SURFACE at Syracuse University SURFACE at Syracuse University

This paper studies racial and ethnic discrimination in discrete choices by real estate brokers using national audit data from the 2000 Housing Discrimination Study. It uses a fixed-effects logit model to estimate the probability that discrimination occurs and to study the causes of discrimination. The data set makes it possible to control for auditors’ actual demographic and socioeconomic characteristics, along with the characteristics assigned for the purposes of the audit. The study finds that discrimination continues to be strong but also documents a downward trend in both the scope and incidence of discrimination since 1989. The estimations also identify both brokers’ prejudice and white customers’ prejudice as causes of discrimination. ( JEL J71, R31)


Introduction
In 2003, the homeownership rate was 75.4 percent for non-Hispanic whites, but only 48.8 percent for blacks and 46.7 percent for Hispanics (U.S. Department of Housing and Urban Development 2004). 1 Many researchers believe that these disparities arise in part because of housing discrimination, which is defined as systematic unfavorable treatment of minorities.
Several studies based upon national audit data from the 1989 Housing Discrimination Study (HDS 1989) find evidence of widespread discrimination. The 2000 Housing Discrimination Study (HDS 2000), which updates and improves upon the earlier study, provides a unique opportunity to determine whether this type of discrimination persists. 2 This study also may shed light on the effectiveness of the 1988 amendments to the Fair Housing Act, which gave the federal government stronger law-enforcement powers to fight discrimination (see Yinger 1995). This paper addresses the following four questions: Does discrimination persist in the housing sales market? If so, how high is the discrimination level? Has discrimination increased or decreased over the last decade? What are the causes of discrimination? Following the literature, a fixed-effects logit model is applied to the HDS 2000 data to study discrete choices by real estate brokers, such as whether to tell the auditor that the advertised unit was available.
HDS 2000 has a feature not found in previous audit studies: it recorded some of the auditors' actual demographic and socioeconomic characteristics, such as income and education. As a result, this paper explicitly controls for an auditor's actual traits, as well as the characteristics assigned for the purpose of the audit.
The paper proceeds as follows. Section 2 presents the existing evidence on housing discrimination based upon the HDS 1989 data. Section 3 describes the HDS 2000 design. The next two sections explain the fixed-effects logit model and data. Sections 6 and 7 test the hypothesis that discrimination exists and measure the discrimination level, respectively. Section -2 -8 compares the HDS 2000 results with the HDS 1989 results (Ondrich, Stricker, and Yinger 1998). The next two sections present and test hypotheses about the sources of discrimination.
The final section summarizes the results.

Existing Evidence on Housing Discrimination
Several recent studies of housing discrimination are based on data from HDS 1989. Ondrich, Stricker, and Yinger (1998) use data from the sales audits to examine discrimination in qualitative actions taken by real estate brokers. This is the first paper to employ the fixed-effects logit model to control for the audit-specific fixed effects. They find evidence of discrimination in the housing sales market and evidence that brokers' prejudice and white customers' prejudice are the causes of this discrimination. 3 Ondrich, Ross, and Yinger (2003) explore brokers' decisions to show a unit to the white auditor, to the black auditor, or to both. They estimate a multinomial logit model and find that agents discriminate and, in suburban areas, practice redlining, defined as withholding units in integrated neighborhoods from all home seekers. They also discover that brokers appear to practice statistical discrimination, defined as treating blacks less favorably based on unfounded stereotypes about their creditworthiness or neighborhood preferences. Page (1995) uses a fixed-effects Poisson model to test for discrimination in the number of housing units shown to whites and minorities by housing agents. She finds that blacks and Hispanics are shown between 10 and 20 percent fewer units than their white teammates. Her analysis also indicates that statistical discrimination and white customers' prejudice are the major sources of discrimination. Yinger (1995) shows that an analysis of discrimination in the number of units shown needs to account for an agent's opportunity to discriminate, defined as (his or) her access to available housing units. 4 With controls for the opportunity to discriminate in the analysis, -3 -Yinger finds evidence that housing agents discriminate to protect their business with prejudiced white clients and on the basis of stereotypes about black and Hispanic customers.
Audit studies provide a powerful test for discrimination, but they are not a controlled experiment. Despite the efforts of project managers, audit teammates might differ on traits that affect their treatment. 5 Heckman and Siegelman (1993) and Heckman (1998) emphasize that unobserved auditor characteristics could bias estimates of discrimination in either direction. To address this problem, this paper explicitly controls for several true auditor characteristics. 6

