Statistics and Maths Revision

0.0(0)
Studied by 0 people
call kaiCall Kai
learnLearn
examPractice Test
spaced repetitionSpaced Repetition
heart puzzleMatch
flashcardsFlashcards
GameKnowt Play
Card Sorting

1/49

flashcard set

Earn XP

Description and Tags

50 Terms

1
New cards

What is a random variable?

A mapping that assigns numerical values to events in a sample space

Sample space = list of possible outcomes (set notation, e.g. {a,b}). A random variable maps those outcomes to numbers, e.g. X = no. of heads in 2 flips: X(H,H)=2, X(H,T)=X(T,H)=1, X(T,T)=0. Outcome is unknown before the event happens.

2
New cards

Discrete vs continuous vs binary random variables?

Discrete = countable values; binary = only 0/1; continuous = infinitely many values in a range.

Discrete: countable values (years of education, pints per week — can be countably infinite {0,1,2,…}). Binary: special discrete case taking 1 or 0 (unemployed yes=1/no=0). Continuous: any value in an interval (monthly income on [0,∞)).

3
New cards

For a DISCRETE random variable, what does the probability distribution function (pdf) give?

Pr(X = xⱼ) = pⱼ for each possible outcome xⱼ

The pdf lists the probability of every possible outcome: Pr(X = xⱼ) = pⱼ, j = 1,…,k (probabilities sum to 1). E.g. fair die: Pr(X = j) = 1/6 for j = 1,…,6.

4
New cards

For a CONTINUOUS random variable, how are the cdf F(x) and pdf f(x) related, and what is Pr(X = x)?

f(x) = dF(x)/dx, F(x) = ∫ f up to x, and Pr(X = x) = 0

cdf: F(x) = Pr(X ≤ x). pdf = derivative of cdf; cdf = integral of pdf. Pr(X = x) = 0 for any single point, so work with ranges: Pr(a < X < b) = F(b) − F(a) = ∫ₐᵇ f(x)dx (area under the pdf).

5
New cards

Normal distribution: what does X ~ N(μ, σ²) mean and what is the standard normal?

X is normal with mean μ and variance σ²; standard normal Z ~ N(0,1)

pdf: f(x) = (1/(σ√(2π))) · exp(−(x−μ)²/(2σ²)), −∞ < x < ∞. σ = SD, σ² = variance. Standard normal N(0,1): f(z) = (1/√(2π)) exp(−z²/2). Notation: Φ = standard normal cdf, φ = standard normal pdf. Standardise with Z = (X − μ)/σ. Widely used because many variables (e.g. ln(income)) are approx. normal.

6
New cards

How is a chi-square (χ²) random variable with n degrees of freedom constructed?

Sum of squares of n independent standard normals: X = Σ Zᵢ²

If Zᵢ ~ N(0,1), i = 1,…,n (independent), then X = Σᵢ₌₁ⁿ Zᵢ² ~ χ²ₙ. Always non-negative.

7
New cards

How is a t random variable with n degrees of freedom constructed?

T = Z / √(X/n), Z ~ N(0,1), X ~ χ²ₙ, Z and X independent

t = standard normal over the square root of an independent χ² divided by its df. As df → ∞ the t-distribution → standard normal (so with large n you can use the normal table).

8
New cards

How is an F random variable with (k, m) degrees of freedom constructed?

F = (X/k)/(Y/m), X ~ χ²ₖ, Y ~ χ²ₘ, independent

Ratio of two independent χ² variables, each divided by its own df. Numerator df = k, denominator df = m. Used for F-tests (Semester 1). Normal, t, F and χ² tables are provided in exams.

9
New cards

Define the joint pdf and the conditional pdf of Y given X (discrete).

Joint: f(x,y) = Pr(X=x, Y=y). Conditional: f(y|x) = f(x,y)/f_X(x)

Joint = probability both happen (person shows up AND travels on business). Conditional = probability of Y given a value of X (analogous to Bayes). The conditional distribution of Y given X contains the most we can know about how X affects Y.

10
New cards

When are X and Y independent, and what does it imply for the conditional distribution?

If f(x,y) = f_X(x)·f_Y(y); then f(y|x) = f_Y(y)

Independence ⇔ joint = product of marginals. Then conditional = marginal: knowing X tells you nothing about Y. Note: Cov = 0 does NOT imply independence (see covariance cards).

11
New cards

Definition of the expected value E(X) for a discrete X?

E(X) = Σ xⱼ f(xⱼ) — probability-weighted average of outcomes

E(X) = x₁f(x₁) + … + xₙf(xₙ). Measure of central tendency; weights are the probabilities.

12
New cards

Key rules for expectations?

