is much smaller for gay respondents (28 respondents vs. 873 overall), which affects the strength of the predictors (though the effect remains significant).
CORRELATIONS OF RATINGS
The second hypothesis relates not to the correlation of the scores with the conditions but rather to the correlation of the scores with each other. These
figure 8.1 Overlapping histogram of rating scores for masculine-feminine, all respondents
figure 8.2 Overlapping histogram of rating scores for straight-gay, gay respondents correlations are informative of how respondents perceived a male voice. While the correlations can be influenced by the conditions (and the other predictors), this is less relevant here.
In order to visualize correlations between ratings, we first look at the correlation matrices in the data set. Respondents gave two kinds of ratings—they rated the stimulus (the sound clip) on twelve scales, and they used the same twelve scales to rate themselves. We will focus on ratings on the stimulus.
figure 8.3 Correlations of ratings on the sound clip with each other, all respondents
Figure 8.3 shows the correlations of the stimuli ratings with each other for all respondents. The size of the circle in each cell indicates the strength of the correlation between the scale indicated in the first column and the scale indicated in the column headers. Only the second attribute of each scale is shown (so that masculine-feminine is feminine). Positive correlations are black, negative correlations are gray. We can see strong patterns in the responses. Tall has a strong negative correlation with feminine. In turn, feminine has a strong positive correlation with gay, ugly, and poor, whereas unfriendly has a positive correlation with unfaithful, indifferent, and, curiously enough, passionate. (Recall that the respondents were rating a male voice.)
Though Figure 8.3 sheds some light on the correlation patterns, the way they cluster together is far from evident. Furthermore, this figure says nothing about the effect of the conditions and the other predictors (like respondent gender or sexual orientation) on these correlations.
We used principal components analysis (PCA) to address these issues. PCA takes a dataset consisting of interrelated variables (like our ratings) and turns it into a dataset consisting of the same number (in this example) of independent variables. These variables are set up in such a way that the first variable explains the most amount of variation in the dataset, the second variable the second largest amount of variation, and so on. These new, independent variables (the principal components) can be input to the original dataset to see how they pattern together with the variables they are based on (our ratings) as well as the subject-level predictors (like the age and gender of the listeners).
A principal components analysis of the stimuli ratings of our respondents yields a set of principal components reflecting the correlations of the original ratings. The first two of these components are responsible for 24% of the variation. They set up the two main dimensions according to which the data are structured. We can visualize these two main dimensions by the means of a biplot (Figure 8.4).
Figure 8.4 shows that ratings pattern together very strongly on two main dimensions. First, component 1 pits indifferent, unfriendly, and passionate against natural and clever. To put it simply, according to the respondents, the person they heard is either typically the former (indifferent, unfriendly, etc.) or the latter (natural and clever). Component 2 shows us something we have seen in the correlation matrix—tall ratings are contrasted with feminine, ugly, gay, and, to a lesser extent, poor ratings (i.e., high ratings on these scales). The introverted scale does not do much on these two dimensions. Our predictors (sound and prime, as well as respondent age, gender, etc.) do not affect these patterns much. If we plot the principal components for female and male respondents, we get almost the same picture. Independently from the stimuli, the prime, and their backgrounds, respondents all agreed on a response pattern that invokes the image of a clever, natural Hungarian man, compared to an unfriendly, passionate, indifferent
figure 8.4 Principal components analysis of stimuli ratings, all respondents
figure 8.5 Principal components analysis of stimuli ratings, gay respondents
one, and that of a tall and rich Hungarian man, compared to a feminine, gay, ugly, and poor one.
One notable exception for the indifference of subject-level variables is the respondents’ sexual orientation. Gay respondents present a different correlation pattern, shown in Figure 8.5. The two dimensions that are typical of all the respondents are less distinct here. Basically, feminine and gay pattern together with not just ugly and poor, but all the negative attributes.