How the quiz works, how party positions are assigned, and how matches are calculated. Dataset v 1.1 , last material change 2026-06-26 . Previous versions of the question set and party positions are preserved and available on request from hello@whichparty.nz . WhichParty NZ scores a respondent’s 33 Likert-scale answers against the published positions of seven parliamentary parties to produce a percentage match per party plus a two-dimensional compass position. The match percentages are computed in the browser and reconstructed from the answer code embedded in the result URL whenever the page is reloaded or shared. That code records one digit per question in a fixed order, using 0 for any question left unanswered, and those 0 entries are dropped when the code is read back, so skipped questions never enter the match or compass. Because the code is positional, the question set and its order have to stay fixed, since changing them would invalidate result links that have already been shared. The compass coordinates are also recomputed on the server when a respondent chooses to submit, and stored alongside the answers for aggregate analysis. Match percentages themselves are never stored. The 33 questions 23 core questions and 10 optional refinement questions, covering 14 policy categories: economy, environment, immigration, retirement, health, energy, cost-of-living, law and order, housing, Treaty, education, social, governance, and digital. Each question is answered on a 1-5 Likert scale (strongly disagree through strongly agree). Questions are drafted to be specific enough that a 1 vs 5 reflects a real policy-level difference rather than an abstract value. Questions about capital gains tax or the 90-day trial, rather than about abstract values like freedom or equality, are what let the quiz tell seven parties apart. Almost every party will nod at a vague statement about helping families or growing the economy, so abstract questions would not separate them. Concrete questions can also be checked, since each party’s position can be sourced to a manifesto line or a recorded vote rather than to our own judgement. The downside is that some questions are tied to debates that are live in 2026 and will date as the cycle moves on, so the question set is reviewed and reworded over time, and any material change is recorded in the dataset changelog below. Every question is tagged with one compass axis (economic, social, or both) and a direction (+1 or -1) describing which end of the Likert scale corresponds to which end of the axis. Questions are also tagged with a weight (default 1) where some questions are structurally more informative than others. Party position assignment Each party is assigned a position from 1 to 5 on each of the 33 questions. The source material for these ratings includes: published manifestos, members’ bills, coalition agreements, voting records on legislation.govt.nz, Hansard contributions, ministerial press releases, leader interviews on RNZ / 1News / Newshub / The Post, and public social-media statements. Where a party has no public position on a question, it is assigned 3 (neutral) with a written note. Every party-position pair carries a written prose justification, published on the per-category breakdown pages under topics . Each topic page lists every question in that category with each party’s position, the source quote or document the rating is drawn from, and a short explanation. Positions are reviewed as new policy is announced between now and election day. Parties that dispute an assigned position can follow the corrections process on the contact page. Importantly, questions on which a party is genuinely neutral (position 3) are excluded from that party’s match calculation for every respondent. A party cannot be credited or penalised for a question on which it has not taken a public position. Match percentage For each answered question where the party has a non-neutral position, we score the match between the respondent and the party on a 0 to 1 scale. The scoring treats which side of neutral you fall on as more important than how strongly you feel. When you and the party are on the same side of neutral, the score is high and stays high even if your intensity differs, so answering agree against a party that answered strongly agree still scores well. Exact agreement scores 1. If you answered neutral on a question, that question scores a flat 0.5 for every party. When you and the party are on opposite sides of neutral the score drops below 0.5 and falls towards 0 as the answers move further apart, with directly opposite extremes scoring 0. Because every same-side score sits above 0.5 and every opposite-side score sits below it, there is a clear jump at the neutral line: two same-side answers that differ only in intensity stay close together, while crossing to the other side of an issue is penalised much more heavily. We weight direction over intensity because on a five-point scale which side of an issue you come down on is a more reliable signal of party fit than the exact strength of the answer. Per-question scores are combined into a weighted average. The questions are not all weighted equally. Each question is weighted by how much the seven parties actually disagree on it, measured as the standard deviation of their assigned positions and then scaled so the weights average one across the 33 questions. A question where every party sits in roughly the same place tells you little about which party fits you, so it counts for less, while a question that splits the parties across the scale counts for more. These weights are computed only from the parties’ published positions, never from respondents’ answers, and the same weight is applied to every party. This spread-based weighting is what drives the match, and it is separate from the per-question weight tag mentioned in the questions section, which