The Design of the 2000 Housing Discrimination Study
HDS 2000 is an audit study aimed at determining whether minority homebuyers continue to encounter discriminatory treatment at the end of the 20 th century. 7 It follows the basic design of HDS 1989 with a few changes in auditors' assigned characteristics and treatment questions. 8 Four minority groups are studied, including black, Hispanic, Asian, and Native American. This paper focuses on the black/white and Hispanic/non-Hispanic-white (henceforth Hispanic/white, for short) audits. In the Hispanic/white audits, all minority auditors have Hispanic surnames, but they differ in skin color and accent.
HDS 2000 sampled 20 metropolitan areas, including 16 sites for the black/white audits and 10 sites for the Hispanic/white audits. 9 Overall, 1,060 black/white audits and 724 Hispanic/white audits were conducted between May 2000 and January 2001. Each audit was conducted by a pair of auditors, one non-Hispanic white and one minority, who were matched by gender and age. 10 They were assigned the same socioeconomic characteristics, such as marital status, family size, and income, to ensure that they were equally qualified for buying an advertised house. 11 Teammates also received the same training on how to behave in front of brokers. To the extent possible, in other words, the white and minority auditors were made to be identical in the agent's eyes except for their race or ethnicity. The audit manager randomly chose -4 -a housing advertisement in a major local Sunday newspaper. 12 Then audit teammates were sent to inquire about this advertisement within a short time period. After her visit, each auditor reported on the information she received from the agent and how she was treated.
Unlike HDS 1989, HDS 2000 collected extensive information on each auditor's true characteristics, such as income, education, current tenure status, and auditing experience. This type of data gives us the ability to control for differences in teammates' true traits.
As in the case of previous audit studies, results from HDS 2000 must be interpreted with care. Because newspaper advertisements form the sampling frame, the results presented here should be interpreted as a measure of discrimination that eligible minorities might come across when they ask about the units advertised in the major local newspapers. Discrimination could be higher or lower under other circumstances (and at different stages of a housing transaction).

The Fixed-Effects Logit Model
This paper uses a fixed-effects logit model as the econometric framework. The model was developed by Chamberlain (1980) as a conditional logit that accounts for the unobserved heterogeneity in a panel setting. Chamberlain's fixed-effects logit method has been used by Whittington (1992), Christian, Gupta, and Lin (1993), Korenman andWinship (1995), Fisman andRaturi (2003), and Anderson and Newell (2004). Yinger (1986) first points out that in fair housing audits, teammates could share some unobserved factors because they are assigned the same socioeconomic characteristics, go through the same training, and visit the same agency to inquire about the same advertised unit. If these common unobserved factors are correlated with right-hand-side variables, estimation results will be biased. Yinger (1998, 1999) show that the fixed-effects logit model can be used to correct the potential bias in the case of qualitative dependent variables, such as whether the auditor was told that the advertised unit was available. Following their studies, this paper focuses exclusively on qualitative dependent -5 -variables and therefore ignores important quantitative broker's actions, such as how many units were shown to each auditor.
A broker's decision about the treatment of a potential homebuyer can be represented by the following equation: where ij Y = 1 stands for favorable treatment, f is assumed to be a logistic distribution function, i is the audit index, j is the visit index, ij W equals 1 for the white auditor and 0 for the minority auditor, i α represents the audit-specific fixed effect, and ij X is a vector of explanatory variables that will be defined later. Although the order in which the white and minority auditors visited the agency was random, it simplifies the discussion to associate visit index value 1 with the white auditor and visit index value 0 with the minority auditor. This association of the visit index values is made throughout the discussion below.
In a fixed-effects logit model, i α is removed from the probability function, conditional on the sum of ij Y . The conditional probability function can be expressed by ( ) In an application of equation (2) to audit data, only a subset of the explanatory variables (identified with an asterisk) appears in , because most of the X 's are assigned to be equal across teammates. The conditioning on the sum of outcomes substantially reduces the sample size. Only the audits in which teammates are treated differently stay in the final subsample for the regression, but Chamberlain (1980) proves that this approach yields consistent -6 -estimates of the population parameters subject to mild restrictions on the rate at which the sequence of i α 's is allowed to become unbounded. Ondrich, Stricker, and Yinger (1998) show that the estimate of discrimination can be interpreted as a national average estimate if the interaction terms are expressed as deviations from the (nationally representative) sample averages. With this refinement, equation (2) becomes: ( ) δ , is to translate it into a probability measure using the assumption that m P falls short of w P by a fixed amount, d, which is called a fixed absolute gap (Ondrich, Stricker, and Yinger 1998). 15 This gap is given by d =

