Is Kaohsiung 2026 a Toss-up? Sampling, Weighting and Member Surveys

ETtoday’s September 11 Kaohsiung release reported 43.8% for Ko Chih-en and 43.5% for Lai Jui-lung, a 0.3-point difference. That readily sounds like a toss-up. But respondents were invited from a proprietary member database to complete an online questionnaire. Generalizing to Kaohsiung voters first requires asking how city residents entered that database.

Whether a member questionnaire supports population-level polling claims is a question about sampling inference. This analysis does not treat it as equivalent to a probability sample of all city residents, nor does it guarantee other surveys merely because they use telephone interviews. As of October 10, 2026, one release with full statutory disclosure illustrates the methods; this is not a ranking of electoral estimates. No original interviews or survey were conducted.

Read methods before candidate rankings

Statutory disclosure under Article 53 of the Civil Servants Election and Recall Act, as published by ETtoday on September 11: survey organization, ETtoday Poll Cloud Co., Ltd.; principal, Hsieh Hui-ling, chief operating officer; fieldwork, August 22–30, 2026; sampling, stratified proportional sampling within the company’s member database, with email and text-message invitations to a closed web questionnaire; population, 406,434 members tagged as registered Kaohsiung residents aged 20 or older across 38 districts; sample, 1,283 valid responses; margin of error, ±2.74%; funding, ETtoday Poll Cloud Co., Ltd.

ETtoday’s original report describes a proprietary database containing 406,434 members tagged as registered Kaohsiung residents aged 20 or older as of August 26, stratified proportional sampling within that database, email and SMS invitations, and a closed online questionnaire. It reports 1,283 valid responses from registered residents aged 20 or older across Kaohsiung’s 38 districts, iterative weighting (raking) by sex, age, district and education, and 12.5% undecided.

Published precision is retained rather than forcing totals to 100%. ETtoday’s 43.8 + 43.5 + 12.5 totals 99.8%. These discrepancies in published totals do not justify rewriting undecided shares or assigning hidden votes to a candidate.

Random selection within a member pool is only one stage

A sampling frame is the list of units or contact records available for selection. Probability sampling uses known, nonzero selection probabilities within the frame; coverage of the target population must also be assessed. Stratified sampling inside a member database improves selection within that pool. It does not establish known probabilities for city residents to enter the pool in the first place.

The release does not establish recruitment from all city residents with known probabilities. It therefore does not demonstrate a citywide probability sample. This does not prove that all members opted in, or establish fabrication; the missing evidence concerns recruitment and population inference. Closed invitations can limit unrestricted participation and differ from an open click-to-vote poll, but being closed alone does not resolve coverage or self-selection.

Telephone surveys face parallel checks. A landline frame may miss residents who use only mobile phones. Selecting a number leaves questions about which household member answers and how that person is chosen. Participation and handling unanswered calls raise nonresponse issues. Without full procedures, interview dispositions and unweighted sample information, “live telephone interviewing” is not a complete quality assessment.

Online surveys can use probability recruitment

Online completion is a mode, not a verdict on sampling quality. A Pew Research Center study published in 2023 compared three U.S. probability-recruited online samples with three opt-in samples. Across 28 benchmark measures, average absolute errors were about 2.6 and 5.8 percentage points, respectively. A panel can recruit through randomly selected addresses and then collect responses online. Online questionnaires are therefore not inherently incompatible with professional polling.

This is a U.S. sample-comparison study, not a correction formula for Taiwanese elections. Its 5.8-point figure cannot be subtracted from Ko’s ETtoday support or used to guarantee telephone accuracy. It supports examining recruitment; it does not identify the most accurate estimate of this Kaohsiung election.

Weighting aligns measured demographics, not every missing preference

Raking adjusts respondent weights so marginal distributions such as sex and age match population targets. When younger people are underrepresented, the younger respondents can receive larger weights. But do they represent peers who never joined the database or never responded? Unmeasured political differences can remain despite demographic alignment, and highly variable weights can reduce effective sample size.

Thus, a nonsignificant post-weighting difference from demographic targets concerns the characteristics tested, not unbiased candidate support. Pew’s weighting research also shows the importance of choosing adjustment variables. Unaddressed factors related to participation and voting preference may persist even as the sample grows.

The meaning of a stated margin of error

For teaching purposes, a simple random sample at a 50% proportion has an approximate maximum 95% sampling margin of 1.96 × √(0.25 / n). At n = 1,283, that is about ±2.74 points. Computing it does not resolve recruitment, nonresponse or weighting issues. Precision for probability surveys should account for design effects. Precision claims for nonprobability samples require disclosure of the model, assumptions, calculation and validation basis.

The ETtoday release reviewed does not provide a complete inferential basis for interpreting ±2.74 as a population-level confidence interval. The table attributes it to the publisher without endorsing that interpretation. Stated telephone-survey margins also warrant checking whether weighting and design effects were incorporated, rather than checking sample size alone.

Another mistake is applying a single-proportion margin to the candidate difference or a small subgroup. A margin for all 1,283 respondents does not directly apply to one age group; the candidate difference also has its own uncertainty. A 0.3-point point-estimate gap does not identify the winner or imply a 50% win probability for each candidate.

Six questions for the next release

These checks make a polling label verifiable. Missing information limits specific conclusions. A member questionnaire without adequate recruitment evidence describes its survey sample, while stronger population claims require more support. Telephone surveys with missing procedures or disclosures deserve the same scrutiny.

Kaohsiung’s 2022 valid-vote result was 58.10% for Chen Chi-mai and 40.16% for Ko, a 17.94-point gap. Different opponents and timing prevent treating it as Lai’s starting margin. A single survey cannot establish a trend for either candidate. Comparable surveys with transparent methods from the same period would be more useful than another toss-up headline.