Two polls published on different dates do not necessarily measure a change in public opinion. They may use different populations, methods, or questions.
A sound comparison starts by identifying what stayed the same.
Use interview dates
Publication dates tell you when a release became available. Fieldwork dates tell you when respondents answered.
Imagine one survey conducted October 1–4 and published October 8. Another was conducted October 5–7 and published October 9. The one-day publication gap conceals a more meaningful difference in interview timing.
If a major event occurred on October 6, the first poll does not measure responses to it. The second may include both pre-event and post-event interviews.
Match the population and question
Check whether both polls cover the same geography and office. Then compare registered voters with registered voters, or likely voters with likely voters.
Read the exact wording. A generic party question and a named candidate question should not be treated as interchangeable measurements.
Look beyond a single difference
A small change can reflect sampling variation. A larger difference can still involve method changes, respondent composition, or other sources of error.
AAPOR recommends considering timing alongside question wording and methods when interpreting differences between surveys. Source: AAPOR's journalist guide.
Rather than declare momentum from one comparison, look for repeated movement in a consistent series and corroboration from other suitable evidence.
Build a comparison table
Include the pollster, sponsor, fieldwork, geography, sample population, sample size, method, candidate estimates, and uncertainty statement.
For an illustrative series, if Candidate A moves from 46% to 47% to 48% under the same design, that deserves attention. It still does not by itself identify why the movement happened or guarantee the election outcome.
Avoid retrospective storytelling
A change appearing after a debate does not prove the debate caused it. Other events or compositional changes may have occurred.
Describe what the evidence supports: a measured change, an uncertain change, or a comparison limited by different methods. Keep explanations about causes separate from the numerical observation.
Related reading: margin of error and polling averages.