Anonymous telescope-time proposals

Why did hiding the applicants' names change which astronomers were awarded telescope time?

Time on the Hubble Space Telescope is one of the most contested resources in science: in a typical cycle, roughly one proposal in five or six succeeds. For most of the telescope's history, panels read proposals with the applicants' names attached. Across many of those cycles, proposals led by men were selected at a slightly higher rate than proposals led by women — a gap of a couple of percentage points, small in any single cycle, but persistent enough across cycles to be hard to dismiss as noise.

In 2018, for Cycle 26, the Space Telescope Science Institute switched to dual-anonymous review: the reviewers did not know who had written the proposal, and the writers did not know who would read it. In that first anonymous cycle the gap closed and slightly reversed, with women-led proposals succeeding at a marginally higher rate than men-led ones. NASA extended the approach to other programmes, and other facilities have followed.

Now, the tempting reading is that a panel of prejudiced astronomers was caught out. I want to resist that reading, because it makes the result smaller than it is. Suppose everyone on those panels was acting in good faith. Would you still expect the change?

Ask first what a reviewer is actually being asked to do. Not simply "is this good science" — also "will this team deliver it?" Feasibility is a legitimate and explicitly stated criterion, because telescope time spent on a programme that cannot be executed is time destroyed. So a reviewer is being asked, in part, to forecast.

Now, forecasting is expensive. A panellist may have thirty or fifty proposals to assess, each a dense technical document in a subfield that is not quite their own. What would you reach for? The cheapest available predictor of whether a team can execute is the team's record — and under named review, that predictor is sitting right there on the front page, free, before you have read a word of the science.

So notice what we have established without invoking any hostility at all. The rational strategy for a time-pressed reviewer, under the criteria they were given, is to lean on the name. The name is not noise; it genuinely carries information. It is simply that it carries the wrong information more often than anyone would like, because a track record is itself the accumulated output of earlier allocations, earlier institutions, earlier access — and it correlates with things nobody intended to select on.

Now ask the sharper question: when you remove the name, what actually changes? Not the reviewer's beliefs. What changes is the evidence available to act on. If the cheap proxy is gone, the only thing left to reward is the thing on the page — the justification, the method, the observing strategy. The reviewer's optimal strategy shifts because their information set shifted, and their behaviour follows.

And this reaches back up the chain, which is the part that gets least attention. Under named review, a proposal writer's effort is well spent on establishing standing: cite yourself, name your collaborators, signal the pedigree of the team. Under anonymous review that effort is wasted — the rules require the science case to be written without self-identifying references — so the same effort has to be redirected into the argument itself. The incentive that changed was not only the reviewer's, but the applicant's.

There is also a quieter finding that supports this reading better than the headline one: first-time and less established proposers fared better under anonymity too. That is what you would expect if the mechanism is "a reputational proxy was removed", and not what you would expect if the mechanism were about a single demographic variable alone.

I should be careful about how much weight the numbers bear. The success-rate gaps involved are of a few percentage points and the cycle-to-cycle scatter is real, so no single cycle proves anything; the case rests on a long pattern before the change and a break at it. And genuine anonymity is imperfect — more on that shortly.

The model to correct is "bias was blocked". What was actually blocked was a proxy. Reviewers were using a legitimate and informative shortcut for a criterion they were told to assess, and that shortcut carried in correlations with gender, seniority and institution that nobody had chosen to select on. Anonymity did not change anyone's values; it changed what the cheapest strategy for a reviewer was, and behaviour followed the incentive.

That framing also explains why the change was not free. Feasibility genuinely does need assessing, so anonymous systems have to reintroduce it elsewhere — separately, after the science ranking, and with the identities revealed only at that point. Removing an informative signal costs you something real; the argument for doing it is that what the signal cost was larger.

• How feasibility is assessed after an anonymous science ranking, and whether reintroducing identity late lets the old effect back in.
• Whether anonymity holds up in very small subfields where a proposal's topic identifies its author.

Every term the collection defines is gathered in the glossary.