Rage Against the Machine Learning
I have several things on my radar this week that I want to share.
The Campaign released a brief on Expanding Federal Work-Study Opportunities for California Community College Students. A few highlights:
- Of the 2.1 or so million community colleges students in California, only around 10,500 are receiving Federal Work-Study (FWS) dollars.
- 70 percent of FWS recipients are under the age of 30.
- Two-thirds of FWS recipients are women.
The Federal Work-Study program is not well-suited to community colleges. There are matching fund requirements, and the distribution formula prioritizes institutions with records of past participation (a bias towards older institutions and Northeastern US colleges and universities), and high-cost schools as judged by FAFSA’s unmet need calculation. Low FAFSA completion rates limit the number of students who might receive an award, and the formula itself has a bias that hurts community colleges. The brief also has some best practices and strategies from Pasadena City College and Long Beach City College, as well as recommendations for policymakers and practitioners. Check it out!
The National Student Clearinghouse Monthly COVID update caught my eye this week. The data has some things that surprise me and some things that do not. Associate’s enrollment was down in the summer, as was enrollment in sub-bac certificates. Black student enrollment took the biggest hit, but Latinx student enrollment was up slightly. The link here is to the NSC’s Tableau dashboard. It’s pretty easy to fiddle around with to see comparisons by race, gender, age, enrollment intensity, and a few other things. I do not think it filters to state levels.
Anthony Carnevale has a piece in Medium that I would highly recommend. He takes a long lens on discriminatory public policy in the United States and discusses the ways in which educational opportunities are part of a larger patchwork of racist public policy. The essay traces issues from slavery, to Jim Crow, to New Deal legislation that excluded Black workers from participating in the welfare state, to redlining, discriminatory mortgage practices, and underinvestment in urban schools. It all culminates in a system that is outrageously unequal. This is a heftier piece, but I urge you to find the time to read it.
One way we might expand access to the baccalaureate degree is through community college programs. In areas that are a little too remote from a public four-year, these have a pretty cool potential upside of allowing access to higher education. Traditional opposition has focused on the potential impact on four-year institutions. A recent study in the American Educational Research Journal looked at community college baccalaureate (CCB) programs to address exactly these questions. These authors found that “local CCB degree programs have a negative effect on overall bachelor’s degree enrollment and bachelor’s degree production at 4-year institutions, but this effect is concentrated primarily within for-profit 4-year institutions.” Pretty cool, right?
Also on the access to four-year degrees front, the folks at PPIC have a report out that highlights the need for additional capacity at the CSU system. They make a number of important points. Among them: the number of eligible applicants who are being denied admission has quadrupled since the great recession; all but two campuses face space constraints under normal operating circumstances; expanding student enrollment requires expanding the faculty ranks; and, even with the additional focus placed on distance education with the pandemic, significant investments in course design and delivery are still required to make sure online course delivery is as effective as it needs to be.
Moving on from four-year degrees, this working paper examines the costs and burdens of FAFSA verification on public institutions. Not surprisingly, these costs are disproportionately born by community colleges and the diverse population they serve – Pell-eligible students are much more likely to undergo verification than their wealthier peers. Verification costs total roughly a half-billion dollars per year. At community colleges, financial aid offices are spending as much as 22% of their budgets on this. At public four-years, it is more like 15%, which is still too high. Furthermore, the verification process results in very few changes to financial aid. I tend to put FAFSA fraud in the same category as voter fraud: it’s a straw man created to keep minorities out. Just think of the ways we could be better purposing those resources…
Financial aid and the cost of college are among several challenges faced by students. . The authors discuss the influence of first-generation status and financial and basic-needs insecurity, as well as students’ genuine drive to succeed and the importance of robust supports. The authors also include policy recommendations that will help ensure more Latinx students are able to not just enroll, but to succeed in their studies.
I really liked this study, both for the fact that it is a randomized control trial, and for the actual content. The authors test out a fairly low-budget intervention that is designed to help increase students’ sense of belonging at a broad-access four-year institution. Students in the treatment group read stories written by upperclassmen about times they felt they didn’t belong, then responded to a writing exercise to help them understand that it is completely normal to question your sense of belonging. Control-group students did a similar reading/writing exercise, but they did not focus on themes of belonging. According to the authors, “The intervention increased the likelihood that racial-ethnic minority and first-generation students maintained continuous enrollment over the next two academic years relative to multiple control groups. This two-year gain in persistence was mediated by greater feelings of social and academic fit one-year post-intervention.”
This last one is a little different, but it is really important to think about in our work. The study looks at the algorithms used by big job-search websites to identify candidates for interviews. I flag this because it is very closely connected to the concept of university admissions and the algorithms used by admissions departments. I will circle back to this below after a brief set up. Stick with me because I think you will enjoy the payoff. (Also, I will acknowledge that I first started learning about algorithms like the ones in this study when I was working at the Community College Research Center with Peter Bergman – a coauthor of this paper. Also, the paper uses the term “upper confidence bound contextual bandit approach”…gotta love the jargon!)
The authors talk about algorithms that employ “supervised machine learning” approaches. These algorithms use data from prior hiring to learn which candidates have a high likelihood of a successful interview. Essentially, everybody gets an estimated score based on the inputs that are chosen for the model, and the highest scores get selected to be interviewed. These estimates, however, also have a confidence range – the model knows there is uncertainty. The confidence ranges will tend to be larger for under-represented groups – there is less data for these groups, so the algorithm is less certain and gives a bigger range. If we just let this thing run, it will produce results that look a lot like prior results, baking in any prior inequities (Still with me? Good. This is where it gets interesting!)
I promised I would bring this back to admissions. Too many capable URM students are excluded from top-tier institutions because we have a system built to reproduce yesterday’s inequities. There are ways to build algorithms that will yield student bodies that are highly likely to succeed and that reflect the diversity of the state’s population. More broadly, I would like to see universities exploring and looking for students who would benefit from the education they provide, not students who will be fine regardless. That’s a conversation for another time.
Vikash Reddy, Senior Director of Policy Research
Read more blog posts