It seems suspicious that no direct causal connection between race and inc was found. Recall, however, that these are the probabilistic relationships among the responses; they are not necessarily the probabilistic relationships among the actual events. There is a problem with using responses on surveys to represent occurrences in nature because subjects may not respond accurately. Let’s as-sume race is recorded accurately. The actual causal relationship between race, inc, and says_inc may be as shown in Figure 11.11. By inc we now mean whether there really was an incident, and by says_inc we mean the survey
Figure 11.10: The hidden node DAG pattern Tetrad III learned from the racial harassment survey at the .01 significance level.
response. It could be that races, which experienced higher rates of harass-ment, were less likely to report the incident, and the causal influence of race on says_inc through inc was negated by the direct influence of race on inc. This would be a case in which faithfulness is violated similar to the situation involving finasteride discussed in Section 2.6.2. The previous conjecture is sub-stantiated by another study. Stangor et al  found that minority members were more likely to attribute a negative outcome to discrimination when re-sponses were recorded privately, but less likely to report discrimination when they had to express their opinion publicly and there was a member of the non-minority group present. Although the survey of military personnel was intended to be confidential, minority members in the military may have had similar feel-ings about reporting discrimination to the army as the subjects in the study in [Stangor et al, 2002] had about reporting it in the presence of a non-minority individual.