Mains PYQs
UPSC IFoS Mains Statistics paper 1 2016 questions
- (i) Show that for arbitrary events $A_1, A_2, \ldots, A_n$,
$$P\left(\bigcup_{1}^{n} A_i\right) \le \sum_{i=1}^{n} P(A_i).$$
(ii) If $P_1$ and $P_2$ are probability measures, prove that $P = \alpha P_1 + (1-\alpha) P_2$ is also a probability measure, $(0 \le \alpha \le 1)$.
- Let $(X_1, X_2)$ denote quantitative scores on test 1, and $(Y_1, Y_2)$ be verbal scores on test 2. If $\mathrm{Cov}(X_1, Y_1) = 5$, $\mathrm{Cov}(X_1, Y_2) = 1$, $\mathrm{Cov}(X_2, Y_1) = 2$ and $\mathrm{Cov}(X_2, Y_2) = 8$, compute the covariance between total quantitative score $(X_1 + X_2)$ and total verbal score $(Y_1 + Y_2)$.
- Derive the lower bound for the variance of an unbiased estimator of $\sigma^2$ of $N(\mu, \sigma^2)$. Hence show that $s^2 = \frac{1}{n-1}\sum (x_i - \bar{x})^2$ does not attain this bound. Why is it still the best unbiased estimator?
- Obtain $100(1-\alpha)\%$ confidence interval for the difference $(p_1 - p_2)$ of success probabilities of two independent Bernoulli distributions.
- Show that the characteristic function (ch. fn.) of a random variable determines its distribution function (d.f.) uniquely.
- If $g$ is a continuous function and $X_n \xrightarrow{P} X$, then show that $g(X_n) \xrightarrow{P} g(X)$.
- For the Pareto distribution with pdf
$$f(x \mid \alpha) = \frac{\alpha}{x_0}\left(\frac{x_0}{x}\right)^{1+\alpha}, \quad x > 0, \alpha > 0,$$
show that the d.f. is $\left(1 - \frac{x_0}{x}\right)^{\alpha}$ and sketch it for $\alpha = \frac{1}{2}$ and $\alpha = 2$. Also show that $\mathrm{Var}(X)$ does not exist for $\alpha \le 2$. (Sketches to be shown on plain paper).
- Given a random sample from
$$f(x \mid \theta) = \theta x^{\theta-1} e^{-x^\theta}, \quad x > 0, \theta > 0.$$
Show that the loglikelihood equation has a unique root and that it provides a strongly consistent estimator.
- Let $X$ have a Poisson distribution with mean $\theta$. Assume that $\theta$ has a $\Gamma(p, \sigma)$ distribution, that is,
$$\pi(\theta) = \frac{\theta^{p-1}\sigma^p e^{-\sigma\theta}}{\Gamma(p)}, \quad \theta > 0, \sigma > 0, p \ge 1.$$
Show that the posterior distribution given $(x_1, \ldots, x_n)$ is Gamma $\left(p + \sum_{1}^{n} x_i, \sigma + n\right)$. Hence obtain the Bayes estimator of $\theta$ with respect to a squared-error loss.
- Derive the likelihood ratio test of $H_0: \theta \le \theta_0$ versus $H_1: \theta > \theta_0$ given a random sample from
$$f(x \mid \theta) = \frac{1}{2}\exp\left\{-\frac{1}{2}(x - \theta)\right\}, \quad x > 0, \theta > 0,$$
considering $\alpha = 0.05$.
- Consider a random sample $X_1, X_2, \ldots, X_n$ from $U(0, \theta)$. Show that the range $R = X_{(n)} - X_{(1)}$ is an ancillary statistic.
- Obtain the level-$\alpha$ UMPU test of $H_0: \sigma^2 \le \sigma_0^2$ against $H_1: \sigma^2 > \sigma_0^2$ of $N(\mu, \sigma^2)$ and its power function ($\mu$ is unknown).
- Using an appropriate pivotal quantity, derive a $100(1-\alpha)\%$ confidence interval for $\mu$ of $N(\mu, \sigma^2)$, where $\sigma^2$ is unknown. Is the C.I. also UMAU?
- Briefly explain the implementation of SPRT and state its optimal property. Illustrate it for testing $\theta = \frac{1}{3}$ versus $\theta = \frac{2}{3}$ for the distribution
$$f(x \mid \theta) = \theta^{\frac{1-x}{2}}(1-\theta)^{\frac{1+x}{2}}, \quad x = -1, 1.$$
- A fair coin is tossed $n$ times and $S_n$, the sum of number of heads, is noted. Use it to find a lower bound for the probability that $\frac{S_n}{n}$ differs from $\frac{1}{2}$ by less than $0.1$ when $n = 100$, and show that $\frac{S_n}{n} \xrightarrow{P} \frac{1}{2}$ as $n \to \infty$.
- Stating the hypothesis, explain the two-sample Wilcoxon-Mann-Whitney test and derive the mean of the test statistic.
- Define convergence in distribution. Show that the maximum of a random sample from $U(0, \theta)$ converges to an exponential distribution.
