Statistics Done Wrong The Woefully Complete Guide (Alex Reinhart)（Z-Library）

# Statistics Done Wrong: The Woefully Complete Guide ## 【One-Line Pitch】 A witty, practical field guide to the most common statistical errors in scientific research, teaching readers how to spot flawed analyses and design studies that actually produce trustworthy results. Essential reading for scientists, researchers, and anyone who wants to critically evaluate claims backed by data. ## 【Book Arc】 - **Opening (~0%–10%)**: Introduces the core problem—scientists routinely misuse statistics, often unintentionally—and sets up the book's mission to expose common blunders. The author shares his origin story as a physics student who discovered that unintentional errors are far more interesting than outright fraud. - **Early (~10%–25%)**: Establishes the foundation of statistical significance, explaining what p values really mean (a measure of surprise, not importance), the historical Fisher vs. Neyman-Pearson debate, and why the conventional p < 0.05 threshold is arbitrary yet entrenched. - **Early–Middle (~25%–40%)**: Dives into statistical power and underpowered studies, showing how small sample sizes doom research before it starts, why "no significant difference" doesn't mean "no difference," and how the winner's curse inflates published effect sizes. - **Middle (~40%–50%)**: Explores pseudoreplication—collecting data that answers the wrong question—using examples like bird song dialects and the famous (and flawed) menstrual synchrony studies to show how non-independent samples fool researchers. - **Middle–Late (~50%–70%)**: Covers data-dredging sins: circular analysis, regression to the mean, stopping rules, needless dichotomization, confounding variables, and model abuse including correlation-vs-causation traps and Simpson's paradox. - **Late–Ending (~70%–100%)**: Addresses researcher freedom and bias, the pressure to publish exciting results, and concludes with practical advice on reporting analyses transparently, publishing data and code, and reforming statistical practice. ## 【Key Takeaways】 - **P values measure surprise, not truth** (Early): A p value tells you how surprising your data would be if the null hypothesis were true—it doesn't measure effect size, importance, or the probability your hypothesis is correct. Treat small p values as a prompt for further investigation, not a verdict. - **Statistical significance ≠ practical significance** (Early): With enough data, you can detect trivially small effects that have no real-world importance; conversely, an insignificant result doesn't prove no effect exists. Always ask whether the magnitude of an effect matters, not just whether it's "significant." - **Underpowered studies are doomed before they start** (Early–Middle): If your sample is too small to detect the effect you're looking for, you're wasting resources and setting yourself up for misleading conclusions. Power calculations should be a prerequisite, not an afterthought. - **The winner's curse inflates published effects** (Middle): When many researchers compete to publish exciting results, the lucky (not the best) studies get published, exaggerating true effect sizes. Early findings in fast-moving fields like genetics are often the most extreme—and the least reliable. - **Pseudoreplication answers the wrong question** (Middle): Treating non-independent observations as independent inflates your effective sample size and can produce confident but false conclusions. The classic bird song dialect studies and menstrual synchrony research both fell into this trap. - **Small samples create spurious patterns** (Middle): Counties with tiny populations show extreme cancer rates purely from random variation; the same logic explains why small schools or small groups often appear at both extremes of any metric. Shrinkage methods help but introduce their own biases. - **Researcher freedom breeds bias** (Late): When analysts have many choices—which variables to include, when to stop collecting data, how to define outcomes—they can unconsciously "double-dip" in the data and find patterns that don't generalize. Pre-registration and transparent reporting are antidotes. ## 【Reading Tips】 - **Skim the preface and introduction** (~0%–10%): They set up the book's philosophy and give you the big picture; the technical details come later. - **Deep-read Chapters 1–2** (~10%–40%): The p value explanation and power analysis are the conceptual core of the book. Work through the coin-flipping and drug trial examples carefully—they make abstract ideas concrete. - **Pay special attention to the pseudoreplication chapter** (~40%–50%): The bird song and menstrual synchrony case studies are memorable and illustrate a subtle error that's easy to commit unknowingly. - **Skim the later chapters on model abuse and researcher freedom** (~50%–80%): These are more applied and less technical; you can move faster while still absorbing the key warnings about correlation, confounding, and bias. - **Don't skip the final chapters on reporting and reform** (~80%–100%): They contain actionable advice on publishing data, sharing code, and changing how you and your colleagues approach statistics. ## 【Coverage Limits】 The excerpts cover roughly the first half of the book in detail (through pseudoreplication and early data-dredging topics), with lighter coverage of the later chapters on model abuse, researcher freedom, and reform. Specific chapter titles beyond Chapter 3 are partially visible in the table of contents but not fully elaborated in the source material. ##