"... In PAGE 10: ... Everything that follows asks in one way or another whether some or all of the coefficients in this simple model are related to the measures of leverage. The impact of leverage Table3 presents a first test of our central hypothesis. We begin with the three-variable specification, and add a single interaction term, given by dI t *DEBT t-1 , where DEBT t -1 is a once-lagged leverage measure.... In PAGE 11: ... This is perhaps easiest to see by comparing the impulse response of house prices to an income shock for cities with different leverage levels, shown in Figure 2. The figure uses the parameter estimates from column (1) of Table3 , and compares a city with the 10th percentile value of HIGHLTV (which is approximately 5%) to a city with the 90th percentile value of HIGHLTV (which is approximately 25%). The figure depicts a dramatic difference in the implied reaction of the two cities to a 1% income shock.... In PAGE 11: ...29% in the high-leverage city, before turning around. As a slight variation on the specifications in columns (1), (3), and (5) of Table3 , we also try including the lagged measure of leverage DEBT t -1 itself in the regression as an additional control variable. This is done in columns (2), (4) and (6) of the table.... In PAGE 11: ...n the tax code, demographics, etc.). 9.In Table3 and those that follow, our standard errors allow for both heteroskedasticity, as well as for correlation within each city-survey cluster. There are a total of 111 of these clusters in our data set.... In PAGE 12: ... However, for our purposes the important point is that including this extra variable in the regression does not materially change the estimated coefficients on the key dI t *DEBT t-1 interaction term. One concern with the regressions in Table3 is that they are very tightly parameterized. First, they allow only the dI t coefficient to vary with leverage, and force the dP t -1 and P t-1 /I t- 1 coefficients to be constant across cities with different leverage.... In PAGE 12: ... For example, the coefficient on dP t -1 is about the same across quartiles when we use HIGHLTV ; is higher in the high- leverage quartile when we use YESLOAN ; and is lower in the high-leverage quartile when we use MEDIAN . Finally, consistent with these first two observations, the regressions in Table 4 yield impulse response functions that look quite similar to those implied by the regressions in Table3 . This is illustrated in Figure 3, which plots the impulse responses for the high and low quartiles according to our HIGHLTV measure of leverage.... In PAGE 13: ... In the interests of brevity, detailed tables are not provided; they can be found in a previous version of this paper (Lamont and Stein, 1997). Moreover, the tests we discuss below represent modifications of our more tightly-parameterized specification from Table3 . We have also examined the analogous modifications of the looser specification in Table 4; as one might expect based on the comparisons above, these yield very similar conclusions .... In PAGE 13: ... We have also examined the analogous modifications of the looser specification in Table 4; as one might expect based on the comparisons above, these yield very similar conclusions . First, we check whether the results in Table3 are due primarily to a few influential outliers. We sort the observations on both dP t and dI t , and discard the top and bottom one percent of the realizations for these two variables.... In PAGE 14: ... The advantage of this approach is that the projected leverage measure at any time t now only contains information available at that time. 12 Next, we re-run the regressions of Table3 , but substitute in our projected leverage measures for the actual stale data . As one might have expected based on the idea that we are fixing a measurement error problem, the coefficients on the key dI t *DEBT t-1 term increase in all six specifications.... In PAGE 14: ... For example, in the first specification using the HIGHLTV measure, the coefficient of interest rises from 2.27 in column (1) of Table3 to 3.03, an increase of approximately 33%.... In PAGE 15: ... 15 We implement this approach in Table 5. The specifications are the same as in Table3 , except that we allow each of the 44 cities to have its own coefficient on dI t. Thus if some cities are more emerging than others over the entire sample, and hence have house prices that are more sensitive to income shocks, this will now be 13.... In PAGE 16: ... As it turns out, this specification does not reduce the interaction coefficients . In fact, in five of six cases, the interaction terms increase relative to Table3 , in some cases by quite a bit. Naturally, by removing all the across-city variation in our leverage measures, we reduce the precision of our estimates.... In PAGE 16: ... One natural such candidate variable is population growth. In Table 6, we run a horse race which effectively asks: are our previous interaction results truly due to leverage effects, or merely to the fact that leverage is correlated with population growth? The regressions are similar to those in Table3 , with the following modifications. In columns (1), (3) and (5), we add a second interaction term, dI t *dPOP t-1 , where dPOP t is defined as a city apos;s population growth in the year from t-1 to t.... In PAGE 18: ...20. Given the success of this first-step regression, we next proceed to run an IV version of the specification in Table3 . Everything is exactly as before, except we use dI t *DUMMY as an instrument for dI t *DEBT t-1 .... ..."

"... In PAGE 36: ... Table3 : Susceptibility to volatility shocks Panel B: Pre-Crisis Asia Mean Median Maximum Minimum Std. Dev.... ..."

"... In PAGE 26: ... Table3 : Susceptibility to volatility shocks Panel B: Pre-Crisis Asia Mean Median Maximum Minimum Std. Dev.... ..."

"... In PAGE 7: ...able 4. Sample music data..........................................30 Table5 .... In PAGE 39: ... RESULTS AND ISCUSION 4.1 Relationship betwen News for the Month and News for the Wek Table5 displays the correlations betwen the results for the news for the month and news for the wek. Table 5.... In PAGE 59: ... If including results les than 41 is more acurate than using al of the data, and predicting the market data is more acurate than predicting the actual election, then these results should be the most acurate of al. Table5 confirms that these are the most acurate data. The web/name data predicted the market data perfectly.... ..."