search

kolesarm / RDHonest

Honest inference in regression discontinuity designs
56 stars 15 forks source link

CI for NPR_lp.R for bias/se < 2? #2

Closed paulgp closed 3 years ago

paulgp commented 4 years ago

Hi Michal,

I'm not sure if this is intentional, but in NPR_lp.R, when constructing the CI, it's coded as:

lower <- r1$estimate - bias - stats::qnorm(1-alpha)*sd upper <- r1$estimate + bias + stats::qnorm(1-alpha)*sd but isn't this only approximately right for bias/sd >= 2? (According to Table Note 1 from your QE paper with Tim).

Thanks!

kolesarm commented 3 years ago

For one-sided confidence intervals, this is correct for whatever the bias-standard deviation ratio is (note upper is the upper endpoint of a one-sided CI of the form (-inf, upper], and similarly for lower).

Two-sided CIs are constructed as c(r1$estimate-r1$hl, r1$estimate+r1$hl). Here the half-length hl takes into account the bias-sd ratio, it is defined as the standard deviation times the critical value from Table 1 in the QE paper