Data
This paper studies a variety of brokers' discrete choices in three broad categories. The first category is related to housing availability, including whether the advertised unit was available, whether similar units were available, whether the advertised unit was shown, whether similar units were shown, whether more units were recommended to an auditor than to her teammate, and whether more units were shown to an auditor than to her teammate. 16 These variables reflect crucial treatments in which discrimination directly blocks minority access to housing. The second category indicates whether the broker made an effort to speed the sale of a housing unit to the auditor. Two variables, namely, whether the broker told the auditor that she was qualified to buy a home and whether the agent made a follow-up contact, belong to this group. 17 The last category is about financing assistance, including whether the agent volunteered to help the auditor find financing, whether the agent discussed downpayment, whether the agent pre-qualified the auditor for financing, and whether the agent suggested lenders. Compared with Ondrich, Stricker, and Yinger (1998), this paper examines three new variables in housing availability (similar units available, more units recommended, and more units shown), three new variables in financing assistance (downpayment discussed, pre-qualified buyer for financing, and lenders suggested), and one new variable in sales effort (qualified auditor for buying). 18 Tables 1 and 2 present the incidence of treatments, with no statistical controls, for the black/white and Hispanic/white audits, respectively. In each table, the first two entries are the weighted shares of audits in which favorable action was taken for whites and for minorities, respectively. The final entry records the difference between these two shares, often called the net incidence of unfavorable treatment. For the black/white audits, all net incidence measures are positive, ranging from 0.007 (or 0.7 percent) for advertised unit available to 0.125 for more units shown. For the Hispanic/white audits, all net incidence measures are positive except for -8 -advertised unit available, advertised unit inspected, and more units shown. The other net incidence values fall between 0.011 for follow-up contact made and 0.111 for financial help offered. Table 3 lists the explanatory variables in our data set. These variables can be classified into four groups: basic variables, auditors' true characteristics, month and site dummies, and neighborhood characteristics for the advertised unit. Each of these groups can be entered as a difference between teammates ( * in equation (3)) or as the value for the white teammate relative to the national average ( 1 in equation (3)). The first versions of these variables insulate the estimates of discrimination from bias due to differences in observable teammate characteristics; the second versions help test hypotheses about the causes of discrimination. The links to such hypotheses are explained in Section 9.
The basic variables include auditor characteristics, such as age, sex, and assigned income; agent characteristics such as race, sex, and age; and agency characteristics, such as whether a multiple listing service was used. True characteristics cover the auditor's actual socioeconomic information along with home seeking and auditing experience. The socioeconomic information includes income, education, employment, and immigration status. Three dummy variables indicate the auditor's home seeking experience, including whether she lived in the audit metropolitan area, whether she was a homeowner, and whether she was actually hunting for a home. Neighborhood characteristics include racial and ethnic composition, median house value, per capita income, and percentage of owner-occupied housing units for the census tract in which the advertised unit was located. 19 Differences between teammates in their actual characteristics might affect differences in their treatment, so an analysis that controls for these differences might paint a different picture of unfavorable treatment than the simple percentages in Tables 1 and 2. An auditor with experience in home seeking might be treated better, for example, because she knows more about the local -9 -housing market and the buying process and can ask better questions. Auditing experience may help the auditor present herself in a more professional way and may also lead to better treatment.
The addition of these variables therefore represents a significant advance over Ondrich, Stricker, and Yinger (1998). Although the explanatory variables other than the auditor's true characteristics are similar to those in Ondrich, Stricker, and Yinger (1998), we also add three new neighborhood characteristics (median house value, per capita income, and percentage of owner-occupied housing units) and one new agency characteristic (whether the agency used the Internet). 20 Audit teammates shared assigned factors but differed in their true characteristics and in the characteristics of the agents they encountered. Tables 4 and 5 provide information on the magnitude of these differences. Table 4 shows that, compared with their teammates, black auditors had higher actual incomes, and higher probabilities of having a job and of actually hunting for a home, while white auditors had more education and were more likely to live in the audit metropolitan area. Table 5 shows that compared with their teammates, Hispanic auditors also had higher actual incomes and a higher probability of being employed, while white auditors had more education and were more likely to be homeowners. It also demonstrates that 20 percent of Hispanic auditors had a discernible accent and that the difference in darkness of skin tone between Hispanics and whites is significant. Overall, white and minority teammates did not encounter agents with significantly different characteristics, but they did differ on several true characteristics-differences that need to be considered in estimating discrimination.

