IGNOU MECE-101 Solved Question Paper PDF Download

IGNOU MECE-101 Previous Year Solved Question Paper in Hindi

IGNOU MECE-101 Previous Year Solved Question Paper in English

Q1. Consider the regression model : Yᵢ = β₁ + β₂Xᵢ + uᵢ where Yᵢ is hourly wage rate and Xᵢ is years of schooling. Also xᵢ = (Xᵢ – X̄) and yᵢ = (Yᵢ – Ȳ). The following data is given : ΣYᵢ = 307.35, ΣXᵢ = 6608.0 ΣXᵢ² = 87040.0, ΣXᵢYᵢ = 440.65 ΣYᵢ² = 25446.29, Σxᵢ² = 4025.43 Σyᵢ² = 760.44, Σxᵢyᵢ = 279.204 Σuᵢ² = 5980.68, n = 526 (a) Compute OLS estimator of β₁ and β₂. Interpret their results. (b) Compute the value of the coefficient of determination (R²). Interpret the results.

Ans. (a) Computation and Interpretation of OLS estimators of β₁ and β₂: For the given model Yᵢ = β₁ + β₂Xᵢ + uᵢ, the OLS estimator formulas are: Estimator for the slope coefficient: β̂₂ = Σxᵢyᵢ / Σxᵢ² Estimator for the intercept: β̂₁ = Ȳ – β̂₂X̄ First, we compute β̂₂ using the given data in deviation form: Σxᵢyᵢ = 279.204 Σxᵢ² = 4025.43 β̂₂ = 279.204 / 4025.43 ≈ 0.06936 Next, we compute the means to calculate β̂₁: X̄ = ΣXᵢ / n = 6608.0 / 526 ≈ 12.5627 Ȳ = ΣYᵢ / n = 307.35 / 526 ≈ 0.5843 Substituting these values into the formula for β̂₁: β̂₁ = 0.5843 – (0.06936 * 12.5627) ≈ 0.5843 – 0.8715 ≈ -0.2872 The estimated regression equation is: Ŷᵢ = -0.2872 + 0.06936Xᵢ Interpretation of the results:

• β̂₂ (Slope coefficient): The estimated value is 0.06936. This means that, holding all other factors constant, an additional year of schooling is expected to increase the hourly wage rate by approximately 0.069 units, on average. This indicates a positive relationship between schooling and wages, which is consistent with economic theory.
• β̂₁ (Intercept): The estimated value is -0.2872. Theoretically, this is the predicted hourly wage rate for an individual with zero years of schooling (Xᵢ=0). A negative wage rate is economically nonsensical. This often indicates that the linear form of the model does not extrapolate well outside the range of the data, particularly near X=0.

(b) Computation and Interpretation of the coefficient of determination (R²):

The coefficient of determination, R², measures the proportion of the variation in the dependent variable (Y) that is explained by the independent variable (X). The formula for its calculation is:

R² = (Σxᵢyᵢ)² / (Σxᵢ² * Σyᵢ²)

Using the given values (Note: The English version had a typo `Ly? न 60.44`, while the Hindi version had `Ly? = 760.44`. We use 760.44 as it yields a plausible R² value):

R² = (279.204)² / (4025.43 * 760.44)

R² ≈ 77954.9 / 3061141.6 ≈

0.0255

Interpretation of the results:

The value of R² is approximately 0.0255, or 2.55%. This means that only

2.55%

of the total variation in the hourly wage rate is explained by the variation in years of schooling. This is a very low R² value, suggesting that years of schooling alone is a very weak predictor of hourly wages. Many other important factors that influence wages (such as experience, gender, occupation, innate ability) are not included in this simple regression model.

Q2. What is meant by autocorrelation ? How is it detected ? Explain one remedial measure to deal with autocorrelation.