E(c)=c; E(aX+b)=aE(X)+b; E(ΣaᵢXᵢ)=ΣaᵢE(Xᵢ); but E(g(X)) ≠ g(E(X)) for nonlinear g

Expectation is linear: constants and sums pass through. Not true for nonlinear functions (Jensen's inequality): e.g. E(X²) ≠ [E(X)]², E(ln X) ≠ ln E(X). E(XY) = E(X)E(Y) only holds under independence.

13
New cards

Definition of Var(X), the shortcut formula, and the SD?

Var(X) = E[(X−μ)²] = E(X²) − μ²; SD = √Var(X)

Variance = expected squared distance from the mean. Derivation: E[(X−μ)²] = E(X²) − 2μE(X) + μ² = E(X²) − μ². Standard deviation σ = √σ². Var(X) = 0 iff X is a constant (P(X=c)=1).

14
New cards

Var(aX + b) = ?

a²Var(X)

Adding a constant shifts but doesn't change spread; scaling by a scales variance by a². Proof: aX+b − (aμ+b) = a(X−μ), so Var = E[a²(X−μ)²] = a²Var(X).

15
New cards

Var(aX + bY) = ?

a²Var(X) + b²Var(Y) + 2abCov(X,Y)

The covariance term disappears only if Cov(X,Y) = 0 (e.g. independence). Also note Var(X − Y) = Var(X) + Var(Y) − 2Cov(X,Y).

16
New cards

Definition of Cov(X,Y), shortcut formula and interpretation of its sign?

E[(X−μₓ)(Y−μᵧ)] = E(XY) − μₓμᵧ; >0: Y tends to be above its mean when X is

Covariance = how X and Y vary together on average. Cov>0: when X above its mean, Y tends to be above its mean; Cov<0: Y tends to be below. Var(X) = Cov(X,X).

17
New cards

Key properties of covariance?

If E(X)=0 or E(Y)=0, Cov=E(XY); independence ⇒ Cov=0; Cov(a₁X+b₁, a₂Y+b₂) = a₁a₂Cov(X,Y)

Additive constants drop out; multiplicative constants multiply. Because of the a₁a₂ scaling, covariance depends on units of measurement → hence correlation.

18
New cards

Independence ⇒ Cov(X,Y) = 0. Is the converse true?

No — zero covariance does not imply independence

Proof: independence ⇒ E(XY) = E(X)E(Y), so Cov = E(XY) − E(X)E(Y) = 0. Converse fails: covariance only measures LINEAR dependence, so Cov = 0 can hide a nonlinear relationship (e.g. Y = X² with X symmetric around 0).

19
New cards

Formula and range of the correlation coefficient ρ(X,Y)?

ρ = Cov(X,Y)/(σₓσᵧ), −1 ≤ ρ ≤ 1, unit-free

ρ = Cov(X,Y)/√(Var(X)Var(Y)) = σₓᵧ/(σₓσᵧ). Doesn't depend on units. If you get |ρ| > 1 you've made an error.

20
New cards

What is a random sample {X₁,…,Xₙ}?

Independent random variables all with the same distribution f(x;θ)

Statistical inference = learning about unknown population parameters (θ) from a sample. Population described by f(x;θ); each Xᵢ is what can be obtained on the i-th draw. Random sample = i.i.d. draws.

21
New cards

Estimator vs estimate?

Estimator = rule/function of the sample, W = h(X₁,…,Xₙ), a random variable; estimate = the number it gives for a particular sample

Because the estimator is a random variable, it has a distribution: the SAMPLING DISTRIBUTION (how estimates vary across samples). True θ is never known. Semester 1 estimation method: Least Squares.

22
New cards

When is an estimator W unbiased for θ?

E(W) = θ

'Right on average': across infinitely many samples, the average estimate equals θ. Key result: sample mean X̄ = ΣXᵢ/n is unbiased for μ regardless of the population distribution (E(X̄) = (1/n)ΣE(Xᵢ) = μ).

23
New cards

Why is unbiasedness alone a weak property?

Silly estimators like 'use only the first observation X₁' are also unbiased

E(X₁) = μ, so X₁ is unbiased, but one draw can be far off (e.g. X uniform on 1–5, mean 3, X₁ could be 1). So we also look at variance → efficiency.

24
New cards

Var(X̄) for a random sample of size n with variance σ²?

σ²/n

Var(ΣXᵢ/n) = (1/n²)ΣVar(Xᵢ) (covariances = 0 by independence) = (1/n²)·nσ² = σ²/n. Standard error of X̄ = σ/√n (estimated by √(S²/n)).

25
New cards

What does it mean for W to be efficient relative to V?