Testing the Hypothesis that Discrimination Exists
Our first major question is: "Does housing discrimination exist?" The answer to this question is affirmative if the estimated w δ in equation (3) is positive and statistically significant.
Tables 6 and 7 present the estimation results for each type of treatment for the black/white and -10 -Hispanic/white audits, respectively. The first column of each table is the number of observations, that is, the number of audits in which teammates were treated differently. Other entries are the estimated values of w δ with different sets of explanatory variables. The first estimates (in the second column) are based on regressions with the basic variables, entered as both the difference in the variable between teammates and the value for the white auditor (relative to the national mean). The estimates in the third column add differences in audit teammates' true characteristics; when compared with the estimates in the second column, these estimates indicate whether the inclusion of these differences alters the estimate of w δ . The estimates in the fourth column add white auditors' true characteristics (expressed as a deviation from the weighted sample mean).
The other columns are based on regressions that successively add month and site dummies, and neighborhood characteristics, all interacted with race variable and expressed as deviations from the national average. Once added, each block of variables is retained in subsequent columns.
Any estimate with a p-value below 5 percent for a two-tailed test is regarded as statistically significant.
First, consider the estimates of w δ in the last column of Table 6, which we believe are the most reliable estimates. For the black/white audits, in the category of housing availability, ˆw δ is positive and significant for similar units inspected, more units recommended, and more units shown. Discrimination is also found in sales effort for qualified auditor for buying and in financing assistance for downpayment discussed and pre-qualified buyer for financing.
Most of these results for the black/white audits are robust to changes in the explanatory variables. The six significant estimates of w δ in the last column also have the p-values below 5 percent in all previous columns. For these variables, the estimated magnitudes are larger in the last column than in the second column for every result except for more units recommended; that is, adding controls tends to raise the estimated level of discrimination. In contrast, the addition -11 -of differences in teammates' true characteristics (that is, moving from the second column to the third column) raises the p-value from below 5 percent to above 5 percent in two cases: similar units available and follow-up contact made. For these variables, a failure to observe these variables could lead to an overstatement of the magnitude and statistical significance of discrimination. 21 For the Hispanic/white audits, we find less evidence of discrimination. No estimate of w δ in the last column of Table 7 is significant for a variable describing housing availability or sales effort. In the category of financing assistance, ˆw δ is positive and significant for financial help offered and lenders suggested. The result for downpayment discussed is significant at the two-tailed 10 percent level, which provides weak evidence of discrimination.
The results for financial help offered and lenders suggested are quite robust across specifications, but some other results are strongly affected by the addition of controls. Adding differences in teammates' true characteristics pushes the p-values for two variables, namely, similar units available and pre-qualified buyer for financing, from below 5 percent to above 5 percent, and the addition of other variables raises these p-values even more. Another variable, follow-up contact made, almost becomes significant when teammate differences in true characteristics are included (with a p-value of 5.1 percent), but it is not close to significant in a regression that includes all the explanatory variables. Moreover, the result for downpayment discussed, which is highly significant with basic variables, is not affected by the inclusion of teammate differences but has a p-value of only 8.2 percent in the final column. These results remind us that controls for differences in auditors' true characteristics can shift the results in either direction and that other controls are needed, as well.
Overall, the results demonstrate the continuing existence of housing discrimination.
Blacks face discrimination in a wide range of agents' actions, whereas Hispanics are treated unfairly with regard to financing assistance. For the most part, brokers do not block minorities -12 -from gaining access to the advertised units, but they continue to take discriminatory actions in recommending and showing similar units. This may reflect the behavior, documented by Ondrich, Ross, and Yinger (2003), that brokers advertise the units they are most willing to sell to minorities while strictly controlling other houses.

The Probability of Discrimination
Our next question concerns the probability that a minority home seeker encounters discrimination. Tables 8 and 9 present the probability measures derived earlier for the subset of agent's actions involving discrimination (as indicated by a positive and significant ˆw δ ) for the black/white and Hispanic/white audits, respectively. In each table, the first entry reports the simple net incidence measure with no controls, which is copied from Table 1 or 2, and the second entry shows the fixed absolute gap measure.
The results for the black/white audits in Table 8 indicate that the estimated fixed absolute gap ranges from 14.6 percent (similar units inspected) to 39.5 percent (pre-qualified buyer for financing). Except in the case of similar units inspected, the estimates are always above 25 percent, which shows that blacks still face a disturbingly high probability of encountering discrimination for a wide range of brokers' actions. These results also indicate that the multivariate estimate of the probability of discrimination significantly exceeds the simple net incidence measure for all agents' actions. Table 9 presents the results for three dependent variables for the Hispanic/white audits.
The fixed absolute gap ranges from 10.8 percent (downpayment discussed) to 33.0 percent (financial help offered). These results indicate that Hispanic home seekers face a probability above 10 percent of encountering discrimination in several aspects of financial assistance provided by brokers. Moreover, as in the case of the black/white audits, adding statistical controls raises the estimated probability of discrimination.

The Trend in Discrimination
Comparing the results for HDS 2000 and HDS 1989 sheds light on the trend in discrimination over the last decade. This paper and Ondrich, Stricker, and Yinger (1998)  in the fixed absolute gap measure is between 6.9 percentage points (follow-up contact made) and 11.2 percentage points (advertised unit available). In one case (similar units inspected), however, the fixed absolute gap measure actually increases by 0.9 percentage points.
As shown in Table 11, Hispanic home seekers also experienced a decline in the probability of discrimination, except in the case of financial help offered. For this financing variable, the increase between 1989 and 2000 is 6.7 percentage points for the simple net incidence and 26.1 percentage points for the fixed absolute gap. This is a substantial increase in the incidence of discrimination. In contrast, the probability of discrimination declines substantially for the other four agents' actions in this table, with declines of 3 to 9 percentage points for the simple net incidence and declines between 10 to 30 percentage points for the fixed absolute gap.

