In a population of men in a city, 3% are observed to suffer form cancer, 15% are found to be smokers, while 10% are noticed to be either smokers or cancer patients. If a man is selected at random, what is the probability that he is a smoker and suffering from cancer ?
Economics · Unit 3
Statistics and Econometrics
Revision notes (27 min read), then 56 real questions from UGC NET papers.
Official syllabus · Unit 3
- Probability theory: concepts of probability, distributions, moments, central limit theorem
- Descriptive statistics: measures of central tendency and dispersion, correlation, index numbers
- Sampling methods and sampling distribution
- Statistical inference, hypothesis testing
- Linear regression models and their properties: BLUE
- Identification problem
- Simultaneous equation models: recursive and non-recursive
- Discrete choice models
- Time series analysis
Verbatim from the UGC NET Bureau syllabus.
How often it is asked
| Paper | Questions |
|---|---|
| December 20257 Jan 2026 | 11 / 100 |
| June 202529 Jun 2025 | 18 / 100 |
| June 2024 (re-exam)28 Aug 2024 | 10 / 100 |
| December 202312 Dec 2023 | 17 / 100 |
Average 14.0 questions (28 marks) per paper. Unit tagging is ours.
Scope: probability, descriptive statistics, sampling, inference, CLRM/OLS and its violations, identification and simultaneous equations, discrete choice, time series — exam-level formulas and matching facts.
1. Probability Theory: Concepts of probability, Distributions, Moments, Central Limit Theorem
Approaches to probability: Classical/a priori (Laplace: favourable/total, equally likely outcomes); Relative frequency/empirical (von Mises); Subjective (Savage, de Finetti); Axiomatic (Kolmogorov, 1933 — P(A) ≥ 0, P(S) = 1, countable additivity).
Rules: P(A∪B) = P(A) + P(B) − P(A∩B). Conditional P(A|B) = P(A∩B)/P(B). Independence: P(A∩B) = P(A)P(B). Mutually exclusive events with non-zero probability are never independent. Bayes' theorem (Thomas Bayes, 1763): P(Aᵢ|B) = P(B|Aᵢ)P(Aᵢ) / Σ P(B|Aⱼ)P(Aⱼ).
Random variables: E(aX + b) = aE(X) + b; Var(aX + b) = a²Var(X); Var(X ± Y) = Var X + Var Y ± 2Cov(X,Y). Cov = 0 does not imply independence (except for jointly normal variables).
| Distribution | Parameters | Mean | Variance | Notes |
|---|---|---|---|---|
| Bernoulli | p | p | pq | — |
| Binomial | n, p | np | npq | variance < mean; symmetric when p = 0.5 |
| Poisson (Poisson, 1837) | λ | λ | λ | mean = variance; limit of binomial as n→∞, p→0, np = λ; positively skewed, β₁ = 1/λ |
| Geometric (trials to 1st success) | p | 1/p | q/p² | memoryless (discrete) |
| Uniform (a, b) | — | (a+b)/2 | (b−a)²/12 | — |
| Exponential | λ | 1/λ | 1/λ² | memoryless (continuous) |
| Normal (Gauss; de Moivre 1733) | μ, σ² | μ | σ² | mean = median = mode; β₁ = 0, β₂ = 3; points of inflexion at μ ± σ |
| Standard normal Z | — | 0 | 1 | — |
| Student's t (Gosset "Student", 1908) | k df | 0 (k>1) | k/(k−2) (k>2) | fatter tails than Z; → Z as k→∞ |
| Chi-square χ² (Karl Pearson, 1900) | k df | k | 2k | sum of k squared independent Z; positively skewed |
| F (Fisher; named by Snedecor) | k₁, k₂ | k₂/(k₂−2) (k₂>2) | 2k₂²(k₁+k₂−2)/[k₁(k₂−2)²(k₂−4)] (k₂>4) | ratio of two independent χ² each divided by its df |
Relationships: Z² ~ χ²(1); sum of independent χ² is χ² (additive property); t = Z/√(χ²ₖ/k); t²ₖ = F(1, k); F = (χ²₁/k₁)/(χ²₂/k₂); 1/F(k₁,k₂) = F(k₂,k₁); k₁F(k₁,k₂) → χ²(k₁) as k₂→∞. Binomial → Normal for large n (de Moivre–Laplace). Poisson → Normal for large λ.
Normal areas: μ ± 1σ ≈ 68.27%, ± 2σ ≈ 95.45%, ± 3σ ≈ 99.73%; ±1.96σ = 95%, ±2.58σ = 99%.
Moments: Raw moment μ′ᵣ = E(Xʳ); central moment μᵣ = E(X−μ)ʳ. μ₁ = 0, μ₂ = variance. μ₂ = μ′₂ − μ′₁²; μ₃ = μ′₃ − 3μ′₂μ′₁ + 2μ′₁³. Central moments do not change when the origin changes but do change with scale. Moment generating function M(t) = E(eᵗˣ); the r-th derivative at t = 0 gives μ′ᵣ.
Pearson's β and γ coefficients: β₁ = μ₃²/μ₂³ (skewness), γ₁ = √β₁; β₂ = μ₄/μ₂² (kurtosis), γ₂ = β₂ − 3. β₂ = 3 mesokurtic (normal), β₂ > 3 leptokurtic (peaked, fat tails), β₂ < 3 platykurtic.
Laws and limit theorems:
- Chebyshev's inequality: P(|X − μ| ≥ kσ) ≤ 1/k², for any distribution.
- Law of large numbers (Bernoulli; weak law by Khinchin): the sample mean converges in probability to μ.
- Central Limit Theorem (Lindeberg–Lévy): for i.i.d. draws with finite σ², √n(X̄ − μ)/σ → N(0,1) whatever the parent distribution. Rule of thumb: n ≥ 30. Liapunov's version drops "identically distributed". The CLT is why the OLS error term (the sum of many small influences) is assumed normal.
Commonly tested
- Mean = variance → Poisson. Mean > variance → Binomial.
- t²(k) = F(1,k); χ² mean k, variance 2k; t variance k/(k−2).
- Normal β₂ = 3; leptokurtic β₂ > 3.
- Axiomatic probability → Kolmogorov; t-distribution → W. S. Gosset; χ² test → Karl Pearson; F → R. A. Fisher.
- The CLT needs finite variance; the population need not be normal.
