Does it matter which fraternity you join?

Members of Congress really did come from the better-ranked houses on their own campuses. Then we checked how big those houses were, and the finding fell apart. Here is the whole thing, including the two checks that break it.

There is a ranking every fraternity system on every campus agrees on and nobody writes down. The top houses, the middle, and the ones nobody rushes. The question is whether that ranking is worth anything — whether getting into the good house does something for you that the bad house doesn’t.

It is a hard question to test, because almost nothing about a career is recorded per chapter. One thing is. If you reach Congress, somebody writes down which fraternity you joined, and it is possible to find out which school you joined it at.

This is the long version: what was built, what it found, and the three separate reasons the finding does not hold up. If you want the files, they are at the bottom.

Why not just compare fraternities?

The obvious version of this question is to rank the national organizations — Sigma Chi against Kappa Sigma against Pike — by how many members of Congress each has produced per 100,000 initiates. We did that first. Across 28 men’s fraternities, 84 members and 5.89 million initiates, the pooled rate is 1.43 per 100,000 and the spread between organizations is statistically indistinguishable from chance: χ² = 25.8 on 27 df, p = 0.53. Monte-Carlo says a real 2× gap between the best and worst organization would have shown up 94% of the time. It didn’t.

That is a well-powered null, and it is the strongest single result here. But it answers a question nobody actually asks. Nobody rushing a house is choosing between brands in the abstract — the letters mean something different on every campus. The house everyone wants at one school is the house nobody wants at another. What people mean by “the top frat” is local status, and a national ranking cannot see it.

The setup

So the test has to happen inside each school. Every member of Congress becomes a point on his own campus’s ranking: zero is the worst house at his school, one is the best. If the good houses do something, those numbers skew high. If they don’t, they average one half.

Doing it within school also removes a confound the national test can’t touch: good universities have both good chapters and good career outcomes. Ranking each house only against the houses across the street cancels the university entirely.

Four pieces had to be joined:

Resolving where somebody went to undergrad is fiddlier than it sounds, because Wikidata records every institution a person attended without saying which was the bachelor’s. Filtering on the name (“drop anything with law in it”) is not good enough — plenty of members hold a JD from a plainly-named university. Instead each education statement is checked for the P512 academic-degree qualifier, and only the ones carrying a bachelor’s count; the name heuristic is a fallback. 118 of the 145 resolved on positive degree evidence.

The join that could have wrecked it

Matching a fraternity to its chapter is where this kind of project usually dies quietly. Greekrank’s URLs use nicknames students wrote themselves — Pikes-Pikas, DSP-Delta-Sigs-D-Sigs-Delt-Sigs, Iotas-Outlaws, FIJI. Fuzzy-matching those against a list of formal organization names produces confident, invisible garbage.

It turns out not to be necessary. Every chapter URL carries a numeric organization id, and that integer is the same for the same fraternity at every school in the country — Pi Kappa Alpha is 84 at Texas A&M and 84 at Duke. Matching on the integer means the nicknames never enter the analysis. And Greekrank’s own rendered name for each id proved 100% consistent across all 7,456 rows, so even the id→name map is derived from the data rather than typed by hand.

Schools are matched by exact normalised name plus a 17-entry hand-checked alias table, because Greekrank shortens flagships to bare city names (Austin, Knoxville, Chapel Hill). Five tempting near-matches are deliberately refused and recorded as such: Amherst College is not UMass Amherst (Amherst banned fraternities in 1984 and isn’t in the data at all), and Xavier University of Louisiana is not Xavier University of Ohio.

Who fell out

70 of the 146 land on a ranked chapter. The attrition is mostly structural, but one line of it matters later:

Writing the test down first

Before the join was run — before there was a number to look at — the analysis was pre-registered: the primary endpoint, the secondary model, the exclusions, how chapter size would be handled, and the requirement to compute power.

That is not ceremony. This is a null-hunting analysis on a small sample, which is exactly the situation where it is easy to try things until one works. Greekrank publishes seven sub-scores besides the overall grade — Reputation, Friendliness, Popularity, Classiness, Involvement, Social Life, Brotherhood. Fixing the overall grade as the primary in advance means there was never an opportunity to search across eight metrics for the one that produced a headline.

