Whoville Statistics: Regression, Probability, and Distributions Summary & Study Notes

📊 Regression and Correlation

Least-squares regression fits a line that minimizes the sum of squared residuals. For two variables $G$ (green marbles) and $IQ$ (WhoIQ score), the slope is computed from the sample correlation $r$ and standard deviations: $b_1 = r \frac{s_{IQ}}{s_G}$. The intercept is $b_0 = \bar{IQ} - b_1 \bar{G}$. Use these to form the prediction equation $\widehat{IQ} = b_0 + b_1 G$.

🔢 Interpreting coefficients

The slope ($b_1$) gives the estimated change in the response (WhoIQ) for a one-unit increase in the predictor (one more green marble), holding other factors implicit in the model. The intercept ($b_0$) gives the predicted WhoIQ when $G=0$; this can be meaningful only if $G=0$ is within the range of observed data. If $G=0$ is outside the observed range, the intercept is primarily a computational anchor, not a real-world guarantee.

🧾 Example values from the Grinch's study

Using $\bar{G}=23.1$, $s_G=6.54$, $\bar{IQ}=135.7$, $s_{IQ}=9.23$, and $r=0.65$ gives $b_1 \approx 0.918$ and $b_0 \approx 114.49$. The model predicts about a $0.918$ point increase in WhoIQ for each additional green marble, and a predicted WhoIQ of about $114.49$ when $G=0$ (interpret with caution).

🧠 Predictions vs. Individual Outcomes

Regression predicts expected (mean) outcomes, not guaranteed individual scores. An individual’s actual score may differ from the prediction because of residual variation (scatter around the line). To claim a specific person's score will change by a precise amount requires assumptions of causation and small residual variability; correlation alone does not establish causality.

🎯 DJ Who scenario (adding 11 marbles)

Arithmetic: adding 11 marbles changes the predicted score by $b_1 \times 11 \approx 0.918 \times 11 \approx 10.1$ points, so the predicted score moves from 113 to about 123.1. Logic: this is a statement about expected change, not a certainty for an individual. Also, unless the study design supports a causal claim (e.g., randomized experiment), we cannot be sure adding marbles causes IQ changes.

🔁 Conditional probability and Bayes' rule

If the population is partitioned into groups (clubs) and group-specific preferences are known, the overall probability of an attribute is a weighted average: $P(\text{green}) = \sum P(\text{club}) P(\text{green}|\text{club})$. To find $P(\text{club} | \text{green})$ use Bayes' theorem: $P(A|B)=\dfrac{P(A\cap B)}{P(B)}$.

🎲 Binomial and Normal Approximations

A Binomial model $Bin(n,p)$ applies when there are fixed $n$ independent trials with common success probability $p$. For large $n$, use the normal approximation with mean $np$ and variance $np(1-p)$; include a continuity correction when approximating discrete probabilities: e.g., $P(X\ge k) \approx P\big(Z \ge \dfrac{k-0.5-np}{\sqrt{np(1-p)}}\big)$.

📏 Finite population note

When sampling without replacement from a finite population, trials are not strictly independent. If the sample size is small relative to the population (common rule: sample size < 5% of population), the independence approximation holds and the Binomial model is reasonable; otherwise apply the finite population correction.

📊 Discrete distributions: expectation and spread

For any discrete random variable $X$ with probabilities $p(x)$, the mean is $\mu = E[X] = \sum x p(x)$ and the variance is $\sigma^2 = E[X^2] - \mu^2$. The standard deviation is $\sigma = \sqrt{\sigma^2}$.

🔎 Normal distribution and z-scores

For a normal $N(\mu,\sigma)$, convert to a standard normal via $z = \dfrac{x-\mu}{\sigma}$. Use tables or software to find tail probabilities. For example, with $\mu=6$ and $\sigma=1.25$, $P(X>7)=1-\Phi(0.8)\approx0.2119$, and $P(5<X<8)=\Phi(1.6)-\Phi(-0.8)\approx0.7333$.

✅ Practical tips

• Always check whether interpretations require extrapolation beyond observed data.
• Distinguish between statistical association and causation. Regression describes association unless the study design allows causal inference.
• Use continuity correction when approximating binomial probabilities with the normal when $n$ is moderately large.
• For conditional probabilities, clearly identify the conditioning event and use Bayes' rule when reversing conditional statements.