2. Descriptive Statistics — Measures of Central Tendency & Dispersion, Correlation, Index Numbers
Central tendency
- AM = Σx/n; GM = (Πx)^(1/n); HM = n/Σ(1/x). AM ≥ GM ≥ HM, with equality only if all values are equal. For two positive numbers, GM² = AM × HM.
- Σ(x − x̄) = 0; Σ(x − A)² is smallest at A = x̄; Σ|x − A| is smallest at A = median.
- Empirical relation (Karl Pearson) for moderately skewed data: Mode ≈ 3 Median − 2 Mean.
- Median and mode are positional averages (used for open-ended classes; the median is used for qualitative data such as intelligence). GM is used for ratios and growth rates and in index numbers; HM is used for speeds and rates.
- Grouped median = L + [(N/2 − cf)/f]·h; mode = L + [(f₁ − f₀)/(2f₁ − f₀ − f₂)]·h. The mode is found graphically from a histogram, the median from ogives (where the less-than and more-than ogives intersect).
Dispersion
- Absolute measures: range, quartile deviation QD = (Q₃ − Q₁)/2, mean deviation, SD σ = √[Σ(x − x̄)²/n].
- Relative measures: coefficient of range, coefficient of QD = (Q₃ − Q₁)/(Q₃ + Q₁), coefficient of variation CV = σ/x̄ × 100 (Karl Pearson; the series with the lower CV is more consistent).
- SD is independent of origin but not of scale: σ(aX + b) = |a|σ.
- For a normal distribution, QD ≈ (2/3)σ and MD ≈ (4/5)σ. In general, SD ≥ MD.
- Combined variance: σ² = [n₁(σ₁² + d₁²) + n₂(σ₂² + d₂²)]/(n₁ + n₂), where dᵢ = x̄ᵢ − x̄.
- Lorenz curve and Gini coefficient measure inequality (Gini = area between the curve and the equality line ÷ the whole triangle; 0 ≤ G ≤ 1).
Skewness and kurtosis
| Measure | Formula | Range |
|---|---|---|
| Karl Pearson (1st) | Sk = (Mean − Mode)/σ | unbounded in theory; usually within ±1 |
| Karl Pearson (2nd) | Sk = 3(Mean − Median)/σ | −3 to +3 |
| Bowley (quartile) | Sk = (Q₃ + Q₁ − 2Md)/(Q₃ − Q₁) | −1 to +1 |
| Kelly (percentile) | (P₉₀ + P₁₀ − 2P₅₀)/(P₉₀ − P₁₀) | −1 to +1 |
| Moment | γ₁ = √β₁ = μ₃/σ³ | — |
Positive skew: Mean > Median > Mode (long right tail). Negative skew: Mean < Median < Mode. Bowley's measure suits open-ended distributions.
Correlation
- Karl Pearson: r = Cov(X,Y)/(σₓσᵧ) = Σxy/√(Σx²Σy²) (deviations from the mean), with −1 ≤ r ≤ 1. r is independent of change of origin and scale (only the sign can flip if one scale factor is negative). r measures linear association only; r = 0 does not imply independence.
- r = ±√(b_yx · b_xy): r is the geometric mean of the two regression coefficients and has their common sign. Both coefficients cannot exceed 1 in absolute value at the same time. The regression lines intersect at (x̄, ȳ); they coincide if r = ±1 and are perpendicular if r = 0.
- Coefficient of determination r² = explained variation ÷ total variation. Probable error PE = 0.6745(1 − r²)/√n; r is significant if r > 6 PE.
- Spearman's rank correlation (1904): ρ = 1 − 6Σd²/[n(n² − 1)]. With ties, add m(m² − 1)/12 to Σd² for each tie group. Example: n = 5, Σd² = 4 gives ρ = 1 − 24/120 = 0.8.
- Partial correlation: r₁₂.₃ = (r₁₂ − r₁₃r₂₃)/√[(1 − r₁₃²)(1 − r₂₃²)]. Multiple correlation R₁.₂₃ lies between 0 and 1.
- Spurious or nonsense correlation (Yule, 1926).
Index numbers
Subscript 0 = base period, 1 = current period.
| Index | Formula | Weights | Bias |
|---|---|---|---|
| Laspeyres (1871) | Σp₁q₀/Σp₀q₀ × 100 | base-year quantities | upward |
| Paasche (1874) | Σp₁q₁/Σp₀q₁ × 100 | current-year quantities | downward |
| Dorbish–Bowley | (L + P)/2 | AM of L and P | — |
| Marshall–Edgeworth | Σp₁(q₀ + q₁)/Σp₀(q₀ + q₁) × 100 | sum of q₀ and q₁ | — |
| Fisher's ideal (1922) | √(L × P) | GM of L and P | free of both biases |
| Kelly (fixed weight) | Σp₁q/Σp₀q | fixed q | — |
| Walsh | Σp₁√(q₀q₁)/Σp₀√(q₀q₁) | GM of the quantities | — |
Tests of adequacy (Irving Fisher):
- Time reversal test: P₀₁ × P₁₀ = 1. Satisfied by Fisher, Marshall–Edgeworth, Walsh, Kelly, the simple GM of relatives and the simple aggregative index. Failed by Laspeyres, Paasche and Dorbish–Bowley (for D–B the product exceeds 1 unless L = P).
- Factor reversal test: P₀₁ × Q₀₁ = Σp₁q₁/Σp₀q₀ (the value ratio). Satisfied only by Fisher's index among the standard formulas. This is why Fisher's index is called "ideal".
- Circular test (an extension of time reversal: P₀₁ × P₁₂ × P₂₀ = 1). Satisfied by the simple GM of price relatives and Kelly's fixed-weight aggregative index. Fisher's index fails it.
- Unit test: satisfied by all indices except the simple aggregative index.
Worked example (good A: p₀ = 2, q₀ = 10, p₁ = 4, q₁ = 8; good B: p₀ = 5, q₀ = 4, p₁ = 6, q₁ = 5): Σp₀q₀ = 40, Σp₁q₀ = 64, Σp₁q₁ = 62, Σp₀q₁ = 41.
- L = 160.0, P = 151.2, Fisher = √(160 × 151.2) = 155.5, M–E = 126/81 = 155.6, D–B = 155.6.
- Check: Fisher P × Fisher Q = 1.55 = 62/40, so the factor reversal test holds. L > P here, as usual.
Other points: base shifting; splicing; deflating (real value = money value ÷ price index × 100); chain base index (link relatives). India's CPI (base 2012) uses a modified Laspeyres formula; the WPI (base 2011-12) is a Laspeyres-type index.
