Table 3.

Effect of excluding “missing data” values on mean treatment effects reported in Krueger and Zhu (2004a)

		Using Mathematica weights	Using our inverse p-score weights
	K&Z estimates	Our replication	Our estimates
	0s included	0s included	0s excluded
Panel A: Effects for full sample, math National Percentile Rankings
Year 1	0.17 (1.34)	0.17 (1.38)	−0.88 (1.05)
Year 2	−0.69 (1.40	−0.69 (1.37)	−1.29 (1.15)
Year 3	0.23 (1.35)	0.23 (1.28)	−0.004 (1.18)
Panel B: Effects for full sample, reading National Percentile Rankings
Year 1	−0.84 (1.25)	−0.84 (1.32)	−1.79 (1.09)
Year 2	0.41 (1.30)	0.41 (1.26)	0.22 (1.16)
Year 3	−0.73 (1.26)	−0.73 (1.32)	0.52 (1.13)

Notes: Table reports original results from panel 3 of Table 3B of Krueger and Zhu (2004a), our replication of these results, and then shows the impact of excluding the 0 percentile values (which are invalid percentiles corresponding to the 99 raw scores) and using inverse propensity score weights as an alternative to the non-response adjusted weights provided with the Mathematica data. Dependent variable is the math (Panel A) or reading (Panel B) National Percentile Ranking scores from the Iowa Test of Basic Skills in the spring for years 1–3 of voucher distribution. Regressions also control for dummies for the strata in the initial sampling. Estimates in column 2 use the Mathematica provided non-response weights, while those in 3 use our inverse propensity score weights. The p-score weights are 1/p̂ for treatment observations and 1/(1 − p̂) for control observations, where p̂ is generated from a logistic regression of treatment status on baseline demographics, dummies for missing demographics, dummies for invalid scores or missing scores, dummies for strata (sample design) and grade at baseline, and baseline test scores. 95% CIs are obtained by bootstrapping families with replacement. Data from the New York City School Choice Scholarships Program evaluation conducted by Mathematica Policy Research.