Hypotheses about the Causes of Discrimination
Three main causes of housing discrimination have been discussed in the literature: brokers' prejudice, white customers' prejudice, and statistical discrimination (see Yinger 1995;Ondrich, Ross, and Yinger 2003). The broker-prejudice hypothesis states that real estate agents discriminate to satisfy their own prejudice. Because measures of brokers' prejudice are not available, this hypothesis cannot be tested directly. Nevertheless, some indirect tests are available because brokers' prejudice is likely to vary with other brokers' characteristics. First, minority agents are likely to have less prejudice against their own racial or ethnic group members and thus, to discriminate less. Second, prejudice increases with the agent's age and is stronger for men than for women, so older, male agents are likely to discriminate more than younger, female agents (see Schuman, Steeh, and Bobo 1985). Third, the broker who has -15 -resources and market power is more likely to act upon her prejudice. This paper uses whether the broker has units similar to the advertised unit and maximum number of people encountered by either auditor at the agency as proxies for the flexibility and the size of the agency.
Brokers also may have stronger prejudice against minorities who have certain characteristics. A higher level of discrimination against younger, black men than against older black men or against black women could be a sign that brokers' prejudice builds on stereotypes about the propensity of younger black men to commit crimes. In addition, brokers may be more likely to feel prejudice (and practice discrimination) against Hispanics with certain traits, such as a heavy accent or a dark skin tone.
The customer-prejudice hypothesis postulates that some brokers discriminate against minority homebuyers in order to satisfy their white clients' prejudice and thereby to preserve their current and future business with their prejudiced white customer base. The share of prejudiced whites among a broker's customers cannot be observed, of course, but we can observe where the advertised unit is located, and the location of this unit provides information about the likely location of the broker's customer base.
Prejudiced whites are opposed to neighborhood integration and may be especially upset about the entry of blacks or Hispanics into their neighborhoods when the areas are at risk of tipping, which will result in neighborhood racial transition. Black neighborhoods and Hispanic neighborhoods may be closer to the tipping point than white neighborhoods. Therefore, the customer-prejudice hypothesis predicts that discrimination against blacks or Hispanics will be higher when the advertised unit is located in a neighborhood that may tip. This hypothesis, however, does not imply higher discrimination against Hispanic home seekers in a neighborhood with many blacks, or black home seekers in a neighborhood with many Hispanics. In the first case, prejudiced whites may be more concerned with new black entries than with new Hispanic entries; in the latter case, prejudiced whites may be more concerned with new Hispanic entries.
-16 -Therefore, the customer-prejudice hypothesis is consistent with lower discrimination against blacks in Hispanic neighborhoods and lower discrimination against Hispanics in black neighborhoods. Ondrich, Stricker, and Yinger (1998) also propose that neighborhoods full of prejudiced white homeowners are more threatened by the entry of blacks than are white renter neighborhoods, so this hypothesis also predicts that discrimination increases with the percentage of housing units that are owner-occupied. We extend this logic to consider neighborhood incomes and house values. In other words, we hypothesize that the concerns of prejudiced white customers about the entry of minorities into their neighborhood increase with house values and incomes, and we test this hypothesis by determining whether discrimination is higher in neighborhoods where average house values and incomes are higher.
The prejudice of a broker's white clients may also depend on the characteristics of the minority homebuyer. White customers may have stronger prejudice against minorities who are younger or who have low incomes, children, a heavy accent, or dark skin. An aversion to minority families with children, for example, might arise because of concerns about school integration. The broker-prejudice hypothesis and the customer-prejudice hypothesis make the same prediction for three of these variables (auditor's age, accent, and skin color), so the estimated coefficients on these three variables cannot be used to distinguish between these two hypotheses.
Finally, agents' incentives to cater to the prejudice of their white clients may depend on their own characteristics. If a broker works for a large real estate agency, for example, she is less likely to be restricted to a set of prejudiced white clients, and an older broker may have a more established reputation that is less threatened by perceptions about a single transaction involving a minority purchaser. These examples suggest that older brokers and brokers in large agencies are less likely to discriminate. Moreover, if a broker works in an agency using a multiple listing -17 -directory or the Internet to serve customers, she can employ these tools to steer minorities away from the neighborhood where the sale may offend hostile white clients and therefore may feel less need to discriminate in other types of actions.
The last hypothesis is that agents practice statistical discrimination, defined as using membership in a certain group as a signal about unobserved preferences or constraints that might influence the broker's profits. Brokers may, for example, presume either that minority home seekers prefer living near people in their own racial or ethnic group instead of in a largely white neighborhood, or that lenders refuse to grant loans to minorities in white neighborhoods. If so, brokers may believe that showing housing in white neighborhoods to minority customers is a poor use of their time.
As explained earlier, our data set contains information on auditors' true characteristics.
These characteristics are difficult to link to hypotheses about the causes of discrimination largely because they cannot be directly observed by agents (with the exception of accent and skin tone for Hispanics). Nevertheless, auditors might send signals about these characteristics during their conversations with brokers, or these characteristics might be proxies for observable traits, such as articulateness or aggressiveness. Because we do not know whether these characteristics, or variables correlated with them, are observed by brokers, we cannot clearly link them to any hypotheses. As a result, we include interactions with these characteristics in our regressions to see whether they are associated with variation in discrimination, but we regard these interaction variables as exploratory and interpret them cautiously and on a case-by-case basis.

