Datasets from an impact evaluation of a targeted prekindergarten program

Given general trends in extant research on the impact of prekindergarten and that the structure and implementation of prekindergarten programs vary by state, researchers, educators, and policymakers are raising questions about what works, for whom, under what conditions, and the cost-benefit of such endeavors. Yet not all states have formally examined program impacts and few datasets have been expressly collected to evaluate effects. The current data article represents empirical examinations in the state of Connecticut on the comparability of treatment and control groups, tests of the robustness of impact estimates, and the psychometric properties of outcome measures. Stata code for replication purposes is included.


Data
The datasets contain raw numeric and text data available through Connecticut's P20 WIN data repository. A summary of data elements contained in the datasets used is presented in Table 1. Table 2 includes information on the comparability of treatment and comparison samples under different assumptions. Tables 3 and 4 include information on the robustness of impact estimates using linear, quadratic, and quintic models. Table 4 includes additional information on the robustness of impact estimates. Fig. 1 provides information on the smoothness and continuity of birthdates around the enrollment date cut-off.

Ethical statements
This study was approved by the University of Connecticut's Institutional Review Board (Protocols H14e193 and H15-247), and carried out in accordance with a Memorandum of Agreement with the Connecticut State Department of Education and Office of Early Childhood and the Family Educational Rights and Privacy Act of 1974.  [1].

Value of the Data
The datasets represent samples from the treatment group and control group students, which can be compared to reveal the overall impact of attending prekindergarten.
The datasets also contain demographic information on both treatment and control group students that can be used to explore variation in treatment effects. Datasets are available in the form of raw numeric and text data that can be further processed by researchers using datasets from other states, their own data, or other datasets available in Connecticut's P20 WIN data repository.

Collection of student educational record data
The Connecticut State Department of Education and Office of Early Childhood maintain monthly and yearly educational records for all students enrolled in public schools (prekindergarten through grade 12). This data includes, for each student, a unique identification number, name, date of birth,  Note. Heteroskedasticity robust standard errors clustered by birthdate are in parentheses. Each coefficient is the reduced form estimate of the relationship between being eligible for statefunded prek based on birthdate and the listed covariate. The coefficients shown are generated by local linear regression using a triangular kernel and specified bandwidth. The center of the specified bandwidth is the cut-off point. Also listed is the mean of the covariate for students just before the threshold for qualifying for prekindergarten. *p < .10, **p < .05, ***p < .01. race/ethnicity, whether the student participates in the federal free and reduced priced lunch program, the school attended, and the school district in which the school is located. These data were provided to the research team through a Memorandum of Agreement. More specifically, datasets are available through Connecticut's P20 WIN. To access datasets, researchers must apply to become an authorized representative. The request is then reviewed by members of the P20 WIN Data Governing Board, and if approved, a Memorandum of Agreement (MOA) specific to the data request is executed. Once executed, the authorized representative receives a matched dataset. All requestors are required to pay a fee. Under special circumstances the fee may be waived or reduced at the discretion of the Data Governing Board. The authorized representative and anyone on their team who has signed Personal Statements of Non-Disclosure is able to work with the data. All publicly shared information as a result of dataset use must be sufficiently aggregated to maintain confidentiality. A detailed articulation of how to access datasets is included in the P20 WIN Data Request Management Procedure manual available at http://www.ct.edu/files/pdfs/P20WIN- Note. The coefficients shown were generated using OLS with the stated bandwidth and a triangular kernel. Models specified with covariates include indicator variables for gender, race/ethnicity, and eligibility for subsidized meals. *p < .10, **p < .05, ***p < .01. Note. *p < .10, **p < .05, ***p < .01.
DataRequestProcedure-Final_01202015.pdf. The full set of restricted use datasets is available as a downloadable excel file at www.ct.edu/files/pdfs/p20win/P20WIN_Data_Dictionary.xlsx.

Collection and measurement of student outcome data
Student outcome data were collected by a team of 58 trained and certified assessors using the Woodcock Johnson IV Tests of Achievement (WJ-IV) and the Peabody Picture Vocabulary Test 4th Edition (PPVT-4). Data from these assessments were then scored and composite measures of student achievement were entered into a datafile by the research team. Four composite measures in total were used and entered into the datafile.
The Basic Reading composite included two subtests from the WJ-IV, specifically Word Attack and Letter-Word Identification. The Letter-Word Identification subtest measures pre-reading skills including letter and word recognition and identification skills. The Word Attack subtest measures phonics and decoding skills. The reliability statistics for individual tests range from 0.90 to 0.94 [2].
The Oral Language composite included two subtests (the Picture Vocabulary and Oral Comprehension) from the WJ-IV. The Picture Vocabulary subtest primarily assesses expressive vocabulary, though early items provide some information about receptive vocabulary skills. The Oral Comprehension subtest measures the ability to understand short, orally presented, passages. Subtest reliability statistics range from 0.82 to 0.88 [2].
The Broad Mathematical Skills composite consisted of three measures from the WJ-IV, specifically Applied Problems, Calculation, and Math Fluency subtests. The Applied Problems subtest was used to measure mathematics problem solving skills. The Calculation subtest measures students' ability to complete items ranging from writing numbers to performing numerical operations. And, the Math Fluency subtest measures students' ability to solve problems containing numerical operations quickly. The reliability statistics for individual tests range from 0.96 to 0.97 [2].

Statistical analysis
All statistical analyses were conducted using Stata. For all student outcomes, the standard score with a mean of 100 with a standard deviation of 15 was used. A multilevel regression model with cluster standard errors was used because it does not force the same parametric assumptions about the distribution of the error terms at either the school or student level, or the correlation among them. The robustness of analysis decisions were evaluated through other specifications, including adding covariates, using an alternate clustering approach, and using fixed effects for each site. Unless otherwise indicated, p < 0.05 was considered statistically significant. The Stata code for analyzing data is included as a supplementary file.