Commonly tested
- AM ≥ GM ≥ HM; GM² = AM·HM for two numbers.
- Bowley skewness ∈ [−1, 1]; Pearson's 2nd measure ∈ [−3, 3].
- Fisher's index satisfies the time reversal and factor reversal tests but not the circular test. Marshall–Edgeworth satisfies time reversal only. Laspeyres and Paasche satisfy neither.
- Laspeyres has an upward bias and Paasche a downward bias.
- r is the GM of b_yx and b_xy; r is unaffected by change of origin or scale.
3. Sampling Methods & Sampling Distribution
- Probability sampling: simple random sampling (SRSWR/SRSWOR; lottery method, random number tables — Tippett); stratified (heterogeneous population → homogeneous strata; proportional allocation, or Neyman's optimum allocation nₕ ∝ NₕSₕ); systematic (every k-th unit, k = N/n); cluster (groups selected whole); multistage (used by NSSO); PPS (probability proportional to size).
- Non-probability sampling: convenience, judgment/purposive, quota, snowball.
- Errors: sampling error falls as n rises. Non-sampling errors (response, coverage, processing) are found in both a census and a sample, and are often larger in a census.
- Sampling distribution of the mean: E(X̄) = μ; standard error σ/√n (with replacement). Without replacement, multiply by the finite population correction √[(N − n)/(N − 1)]; this can be ignored when n/N < 0.05.
- SE of a proportion = √(PQ/n). SE of a difference of means = √(σ₁²/n₁ + σ₂²/n₂).
- If the population is normal, X̄ is exactly normal. Otherwise X̄ is approximately normal for large n (CLT).
- (n − 1)s²/σ² ~ χ²(n − 1) for a normal population; s² = Σ(x − x̄)²/(n − 1) is unbiased for σ².
Commonly tested: The SE halves when n is quadrupled. Neyman allocation → strata with larger size and larger SD get more units. The NSS uses stratified multistage sampling. Quota sampling is non-random.
4. Statistical Inferences, Hypothesis Testing
Properties of estimators (R. A. Fisher, 1922):
- Unbiasedness: E(θ̂) = θ.
- Consistency: plim θ̂ = θ; a sufficient condition is that bias → 0 and variance → 0.
- Efficiency: minimum variance among unbiased estimators; the Cramér–Rao lower bound = 1/(information).
- Sufficiency: uses all the information in the sample. Fisher–Neyman factorisation; Rao–Blackwell theorem.
- MSE = Variance + Bias².
Methods of estimation: method of moments (Karl Pearson); maximum likelihood (Fisher). MLEs are consistent, asymptotically normal and asymptotically efficient, and they are invariant. The MLE of σ² in regression, RSS/n, is biased.
Interval estimate: x̄ ± z_(α/2)·σ/√n. Use t(n − 1) with s when σ is unknown and n is small.
Hypothesis testing (Neyman–Pearson framework, 1933; Neyman–Pearson lemma → the likelihood ratio gives the most powerful test):
| H₀ true | H₀ false | |
|---|---|---|
| Reject H₀ | Type I error (α) = level of significance, "producer's risk" | correct (power 1 − β) |
| Accept H₀ | correct (1 − α) | Type II error (β), "consumer's risk" |
- Power = 1 − β = P(reject H₀ | H₀ false). For a fixed n, lowering α raises β. Increasing n reduces both.
- The p-value is the smallest α at which H₀ is rejected.
- One-tailed 5% critical value z = 1.645; two-tailed z = 1.96.
Tests: z (large sample, σ known); t (small sample — mean, difference of means, paired t, significance of r: t = r√(n − 2)/√(1 − r²) with n − 2 df); χ² (goodness of fit and independence, df = (r − 1)(c − 1); variance test); F (equality of two variances, ANOVA — Fisher). Non-parametric tests: sign, Wilcoxon, Mann–Whitney U, Kruskal–Wallis, runs test, Kolmogorov–Smirnov.
Commonly tested: α = P(Type I error); power = 1 − β. ANOVA → F test. χ² test of independence df = (r − 1)(c − 1). MLE → Fisher; Neyman–Pearson lemma → most powerful test. An unbiased estimator need not be consistent, and a consistent estimator can be biased.
5. Linear Regression Models and their Properties — BLUE
Two-variable model: Yᵢ = β₁ + β₂Xᵢ + uᵢ. Lowercase letters are deviations from the mean.
- β̂₂ = Σxᵢyᵢ/Σxᵢ²; β̂₁ = Ȳ − β̂₂X̄.
- Var(β̂₂) = σ²/Σxᵢ²; Var(β̂₁) = σ²ΣXᵢ²/(nΣxᵢ²).
- σ̂² = Σûᵢ²/(n − k), with n − 2 df in the two-variable case.
- Matrix form: β̂ = (X′X)⁻¹X′Y; Var(β̂) = σ²(X′X)⁻¹.
- Properties of OLS: the fitted line passes through (X̄, Ȳ); Σûᵢ = 0; ΣûᵢXᵢ = 0; ΣûᵢŶᵢ = 0; the mean of Ŷ equals Ȳ.
CLRM assumptions (Gujarati & Porter):
- Linear in parameters.
- X fixed in repeated samples (or independent of u).
- E(uᵢ|X) = 0.
- Homoscedasticity: Var(uᵢ) = σ².
- No autocorrelation: Cov(uᵢ, uⱼ) = 0.
- n > number of parameters.
- Variation in X.
- No exact multicollinearity.
- Correct specification.
The CNLRM adds uᵢ ~ N(0, σ²), which is needed for the t, F and χ² tests in small samples.
Gauss–Markov theorem: under the CLRM assumptions (normality is not needed), OLS is BLUE — the Best (minimum variance) Linear Unbiased Estimator. Under normality, OLS = MLE for β, and OLS is BUE (best among all unbiased estimators, linear or not).
Goodness of fit:
- TSS = ESS + RSS. R² = ESS/TSS = 1 − RSS/TSS, with 0 ≤ R² ≤ 1. R² never falls when a regressor is added.
- Adjusted R²: R̄² = 1 − (1 − R²)(n − 1)/(n − k) (Theil). R̄² ≤ R², and R̄² can be negative. R̄² rises when a variable is added only if that variable's |t| > 1.
- Comparing R² across models is valid only if the dependent variable and n are the same.
- Information criteria for model selection: AIC (Akaike), SIC/BIC (Schwarz). Lower is better; SIC penalises extra parameters more heavily.