Testing Hypotheses about the Causes of Discrimination
We test hypotheses about the causes of discrimination through the interaction terms in equation (3). Our results are presented in Tables 12 (black/white audits) and 13 (Hispanic/white audits). These tables only present results for agents' actions that appear to involve -18 -discrimination. In addition, the results in these tables are based on regressions with a complete set of interaction variables. Virtually all of the significant results in these tables are also significant if they are included in more parsimonious regressions.
The six regressions in Table 12  As shown in Table 13, only a few interaction terms are significant for the Hispanic/white audits. For financial help offered, we find that older Hispanic home seekers encounter less discrimination than younger ones, which is consistent with both the broker-prejudice and customer-prejudice hypotheses. For downpayment discussed, a Hispanic auditor with a heavy -19 -accent faces a higher probability of discrimination than one with no accent, which supports both the broker-prejudice and customer-prejudice hypotheses, as well. Finally, for lenders suggested, the negative sign on the variable indicating that the broker uses the Internet reinforces the comparable result for the black/white audits and is consistent with incentives linked to white customers' prejudice.
In both Tables 12 and 13, several true auditor characteristics have significant coefficients.
For the black/white audits, actual homeowners encounter less discrimination for both similar units inspected and more units shown. Although information on actual homeownership was not directly observed by brokers, it may have been indirectly revealed to them during the interview.
Compared with renters, for example, homeowners are likely to be more familiar with the process of buying a house and to know more about the local housing market. It is important to note, however, that this variable is not picking up a difference in true homeownership between audit teammates; instead, it indicates that differences in treatment between blacks and whites are not as great when both auditors seem to know more about the local housing market (or otherwise reveal their actual homeownership). To put it another way, the signals that come from being an actual homeowner, whatever they are, have a larger impact on the treatment of blacks than on the treatment of whites.
In addition, Table 12 reveals that the level of discrimination for both downpayment discussed and pre-qualified buyer for financing declines when both teammates were foreign-born.
Agents may be able to infer an auditor's foreign birth from something that is said in the conversation-or by hearing an accent. This result suggests that they can make this inference and that it has a more favorable impact on black than on white auditors. One possible explanation of this finding is that brokers' prejudice (or their anticipated prejudice from white customers) is linked to black people who grow up in America, not to foreign-born blacks. Table 13 indicates that, for the Hispanic/white audits, auditing experience is associated with more discrimination for both financial help offered and lenders suggested. The literal interpretation of this result is that when teammates both had auditing experience the Hispanic auditor was more likely to be discriminated against. However, we do not believe that is the story.
This result is difficult to interpret because neither the conversation between the auditor and the broker nor the auditor's behavior during the audit can reveal anything about the auditor's auditing experience information to the broker. As a result, there is no reason to believe that auditing experience itself triggered more unfavorable treatments for Hispanics than for whites.
Instead, auditing experience might be linked to the accuracy with which the auditors filled out the survey forms. If so, this result might indicate that improved reporting quality as a result of previous auditing experience may be more significant for Hispanics than for whites, which could help uncover more cases of discrimination. 23 Finally, we find significant interaction terms for three true auditor characteristics, but we cannot explain what these findings mean. Specifically, for the black/white audits (Table 12), we find that discrimination in pre-qualified buyer for financing increases with the level of auditors' education. In the case of the Hispanic/white audits (Table 13) discrimination in financial help offered increases with auditors' true incomes, and discrimination in downpayment discussed decreases when both auditors actually lived in the audit metropolitan area.
In summary, our results support the hypotheses that both brokers' prejudice and the prejudice of brokers' white customers are the causes of housing discrimination, but we do not rule out the possibility that other causes are at work. Moreover, these results also indicate that the causes of housing discrimination may vary from one type of brokers' behavior to the next and are not necessarily the same for blacks and Hispanics. These findings are consistent with those of Ondrich, Stricker, and Yinger (1998) for HDS 1989.

