Polling margin of error vs. the poll's expected voter turnout model.
Maybe someone can explain the difference? I tend to think that if the poll's turnout model is wrong then the margin of error doesn't mean anything. But I could be wrong.
Polling margin of error is the statistical uncertainly that is created by sampling a smaller group that is intended to be representative of a larger group. The MOE is based on sample size versus population size and is done via a mathematical calculation. You can play with the sample size and population size inputs via a MOE calculator to see the effects:
Margin of Error Calculator
On a related note, MOE is why, if you dive into the crosstabs of a poll where they break the results down by voter characteristics (gender, race/ethnicity, party ID, etc) there will often be no results provided for groups that aren't reasonably large within the polled group. If you don't have enough respondents from a particular group, the MOE is so large that the polled group isn't considered "representative" at that level and therefore results typically aren't provided.
Expected voter turnout is something that pollsters have to model (aka, make an educated guess at) as part of their polling methodology. Note: This is only true for "Likely Voter (LV)" polls, "Registered Voter (RV)" polls do not have to model turnout because they use the party ID percentages of how voters are registered in whatever area (typically either region or state, occasionally congressional district) they're polling. Turnout used to be relatively stable, especially on the Pub side, and so most pollsters were in reasonably close agreement of what turnout would look like (not necessarily in a colluding sort of way, more that with turnout being relatively stable it was far easier to get a reasonably correct answer and so most had similar answers).
Two things have happened regarding turnout & polling...
The first is that turnout has become much more unpredictable as Trump has brought new folks into his base who are unreliable voters and as the extremism in our political system has brought out more unreliable voters on both sides of the aisle. Trump introducing unreliable voters into the Republican Party seems to have played a significant role in Pubs under-performing in midterm elections in 2018 and 2022, as many of these "Trump voters" don't seem to vote when Trump isn't one the ballot. For the purposes of polling, it has made things much, much tougher for legitimate polling firms doing Likely Voter polls because it has made modeling turnout much more difficult. Do you expect unreliable Pubs to turn out? Unreliable Dems? Both? Neither? How polling firms estimate likely turnout determines how they weight the responses they've received from doing polling which directly impacts the results of their polls. After polling firms missed in 2016, many adjusted moving forward under the idea that there were "silent Trump voters" who either wouldn't admit they would vote for Trump/Republicans or polling largely missed altogether in their samples. But then in later elections that caused many polls to miss results by overstating Pub support and so now pollsters are divided on how to properly model turnout, leading to more varied results than in pre-2018 elections.
The second is that biased pollsters, namely right-wing polling firms who want to create biased polls showing better results for Trump and Pubs, have figured out that their turnout model is an easy way to put their thumb on the scale for Trump and other Pubs. Simply have your turnout model show more Pub-leans and less Dem-leans and, voila, a poll where the results instantly move toward Pub success. Because of this use of turnout modeling, modeling across various polls have become even more varied/uneven than among neutral pollsters, creating more uncertainty in terms of polls.
I hope that gives a good overview of MOE and voter turnout modeling. Polling is an interesting subject, if you take the time to look into it.