Tests:
- t = β̂ⱼ/se(β̂ⱼ), with n − k df.
- Overall significance: F = [R²/(k − 1)] / [(1 − R²)/(n − k)] with (k − 1, n − k) df.
- Restricted vs unrestricted model: F = [(RSS_R − RSS_UR)/m] / [RSS_UR/(n − k)].
- Chow test (1960) for structural stability.
- Ramsey RESET for specification error.
- Jarque–Bera normality test: JB = n[S²/6 + (K − 3)²/24] ~ χ²(2).
Worked example: X = 1, 2, 3, 4, 5; Y = 2, 4, 5, 4, 5.
- Σxy = 6, Σx² = 10, so β̂₂ = 0.6 and β̂₁ = 4 − 0.6 × 3 = 2.2.
- TSS = 6, ESS = 3.6, so R² = 0.6.
- RSS = 2.4, so σ̂² = 2.4/3 = 0.8.
- se(β̂₂) = √0.08 = 0.283, giving t = 2.12 with 3 df.
Functional forms:
- Log-log: the slope is the elasticity.
- Log-lin (ln Y on X): 100 × β₂ is the % growth rate (semi-elasticity).
- Lin-log: β₂/100 is the absolute change in Y for a 1% change in X.
- Reciprocal model: used for the Phillips curve and Engel curves.
Violations of the assumptions
| Problem | Consequence for OLS | Detection | Remedy |
|---|---|---|---|
| Multicollinearity (Frisch, 1934) | Still BLUE; large variances and SEs; wide confidence intervals; high R² but few significant t's; estimates sensitive to the data. With perfect multicollinearity, the coefficients cannot be determined. | High pairwise r (> 0.8); VIF = 1/(1 − R²ⱼ), where VIF > 10 is serious (tolerance = 1/VIF); condition index > 30 (severe); Klein's rule; Farrar–Glauber test | Drop a variable, add data, transform (first differences, ratios), principal components, ridge regression |
| Heteroscedasticity (common in cross-section data) | Unbiased and consistent but not efficient; the usual SEs are biased, so t and F tests are invalid | Graphs of residuals; Park (1966: ln û² on ln X); Glejser (1969: |û| on various functions of X); Goldfeld–Quandt (1965: order by X, omit c central observations, F = RSS₂/RSS₁); Breusch–Pagan–Godfrey (1979); White general test (1980: û² on the X's, their squares and cross-products; nR² ~ χ²); Spearman rank test | WLS/GLS; log transformation; White's heteroscedasticity-consistent (robust) SEs |
| Autocorrelation (common in time series) | Unbiased and consistent but not efficient; with positive autocorrelation, σ² and the SEs are underestimated, so t and R² are inflated | Graphs; runs (Geary) test; Durbin–Watson d; Durbin h; Breusch–Godfrey LM | GLS; Cochrane–Orcutt (iterative); first differences; Newey–West (HAC) SEs |
Durbin–Watson (1950/51): d = Σ(ûₜ − ûₜ₋₁)²/Σûₜ², with 0 ≤ d ≤ 4 and d ≈ 2(1 − ρ̂).
- ρ̂ = 0 → d = 2. ρ̂ = +1 → d = 0. ρ̂ = −1 → d = 4.
- Decision zones:
- 0 to d_L: reject H₀ (positive autocorrelation).
- d_L to d_U: inconclusive.
- d_U to 4 − d_U: no autocorrelation.
- 4 − d_U to 4 − d_L: inconclusive.
- 4 − d_L to 4: negative autocorrelation.
- Assumptions of the test: the model has an intercept; X is non-stochastic; errors follow AR(1) only; there is no lagged dependent variable among the regressors; there are no missing observations.
Durbin h (1970), for models with a lagged Y among the regressors: h = ρ̂ √[n / (1 − n·Var(α̂))], where α̂ is the coefficient of Yₜ₋₁. h is asymptotically N(0,1). It cannot be computed if n·Var(α̂) ≥ 1.
Breusch–Godfrey (1978) LM test: allows AR(p) errors, lagged Y and other stochastic regressors. (n − p)R² ~ χ²(p).
Other specification issues:
- Omitting a relevant variable → bias and inconsistency.
- Including an irrelevant variable → still unbiased, but inefficient.
- Errors in X → bias and inconsistency (attenuation bias); errors in Y alone → OLS stays unbiased.
Dummy variables:
- A qualitative variable with m categories needs m − 1 dummies when the model has an intercept. Using all m dummies together with the intercept gives perfect multicollinearity — the dummy variable trap. (Alternatively, use m dummies and drop the intercept.)
- An intercept dummy shifts the line; a slope (interaction) dummy changes the slope.
- The dummy-variable approach to structural change is an alternative to the Chow test.
- Seasonal adjustment with quarterly data uses 3 dummies.
- For a dummy D in a semi-log model, the % effect is 100(e^β − 1).
Commonly tested
- Gauss–Markov → BLUE without normality.
- VIF > 10; condition index > 30.
- Goldfeld–Quandt (1965), Park (1966), Glejser (1969), Breusch–Pagan (1979), White (1980) → heteroscedasticity.
- DW, h, BG, runs test → autocorrelation.
- Multicollinearity leaves OLS BLUE; heteroscedasticity and autocorrelation leave it unbiased but inefficient.
- d ≈ 2(1 − ρ); the DW test is invalid with a lagged Y → use Durbin h.
- Adjusted R² can be negative.
- Overall F = [R²/(k − 1)]/[(1 − R²)/(n − k)].
6. Identification Problem
- Structural vs reduced form: reduced-form coefficients are functions of the structural coefficients. Identification asks whether the structural parameters can be recovered from the reduced form.
- Categories: under-identified (cannot be estimated); exactly (just) identified (a unique solution); over-identified (more than one solution).
- Working (1927) — "What do statistical demand curves show?" Identification theory: Koopmans (1949); the Cowles Commission.
Order condition (necessary, not sufficient) — notation from Gujarati & Porter:
- M = endogenous variables in the model; m = endogenous variables in the equation.
- K = predetermined variables in the model; k = predetermined variables in the equation.
- Condition: K − k ≥ m − 1. Equality → just identified; > → over-identified.
- Equivalent form: the number of variables excluded from the equation ≥ M − 1.
Rank condition (necessary and sufficient): in a model of M equations, an equation is identified if and only if at least one non-zero determinant of order (M − 1) × (M − 1) can be formed from the coefficients of the variables excluded from that equation but included in the other equations.