Conclusions
Our analysis of the data from HDS 2000 indicates that black and Hispanic home seekers still encounter discrimination in the housing sales market. Indeed, for some types of brokers' behavior, the probability of discrimination is still disturbingly high. Nevertheless, we also find that both the scope of discrimination and the probability that it will be encountered in any particular agent's action have diminished sharply since 1989. This finding indicates that the housing market situation has improved for black and Hispanic buyers over the last fifteen years.
One possible explanation for this improvement is the enactment of the 1988 amendments to the Fair Housing Act, which significantly boosted the federal government's enforcement powers.
We also find that discrimination still appears to be caused by both brokers' prejudice and white customers' prejudice, although we cannot rule out other possible explanations. This finding indicates that there is an ongoing role both for education, which may help to eliminate brokers' prejudice, and active anti-discrimination enforcement, which may help to offset the economic incentives that apparently lead some brokers to discriminate. Finally, we find that the addition of auditors' true characteristics sometimes has a significant impact on the estimated probability of discrimination and that these characteristics are correlated with some types of broker discrimination. Further investigation of these findings is clearly warranted.     a. p-value is the level of significance for difference of means for whites and blacks. b. Auditor's true annual income is coded as 1=under $10,000, 2=$10,000-19,999, 3=$20,000-29,999, 4=$30,000-39,999, 5=$40,000-49,999, 6=$50,000-74,999, 7=$75,000-100,000, and 8=over $100,000. c. Agent's age is coded as 1=18-30, 2=31-45, 3=46-65, and 4=over 65. a. p-value is the level of significance for difference of means for whites and Hispanics. b. Auditor's true annual income is coded as 1=under $10,000, 2=$10,000-19,999, 3=$20,000-29,999, 4=$30,000-39,999, 5=$40,000-49,999, 6=$50,000-74,999, 7=$75,000-100,000, and 8=over $100,000. c. Auditor's skin tone is coded as an integer between 0 and 3, where 0=white and 3=the highest darkness degree. d. Agent's age is coded as 1=18-30, 2=31-45, 3=46-65, and 4=over 65. -0.061 (0.762) Notes: a. The first row of Columns 2-6 lists the specifications of the explanatory variable set. Each specification includes the ones designated by all previous columns and the explanatory variable block designated by the current column. See Table 3 for details of each explanatory variable block. "Differences" are differences between teammates. "Interactions" are values for the white auditor (relative to the national average). See equation (3). b. The cells of Columns 2-6 give the estimated values of δ w from equation (3). p-values are in parentheses. c. Number of observations=number of audits in which teammates were treated differently.  Table 3 for details of each explanatory variable block. "Differences" are differences between teammates. "Interactions" are values for the white auditor (relative to the national average). See equation (3). b. The cells of Columns 2-6 give the estimated values of δ w from equation (3). p-values are in parentheses. c. Number of observations=number of audits in which teammates were treated differently. Notes: a. Net incidence=the share of audits in which favorable action was taken for white auditors minus the share of audits in which favorable action was taken for black auditors. b. Fixed absolute gap=fixed amount by which the probability of favorable treatment of blacks falls short of the probability of favorable treatment of whites.  Notes: a. * stands for significance at the two-tailed 5 percent level. b. Net incidence=the share of audits in which favorable action was taken for white auditors minus the share of audits in which favorable action was taken for black auditors. c. Fixed absolute gap=fixed amount by which the probability of favorable treatment of blacks falls short of the probability of favorable treatment of whites. d. The HDS 1989 results come from Table 3 of Ondrich, Stricker, and Yinger (1998), with a correction of typographical error of the fixed absolute gap for advertised unit available. The value in their table was 0.134, whereas the correct value is 0.119. Notes: a. * stands for significance at the two-tailed 5 percent level. b. Net incidence=the share of audits in which favorable action was taken for white auditors minus the share of audits in which favorable action was taken for Hispanic auditors. c. Fixed absolute gap=fixed amount by which the probability of favorable treatment of Hispanics falls short of the probability of favorable treatment of whites. d. The HDS 1989 results come from Table 4 of Ondrich, Stricker, and Yinger (1998).  Table 3, both in the form of teammate differences (if they exist) and in the form of values for the white auditor (expressed as a deviation from the weighted national mean). See equation (3). Only the second form of the variable is reported here, since it is the one that tests hypotheses about the causes of discrimination. b. The white auditor/similar units available interaction is dropped out of the regression for similar units inspected because there is no variation in the variable when the value of the dependent variable is 1.  (2002). This paper builds on Chapter 7 of that report and goes beyond it thanks to additional data cleaning and the use of more explanatory variables.
3. Ondrich, Stricker, and Yinger (1999) apply the fixed-effects logit model to the rental audit data. They study landlords' discrete choices and find evidence of discrimination.
Their work indicates that landlords discriminate against black and Hispanic renters based upon their own prejudice and white tenants' prejudice. Yinger (1995) also provides a survey of the audit-based literature on discrimination in housing before HDS 1989.

