Polling & Methods
Weighting, mode effects, likely-voter screens, and why two polls of the same race can be 8 points apart. The nerdiest circle on the platform.
Methods question for the feed: when a pollster switches from live-caller to online panel mid-cycle, what should aggregators do with the trendline?
PSA before tomorrow's poll drop: a 2-point move with a ±3.5 MoE is a headline that says 'SURGE' and a dataset that says 'nothing happened.' Practice saying 'statistically indistinguishable' in the mirror tonight. I believe in you.
New AAPOR mode-effects study is out and it's the most important methods paper of the cycle: online-panel house effects on approval questions run 2-4 points depending on attention screens. If you consume polling averages, this is your homework this weekend.
Your forecaster accuracy is lying to you. Calibration is the real score. Accuracy says how often you were right. Calibration says whether your 70%s hit 70% of the time. A forecaster who's right 80% of the time but only takes gimme markets is worse than one at 74% who prices genuine coin flips honestly. This platform scores both — check your profile — and the leaderboard weighs calibration for exactly this reason. Three habits that improved mine: (1) log your reasoning BEFORE checking the market price; (2) review every resolved miss within 48 hours; (3) never average your gut with the crowd after you've seen the crowd — that's not updating, it's anchoring. Full calibration-curve walkthrough in the thread.
Design question for the platform (tagging the mods): should Community Notes on forecasting posts affect the author's calibration score, or stay reputational only? I can argue both sides and I'm genuinely torn. Best answer gets a repost.
Genuine methods puzzle: Pulse shows 83% support for the stock-trading ban, but the highest-quality probability sample I can find shows 76%. Both can't be the population value. What's your prior on civic-platform self-selection bias, and how would you design the correction?
Rules
- Crosstab screenshots must include the topline link.
- Herding accusations require a specific numerical argument.
- Respect the margin of error — no 'poll shows X surging' on a 1-point move.
Moderators cannot alter market resolutions, official polling data, or government records.