General rule:
- Rank condition holds and K − k > m − 1 → over-identified.
- Rank condition holds and K − k = m − 1 → exactly identified.
- Rank condition fails → under-identified.
Example: a demand–supply system with no exogenous variables has both equations under-identified. Adding income to the demand equation identifies the supply equation, because a shift in demand traces out the supply curve.
Commonly tested: The order condition is necessary; the rank condition is necessary and sufficient. An excluded exogenous variable in one equation identifies the other equation.
7. Simultaneous Equation Models — Recursive and Non-recursive
- Simultaneity bias: an endogenous regressor is correlated with u, so OLS is biased and inconsistent (Haavelmo, 1943).
- Hausman specification test (1978) checks for simultaneity or endogeneity.
Recursive (triangular / causal-chain) models (Wold):
- Y₁ depends only on exogenous variables; Y₂ depends on Y₁ and exogenous variables; Y₃ on Y₁, Y₂ and exogenous variables; and so on.
- The coefficient matrix of the endogenous variables is triangular, and the errors are uncorrelated across equations (diagonal covariance matrix).
- Under these conditions OLS can be applied equation by equation and is consistent, so there is no identification problem.
Non-recursive (interdependent) models need system methods.
Estimation methods:
| Method | Applies to | Notes |
|---|---|---|
| ILS (indirect least squares) | exactly identified equations | OLS on the reduced form, then solve for the structural coefficients; consistent |
| 2SLS (Theil 1953; Basmann 1957) | exactly or over-identified | Stage 1: regress the endogenous regressor on all exogenous variables. Stage 2: replace it with its fitted value. Single-equation, limited information. Same as ILS when the equation is exactly identified. A special case of IV. |
| LIML (Anderson–Rubin, 1949) | single-equation | limited-information maximum likelihood |
| 3SLS (Zellner & Theil, 1962) | the whole system | full information; 2SLS + SUR (GLS across equations); more efficient than 2SLS |
| FIML | the whole system | full-information maximum likelihood |
| IV | any | the instrument must be correlated with the regressor and uncorrelated with u |
Commonly tested: ILS → exactly identified; 2SLS → over-identified; 3SLS → Zellner–Theil (full information). Recursive model → OLS is valid (Wold). Simultaneity → OLS is inconsistent. The Klein model is a classic simultaneous-equation system.
8. Discrete Choice Models
Linear probability model (LPM): Yᵢ = β₁ + β₂Xᵢ + uᵢ with Y ∈ {0, 1}, so E(Y|X) = P(Y = 1). Problems:
- The errors are non-normal (binary).
- The errors are heteroscedastic: Var(u) = Pᵢ(1 − Pᵢ).
- Predicted probabilities can fall outside [0, 1].
- R² is questionable.
- The marginal effect is constant.
Logit (Berkson, 1944):
- Pᵢ = 1/(1 + e^(−Zᵢ)), where Zᵢ = β₁ + β₂Xᵢ. Uses the logistic CDF.
- Log of the odds: Lᵢ = ln[Pᵢ/(1 − Pᵢ)] = β₁ + β₂Xᵢ. The log odds are linear in X, while P is non-linear in X.
- β₂ is the change in the log odds per unit of X. The marginal effect on P is β₂P(1 − P).
- e^β gives the odds ratio.
Probit / normit (Bliss, 1934; Finney):
- P = Φ(Zᵢ), using the standard normal CDF.
- The marginal effect is β·φ(Z).
Comparing logit and probit:
- Both are estimated by maximum likelihood.
- The logistic distribution has fatter tails; otherwise results are similar.
- Logit coefficient ≈ 1.81 × probit coefficient (= π/√3; Gujarati & Porter). Amemiya suggests ≈ 1.6.
- Goodness of fit: McFadden's pseudo-R², count R². Joint significance: the likelihood ratio statistic ~ χ².
Related models: Tobit (Tobin, 1958) for censored dependent variables; multinomial and ordered logit/probit for several categories; Heckman (1979) two-step for sample selection.
Commonly tested: LPM → probabilities outside 0–1 and heteroscedasticity. Logit = log of the odds ratio. Probit = normal CDF; logit = logistic CDF. Both are estimated by ML. Tobit → censored data (James Tobin).
9. Time Series Analysis
Components (classical): trend (T), seasonal (S), cyclical (C), irregular (I). Multiplicative model Y = T × S × C × I; additive model Y = T + S + C + I. Trend is measured by moving averages or least squares; seasonality by the ratio-to-moving-average or link relatives.
Stationarity:
- Weak (covariance) stationarity: constant mean, constant variance, and an autocovariance that depends only on the lag k, not on t.
- Strict stationarity: the whole joint distribution is invariant over time.
- White noise: mean 0, constant variance, no autocorrelation.
Random walk:
- Without drift: Yₜ = Yₜ₋₁ + uₜ. E(Yₜ) = Y₀ but Var = tσ², which grows over time, so the series is non-stationary.
- With drift: Yₜ = δ + Yₜ₋₁ + uₜ.
- A random walk is difference stationary (a stochastic trend, I(1)). A trend-stationary series (deterministic trend) becomes stationary by detrending.
- Integration: a series is I(d) if it must be differenced d times to become stationary.
Unit root tests:
- Dickey–Fuller (1979): ΔYₜ = δYₜ₋₁ + uₜ, with H₀: δ = 0 (unit root, ρ = 1). The test statistic τ (tau) does not follow the t-distribution; critical values come from DF tables or MacKinnon. There are three forms: no constant, constant, and constant + trend.
- Augmented DF (Said & Dickey, 1984; Dickey & Fuller, 1981): adds lagged ΔY terms to remove serial correlation.
- Phillips–Perron (1988): a non-parametric correction of the DF statistic for serial correlation and heteroscedasticity.
- KPSS (1992): H₀ = stationarity (the reverse of DF).
- The correlogram (ACF) of a non-stationary series decays very slowly.
Spurious regression:
- Regressing one I(1) series on another unrelated I(1) series gives a high R², significant t values and a very low DW.
- Granger & Newbold (1974) rule of thumb: R² > d suggests a spurious regression.
- Earlier: Yule (1926), "nonsense correlations".
Cointegration (Engle & Granger, 1987; Nobel Prize 2003):
- Two I(1) series are cointegrated if a linear combination of them is I(0), i.e. they share a long-run equilibrium.