5.
For a review of steps to minimize this possibility, see Yinger (1995) or Turner et al. (2002).

6.
Controlling for the auditor's true characteristics does not completely eliminate the problem because there still exist other unobservable characteristics. For an alternative approach to this problem, see Ondrich, Ross, and Yinger (2003).
-38 -7. HDS 2000 is composed of rental audits and sales audits, which are aimed at studying discrimination in the rental housing market and the sales housing market, respectively.
This paper focuses on the sales part.

8.
In HDS 1989, for example, some auditors were assigned to be homeowners while others were assigned to be renters. In HDS 2000 all auditors' tenure status was assigned to be renter.

9.
The black/white audits were conducted in Atlanta, Austin, Birmingham, Chicago, Washington D.C., Denver, Dayton, Detroit, Houston, Los Angeles, Macon County, New Orleans, New York City, Orlando, Philadelphia, and Pittsburgh. The Hispanic/white audits were conducted in Austin, Chicago, Denver, Houston, Los Angeles, New York City, Pueblo, San Antonio, San Diego, and Tucson. These sites were also used in HDS 1989.

10.
The ages of the white and minority auditors are close but not identical, so we control for age differences in our regressions.

11.
The minority auditor was always assigned a slightly higher income to avoid the real estate agent's suspicion and to ensure that unfavorable treatment received by the minority auditor did not result from the lower income. The non-random income difference across teammates is not controlled for in our estimations, which might result in understating discrimination, but we think that the impact would be small because the income differences are small and brokers usually did not ask about auditors' incomes.
12. HDS 2000 also conducted some audits that were not based on an advertisement in a major newspaper. These audits are not considered here, but they exhibit similar levels of discrimination. See Turner et al. (2002).
A weighing system was used to account for the HDS 2000 sampling design. See Turner et al. (2002). 14.
Because it compares the average treatment of white and minority auditors, δw is called a "net" measure of discrimination. As shown by Ondrich, Ross, and Yinger (2000), a net measure may understate discrimination, but can be interpreted as a lower bound estimate.
15. Ondrich, Stricker, and Yinger (1998) also derive an identical measure of the probability of discrimination by assuming that there is a fixed percentage gap between Pw and Pm.

16.
A similar unit is defined as a housing unit that has the same number of bedrooms as the advertised unit.

17.
A follow-up contact can be a telephone call to the auditor at home, a telephone message left at the auditor's home, a voicemail message, a postal mail, or an E-mail.
18. Ondrich, Stricker, and Yinger (1998) also study whether the broker asked about the auditor's income, whether the broker asked about the auditor's housing needs, and whether the broker invited the auditor to call back. This paper does not examine these dependent variables for the following reasons: asking the auditor about income, defined as unfavorable treatment, can be argued to be a nondiscriminatory routine action by brokers; there is no specific question about whether the auditor was asked about housing needs in the HDS 2000 report forms; an invitation to call back may not be as important as other dependent variables and also the 1989 study does not find evidence of discrimination against blacks in that behavior.

19.
We define a black neighborhood and a Hispanic neighborhood as a census tract with a population more than 15 percent black and Hispanic, respectively. We have also tried 20-, 25-, and 30-percent dividing lines, black and Hispanic percentages, and whether the -40 -combined black and Hispanic percentage is above 5, 15, or 30-all of which turn out to have no or weak explanatory power.
20. Ondrich, Stricker, and Yinger (1998) also consider whether the broker located the office in a white neighborhood and whether the broker advertised a unit located in a central city.
They argue that brokers' prejudice may simultaneously determine that she took discriminatory action, located the office in a white neighborhood, and did not advertise a unit from a minorities-concentrated central city. These explanatory variables, however, are not included in this paper for the following reasons: first, the tract number of the broker's office is not coded in HDS 2000, which makes measuring racial and ethnic composition of the agency's neighborhood impossible; second, the 1989 study does not find any indication that brokers reveal their prejudice through their office-setting and advertising choices. Hence, missing brokers' office-setting and advertising variables are not believed to affect our results.

21.
An alternative way to think about this issue is to ask whether estimations that add month and site dummies and neighborhood characteristics (which were available in 1989) to the basic variables lead to the same inferences as the "full-information" estimations in the last columns of Tables 6 and 7. The answer to this question is affirmative for follow-up contact made (black/white audits) and pre-qualified buyer for financing (Hispanic/white audits), but not for similar units available (both types of audit).

22.
It is also consistent with the view that non-Hispanic residents in a neighborhood with some Hispanics do not want their community to be predominated by Hispanics and thus, they would like to accept people from other minority groups, so their community could become more diverse and balanced.
-41 -23. Yinger (1995, p. 24) indicates "some early audit studies discovered that minority auditors who encountered blatant unfavorable treatment became upset and were unable to complete their audit forms in an accurate manner, thereby invalidating some audits in which discrimination was the most severe." So a Hispanic auditor with auditing experience might be better able to complete the report accurately, which could help find more discrimination.