Half a Standard Deviation: What a Graduate Recommendation Gap Can and Cannot Prove

On one page, two numbers sit side by side: 36.8% for Huazhong University of Science and Technology and 19.9% for Wuhan University of Technology. Same city. Both science and engineering schools. Their program mix overlaps enough that the comparison almost asks to be made.

Recommendation quotas come from education authorities. The rate reflects student quality and training capacity to some degree. That instinct is not foolish. But an instinct becomes a claim only after you name the assumptions.

Here is the version that makes the comparison work. Assume students at both schools have academic levels that follow a normal distribution. Assume the two distributions have the same spread. Assume the recommendation cutoff is the same absolute line for both. Under those three assumptions, the gap between the two schools’ averages comes out to about half a standard deviation. Huazhong’s average sits roughly 0.5 standard deviations above Wuhan University of Technology’s average.

Half a standard deviation sounds small. On a percentile scale, it is not. Huazhong’s median would land near the 69th percentile of the other school’s distribution. That is about the top 30.6%. Not the top quarter. Inside the top third. Pick one Huazhong student and one Wuhan University of Technology student at random. The chance the Huazhong student has the higher academic level is about 64%.

A second adjustment uses top five graduate destinations. Suppose Huazhong sends about 300 students to top five programs, roughly the top 4%. Suppose Wuhan University of Technology sends about 80, roughly the top 1%. That adjustment pushes the gap to about 0.6 standard deviations. Huazhong’s median then lands near the top 27.4% of the other distribution. The random comparison probability rises to about 66.4%.

That gives the familiar magnitude. A mid-tier student at a 985 university would sit somewhere in the top third of a typical 211 university, but not the top quarter. In a random academic comparison, the 985 student wins about six or seven times out of ten.

The arithmetic works. The assumptions carry the weight.

First assumption: the same absolute recommendation cutoff. The recommendation rate is an internal quota ratio. It is not a national cutoff score. A 985 school may have a higher rate because its students are stronger. It may also have more quota slots, more honors programs, or a more favorable set of disciplines. Treating the whole gap as student quality writes institutional advantage into the language of ability.

Second assumption: equal spread. After the college entrance exam sorts students, the range inside a 985 school may not match the range inside a 211 school. If the spreads differ, the calculation changes. The 64% figure changes too.

Third assumption: normality. The shape of the tails matters near the cutoff. With recommendation rates around 20% and 37%, the tails are not a minor detail.

A safer sentence goes like this. Under the strong assumptions of normality, equal spread, and an identical cutoff, Huazhong’s academic distribution sits about 0.5 to 0.6 standard deviations above Wuhan University of Technology’s. Huazhong’s median corresponds to roughly the top 27% to 31% of the other distribution. The random academic win rate is about 64% to 66%. That is a statement about group averages. It is not a ruling on any one person.

The “not a ruling on any one person” line needs a number behind it. When two normal distributions have the same spread and their averages differ by half a standard deviation, the overlap is still about 80%. At 0.6 standard deviations, the overlap is about 76%. Many students in the top 30% at Wuhan University of Technology could sit above Huazhong’s median. Students in the lower part of Huazhong’s distribution could sit below Wuhan University of Technology’s median. The center has moved. The overlap remains large.

Salary shrinks the gap. Suppose Huazhong’s average salary is 10% to 15% higher. Suppose salaries inside each school follow a log-normal distribution with a log standard deviation of 0.4. The difference in log averages works out to about 0.24 to 0.35 standard deviations. Pick two graduates at random. The chance the Huazhong graduate earns more is about 57% to 60%. Still above a coin flip. No longer six or seven out of ten. Program mix, industry cycles, and job choices dilute the gap further.

Push the comparison past salary and it stops working. The recommendation rate measures academic selection. It says little about organizing a team, playing music, playing sports, public speaking, social ease, or appearance. Those traits correlate weakly with academic level. The two schools may sit near fifty-fifty on those dimensions. If some selection or training environment favors a particular trait, the result may move. Without data, no ratio should be given. “Almost exactly fifty-fifty, or even reversed” has a nice sound. As evidence, it goes too far.

Return to the table. The two numbers are still there: 36.8% and 19.9%. Next to them is a new column: assumptions. If the assumptions hold, half a standard deviation holds. If the assumptions fall, the numbers move. The recommendation rate can explain something. It cannot explain everything. A distribution can describe a group. It cannot judge a person.

The table is still on the desk. A cup presses one corner down. A ring of water marks the bottom.