Data 101 - Midterm 2

P-value

Probability of results this extreme if null hyphotesis is true.

Prosecutor fallacy

Confusing P(Evidence|Innocent) with P(Innocent|Evidence).

Defense fallacy

Claiming evidence is meaningless because false positives exist.

Bayesian posterior formula

P(H|E) = (P(E|H)P(H)) / P(E).

t-test

Unknown population variances, moderate sample size.

Permutation test

When no parametric assumptions can be made.

k-anonymity

Each record indistinguishable from at least k-1 others.

Limitation of k-anonymity

Vulnerable to linkage attacks.

Differential privacy

Adding noise so individuals cannot be identified.

ε in differential privacy

Privacy budget; lower ε = stronger privacy.

Disparate impact

Neutral rule disproportionately harms protected groups.

Disparate treatment

Explicit unequal treatment based on protected attribute.

Redundant encoding

A non-protected feature acts as a proxy (ZIP → race).

Individual fairness

Similar individuals should be treated similarly.

Conjunction fallacy

Judging a detailed scenario as more likely than the general one.

Survivorship bias

Only observing who survives, missing failures.

Ecological fallacy

Inferring individual traits from group averages.

Goodhart's Law

When a measure becomes a target, it stops being a good measure.

Binomial approximate normal

Large n and p not near 0 or 1.

68-95-99.7 rule

For normal: ±1σ (68%), ±2σ (95%), ±3σ (99.7%).

Bonferroni correction

α_new = α / (#tests) controls familywise error.

High accuracy unfairness

Because underlying data may encode bias.

Truncated y-axis problem

Exaggerates small differences.

Permutation distribution

Distribution of statistics across label-shuffled datasets under H₀.

Simpson's paradox

Trend reverses when data is aggregated vs subgrouped.

Standard deviation

Spread around the mean.

Low prevalence false-positive issues

False positives may outnumber true positives.

Practical significance

Whether effect size matters in the real world.

Central Limit Theorem (CLT)

Means of samples → normal as n increases.

Proportional ink

Graphical area should match magnitude.