- Engle–Granger two-step method: (1) estimate the levels regression by OLS and test the residuals for a unit root using EG/AEG critical values; (2) estimate the error correction model ΔYₜ = α₀ + α₁ΔXₜ + α₂ûₜ₋₁ + εₜ, where α₂ < 0 is the speed of adjustment. (Granger representation theorem: cointegration ⇔ an ECM exists.)
- The ECM originates with Sargan (1964) and was popularised by Davidson, Hendry, Srba & Yeo (1978).
- With more than two variables, use Johansen's (1988) trace and maximum-eigenvalue tests.
- Cointegrating regression DW (CRDW) test.
Granger causality (1969):
- X Granger-causes Y if past values of X improve the prediction of Y beyond Y's own past values. Tested with an F test on the lagged X terms.
- Possible outcomes: unidirectional, bidirectional (feedback), or independence.
- The variables must be stationary. "Precedence" is not true causality.
- Sims (1972, 1980) proposed VAR models as an alternative.
ARIMA / Box–Jenkins (1970):
- AR(p): Yₜ = φ₁Yₜ₋₁ + … + φₚYₜ₋ₚ + uₜ. An AR(1) is stationary if |φ| < 1.
- MA(q): Yₜ = μ + uₜ + θ₁uₜ₋₁ + … + θ_q uₜ₋_q. A finite-order MA is always stationary.
- ARMA(p, q). ARIMA(p, d, q) = ARMA on a series differenced d times.
- Four steps: identification (ACF, PACF, correlogram) → estimation → diagnostic checking (are the residuals white noise? Box–Pierce Q = nΣρ̂ₖ² and Ljung–Box Q ~ χ²(m)) → forecasting.
| Process | ACF | PACF |
|---|---|---|
| AR(p) | decays exponentially or as a damped sine wave | cuts off after lag p |
| MA(q) | cuts off after lag q | decays |
| ARMA(p, q) | decays | decays |
ARCH (Engle, 1982) and GARCH (Bollerslev, 1986) model volatility clustering.
Commonly tested
- Random walk → non-stationary, I(1), variance tσ².
- DF τ ≠ t; ADF adds lagged differences; PP is non-parametric; KPSS has stationarity as H₀.
- Spurious regression → R² > d (Granger–Newbold 1974).
- Cointegration → Engle–Granger 1987 + ECM. Granger causality → 1969. Box–Jenkins → 1970.
- PACF cuts off for an AR process; ACF cuts off for an MA process.
Last-minute 20 facts
- Poisson: mean = variance = λ. Binomial: mean np > variance npq.
- χ²(k): mean k, variance 2k. t(k): variance k/(k − 2). t²(k) = F(1, k).
- Normal: β₁ = 0, β₂ = 3. β₂ > 3 → leptokurtic.
- AM ≥ GM ≥ HM; GM² = AM × HM (two values).
- Mode = 3 Median − 2 Mean (moderate skew).
- Bowley skewness = (Q₃ + Q₁ − 2Md)/(Q₃ − Q₁) ∈ [−1, 1].
- Spearman ρ = 1 − 6Σd²/[n(n² − 1)].
- r = GM of the regression coefficients; r is independent of origin and scale.
- Fisher's index passes time reversal and factor reversal but fails the circular test. The circular test is passed by the simple GM of relatives and Kelly's index.
- Laspeyres (base weights) has an upward bias; Paasche (current weights) a downward bias.
- α = Type I error; power = 1 − β; Neyman–Pearson lemma → most powerful test.
- Gauss–Markov: OLS is BLUE under the CLRM; normality is not needed.
- R̄² = 1 − (1 − R²)(n − 1)/(n − k); it can be negative.
- VIF = 1/(1 − R²ⱼ) > 10 → serious multicollinearity; OLS is still BLUE.
- d ≈ 2(1 − ρ̂), range 0–4; use Durbin h (or BG) when a lagged Y is a regressor.
- Order condition K − k ≥ m − 1 is necessary; the rank condition (an (M − 1) × (M − 1) non-zero determinant) is necessary and sufficient.
- ILS → just identified; 2SLS → over-identified; 3SLS → Zellner–Theil 1962; recursive model → OLS.
- Logit = ln odds, linear in X; logit coefficient ≈ 1.81 × probit coefficient; LPM gives probabilities outside [0, 1].
- Spurious regression: R² > DW (Granger–Newbold 1974). Cointegration and ECM: Engle–Granger 1987.
- AR(p): PACF cuts off at p. MA(q): ACF cuts off at q. Box–Jenkins steps: identify → estimate → diagnose → forecast.
Sources
- Gujarati, D. N. & Porter, D. C., Basic Econometrics, 5th ed. (McGraw-Hill) — chapters on CLRM, multicollinearity, heteroscedasticity, autocorrelation, dummy variables, qualitative response models, simultaneous equations and identification, time series.
- Wooldridge, J. M., Introductory Econometrics: A Modern Approach — Gauss–Markov, BG test, unit roots.
- Gupta, S. C. & Kapoor, V. K., Fundamentals of Mathematical Statistics (Sultan Chand) — distributions, moments, skewness, index-number tests.
- Wikipedia, "F-distribution" — mean and variance (cross-checked): https://en.wikipedia.org/wiki/F-distribution
- SAS/ETS documentation, AUTOREG — Durbin h formula: https://support.sas.com/documentation/cdl/en/etsug/60372/HTML/default/etsug_autoreg_sect024.htm and https://www.sfu.ca/sasdoc/sashtml/ets/chap8/sect24.htm
- gretl-users mailing list (2014) — Durbin h implementation, cross-check: https://gretlml.univpm.it/hyperkitty/list/[email protected]/message/4A7XTRLSPQHCOSK2KAKPXSJHTMBCFWG3/
- Numeric examples (index numbers, regression, Spearman) recomputed by hand and in Python.
Practice: 56 real questions
Newest papers first. Pick an option to check yourself; your progress is saved on this device.
Given below are two statements : one is labelled as Assertion (A) and the other is labelled as Reason (R).
If X is a given series of observations, a and b are any two constants, then with regard to Standard Deviation (SD).
Assertion (A) : SD(X + a) = SD(X − b) = SD(X) and SD(aX) = a².SD(X)
Reason (R) : SD is independent of change of origin, but not of scale.
