Study of the Week: More Bad News for College Remediation

Today’s Study of the Week combines two subjects we’ve talked about recently on the ANOVA, college remediation and regression discontinuity design. The study, by the University of Warwick’s Emma Duchini, throws even more cold water on our efforts to fix gaps in college student readiness with remediation – and leaves us wondering what to do instead.

One of the basic difficulties in improving educational outcomes lies in the chain of disadvantage. Students who start out behind tend to stay behind, and it’s not productive to ask teachers to make up for the gaps that have been opened over the course of a student’s life. As I’ve said on this blog many times, most students tend to sort themselves into fairly stable academic rankings early in life, and though individuals move between those rankings fairly often, at scale and in numbers this hierarchy is remarkably persistent. So third grade reading group serves as a good predictor of high school graduation rates, which in turn obviously predicts college completion rates. Meanwhile, the racial achievement gap appears to exist before students ever show up in formal schooling at all. It’s discouraging.

This study comes from the United Kingdom, but it concerns a question of great interest on this side of the Atlantic: do college remediation classes work? We know that college student populations are profoundly different in incoming ability. The college admissions process makes sure of that. That means that institutions like mine, the City University of New York, face profoundly higher hurdles in getting students to typical levels of ability, as our admissions data tells us that many of our students are unprepared. Typically, this results in remedial classes, to the tune of $4 billion a year for public universities. But as Duchini notes, evidence for the effectiveness of remediation is thin on the ground. Her study takes another look.

Duchini’s study draws its data from the economics department of a public Italian university. This university implemented an entrance exam for potential students, consisting of a math section, a verbal section, and a logic section. The results of this test, combined with high school grades, determines whether students are admitted to the program. However, the math section alone is used to determine whether students need to take a remedial program. Because this involves using a cut score, the cut score is fairly close to the mean, and there are no other systematic differences between students placed in or out of the remediation program, this is an ideal situation for a regression discontinuity design, as I explained in this previous post.

I can teach you regression discontinuity design in two images

Ultimately Duchini considers the exam scores and educational and demographic data of 2,682 students, sorted into descriptive categories like gender, immigrant or domestic, vocational or general track, or similar. Importantly for a regression discontinuity design, there is no evidence of student groupings tightly on either side of the cut score, which can indicate that there is student manipulation of placement that would invalidate the design.

There’s an interesting dynamic in the data set, perhaps an example of Berkson’s paradox. Students who perform better on the entrance test are actually less likely to enroll in the program, even though doing well on the test is a requirement for attendance. Why? Think about what it means to do well on the test: those students are more academically prepared over all, and thus have more options for majors to take, meaning that more of them will choose to enroll in a different program.

In any event, Duchini uses a regression discontinuity design to see if there is any meaningful difference between students on students on either side of the cut score and how the trend line changes, looking at outcome variables like odds of dropping out, passing college-level math, and credits accumulated. The results are not encouraging. In particular, the real nut is here, how remediation affects the odds of passing college-level math. Note that the sample is restricted here to edge cases, as we don’t want to get a misleading picture from looking at students too far from the cutoff – this is a last in/last out style model, after all – and bear in mind that because this is a remediation test, the treatment is assigned to those on the left hand side of the cut line.

The upward-sloping trend is no surprise; we should expect student performance on an entrance exam to predict the likelihood that they’ll get through a class in the test subject. What we want to see here is a large break in the performance of the groups at the cut score, with a corresponding shift in the trend line, to suggest that the remediation program is meaningfully affecting outcomes – that is, that it’s bringing students below the cutline closer to the performance of those well above it. Neither eyeballing this scatterplot nor the statistical significance checks Duchini describes provides any such evidence. I find that fact that the data points are more tightly grouped on the left side of the cutline than on the right interesting, but I’m guessing it’s mostly noise. Look in the PDF for more scatterplots with similar trend lines as well as the model and threshold for significance.

Duchini goes into a lot of extra detail, breaking the data set down by demographic groupings and educational factors, though in every case there is little evidence of meaningful gains from the remediation program. Duchini also speaks at length about potential reasons why the program failed to meaningfully prepare students to pass college-level math, including wondering if being assigned remediation might discourage students by making them feel like the work of getting their degree will be even harder than they thought. It’s interesting stuff and worth reading, but for our purposes the conclusion is simple: this remediation program does not appear to meaningfully help students succeed in later college endeavors. It’s only one study from a particular context. But given similar studies that also find little value in remediation, this is more reason to question the value of such programs. More study is needed, but it’s not looking good.

Clearly, if remedial classes don’t work, and they cost students time and money, they should be scrapped. But scrapping them won’t solve the underlying problem: students are arriving at college without the necessary academic skills to ensure that they succeed. College educators will typically lament that they’re trying to solve the deficiencies of high school education, but of course high school teachers can fairly look back as well. Ultimately the dynamic is applicable to the whole system: students are profoundly unequal in their various academic talents from a very early age, and we’re all searching for ways to serve them better. Perhaps the conversation needs to turn to whether we should be pushing so many students into college in the first place, and whether we need to look for answers to economic woes outside of the education system entirely. But for now, we as college educators are left with a sticky problem: our students come to our schools unprepared, but our programs to fill those gaps show little sign of working.