⏱️ Slide Timing: 1 min - Welcome and frame the session: a classic 1973 statistics puzzle with a very modern lesson - Promise: by the end, everyone will distrust a bare set of summary statistics - Bridge: start with a setup that looks too good to be true

⏱️ Slide Timing: 3 min - Statistician Francis Anscombe built these four datasets in 1973 specifically to make this point - Every summary number a regression report would show you matches, exactly, across all four - Ask the audience to hold their answer to the blockquote question — the next slide reveals it ❓ Ask: "Based on these six matching statistics, would you expect these four datasets to look the same on a chart?"

⏱️ Slide Timing: 4 min - Reveal: identical statistics, four wildly different shapes - Point out each quadrant briefly — clean line, curve, outlier-driven line, vertical cluster with one flier - This single image is the entire argument for the rest of the deck — everything after is elaboration

⏱️ Slide Timing: 3 min - This is the textbook case — the one every stats course example secretly assumes - Real scatter around the line is normal and expected, not a sign anything is wrong - Worth noting: even here, no individual point sits exactly on the regression line

⏱️ Slide Timing: 4 min - Analogy: fitting a straight ruler to a rainbow — it touches at a couple points, misses everywhere else - The danger here is subtle because the fit still "looks" reasonable on a summary printout - Foreshadow the residual-plot lesson later in the deck — this is exactly the case it's designed to catch

⏱️ Slide Timing: 4 min - Without that one point, this dataset is a near-perfect line — the outlier is doing all the damage - Real-world example: a sensor misread, a fat-fingered data entry, or a genuinely unusual event worth studying - Key discipline: investigate before you delete — the outlier could be the most informative row in the table

⏱️ Slide Timing: 4 min - Ten of eleven points carry zero information about the x-y relationship — one point manufactures the entire correlation - This is the most dangerous pattern of the four because the regression output gives no visible warning - Practical test: does the conclusion survive if that single point is removed? If not, don't trust it

⏱️ Slide Timing: 4 min - Translate the abstract math into a boardroom decision: same broken statistics, real budget on the line - Dataset II failure mode: overspending past the point of diminishing returns because the model assumed linearity - Dataset IV failure mode: a lucky quarter with one big campaign gets mistaken for a durable trend

⏱️ Slide Timing: 2 min - Transition slide — moving from "here's the puzzle" to "here's what to do about it" - Six lessons follow, each one a direct, actionable takeaway from what was just shown

⏱️ Slide Timing: 2 min - A scatter plot takes seconds to make and catches problems no summary table will surface - Encourage a habit: never run a regression on data that hasn't been plotted first

⏱️ Slide Timing: 3 min - These five assumptions are the fine print behind every linear regression coefficient - Dataset II broke linearity, Dataset III and IV broke the outlier assumption — same model, different failure modes - Not all violations are fatal, but each one should be checked and reported, not assumed away

⏱️ Slide Timing: 3 min - Residual plots are the diagnostic tool that would have caught Dataset II immediately, even without seeing the scatter plot - Rule of thumb: if you can see a shape in the residuals, the model is missing something > heteroscedasticity - spread of data points is not the same every where

⏱️ Slide Timing: 3 min - Leverage differs from a simple large residual — it's about how unusual the x-value is, not just how far off the prediction is - The "don't auto-delete" rule matters most in fraud detection and anomaly work, where the outlier is often the whole point

⏱️ Slide Timing: 3 min - Ties the whole deck together: identical R² across all four datasets, radically different reliability - A single metric is a compression of the data — and compression always loses information

⏱️ Slide Timing: 2 min - This is the workflow version of everything covered so far — a checklist to run before trusting any model - Step 7 is deliberately last: skepticism should persist even after a model passes every earlier check

⏱️ Slide Timing: 3 min - Grounding the four abstract patterns in domains the audience will recognize - Dose-response curves in medicine are a classic Dataset II case; fraud detection often lives in Dataset III territory

⏱️ Slide Timing: 2 min - This list doubles as a code-review checklist for anyone signing off on someone else's regression work - Every item on this list is a mistake that Anscombe's Quartet, by construction, would sail right through

⏱️ Slide Timing: 3 min - Technical skill alone would not have caught these four datasets — critical thinking and communication close the gap - "Don't fall in love with your models" is worth repeating out loud — it's the hardest habit on this slide to build

⏱️ Slide Timing: 2 min - Tukey and Anscombe were contemporaries — this quote could have been written about this exact dataset - Land the "guilty until proven innocent" framing as the one-line mental model to carry forward

⏱️ Slide Timing: 1 min - Closing summary before references — the three lines here are the entire deck compressed - If nothing else survives from this session, this slide should

⏱️ Slide Timing: 1 min - Good slide to share on-screen for anyone who wants the visual recap after the session - Worth revisiting before presenting any regression result to a non-technical audience

⏱️ Slide Timing: 1 min - Anscombe (1973) is the primary source — worth bookmarking for anyone who wants the original four-dataset construction - Tukey (1977) is the broader philosophical grounding for "always plot first"