The finding

They came from the 58.5th percentile.

Both pre-registered tests agree, and the result survives every threshold we tried for how many ranked houses a school needs to qualify (p ranges 0.010–0.043 from a 3-house floor to a 15-house floor). On its face: the tier list is real.

It isn’t. And the reason is the most boring one available.

1 · The top houses are the big houses

A chapter with 130 men produces more congressmen than a chapter with 35, for the same reason a bigger city produces more of everything. That is not an advantage to any individual member. It is more lottery tickets.

This is the confound the national per-100,000 test solves by construction — and the within-school test does not get it for free. Nobody publishes chapter rosters, so there is no denominator to divide by. What Greekrank does publish is how many people rated each chapter, which tracks size and notoriety together. Across all 6,739 ranked chapters it correlates with prestige at ρ = +0.450:

Put that into the model and the ranking stops predicting anything:

The size term is highly significant on its own (p = 0.0017). The prestige term is not. Whatever the tier list looked like it was doing, the size of the house already explains.

2 · The chapters that died are missing

Those eight members from earlier — the ones whose chapter isn’t on Greekrank at all. A chapter absent from a current ratings site has usually folded, and chapters that fold are disproportionately the bottom-tier ones. Dropping those men quietly inflates the average, and it is the one bias here capable of manufacturing a false positive.

If the dead chapters sat around the 33rd percentile of their campuses — a mild assumption for houses that no longer exist — there is no significant finding left.

3 · Seventy men is not many

Monte-Carlo power at this sample size is 38% for a true doubling of the odds between the best and worst house. Which cuts both ways, and the second way is the one people forget:

Compare that with the national test, which had 94% power at 2×. There, “we found nothing” is a real statement about the world. Here it isn’t. The within-school null cannot be sold as “ruled out” any more than the positive can be sold as real.

What we are actually willing to claim

The campus tier list does not survive the obvious confound. And we cannot prove the confound is the whole story.

That second sentence is the honest limit, not modesty. The size adjustment is imperfect in the other direction: ratings volume is driven by prestige as well as by size, so controlling for it also controls away part of the thing being measured. The unadjusted estimate is biased up; the adjusted one is biased down. The true effect is somewhere between 1.24× and 2.70×, and seventy men cannot narrow it further.

Other things worth knowing about this data

Greekrank measures reputation today. These men pledged decades ago and chapter reputation drifts. That is measurement error, and measurement error pushes toward the null — so it makes the negative result weaker evidence than it looks, not stronger.

The raters are self-selected undergraduates, which sounds like a flaw and isn’t. Perceived prestige is exactly the construct the question is about. We do not want an objective quality score. We want the tier list people argue about, and that is precisely what a site where students rate their own campus’s houses produces.

Congress is a strange outcome. Disproportionately lawyers, disproportionately from political families, in a line of work where joining clubs is itself a qualification. It is used here because it is the only elite outcome recorded per chapter — not because it represents careers.

The question this doesn’t answer

Whether joining at all is worth something. That has a real causal answer and it is not ours.

Mara, Davis & Schmidt (2018), in Contemporary Economic Policy, instrumented fraternity membership using three exogenous residential policy changes at one Northeastern college — the college going co-ed, theme houses, and Minerva houses — which shifted who joined for reasons unrelated to the students themselves. They find roughly +36% later income, alongside a GPA about a quarter-point lower.

Three limits travel with that number and all three matter: it is one college, 3,762 alumni, and it is a local average treatment effect — it describes the men who were on the fence about joining, not everybody. For national coverage there is Routon & Walker (2018) in LABOUR, on a nationally representative sample: about +10 percentage points of self-employment and a sizable wage premium, but built on controls rather than a causal design.

One good causal estimate at one school, one national estimate that isn’t causal, and nobody has both. We cannot close that gap ourselves either: the microdata carrying a Greek-membership variable (NCES Baccalaureate & Beyond, HERI CIRP) sits behind a restricted-use licence.

So: joining appears to do something. Which house you join is a ranking that, once you account for how many people are in the room, stops predicting anything at all.