Day+8+-+July+8

Simulation‐Based Null Distributions and Their Bell Shape

• In the dolphin communication and community recycling studies, researchers simulated data under the null hypothesis to create null (chance) distributions.
  • Dolphin study: simulated counts of correct guesses.
  • Recycling study: simulated proportions of people answering “yes.”
• Despite different statistics (count vs. proportion), both simulated histograms looked bell-shaped.
• This similarity is NOT a coincidence; it foreshadows the Central Limit Theorem (CLT) and the normal model.

From Hands-On Simulations to Theory-Based Models

• Why simulations matter:
  • Build intuition about sampling variability and the role of chance.
  • Provide a visual, concrete sense of the distribution of a statistic.
• Practical challenges of simulations:
  • Complex scenarios (e.g., multiple regression with many predictors) may require huge numbers of resamples → computationally expensive or impossible in practice.
• Solution: theory-based models (especially the normal distribution) allow:
  • Faster probability calculations without resampling.
  • Generalization to a wide range of statistical techniques.
  • A bridge from exploratory simulation to formal inference (confidence intervals, hypothesis tests).

Central Limit Theorem (CLT)

• Fundamental mathematical result explaining why sampling distributions are (approximately) normal.
• Statement (informal):
  • For many statistics (sample mean, sample proportion, etc.), as sample size $n$ increases and certain conditions hold, the sampling distribution of the statistic approaches a normal distribution, regardless of the population’s shape.
• Importance:
  • Makes the normal curve the workhorse for probability, confidence intervals, and hypothesis testing.
  • Removes the need for repeated simulations once the conditions are verified.
• Key conditions (implied, though not fully listed in slides):
  • Independence of observations.
  • Sample size large enough (often phrased as “large $n$” or “success–failure condition” for proportions).

The Normal Distribution: Overview

• Appears naturally in many phenomena and as the CLT limit.
• Five universal features:
  1. Symmetric about the mean.
  2. Unimodal (one peak).
  3. Bell-shaped curve.
  4. Centered at the mean $\mu$.
  5. Total area under the curve equals $1$ (interpretable as probability).
• Notation: $N(\mu, \sigma)$ where $\mu$ = mean, $\sigma$ = standard deviation.

How Mean and Standard Deviation Shape the Curve

• Changing $\mu$ (keeping $\sigma$ fixed):
  • Translates the curve left or right; shape does not change.
• Changing $\sigma$ (keeping $\mu$ fixed):
  • Larger $\sigma$ → wider, flatter curve (greater variability).
  • Smaller $\sigma$ → narrower, sharper curve (less variability).
• Graphical example (slide 10) illustrated three curves: $N(70,5)$, $N(80,5)$ (shift) and $N(70,10)$ (stretch).

Standard (Z) Scores

• Purpose: place observations from different distributions onto a common scale (number of SDs from mean).
• Formula:
  $z = \frac{x - \mu}{\sigma}$
• Interpretation:
  • z > 0 ⇒ observation above mean.
  • z < 0 ⇒ observation below mean.
  • $z = 0$ ⇒ observation equals mean.
• Standardizing a normal variable $X\sim N(\mu,\sigma)$ always yields $Z \sim N(0,1)$ (the standard normal).

Visualizing the Standard Normal

• The horizontal axis is measured in SD units.
• Regions marked at $-3,-2,-1,0,1,2,3$ show typical to extreme values.

Empirical Rule (68–95–99.7 Rule)

• For any nearly normal distribution $N(\mu,\sigma)$:
  • About $68\%$ of observations lie in $[\mu-\sigma,\mu+\sigma]$ (|z|≤1).
  • About $95\%$ lie in $[\mu-2\sigma,\mu+2\sigma]$ (|z|≤2).
  • About $99.7\%$ lie in $[\mu-3\sigma,\mu+3\sigma]$ (|z|≤3).
• Rare events: |z|>3 (~0.3% of data) are highly unusual.
• Graphical depiction shows shaded areas centered at $\mu$, widening with each additional $\sigma$.

Statistical Significance & the 2-SD Benchmark

• A rule of thumb: statistics more than 2 SDs from the mean of the null distribution are “statistically significant.”
• Caveats:
  • Only in a formal hypothesis-testing setting.
  • Requires that the sampling distribution of the statistic is (approximately) normal.
  • Significance level assumed to be $\alpha = 0.05$ (two-tailed).

Worked Examples and Connections

1. Elephant Heights (Slide 15)

• Asian males: $\mu=7.9\text{ ft},\;\sigma=0.4\text{ ft}$.
• African males: $\mu=11.5\text{ ft},\;\sigma=0.5\text{ ft}$.
• Heights: Kavi (Asian) $x=8.4\text{ ft}$, Temba (African) $x=10.8\text{ ft}$.
• Z-scores:
  • Kavi: $z=\frac{8.4-7.9}{0.4}=1.25$.
  • Temba: $z=\frac{10.8-11.5}{0.5}=-1.4$.
• Interpretation: |z| larger for Temba (1.4 vs 1.25) ⇒ Temba’s height is a bit more unusual relative to African males than Kavi’s relative to Asian males.

2. Pelican Wingspans (Slides 22–25)

• Population model: $X\sim N(6.5,0.5)$ (feet).
• (a) Probability wingspan > 7.0 ft:
  • $z=\frac{7.0-6.5}{0.5}=1$.
  • From standard normal table: P(Z>1)\approx0.1587 (≈16%).
• (b) Probability 5.5 ft < wingspan < 7.0 ft:
  • Lower boundary: $z=\frac{5.5-6.5}{0.5}=-2$.
  • Upper boundary: $z=1$ (as above).
  • P(-2<Z<1)=P(Z<1)-P(Z<-2)\approx0.8413-0.0228=0.8185 (≈81.9%).
• Demonstrates conversion from raw $X$ to $Z$, then use of normal table/graph.

3. Dolphin & Recycling Simulations (Slides 2–3)

• Although underlying variables differed (count vs proportion), CLT predicts their sampling distributions approach normal → bell-shaped histograms confirmed visually.

Visual & Graphical Takeaways

• Multiple plots illustrated:
  • Simulated null distributions exhibiting bell curves.
  • Overlays of different $N(\mu,\sigma)$ to show mean shift and variability change.
  • Standard normal curve with SD tick marks, highlighting 68–95–99.7 rule.

Practical, Ethical, and Real-World Connections

• CLT and normal approximations allow practitioners to:
  • Quickly compute $p$-values and confidence intervals in business, health, and environmental studies without exhaustive simulation.
  • Communicate probabilistic results (e.g., “only 2% chance of observing a wingspan that large”).
• Ethical edge: choosing an inappropriate model (normal when conditions fail) may mislead conclusions → underscores importance of checking assumptions.

Key Equations & Numeric Benchmarks

• Z-score: $z=\frac{x-\mu}{\sigma}$.
• Area benchmarks (standard normal):
  • $$P(|Z|>1