Both unbiased and Var(W) ≤ Var(V) for all θ, strictly < for at least one θ

Compare unbiased estimators by variance — smaller is better. An estimator is efficient (absolutely) if it has the smallest variance of all possible (unbiased) estimators.

26
New cards

Null and the three possible alternative hypotheses?

H₀: θ = θ₀ vs H₁: θ > θ₀ (upper one-sided), θ < θ₀ (lower one-sided), θ ≠ θ₀ (two-sided)

ALWAYS write both H₀ and H₁. Hypotheses are about the population PARAMETER θ, never the estimate θ̂.

27
New cards

Type I vs Type II error?

Type I: reject H₀ when it's true. Type II: fail to reject H₀ when it's false

Type I = false positive (finding an effect that isn't there). Type II = false negative (missing a real effect).

28
New cards

What is the significance level α?

P(reject H₀ | H₀ true) = P(Type I error), chosen before the test

Your tolerance for Type I error, chosen in advance (usually 5% or 1%). At 5%, if H₀ is true and you test 100 samples, expect ~5 false rejections.

29
New cards

What is a test statistic and which distribution do we use for it?

A known function of the sample (so a random variable); we use its distribution under H₀

T = f(sample). Usually built from θ̂ and the estimated variance of θ̂. Always use its distribution assuming H₀ is true — that's where critical values come from.

30
New cards

Standard normal critical values: 5% one-sided, 5% two-sided, 1% one-sided?

1.645, 1.96, 2.33

Critical value c depends on the distribution, α, and the alternative. Two-sided puts α/2 in each tail, so needs a bigger |t| (1.96 vs 1.645) — a stronger test. (1% two-sided ≈ 2.576.) Reading the normal table: rows = 1st decimal, columns = 2nd decimal.

31
New cards

Rejection rules for the three alternatives (test stat t, critical value c)?

Upper: reject if t > c. Lower: reject if t < −c. Two-sided: reject if |t| > c

Rejection region is in the tail(s) the alternative points to. Equivalent: reject if p-value < α.

32
New cards

What is the p-value?

Probability (under H₀) of a test stat at least as extreme as the one observed; the largest α at which you'd fail to reject

Two-sided: p = P(|T| > |t| | H₀). Reject H₀ if p < α. Small p = strong evidence against H₀. p is never exactly 0 (report e.g. p < 0.0001). Software usually reports two-sided p-values.

33
New cards

95% confidence interval for θ (normal sampling distribution) and its interpretation?

[θ̂ − 1.96·se, θ̂ + 1.96·se]; in 95% of repeated samples the interval contains the true θ

se = √S², where S² is the estimated variance of the estimator. Uses 1.96 because a CI is two-sided. The interval is random (varies by sample); θ is fixed.

34
New cards

Petrol tax example: n = 1256, Ȳ = 12.8, S² = 553.7, H₀: μ = 0 vs H₁: μ > 0. t-stat and conclusion at 5%?

t = 12.8/√(553.7/1256) ≈ 19.28 > 1.645 → reject H₀

t = (θ̂ − θ₀)/√(S²/n). Large n so use standard normal. Also rejects at 1% (c ≈ 2.33); p < 0.0001. Conclusion: mileage significantly decreased. Caveat: assumes nothing else (e.g. cheaper trains) changed — confounders, which econometrics controls for.

35
New cards

Same test but n = 100: which table, which df, and roughly what critical value (5% one-sided)?

t-table, df = n − 1 = 99, c ≈ 1.660

t-table: column = significance level (one/two-sided), row = df. df = sample size − no. of restrictions/parameters estimated = n − 1 here. When df = ∞, t critical value = normal value (1.645). Know how to read t, F and χ² tables.

36
New cards

The 5-step hypothesis testing 'cookbook'?

1 State H₀/H₁; 2 test stat + its null distribution; 3 compute t; 4 choose α + decision rule; 5 compare to critical value (or p-value) and conclude in words

E.g. H₀: θ = 5, H₁: θ ≠ 5; t = (θ̂ − θ₀)/se ~ t₁₀₀ under H₀; reject if |t| > c₀.₀₅. Final step: state conclusion in words, e.g. 'Reject H₀ at 5%: smoking has a significant effect on birthweight.'

37
New cards

What does log_b(x) = y mean, and what is ln?

bʸ = x; ln = log base e (e ≈ 2.718)

Log is the inverse of the exponential. Econometrics uses natural logs. If ln(c) = d then c = eᵈ, so ln(1) = 0 (e⁰ = 1).