In the light of the above statements, choose the most appropriate answer from the options given below :
Given samples of sizes 20 and 15 with respective sample variances 35.53 and 34.92, the calculated value of F-statistic is :
Coefficient of Determination (R²) :
A. is an increasing function of the number of regressors
B. is a decreasing function of the number of regressors
C. is unaffected by increasing or decreasing the number of regressors
D. is always positive
Choose the correct answer from the options given below :
Given the structural model
y₁ = 3y₂ − 2x₁ + x₂ + u₁
y₂ = y₃ + x₃ + u₂
y₃ = y₁ − y₂ − 2x₃ + u₃
Based on the order condition, the second equation is :
If P(A) = 1/5 and P(B) = 1/3 and A and B are two mutually exclusive events, then the probability of neither A nor B will occur is :
Given below are two statements : one is labelled as Assertion (A) and the other is labelled as Reason (R).
Assertion (A) : The presence of heteroscedasticity problem in regression analysis implies that the least square estimators are still unbiased but inefficient.
Reason (R) : The estimates of the variances are also unbiased.
In the light of the above statements, choose the most appropriate answer from the options given below :
Given below are two statements : one is labelled as Assertion (A) and the other is labelled as Reason (R).
Assertion (A) : If X and Y are any two variables and are transformed to the new variables U and V defined by U = (X − A)/h, V = (Y − B)/K where A, B, h, K are any constants, h and K > 0 then, rxy = ruv
Reason (R) : Correlation coefficient(r) is independent of change of origin and scale.
In the light of the above statements, choose the most appropriate answer from the options given below :
Match List - I with List - II.
| List - I | List - II |
|---|---|
| A. Type I error | I. Accept H0 when it is false |
| B. Type II error | II. H0 : μ = μ0 and σ² = σ0² |
| C. Simple Hypothesis | III. Reject H0 when it is true |
| D. Composite Hypothesis | IV. H0 : μ = μ0 and σ² is unknown |
Choose the correct answer from the options given below :
Based on Beta (β) and Gama (γ) coefficients, if β2 < 3 or γ2 < 0 then the curve is said to be :
Based on {εt} is a purely random process with mean zero and variance σ² :
A. A process {xt} defined by xt = α1xt−1 + α2xt−2 + ...... + αrxt−r + εt is called an autoregressive process of order r
B. A process {xt} defined by xt = β0εt + β1εt−1 + ...... + βmεt−m is called a moving average process of order m
C. A process {xt} defined by xt = β0εt + β1εt−1 + ...... + βmεt−m is called an autoregressive process of order r
D. A process {xt} defined by xt = α1xt−1 + α2xt−2 + ...... + αrxt−r + εt is called a moving average process of order m
Choose the correct answer from the options given below :
Bootstrapping technique is use to ___________.
According to Granger and Newbold, a good rule of thumb to suspect that the estimated regression is spurious is given by :
The indirect least square is applied to estimate the coefficient of the :
The management of a manufacturing firm wishes to determine the average time required to complete a certain manual operation. There should be 0.95 confidence that error in the estimate will not exceed 2 minutes. What sample size is estimated by a time and motion study expert as 16 minutes?
If x is a poisson variate with mean λ, then P(x+1) is :
A candidate is selected for interview for three posts. For the first post, there are 3 candidates, for the second 4 candidates and for the third 2 candidates. What is the probability that the candidate is selected for at least one post ?
The data about the sales and advertisement expenditure of a firm is given below:
| Sales (In crore of Rs.) | Advertisement expenditure (In crore of Rs.) | |
|---|---|---|
| Mean | 40 | 6 |
| Standard deviation | 10 | 1.5 |
Coefficient of correlation=r=0.9
Estimate the likely sales for the proposed advertisement of Rs. 10 crore of rupees.
If x̄ = 32, ȳ = 38, the regression coefficients b_xy = −0.2337, b_yx = −0.6643. Find the equation of the line of regression of y on x.
The null hypothesis that all slope coefficients are simultaneously equal to zero is tested in logit model by.
Which of the followings are correct in case of Normal distribution?
A. μ₃ for Normal distribution is 0.
B. β₂ for Normal distribution is 0.
C. Mean deviation (about mean) for Normal distribution is (4/5)σ
D. μ₄ for Normal distribution is 2σ⁴
E. Characteristic function for Normal distribution is e^(iμt − t²σ²/2)
Choose the correct answer from the options given below :
By using the uniqueness property of moment generating function (M.G.F.), which of the followings are correct distribution if M.G.F. is as follows?
A. M(t) = (1/2 + (1/2)eᵗ)⁶ is a poisson distribution.
B. M(t) = (1 + eᵗ)⁵/32 is a binomial distribution.
C. M(t) = e^(3(eᵗ−1)) is poisson distribution
D. M(t) = e^((eᵗ−1)/4) is poisson distribution
E. M(t) = 4(3e^(−t) − 1)^(−2) is binomial distribution
Choose the correct answer from the options given below :
Match List-I with List-II :
| List-I | List-II |
|---|---|
| A. β₂ for Chi-square distribution | I. (n-2) |
| B. Skewness for Chi-square distribution | II. 48n+12n² |
| C. Mode for Chi-square distribution | III. (12/n)+3 |
| D. μ₃ for Chi-square distribution | IV. √(2/n) |
Choose the correct answer from the options given below :
Match List-I with List-II :
| List-I | List-II |
|---|---|
| A. The mean of a Hypergeometric distribution | I. q/p² |
| B. The moment generating function of negative binomial distribution | II. (Q − Peᵗ)^(−r) |
| C. The coefficient of Kurtosis of a binomial distribution | III. nM/N |
| D. The variance of geometric distribution | IV. (1−6pq)/(npq) |
Choose the correct answer from the options given below :
Passage / data
Based on the information given below, answer the questions :
Two fair dice are thrown independently. Three events A, B, and C are defined as follows;
A : odd face with first dice
B : odd face with second dice
C : sum of points on two dice is odd
Find the P(C).
Same passage / data as the previous question
Based on the information given below, answer the questions :
Two fair dice are thrown independently. Three events A, B, and C are defined as follows;
A : odd face with first dice
B : odd face with second dice
C : sum of points on two dice is odd
Find the P(A∩B).
Same passage / data as the previous question
Based on the information given below, answer the questions :
Two fair dice are thrown independently. Three events A, B, and C are defined as follows;
A : odd face with first dice
B : odd face with second dice
C : sum of points on two dice is odd
Find the P(A∩C).
Same passage / data as the previous question
Based on the information given below, answer the questions :
Two fair dice are thrown independently. Three events A, B, and C are defined as follows;
A : odd face with first dice
B : odd face with second dice
C : sum of points on two dice is odd
Find the P(B∩C).