is currently 1 for every question and is used only in the compass. The weighted average gives a raw 0 to 100 figure, which is then passed through a fixed non-linear stretch before it is displayed. The stretch pushes scores above 50 percent a little higher and scores below 50 percent a little lower, with the largest adjustment, a little under ten points, occurring roughly a quarter and three-quarters of the way along the scale. The points 0, 50, and 100 are left unchanged. Because the stretch only ever increases as the raw score increases, it never changes the ordering of parties or which party a respondent matches best. Its only purpose is to widen the visible gap between a respondent’s stronger and weaker matches so a crowded result list is easier to read, and it is applied identically to every party. The displayed figure is finally rounded to the nearest whole number, so two parties shown with the same or adjacent percentages may have slightly different underlying scores. Questions where a party is genuinely neutral (position 3) are excluded from that party’s match calculation, as noted in the party-position assignment section above. This means a respondent who skips the optional refinement questions is simply scored on the questions they did answer, and a party can show zero percent when every question the respondent answered happens to be one that party is neutral on. If a respondent has answered no questions at all, every party match is reported as zero percent rather than the page showing an error. Political compass Two axes. Economic x, from −1 (left, more government) to +1 (right, free markets). Social y, from −1 (progressive) to +1 (traditional). Each axis is computed independently by summing the respondent’s deviation from neutral on every question tagged to that axis, multiplied by the question’s direction and weight, and dividing by the maximum possible absolute score. Questions tagged as relevant to both axes are counted at half weight on each. All question weights are currently set to 1, so in practice every question contributes equally apart from that half-weighting of both-axis questions. The resulting (x, y) is mapped to one of fourteen ideology labels (Pragmatic Centre, Social Democrat, Classical Liberal, Green Left, Social Conservative, Libertarian, and so on) using fixed thresholds. A user whose coordinates fall within 0.25 of the origin on both axes is labelled Pragmatic Centre regardless of which quadrant they sit in; outer quadrants are named more specifically as the coordinates move further out. The compass is an illustrative two-dimensional summary that projects the answers onto two chosen axes. The party match percentages are computed separately across all of the questions and do not use these two axes, so how close a respondent sits to a party on the compass is not the same thing as their match percentage with that party. Read the compass as a rough placement of where someone sits, and the match percentages as the actual measure of party fit. Coalition builder A coalition’s blended position on any question is the arithmetic mean of its member parties’ positions on that question. This mean is unweighted, so every member party counts equally regardless of its size or poll share, and a party’s position is kept even when it is a genuine neutral (position 3). This differs from the single-party match, where a party’s own neutral positions are dropped, so here a member party that is neutral on a question pulls the blended position towards the centre rather than being ignored. The coalition’s compass point is derived from the same axis formula applied to that blended platform. The match percentage between a respondent and a coalition uses a distance-based metric, 1 minus the mean per-question absolute distance divided by 4, scaled to 0-100, which is a different calculation from the single-party match described above. Seat counts in the coalition builder use the NZ-modified Sainte-Laguë formula (first divisor 1.4, subsequent divisors 3, 5, 7...) with the 5 percent party-vote threshold waived for parties that are modelled as likely to win at least one electorate. Party-vote inputs come from the latest published polling aggregate sourced from the Wikipedia rolling average for the 2026 election and refreshed roughly monthly; the snapshot date is recorded in the dataset. The simulator ignores overhang seats and assumes a 120-seat Parliament. The builder also flags a selected bloc as a coalition of losers when its combined polled party vote is lower than the polled party vote of the single largest party left out, comparing the raw party-vote percentages from that same snapshot. The seat tally treats a governing majority as 61 of the 120 seats, so any combination reaching 61 is shown as forming a majority. The majority check first converts the polling snapshot into seats using the Sainte-Laguë allocation described above, while the coalition-of-losers flag works straight off the party-vote percentages. This section is a methodology-level summary. The full privacy policy is the authoritative document for what is collected, how it is used, retention, and your rights under the Privacy Act 2020. Each quiz submission is stored with a random session ID, the answers as integers 1-5, the computed compass coordinates, a SHA-256 hash of the IP address (salted with a secret), a coarse device-type bucket (mobile, tablet, desktop, or bot derived from the user agent rather than the raw user-agent string), and the total time spent on the quiz. Nothing personally identifying is stored alongside the answers. Submissions completed implausibly fast are treated as bot traffic and discarded before reaching the database. This particular check reads a timer reported by the browser, so it is a soft filter that a determined script could bypass, and it works alongside other anti-automation checks rather than standing on its own.