- Given the Gauss-Markov linear model $(\mathbf{Y}, \mathbf{A}\boldsymbol{\theta}, \sigma^2 \mathbf{I})$ with $E(Y_1) = \theta_1 + \theta_2 - \theta_3$, $E(Y_2) = 2\theta_1 - \theta_2 - \theta_3$, $E(Y_3) = \theta_1 - 2\theta_2$ and $E(Y_4) = 3\theta_1 - 2\theta_3$. Find a necessary and sufficient condition for $\mathbf{a}'\boldsymbol{\theta} = a_1\theta_1 + a_2\theta_2 + a_3\theta_3$ to be estimable. Hence examine the estimability of $\theta_1 + \theta_2 - 2\theta_3$.
- Given $\mathbf{X} \sim N_3\left(\boldsymbol{\mu}' = (2, 4, 3),\ \Sigma = \begin{bmatrix} 8 & 2 & 3 \\ 2 & 4 & 1 \\ 3 & 1 & 3 \end{bmatrix}\right)$,
(i) find the regression function of $X_1$ on $X_2$ and $X_3$, and
(ii) compute the conditional variance of $X_1$ given $X_2$ and $X_3$.
- Define connectedness of a block design. Examine if the following design is connected:
$B_1: (T_1, T_2, T_3)$, $B_2: (T_3, T_4, T_5)$, $B_3: (T_3, T_4)$,
$B_4: (T_1, T_2, T_2)$, $B_5: (T_4, T_4, T_5)$.
- Describe two-stage sampling with a real life example (stating clearly the first and second stage units) and compare it with cluster sampling.
- What is the importance of confounding in factorial experiments? Give a layout of a $2^3$ factorial experiment with 2 replicates so as to confound partially the interaction effects $ABC$ and $BC$.
- Define multiple and partial correlation coefficients. Suppose that the correlation matrix of $(X_1, X_2, X_3, X_4)'$ is
$$R = \begin{bmatrix} 1 & 0.6 & -0.36 & -0.53 \\ & 1 & -0.46 & -0.37 \\ & & 1 & 0.37 \\ & & & 1 \end{bmatrix}.$$
Find
(i) the multiple correlation coefficient between $X_1$ and the other 3 variables, and
(ii) the partial correlation coefficient $\rho_{X_1 X_2 \cdot X_3 X_4}$.
- Suppose that $Y_i$, $i = 1, \ldots, 6$ is uncorrelated with common variance $\sigma^2$ and $E(Y_1) = \beta_1 = E(Y_2)$, $E(Y_3) = \beta_2 = E(Y_4)$ and $E(Y_5) = \beta_1 + \beta_2 = E(Y_6)$. Obtain four mutually orthogonal contrasts of $(Y_i)$ belonging to the error space. Find BLUE of $\beta_1$ and $\beta_2$.
- If $\mathbf{Q}$ denotes the vector of adjusted treatment totals in a block design, find (i) $\mathbf{Q}'\mathbf{J}$ ($\mathbf{J}$ is a vector of 1's), (ii) $E(\mathbf{Q})$, and (iii) covariance matrix of $\mathbf{Q}$.
- Define a regression estimator of population mean. Show that this estimator is more efficient than the sample mean under SRSWOR using large sample approximation.
- Discuss the main effects and interaction effects of a $2^3$ factorial experiment as orthogonal contrasts. Also describe how to compute the sum of squares due to these contrasts.
- Explain the concept and utility of principal components. Find the first principal component and its variance of the random vector with covariance matrix $\Sigma = \begin{bmatrix} 9 & 3 & 3 \\ & 9 & 3 \\ & & 9 \end{bmatrix}$.
- Discuss the need for PPS Sampling illustrating with an example from real life. Define Horvitz-Thompson estimator and show that it is unbiased for the population total.
- Show that a split-plot design in an RBD with $p$ main plot treatments and $q$ sub-plot treatments provides more precise comparisons among sub-plot treatments relative to an RBD with $pq$ plots.
- Suppose that a variable of interest is distributed as $U(b, b+h)$, $b > 0$, $h > 0$. Let the range be divided into $L$ strata of equal sizes. An SRS of size $\frac{n}{L}$ is selected from each stratum. Denoting by $V$ and $V_1$, the variances under an SRS of size $n$, and a stratified random sample as described above, respectively, show that $\frac{V_1}{V} = \frac{1}{L^2}$. Interpret this result.
- Define multiple linear regression model. Find the unbiased estimators of the parameters of this model. State the sampling distribution of these estimators under the assumption of normality.
- Explain how one missing value in an RBD is estimated so as to minimise the error sum of squares. Set up the ANOVA table for analysis in this situation.
- Suppose that a sample of size $N = 20$ is drawn from $\mathbf{X} \sim N_3(\boldsymbol{\mu}, \Sigma)$ which yields the sample mean vector and sample covariance matrix as
$$\bar{\mathbf{x}} = (21.05,\ 21.65,\ 28.95)'$$
$$S = \begin{bmatrix} 2.2605 & 2.1763 & 1.6342 \\ & 2.6605 & 1.8237 \\ & & 2.4710 \end{bmatrix}.$$
Construct the Likelihood Ratio Test of the hypothesis
$$H_0: \begin{pmatrix} \mu_2 - \mu_1 \\ \mu_3 - \mu_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix}.$$