There’s a well-known pattern in conjoint experiments on candidate choice: when you don’t tell people the candidate’s party, effects like gender and race often get bigger. The standard explanation is “masking” — people are missing information they’d like to have (party), so they use whatever’s on offer (gender, race) to guess at it. Show a female candidate, assume she’s a Democrat, vote accordingly.

It’s a tidy story. According to our new paper (with Simon Calmar Andersen), it’s probably not the whole story.

A second explanation

There’s an older idea from decision theory called lexicographic preferences. Some people don’t weigh attributes against each other at all — they have one dominant attribute (say, party) and only start caring about anything else once two candidates are tied on that one. If party info isn’t shown, it’s as if every pair is tied, so people fall straight through to gender, race, and so on.

The annoying part: masking and lexicographic preferences predict the same basic pattern. Bigger gender effect when party is hidden versus shown — both theories are consistent with that. So how do you tell them apart?

Using the ties already built into conjoint designs

Conjoint experiments generate plenty of choice tasks where both candidates happen to land on the same party, just from random assignment. These “same-party” tasks aren’t typically used to separate the two theories — that’s what we do here.

If masking is doing the work, it should only show up when party is genuinely missing. Tell people party — even if it’s identical for both candidates — and there’s nothing left to infer, so masking should disappear. Lexicographic preferences don’t care about that distinction: if party is tied, second-order attributes kick in regardless of whether party was ever hidden.

So we split the analysis three ways:

1. Candidates have different parties
2. Candidates have the same (but shown) party
3. No party info at all

Comparing (3) to (1) gives the traditional masking estimate. Comparing (3) to (2) — no info vs. same-but-known — isolates what we call pure masking: the part that can only be explained by inference, with lexicographic preferences ruled out.

What we found

Applying this to Kirkland and Coppock’s (2018) candidate data, and then replicating it in our own preregistered MTurk study, the pattern was clear: once the lexicographic-preference explanation is accounted for, the “masking” effect for gender and race on candidate choice largely disappears. What looked like people inferring party from gender was, to a substantial extent, people not weighing gender at all until party was settled.

That changes the underlying story — from “voters use stereotypes to fill information gaps” to “voters have a strict order of priorities, and only get to gender or race once party is out of the way.”

Why it matters beyond candidates

We only tested this empirically on candidate choice, but the same ambiguity shows up anywhere researchers see a bigger effect under missing information and attribute it to masking: anti-immigrant attitudes and rival countries, police discrimination and caste, wartime preferences and territorial integrity. In each case, what looks like filling in blanks with stereotypes could just as well be a first-order preference that people only move past once it’s satisfied.

The recommendation for conjoint studies: don’t just compare “info” vs. “no info” conditions. Build in same-value ties deliberately, and use them to separate the two explanations — the ties are already part of the classic design, just not usually put to this purpose.

The full paper — figures, formal decomposition, and caveats on what kind of information processing might underlie all this — is out now in Political Science Research and Methods. Data and replication code are on Harvard Dataverse if you want to dig in yourself.