without our full set of covariates (including schooling, university education, occupation,

gender, year of birth, and year of death). Just as in a randomized experiment, we expect

the inclusion of covariates to have only a small eﬀect on the estimate of τ because, in

the close neighborhood of the threshold, all observed and unobserved covariates should be

independent of W . Covariates may help to eliminate some residual bias that is the result

of the inclusion of observations further away from the threshold, and they may improve

precision to the extent that they are predictive of the outcome.

Our RD estimates, presented in Table 4, mirror the graphical and matching results

presented above: we ﬁnd that Conservatives roughly doubled their wealth by winning a

seat in Parliament, while Labourites did not ﬁnancially gain from oﬃce. We again reject

the null at conventional levels but the standard errors, as expected, are slightly larger than

in the matching analysis because the RD approach focuses on the neighborhood of the

threshold, where there are fewer observations. Also as might be expected, the inclusion of

covariates does not change the substantive conclusions.

C. Robustness Tests for RD Estimation

C.1.

Test for Wealth Jumps at Non-discontinuity Points

Following the proposal by Imbens and Lemieux (2007), we test for jumps in wealth at

points other than the threshold at which oﬃce was assigned. We produce RD estimates at

several points along the range of the vote share variable, in each case limiting analysis to

either the winning or losing candidates.

23

Figure 7 compares these placebo eﬀect estimates

with our estimate of the eﬀect of winning oﬃce on wealth. (We focus on Conservative

candidates, since we did not ﬁnd an eﬀect for Labour.) The upper panel presents the point

estimates for each of the placebo runs contrasted with the estimate at the true threshold;

the lower panel presents the corresponding t-values. The true eﬀect estimate clearly stands

of bandwidth, although obviously the standard errors tend to increase as the bandwidth is decreased due

to the smaller number of observations. For example, for the Conservatives the estimated treatment eﬀect

(including all covariates) is .98 (.64) when we use half the optimal threshold (i.e. 7.5 percentage points)

and .54 (.28) when double the optimal bandwidth (i.e. 30 percentage points) is used.

23

By focusing on each subsample separately, we follow Imbens & Lemieux (2007, pg. 27), who note that

otherwise our regression function would assume continuity at a point where we know there is a break.

17