Same passage / data as the previous question
Based on the information given below, answer the questions :
Two fair dice are thrown independently. Three events A, B, and C are defined as follows;
A : odd face with first dice
B : odd face with second dice
C : sum of points on two dice is odd
Find the P(A∩B∩C).
ANCOVA models include regressors that are :
A and B are two events such that P(Ā) = 0.4 and P(A ∩ B) = 0.2, then P(A ∩ B̄) is :
You have carried out a regression analysis but after thinking about the relationship between the two variables, you have decided that you must swap the explanatory and response variables. After refitting the regression model to the data, you expect that :
The lowest value of a set of observations is 4.5 and their range is 10.9. Artithmetic mean is found to be 19.9 and median found to be 15.6. Which of the following statements is correct ?
Estimation of regression coefficients in the presence of high but not perfect multicollinearity may result in all of these except :
Testing for cointegration is performed by :
Heteroscadesticity may arise due to various reasons. Which one of the following is not a reason ?
To test H0 : μ = μ0, when both μ, and σ are unknown, the relevant test statistic is :
Match List - I with List - II.
| List - I | List - II |
|---|---|
| (A) Endogenous variable | (I) Significance of Autocorrelation coefficient |
| (B) Count R² | (II) Simultaneous equation |
| (C) Ljung Box Statistic | (III) Panel data regression |
| (D) Least square dummy variable model | (IV) Logit regression |
Choose the correct answer from the options given below :
Match List - I with List - II.
| List - I | List - II |
|---|---|
| (A) General Heteroscadesticity test | (I) Ramsey |
| (B) General test for auto correlation | (II) Hausman |
| (C) Test of simultaneity | (III) White |
| (D) General test of specification error | (IV) Breusch-Godfrey |
Choose the correct answer from the options given below :
In a regression of a dependent variable on intercept and k other independent variables using n observations, what are respectively the numerator and denominator degrees of freedom for the F test for overall significance of the regression ?
Suppose Xi is income of an individual i and Yi is the event that this individual owns a house, Pi = prob (Yi = 1 | Xi) and Zi = β1 + β2 Xi, which of the following is true for the logit model ?
If the estimated slope coefficient in the regression of Y on X is β̂, what will be the estimated slope coefficient in the regression of 5Y on 10X ?
Which of the following tests has stationarity as the null hypothesis ?
For a VAR model estimated using two variables X and Y, Lags of Y are jointly significant in the equation for X but Lags of X are jointly insignificant in the equation for Y. What can you conclude from this ?
In the model ln Yi = α + βXi + ui Where ln is natural log, the elasticity of Y w.r.t. X is given by :
Suppose you are testing a Null Hypothesis about the population mean, Ho : μ = μ0 against H1 : μ > μ0, with population variance given. The computed value of the test statistic is Zc. What is the p-value for this test, where Φ( ) is the cumlative distribution function for the standard normal distribution ?
The power of a test refers to :
Given a random sample of size n from a population with mean μ and variance, σ², the sample mean has the following expected value and variance :
Which of the following statements are not true for the Linear Probability Model (LPM) ?
(A) Conditional probability of occurence of an event conditional on the value of explanatory variable X is a linear function of X
(B) Estimated value of the conditional probability of occurence of the event lies between 0 and 1
(C) Error term is homoscedastic
(D) R² is a good measure of goodness of fit
Choose the correct answer from the options given below :
Given that Zt is a zero-mean white noise process with variance σ², which of the following are I(0) processes ?
(A) Xt = Xt−1 + Zt
(B) Xt = αXt−1 + Zt where 0 < α < 1
(C) Xt = Zt + Zt−1
(D) Xt = 8 + Xt−1 + Zt
Choose the correct answer from the options given below :
Which of the following are features of a weakly stationary stochastic process ?
(A) Mean is time-varying
(B) Mean is constant
(C) Variance is constant
(D) Covariance is time-invariant
Choose the correct answer from the options given below :
The OLS regression output of wages (Rs in thousand) on education (years of schooling S) and cognitive ability scores (CA) comes out to be :
wages^ = −3.36 + 0.64 S + 0.15 CA
(1.98) (0.12) (0.05)
with figures in parenthesis giving standard error.
The explained sum of squares is 4745.75 with degrees of freedom 2 and residual sum of squares is 33651.28 with degrees of freedom 567, which of the following statements is/are true ?
(A) For one year increase in schooling, wage per hour increases by Rs. 640 on the average, when CA is kept constant.
(B) The value of t statistic of CA is 3.
(C) The F statistic is given by F = (4745.75/2)/(33651.28/567)
(D) The F statistic is given by F = (33651.28/567)/(4745.75/2)
Choose the correct answer from the options given below :
Match List - I with List - II.
| List - I (Model/method) | List - II (Feature) |
|---|---|
| (A) Logit model | (I) elasticities are constant |
| (B) Log-log model | (II) binary dependent variable |
| (C) Autoregressive model | (III) corrects for heteroscedasticity |
| (D) Weighted least squares | (IV) includes lags of dependent variable |
Choose the correct answer from the options given below :
Match List - I with List - II.
| List - I (Variable) | List - II (Distribution) |
|---|---|
| (A) Sum of squares of n independent standard normal variates | (I) t distribution |
| (B) Ratio of a standard normal variate to square root of an independent, chi square variable (divided by its degrees of freedom) | (II) F distribution |
| (C) Ratio of 2 independent chi-square variables each divided by their respective degrees of freedom | (III) Standard normal distribution |
| (D) Normal variate minus its mean, divided by its standard deviation | (IV) chi square distribution |
Choose the correct answer from the options given below :
Chow Test is normally run to understand if a single regression run on pooled data is more efficient to describe a relation between a dependent variable and a set of independent variables than the same regression being run separately on two subgroups of sample. The steps involved in running the Chow Test are :
(A) Run the regression model on pooled sample and calculate RSSp
(B) Run the same regression model on the two sub-samples separately and calculate RSS1 and RSS2
(C) Calculate the F statistic
(D) Check the critical F with related degrees of freedom and compare it with F calculated
(E) Do not reject H0 if Fcalculated < Fcritical
Choose the correct answer from the options given below :
Given below are two statements : one is labelled as Assertion (A) and the other is labelled as Reason (R).
Assertion (A) : Given a random sample, the distribution of the sample mean tends to a normal distribution, as the sample size becomes very large.
Reason (R) : Central limit theorem
In the light of the above statements, choose the most appropriate answer from the options given below :