By the Wavefield Research team · Published Sep 14, 2026
The standard gender survey question: “How do you describe your gender?” with Woman, Man, Non-binary, a self-describe box, and Prefer not to say. That wording is the easy part. The hard parts — weighting five answer options against a binary census frame, and reporting small categories honestly — are what this guide actually covers.
In order: the recommended question and why each option earns its place, sex versus gender as two separate questions, three graded option sets, and then the three problems every other guide stops short of — weighting, small bases, and quotas.
“How do you describe your gender?” asks about identity in the respondent's own frame, which is why it outperforms “What is your gender?” — describe signals that the survey will accept the answer given rather than validate it against a list. Pew Research Center's methodology work on adapting its own gender question is the load-bearing reference here: their tested phrasing — “Do you describe yourself as a man, a woman, or in some other way?” — reaches the same design from a different direction, a closed question with a genuine exit.
Each option has a distinct job. Woman and Man carry most of the base. Non-binary is the largest identity the binary misses, and listing it matters more than the write-in alone — an identity that must be typed rather than ticked gets systematically undercounted. The self-describe box (on Wavefield, a specify option: tick it and an inline text box opens, storing the answer in the respondent's own words) catches everything the list missed. And Prefer not to say is a refusal, not an identity — keep it last, keep it always, and never pool it with self-described answers in a report. How to ask gender in a survey is, in the end, mostly a matter of respecting those distinctions mechanically.
Sex is assigned at birth; gender is current identity. Health research, government statistics, and any study where the biological variable is analytically load-bearing use the two-step: “What sex were you assigned at birth?” followed by “How do you describe your gender?” — the pattern that also identifies transgender respondents without ever asking “are you transgender?” directly. Statistics Canada's 2021 Census ran exactly this design at national scale — the first census anywhere to publish data on transgender and non-binary populations.
Most market research does not need the two-step. A brand tracker, a concept test, a customer satisfaction study — these use gender as a profiling and weighting variable, and the single question suffices. The real rule is negative: never blend the constructs into one question. “Gender: Male / Female / Transgender” is a category error — transgender women are women, and the option list forces an impossible choice — and it belongs in the same bin as the double-barreled items in our bad survey questions collection: questions that fail silently and produce clean-looking, wrong data.
Woman · Man · Non-binary · Prefer to self-describe: ____ · Prefer not to say
The standard five-option gender set fits most market research: respectful, quick to answer, and analyzable. The self-describe box and the opt-out do different jobs — one is an identity answer, the other is a refusal — so both stay, always, and they are never merged in reporting.
Woman · Man · Non-binary · Transgender woman · Transgender man · Two-spirit · Another identity: ____ · Prefer not to say
An extended gender list belongs in studies that are about gender identity, where the distinctions are the findings. In general-population work it splits already-small categories into unreportable slivers — more precision at collection, less usable data at analysis.
Woman · Man
A binary-only gender question is occasionally defended for tracker comparability with old waves. The defense is weak: adding options does not break the trend for the two original categories, and the cost is answered-wrong or abandoned surveys from everyone the list excludes. If a legacy client insists, log the objection and add at least a self-describe box.
Placement follows the same logic as the rest of the demographic block: near the end of the survey, after the substantive questions, unless a quota needs it up front — the full argument, including when the ask-demographics-last advice breaks, is in how to ask about age and demographics. Survey questions about gender differ from the rest of that block in one way only: the option list itself is the sensitive surface, so the respect work happens in the options, not the placement.
Here is the problem every wording guide walks past. You asked gender with five options, because that is the right question. Now you need to weight the sample to population benchmarks — and the benchmark frame is almost certainly binary. The US Census still asks sex with two categories; most syndicated frames follow it. Canada's 2021 Census publishes a three-category gender distribution, and even there the third category is a third of a percent — far too small to serve as a raking target for a 400-person study.
The working convention: weight on a collapsed variable, report the full one. Rake against the frame's two categories using the respondents who fall into them, and leave everyone else out of that raking dimension — on Wavefield this is the default mechanical behavior, because respondents whose answer has no weighting target simply stay at weight one, and the output tells you how many were left unraked. What the convention forbids is the tempting shortcut: asking the binary question because the weighting frame is binary. That order of operations runs the survey backwards — the frame is a constraint on the correction step, not on what you are allowed to learn. The mechanics of raking, trimming, and effective bases are covered in survey weighting.
In a 400-complete general-population study, expect single-digit respondents outside Woman and Man. That is not a flaw in the question — it is the population's actual shape reaching your data. The failure mode is what reports do next: quoting “non-binary respondents were twice as likely to…” on a base of six. Our crosstabs enforce the floor mechanically — columns under 30 effective base appear in tables but are marked untested and excluded from significance letters, so the claim cannot dress itself in statistical clothing it does not have.
When a comparison genuinely matters, combine categories for the statistical test without erasing anyone in the labeling — “Non-binary and self-described (n=14)” is honest; silently folding those respondents into a category they did not choose is not. And in small populations — a 200-person company, a niche panel — small gender categories become a privacy problem before they become a statistics problem: a table cell of two is an identification, not a finding. Suppress counts below five in anything published internally, the same discipline customer feedback and employee reporting already apply to verbatims.
Gender quotas inherit the weighting collision: the quota targets come from a binary frame, the question offers five answers. The workable design is two cells matched to the frame — Woman and Man at their census shares — with all other responses flowing through unquoted. Capping the categories the frame cannot see would screen out real respondents to satisfy a spreadsheet. On Wavefield, quota cells attach to specific answer options, so this design is declarative: two capped cells, everything else open, and a full cell screens matching respondents out mid-survey automatically — the mechanics survey screening questions covers in full.
As a screener, gender is rarer than beginners expect: unless the product or sample design is gender-specific, gender belongs in the demographics block, not the screener. Asking it early costs goodwill at the exact moment respondents are deciding whether the survey deserves fifteen minutes — and a screener that terminates on gender without a gender-specific reason reads, correctly, as a quota dressed up as eligibility.
The working standard is five: Woman, Man, Non-binary, a self-describe write-in, and Prefer not to say. The write-in catches identities the list missed and stores the respondent's own words; Prefer not to say is a refusal, not an identity, and absorbs opt-outs that would otherwise become abandonment or fiction. Extend the list only when the study is about gender identity itself.
Yes, functionally: it should always carry Prefer not to say, and nothing in the survey should force a real answer. Refusal rates on gender are low — far lower than income — but the option's presence is doing quiet work for data quality: respondents who would refuse pick the honest opt-out instead of a random category, which keeps the categories you analyze clean.
Sex is assigned at birth and recorded on documents; gender is current identity. They are separate constructs and separate questions — a survey that needs both asks both, in the two-step pattern health research uses (“What sex were you assigned at birth?” then “How do you describe your gender?”). Most market research needs only gender, and mixing the two constructs in one question produces data that measures neither.
Two disciplines. Weighting: census frames are still mostly binary, so rake on a collapsed two-category variable while reporting the full question — respondents outside the collapsed categories stay at weight one rather than being forced into a frame that has no target for them. Significance: categories under roughly 30 effective base appear in tables but cannot be significance-tested, so plan reporting groups before fielding, not after.
Related: how to ask about age · survey weighting · bad survey questions, fixed · survey question types
Brief the agent and it writes the questions, attaches the specify boxes, sets the quota cells, and leaves untargeted categories honestly unraked. From $99 per study.
Start a study