The submission endpoint is rate-limited per IP hash and protected by standard anti-automation checks. Specific thresholds are not published to keep them effective. The demographics section is required and is saved against the same random session ID. The fields collected are gender, age band, electorate, ethnicities, enrolment status, 2023 vote, current vote intention, certainty, parties under consideration, and top issues. Email addresses are collected only via the explicit opt-in on the results page and are stored in a separate table from the quiz answers. To remember which result was shared with a subscriber and to prevent duplicate sign-ups, the opt-in row also records the answer code that produced the result and a salted IP hash. In principle this means an email address could be associated with the answer set that was on screen when the user opted in, and could be matched to a quiz session from the same IP. Email subscribers can unsubscribe at any time using the link in every email. The site exposes per-question average answers across all submissions as an aggregate trends feed; individual responses are never exposed publicly. A separate group-comparison feature lets a respondent share a code with friends and view everyone’s compass points and top party on a single page. Group membership stores a display number, optional nickname, compass coordinates, top party, and answer code against the group ID, with no personally identifying information. See the full privacy policy for retention, access rights under the Privacy Act 2020, and the complete list of what is and isn’t collected. WhichParty NZ is a self-directed project and is not under a university ethics board review. In the absence of that external oversight, the project operates by the following principles. Informed consent. Before submitting the quiz, respondents can read the privacy policy linked from every page. The email opt-in is explicit and opt-in rather than opt-out: respondents choose whether to sign up for the WhichParty 2026 Voter Guide, which is never shared with any political party or other third party. Anonymity and data minimisation. Quiz submissions carry no name, account, or other identifier. The IP address is hashed one-way with a secret salt. The user-agent string is bucketed to a coarse device type rather than stored verbatim. No individual-level sharing. Data shared with third parties (media, researchers, parties) is aggregate only. Individual quiz responses are not for sale, transfer, or inspection by anyone outside the project. Right to contact. Any respondent can contact the project to ask what is held about them, or to request correction or deletion of any identifiable record. Details and response timing are on the privacy policy. Collaboration. Where the project provides data to academic researchers, the researcher’s own institutional ethics process applies. The project is happy to support any ethics documentation a collaborator needs. The quiz is a useful pointer rather than a definitive verdict. 33 Likert questions cannot capture every dimension of a party platform or a voter’s views. Match percentages reflect alignment on the questions covered, not a comprehensive policy audit. Where the difference between a respondent’s top two or three matches is a few percentage points, the signal should be treated as noise. The respondent pool is self-selected. WhichParty NZ is not a random-sample survey and its aggregate figures should not be treated as polling. Public figures derived from respondent data are directional and should be interpreted as such. Party positions reflect our best reading of public material as of the most recent review. Where we have misread a position, parties are encouraged to follow the corrections process on the contact page. Dataset changelog Material changes to the question set and party positions are versioned so that papers and articles can pin to a specific revision. Polling refreshes are not versioned here; the polling snapshot carries its own date in the dataset. v 1.1 · 2026-06-26 ACT's position on Q3 (prioritise training locals over importing workers) moved from strongly disagree to disagree, reflecting its May 2026 immigration policy where skilled-visa categories default to closed unless ongoing demand is re-proven, alongside the June 2026 Rural Workforce Visa. Explanation updated. v 1.1 · 2026-06-26 ACT's position on Q3 (prioritise training locals over importing workers) moved from strongly disagree to disagree, reflecting its May 2026 immigration policy where skilled-visa categories default to closed unless ongoing demand is re-proven, alongside the June 2026 Rural Workforce Visa. Explanation updated. v 1.0 · 2026-04-23 Initial published dataset: 33 questions across 14 categories, positions for the seven parliamentary parties. v 1.0 · 2026-04-23 Initial published dataset: 33 questions across 14 categories, positions for the seven parliamentary parties. Questions about methodology For questions about any part of this document, to suggest improvements, or to request more detail on a specific party position, email hello@whichparty.nz or see the contact page .