the distribution that is truncated rises, the racial gap in the saving rate and capital gains tends to narrow. We now turn to multivariate analysis to explore whether the
racial differences remain after controlling for other characteristics.
D. Multivariate Analysis
It is well known that multivariate analysis using wealth data poses special chal- lenges because of the skewness of the wealth distribution. In order to avoid the undue
inuence of extreme observations, researchers have sometimes transformed the data by taking logarithms. Analysts who take this step must then make a decision with
respect to the large proportion of observations that cannot be transformed because they are zero or negative, with some electing to truncate the sample, and others
arbitrarily assigning a value, such as zero, for the log of zero or negative wealth Carroll, Dynan, and Krane 1999. Irrespective of one’s view of such an approach,
the use of a log transformation is not practical for an analysis of changes in wealth because of an even larger proportion of the sample with zeroes or negative numbers.
For instance, for the 1984– 94 longitudinal sample, about 15 percent of the observa- tions have zeroes or negative numbers for the level of wealth in 1984, while the
corresponding share for changes in wealth is more than twice that.
Instead of a log transformation we make use of two complementary approaches: The rst is quantile regression, which is not sensitive to outliers and thus can be
used on the full longitudinal samples to examine the determinants of the rates of saving, capital gains and wealth appreciation at their medians. The second is OLS,
which will be employed on the longitudinal samples with the top and bottom percen- tiles trimmed for an examination of tendencies at the means. For the analysis of
saving rates, we rst run a baseline regression where a dummy variable for African Americans is the only independent variable, then add controls for income to see if
any racial differences can be accounted for by income differences, and then include additional controls for age, education and sex of head, change in marital status, and
number of children. The same controls are used for an analysis of inheritances. When the rate of return to capital and the rate of wealth appreciation are the dependent
variables, controls for wealth and its square at the start of the period are also included.
In Columns 1 and 4 of Table 6, we rst reproduce on the basis of OLS and median regressions, respectively, the earlier nding that with no controls whites save a higher
share of their income than do African Americans. However, after accounting for income, we nd that this relationship no longer holds, as shown in Columns 2 and
5. In fact, for two of the three median regressions, the coefcients for African Ameri- cans become positive and signicant. As already noted, these results are consistent
with those of Dynan, Skinner, and Zeldes 2000, who use the Survey of Consumer Finances and the Consumer Expenditure Survey in addition to the PSID to document
a positive relationship between personal saving rates and lifetime incomes. As shown in Columns 3 and 6, adding the remaining demographic controls leaves none of the
coefcients signicant, with the changes in the median regressions largely attribut- able to the fact, that conditional on income, saving rates rise somewhat faster with
age for African Americans than for whites.
Regression analysis of racial differences in mean rates of return on capital, sum- marized in Table 7, is hampered by the fact that returns to capital gains can be
Table 6 Racial Differences in Saving Rates African American2White, 1984– 94
Type of regression
OLS OLS
OLS Median
Median Median
Period 1984– 89
20.052 20.012
0.003 20.034
0.003 0.005
0.029 0.029
0.029 0.007
0.005 0.005
1989– 94 20.024
0.027 0.008
20.022 0.013
20.007 0.026
0.026 0.026
0.008 0.006
0.007 1984– 94
20.037 0.003
0.004 20.031
0.010 0.003
0.020 0.020
0.021 0.008
0.005 0.007
Other Income 1
Income 1 regressors None
Income others
None Income
others
Notes: The saving rate is dened as the ratio of saving over the period divided by family income over the same
time span. The racial difference is the coefcient on a dummy variable for African Americans. Standard errors are in parentheses. Total family income over the time span and its square are the income variables,
while the other regressors are age of head and its square, sex and education of head, marital status at start and end of period, and number of children. The trimmed longitudinal samples are used for OLS and the
untrimmed ones for median regressions for details, see the appendix on the JHR website. Regressions use PSID family weights multiplied by income at start of period.
signicant at 10 percent level signicant at 5 percent level
signicant at 1 percent level
calculated only with respect to positive wealth. Thus, we are forced to exclude from the OLS analysis households with nonpositive wealth—about 15 to 17 percent
of the sample, depending upon the period.
13
Calculations not shown here reveal that the racial difference in mean rates of return to capital is somewhat smaller when
only positive wealth holders are included. With this caveat in mind, we nd that the coefcient for African Americans is positive in the baseline OLS regressions, though
the estimate is too imprecise to be signicant at conventional levels. Likewise, as we found in Table 5, signicant differences in rates of return are not evident at the
median. This situation scarcely changes when additional controls are added, as the race coefcient is almost always positive but almost never estimated precisely
enough to achieve statistical signicance.
For our multivariate analysis of inheritances, we forego median regressions, be- cause only a small portion of the sample received an intergenerational transfer during
the periods. We use tobit regressions instead of OLS because inheritances are a censored variable, not at zero in this case but at 10,000 in nominal currency, the
threshold used in the PSID questionnaire. In light of substantial differences by race in the share who receive inheritances 12 percent for whites and 1 percent for African
13. These observations are also excluded in the median regressions and in the median calculations for Table 5, so that there is no inconsistency in comparing the two sets of regression results.
