Why did three Taipei post-debate polls disagree? One weekend, different samples

Three Taipei mayoral polls interviewed in overlapping periods after the October 31 debate. Subtracting Puma Shen from Chiang Wan-an yields +4 points for TVBS, +11.3 for ETtoday and −17.3 for the Taiwan Strategy and Policy Institute. The 28.6-point span compares published margins. It does not mean that 28.6 points of voters switched over the weekend.[1][4][6]

Ask who could be reached, who entered each denominator and what each respondent was asked. Similar field dates do not make the samples interchangeable. A simple average would assign weights without validating them.

Margins are calculated from published percentages. Each row has its own selection process; none is a count of votes cast. The disclosed information does not support a common significance test across rows.

A residential phone survey depends on household reachability and willingness to answer. A dual frame adds mobile coverage but still has to handle overlapping coverage and incomplete interviews. A member survey first depends on platform membership and then on invitation and response. These mechanisms can select different people without establishing the direction of any bias.

Weighting by age, sex or district cannot ensure that respondents and nonrespondents within those groups have the same political preferences. A larger sample reduces some random variation but cannot by itself correct coverage or response differences.

The table cannot identify an AI interviewing effect. A more useful design would randomly assign human and automated interview modes within a shared recruitment frame and compare completion and response patterns. These polls do not provide that experiment.

Mirror Media and NOWnews report the same ETtoday preferences. Another Mirror Media report and Newtalk report the opposite leader in the institute survey. Media Corner compares all three and identifies the restricted TVBS preference denominator. Five reports are traceable accounts, not five independent measurements.[1][2][3][4][5]

Expecting a candidate to win differs from intending to vote for that candidate. ETtoday asked respondents exposed to debate content to score the candidates; the other polls compared who performed better. A mean score and a share choosing one candidate use different scales. Seeing clips also differs from viewing the whole debate.

Mobilization and willingness to state a preference may differ across samples. Without repeated observations of the same people and the complete questionnaires for the two additional waves, the differences cannot be decomposed into mobilization, selection and wording. That limitation does not establish any sponsor’s motive.

To assess movement, compare successive waves from the same pollster with similar wording, denominators and mode, then check field dates, response and weighting. To assess whether the debate changed individuals’ preferences, track the same people and address differences that made some people more likely to watch in the first place.

Before November 28, treat the three results as conditional observations. Publishing wording, distributions before and after weighting, sample bases and treatment of undecided responses would make later comparisons more diagnostic. The reported measurements differ; the evidence does not yet identify why.