38
New cards

The four log rules?

ln(cd)=ln c+ln d; ln(c/d)=ln c−ln d; ln(cᵈ)=d·ln c; ln(c^(1/d))=ln(c)/d

Last rule = d-th root: ln(ᵈ√c) = ln(c)/d. Example: solve (1/2)^(2x+1) = 1 → take logs: (2x+1)ln(1/2) = ln 1 = 0 → 2x+1 = 0 → x = −1/2. Note ln(c+d) ≠ ln c + ln d.

39
New cards

What does ∂y/∂x tell you for y = f(x,z), and what is it for y = ax + b vs y = ax²?

Marginal effect of x on y holding z fixed; a (constant) vs 2ax (varies with x)

'If x changes by a tiny amount, how does y change?' (e.g. effect of smoking on birthweight). Linear → constant marginal effect; quadratic → marginal effect depends on the level of x.

40
New cards

Summation rules: Σcxᵢ and Σxᵢyᵢ?

Σcxᵢ = cΣxᵢ; but Σxᵢyᵢ ≠ (Σxᵢ)(Σyᵢ)

Σᵢ₌₁ⁿ xᵢ = x₁ + … + xₙ (limits at bottom/top; Σᵢ₌₂⁴ xᵢ = x₂+x₃+x₄). Constants come out; products of variables can't be split. Also Σᵢ₌₁ⁿ c = nc.

41
New cards

How do you evaluate a double sum Σⱼ₌₁ᵐ Σᵢ₌₁ⁿ xᵢyⱼ?

Fix j, sum over all i, then move to next j: = y₁(x₁+…+xₙ) + … + yₘ(x₁+…+xₙ) = (Σxᵢ)(Σyⱼ)

Work from the inside out. Because yⱼ doesn't depend on i, it factors out of the inner sum.

42
New cards

Link between the mean and the sum?

x̄ = (1/n)Σxᵢ, so Σxᵢ = n·x̄

Used constantly to simplify sums. Consequence: Σ(xᵢ − x̄) = nx̄ − nx̄ = 0 (deviations from the mean sum to zero).

43
New cards

Σ(xᵢ − x̄)² simplifies to?

Σxᵢ² − n x̄²

Expand: Σxᵢ² − 2x̄Σxᵢ + Σx̄² = Σxᵢ² − 2x̄(nx̄) + nx̄² = Σxᵢ² − nx̄².

44
New cards

Σ(xᵢ − x̄)(yᵢ − ȳ) simplifies to?

Σxᵢyᵢ − n x̄ȳ

Expand: Σxᵢyᵢ − ȳΣxᵢ − x̄Σyᵢ + nx̄ȳ = Σxᵢyᵢ − nx̄ȳ − nx̄ȳ + nx̄ȳ = Σxᵢyᵢ − nx̄ȳ. Also = Σ(xᵢ − x̄)yᵢ. (This is the numerator of the OLS slope.)

45
New cards

An m × n matrix: what are m and n, and what is aᵢⱼ?

m rows, n columns; aᵢⱼ = element in row i, column j

Matrices: bold upper case (A), elements lower case. E.g. A = [2 −1 7; 0 5 8; −4 6 0]: a₁₃ = 7, a₃₂ = 6.

46
New cards

Square, diagonal, identity and zero matrices?

Square: rows = columns. Diagonal: square, zeros off the diagonal. Identity Iₙ: diagonal with 1s. Zero: all elements 0

Diagonal and identity matrices must be square (otherwise no full diagonal). Iₙ is the matrix equivalent of 1: AI = A.

47
New cards

What is a vector?

A matrix with one row (1 × n row vector) or one column (n × 1 column vector)

Transposing a row vector gives the corresponding column vector and vice versa.

48
New cards

Rules for matrix addition and scalar multiplication?

Addition: same dimensions only, add element by element. Scalar cA: multiply every element by c

(A+B)ᵢⱼ = aᵢⱼ + bᵢⱼ. (cA)ᵢⱼ = c·aᵢⱼ.

49
New cards

When can you compute AB, and what is element (AB)ᵢⱼ?

No. of columns of A = no. of rows of B; (AB)ᵢⱼ = Σₖ aᵢₖbₖⱼ (row i of A · column j of B)

(m × n)(n × p) → m × p. Order matters: AB ≠ BA in general (BA may not even exist). Matrix × vector works if the inner dimensions match.

50
New cards

A = [−4 0; 1 3] (2×2), B = [−4 5 0; 1 2 3] (2×3). AB = ?

[16 −20 0; −1 11 9]

2×2 times 2×3 → 2×3. Row 1: (−4)(−4)+0·1 = 16; (−4)(5)+0 = −20; 0. Row 2: 1(−4)+3(1) = −1; 1(5)+3(2) = 11; 0+3(3) = 9. (BA is not defined: 2×3 times 2×2.)