DC School Growth Scores And Poverty

As noted in a nice little post over at Greater Greater Washington's education blog, the District of Columbia Office of the State Superintendent of Education (OSSE) recently started releasing growth model scores for DC’s charter and regular public schools. These models, in a nutshell, assess schools by following their students over time and gauging their testing progress relative to similar students (they can also be used for individual teachers, but DCPS uses a different model in its teacher evaluations).

In my opinion, producing these estimates and making them available publicly is a good idea, and definitely preferable to the district’s previous reliance on changes in proficiency, which are truly awful measures (see here for more on this). It’s also, however, important to note that the model chosen by OSSE – a “median growth percentile," or MGP model, produces estimates that have been shown to be at least somewhat more heavily associated with student characteristics than other types of models, such as value-added models proper. This does not necessarily mean the growth percentile models are “inaccurate” – there are good reasons, such as resources and more difficulty with teacher recruitment/retention, to believe that schools serving poorer students might be less effective, on average, and it’s tough to separate “real” effects from bias in the models.

That said, let’s take a quick look at this relationship using the DC MGP scores from 2011, with poverty data from the National Center for Education Statistics.

The simple table below presents correlation coefficients between MGPs (for math and reading) and the percent of schools’ students eligible for free lunch, as well as the percent eligible for free or reduced-price lunch (FRL).

The correlations between school lunch eligibility and growth scores are solidly moderate, particularly when using free instead of free/reduced-price lunch.

Let’s quickly present one of these associations (free lunch and reading) visually, in a scatterplot.

This is far from a clean relationship, but you can see a discernible downward slope in the dots (the blue line represents the average relationship between these two variables) - schools with lower poverty rates tend to get higher MGPs, and vice-versa. For instance, virtually none of the schools with free lunch rates below 50 percent receives an MGP below the median (50th percentile).

Again, this is not at all surprising, and it confirms what you’ll find in the research (see this post for more discussion of this issue and what it might mean for school accountability policies).

Nevertheless, the relationship is noteworthy, and parents, administrators and other stakeholders using these results should bear that in mind.

- Matt Di Carlo


The REAL question is whether teachers and school leaders will be fired on the flawed metric of MGPs? Given no controls for student demographics or other observed and unobserved characteristics of schools, would it be fair to fire educators in schools with low MGPs? Most definitely not. We seem to be in a world where cheap and easy to understand data trumps expensive and accurate data when making policy decisions. Better to collect no data at all and go with gut instincts than rely on bad data.


Interesting post. What do you think accounts for the difference between the MGP in the DC model vs. the Colorado model? The scatterplot graph above looks very different than the one you did here: http://shankerblog.org/?p=6090