Same Variable, Opposite Direction
We grouped institutions by their share of distance education enrollment and split each group by non-profit vs. for-profit control. For non-profits, going online is protective. For for-profits, it's catastrophic.
For non-profits, shifting online reduces fixed costs while serving students flexibly. For for-profits in the 50–90% range, heavy online enrollment was often a sign of aggressive growth strategies that attracted regulatory scrutiny and left institutions vulnerable when enrollment rules tightened.
The Case for Interaction Effects
This is where the ML model actually helps. A simple threshold rule can't capture the fact that online share means opposite things for different institution types. A gradient-boosted model can learn this automatically from the data. ML helps here. It doesn't help everywhere in this study.
The "fully online" bucket (>90%) has zero closures in both sectors, but it contains only 5 institutions, all for-profit. The sample is too small to draw conclusions. The real insight is in the 50–90% band, where the divergence between sectors is enormous and statistically meaningful.
The hybrid band (10–50% online) is where most institutions live: 3,048 total. Even here, the for-profit closure rate is 13.1% vs. 1.9% for non-profits. The gap is consistent across all online share levels, but it's widest in the primarily-online category.
Online share from IPEDS distance education variable EFDESOM (College Scorecard data). "Online share" = fraction of total enrollment in exclusively-online programs.