Table 7 Racial Differences in the Rate of Return on Capital African American2White,
1984– 94
Type of regression
OLS OLS
OLS Median
Median Median
Period 1984– 89
0.048 0.043
20.043 0.000
0.014 0.014
0.121 0.121
0.121 0.003
0.010 0.019
1989– 94 0.106
0.082 0.056
0.000 0.001
0.007 0.142
0.142 0.143
0.000 0.000
0.008 1984– 94
0.107 0.102
0.022 20.003
0.008 0.005
0.264 0.263
0.265 0.024
0.043 0.037
Other Income 1
Income 1 regressors
None Income
others None
Income others
Notes: The rate of capital gains is dened as capital gains divided by initial wealth. Rates are calculated only if the
denominator is positive. The racial difference is the coefcient on a dummy variable for African Americans. Standard errors are in parentheses. Total family income over the time span and its square are the income
variables, while the other regressors are initial wealth and its square, age of head and its square, sex and education of head, marital status at start and end of period, and number of children. The trimmed longitudi-
nal samples are used for OLS and the untrimmed ones for median regressions for details, see the appendix on the JHR website. Regressions use PSID family weights multiplied by wealth at the start of period.
signicant at 10 percent level signicant at 5 percent level
signicant at 1 percent level
Americans and the higher average inheritance among those who receive one 75,236 versus 48,946, it is no surprise that there is a large difference by race
evident from the baseline analysis, shown in Table 8. We use the parameters of the tobit regression to predict the average difference by race if inheritances had not been
censored at 10,000, and for the sake of comparison, show this prediction under the assumption that there were, in fact, no inheritances below the 10,000 threshold.
Though the estimation is not particularly precise because of the small number of positive values and their high level of skewness, the results suggest that little bias
arises from assuming that there are no transfers below 10,000. The results reported here are, moreover, broadly similar to those reported in Wolff 1998 on the basis
of the 1995 Survey of Consumer Finances SCF. According to the SCF, 24 percent of white households had reported an inheritance on or before 1995, compared to 11
percent of African American households, and the average bequest for inheritors was 115,000 among whites and 32,000 for African Americans. The results of the other
tobit regressions indicate that racial differences in income and various demographic characteristics account for only a small portion of racial differences in inheritances.
In Table 9, we consider racial differences in the overall rate of wealth appreciation. As in our analysis of the rate of return on capital, we are forced to exclude households
with nonpositive wealth from the analysis. This truncation lowers the rate of wealth
G ittle
m an
and W
olff
217
Table 8 Racial Differences in Inheritances, Tobit Regressions African American2White, 1984–94
Predicted Predicted
Predicted Predicted
Predicted Predicted
Average Average
Average Average
Average Average
Difference, Difference,
Difference, Difference,
Difference, Difference,
Censored Censored
Censored Censored
Censored Censored
Period Coefcient
at Zero as in Survey
Coefcient at Zero
as in Survey Coefcient
at Zero as in Survey
1984–89 2748,015 28,726
28,647 2701,255 27,506
27,431 2628,738 25,161
25,098 56,307
1,169 1,173
53,926 1,077
1,080 49,997
786 786
1989–94 2830,827 29,031
28,977 2784,920 27,658
27,607 2738,266 26,798
26,749 60,468
1,502 1,510
58,472 1,310
1,315 56,311
1,165 1,167
1984–94 2761,886 217,204 217,128
2707,897 215,114 215,040 2643,932 213,032 212,961
50,996 2,304
2,312 48,996
1,999 2,007
46,697 1,788
1,792 Other
Income 1 regressors
None Income
others
Notes: Respondents are asked about dollar amount of inheritances only if they have received an amount of 10,000 or more in nominal currency. The coefcient shown in
the table is for a dummy variable for African Americans from a Tobit regression, and reects differences in latent variables. Standard errors are in parentheses. “Predicted Average Difference” uses the Tobit results to predict differences by race in the amount of inheritances in 1998 dollars, with censoring at zero and at 10,000 as in
the actual survey. Standard errors are calculated by bootstrapping. Total family income over the time span and its square are the income variables, while the other regressors are age of head and its square, sex and education of head, marital status at start and end of period, and number of children. Trimmed longitudinal samples
are used for details, see the appendix on the JHR website. Regressions use PSID family weights. signicant at 10 percent level
signicant at 5 percent level signicant at 1 percent level
Table 9 Racial Differences in the Rate of Wealth Appreciation African American –
White, 1984– 94
Type of regression
OLS OLS
OLS Median
Median Median
Period 1984– 89
20.057 0.007
20.082 20.164
20.087 20.149
0.142 0.142
0.139 0.057
0.066 0.051
1989– 94 0.120
0.168 0.093
20.088 20.025
0.033 0.341
0.343 0.345
0.058 0.055
0.059 1984– 94
0.041 0.174
0.015 20.293
20.044 20.066
0.385 0.385
0.383 0.145
0.103 0.080
Other Income 1
Income 1 regressors None
Income others
None Income
others
Notes: The rate of wealth appreciation is dened as the change in wealth divided by initial wealth. Rates are
calculated only if the denominator is positive. The racial difference is the coefcient on a dummy variable for African Americans. Standard errors are in parentheses. Total family income over the time span and
its square are the income variables, while the other regressors are initial wealth and its square, age of head and its square, sex and education of head, marital status at start and end of period, number of children
and inheritances over the period. The trimmed longitudinal samples are used for OLS and the untrimmed ones for median regressions for details, see the appendix on the JHR website. Regressions use PSID
family weights multiplied by wealth at the start of period. signicant at 10 percent level
signicant at 5 percent level signicant at 1 percent level
appreciation of African Americans relative to whites. Because of this reduced gap and to the imprecision of the estimates, the OLS results do not show any statistically
signicant differences by race. For the median regressions, the baseline regressions match the results of Table 5, where whites have a substantially higher rate of wealth
appreciation for 1984– 89 and 1984– 94. The statistical signicance of this difference remains when other controls are added for 1984– 89, but not for 1984– 94. As noted
above, in their multivariate analysis of wealth accumulation, Hurst, Luoh, and Stafford 1998 do not usually nd the coefcient for the race dummy to be statisti-
cally signicant.
IV. Simulations and Sensitivity Analysis