Raising the Bar: Impacts and Implementation of the New Heights Program for Expectant and Parenting Teens in Washington, dc

Impact analysis methods 
Impacts are estimated using the treatment and comparison groups described previously. We 
use regression adjustment to control for differences in students’ characteristics among these 
groups. Our primary analysis consists of three steps. The first step in calculating impacts is to 
calculate the difference in outcomes between the post-2011 treatment group and the pre-2011 
comparison group. This post-pre difference is an accurate estimate of program effects only if 
nothing else changed in the schools or the surrounding areas since the New Heights expansion, 
which would have similarly affected the outcomes of interest (such as school policies aimed at 
academic engagement and completion). If any other changes affecting outcomes for parenting 
females occurred, then the observed impacts could not be attributed solely to the program. To 
address this concern, in the second step we calculate post-pre difference in outcomes of interest 
over the same time period for nonparenting females in the same schools, because New Heights 
should not affect nonparenting females. To better estimate the true New Heights impact, in the 
third step we subtract the change in outcomes for nonparenting females from the change in 
outcomes for parenting females. This approach to estimating impacts is sometimes called a 
difference-in-differences approach. 
Our primary analyses examine the impact of the offer of New Heights, akin to an intent-to-
treat analysis for an RCT. That is, the treatment group includes both parenting females who 
participate in New Heights and parenting females who do not participate; 75 percent of the 
14 

---

RAISING THE BAR: IMPACTS AND IMPLEMENTATION OF THE NEW HEIGHTS PROGRAM
 
parenting females between fall 2011and spring 2015 have a record of New Heights program 
participation. We use linear regression to calculate program impacts, adjusting for differences 
between the treatment and comparison groups with respect to students’ age and race and 
ethnicity. We adjust standard error estimates to account for the fact that there are multiple 
semesters of data for each student. We also account for multiple hypothesis testing within the 
school engagement outcome domain due to the multiple outcomes in that domain. The appendix 
reports our impact equations and other technical details. 
Though our primary analyses examine the impact of the offer of New Heights, we would 
logically expect the impact of the offer to be the result of program participation. To calculate the 
impact of participation in the New Heights program, we divide our intent-to-treat or primary 
impact of the offer of New Heights by the proportion of parenting females who participate in the 
program (75 percent); this is the impact of treatment on the treated (Bloom 1984; the appendix 
describes the technical details of this method and its assumptions). 
A retrospective design such as this, drawing from three administrative data sources, 
involved numerous decisions on sample construction and outcome specification. In addition, a 
non-experimental, multiple comparison group design such as this involves many more analytic 
decisions than a typical experimental evaluation that estimates point-in-time impacts for two 
easily defined groups. We conducted numerous sensitivity tests, described in the appendix, to 
understand whether our findings depended on sample construction, outcome specification, and 
analytic approaches. Findings that are highly sensitive to research methods are considered less 
credible (Leamer 1985). The appendix describes these analyses in detail.
15 

---