Ans. Meaning of Autocorrelation: Autocorrelation, also known as serial correlation, is a violation of the assumptions of the Classical Linear Regression Model (CLRM). It refers to the presence of a correlation between the error terms of a regression model across different observations, typically in time-series data. Mathematically, autocorrelation means that the covariance of the error terms is not zero: Cov(uᵢ, uⱼ) ≠ 0 for i ≠ j This violates the CLRM assumption that the error terms are independent of one another. The consequences of autocorrelation are that OLS estimators are no longer BLUE (Best Linear Unbiased Estimator); while they remain unbiased, they are inefficient (not ‘Best’), and their standard errors are incorrect, leading to unreliable hypothesis tests. Detection of Autocorrelation: There are several methods to detect autocorrelation:

• Graphical Method: This is an informal method where the residuals (ûᵢ) from an OLS regression are plotted against time. If the plot shows a clear pattern (e.g., a wave-like or cyclical pattern), it is indicative of autocorrelation.
• Durbin-Watson (DW) d-statistic: This is the most famous test for first-order autocorrelation (AR(1)). The d-statistic is calculated and compared to two critical values, dₗ (lower bound) and dᵤ (upper bound). The outcome depends on where the statistic falls: positive autocorrelation, no autocorrelation, inconclusive region, or negative autocorrelation. Its major limitation is that it’s not applicable for models with lagged dependent variables.
• Breusch-Godfrey (BG) Test: This is a more general and powerful test. It can test for higher-order autocorrelation and is also valid for models that include lagged values of the dependent variable, where the DW test is inappropriate. It is an LM (Lagrange Multiplier) test based on an auxiliary regression of the residuals on the original regressors and lagged residuals.

Remedial Measure: The Cochrane-Orcutt Procedure

When autocorrelation is detected, a common remedial measure is to use a procedure that transforms the model so that the new error term is free of autocorrelation. The Cochrane-Orcutt iterative procedure is one such method, which is a form of Generalized Least Squares (GLS).

1. Estimate the original model: Apply OLS to the original model Yₜ = β₁ + β₂Xₜ + uₜ and obtain the residuals, ûₜ.
2. Estimate ρ (rho): Assuming the errors follow a first-order autoregressive (AR(1)) scheme (uₜ = ρuₜ₋₁ + εₜ), estimate the autocorrelation coefficient (ρ) by regressing the residuals on their one-period lagged values: ûₜ = ρûₜ₋₁ + vₜ This gives an estimate of ρ, denoted as ρ̂.
3. Transform the variables: Use ρ̂ to create quasi-differenced or transformed variables: Y*ₜ = Yₜ – ρ̂Yₜ₋₁ X*ₜ = Xₜ – ρ̂Xₜ₋₁
4. Estimate the transformed model: Run an OLS regression on the transformed variables: Y ₜ = β₁ + β₂X*ₜ + εₜ where β₁ = β₁(1 – ρ̂). The estimator of β₂ from this regression is now BLUE, and an estimate of β₁ can be recovered from β̂₁ / (1 – ρ̂). The new error term, εₜ, is now serially uncorrelated, allowing for reliable hypothesis testing. This procedure can be iterated until the estimates of ρ̂ converge in successive iterations.

Q3. Consider the regression model : Y = β₁ + β₂X₂ + β₃X₃ + u where u follows N(0, σ²). The sample size is 25. Suppose you have to estimate the hypothesis β₂ = β₃ = 0. Explain the steps to test the hypothesis.

Ans. The hypothesis H₀: β₂ = β₃ = 0 is a joint hypothesis . It tests whether the independent variables X₂ and X₃ are simultaneously or jointly statistically significant in explaining the dependent variable Y. The standard method to test such a hypothesis is the F-test , also known as the test for overall significance or the test for restrictions on a regression. The steps to test the hypothesis are as follows: Step 1: Estimate the Unrestricted Model First, we estimate the full or unrestricted model, which includes all the variables: Y = β₁ + β₂X₂ + β₃X₃ + u After applying OLS to this model, we obtain the Unrestricted Residual Sum of Squares (RSS_UR) . The degrees of freedom for this model is df_UR = n – k , where n is the sample size and k is the number of parameters estimated. In this case, n = 25 and k = 3 (for β₁, β₂, β₃). So, df_UR = 25 – 3 = 22 . Step 2: Estimate the Restricted Model Next, we impose the restrictions given by the null hypothesis (H₀), which are β₂ = 0 and β₃ = 0. Substituting these restrictions into the original model gives us the restricted model: Y = β₁ + u We estimate this simple model and obtain the Restricted Residual Sum of Squares (RSS_R) . The degrees of freedom for the restricted model is df_R = n – k_R , where k_R is the number of parameters in the restricted model. In this case, k_R = 1 (only β₁). So, df_R = 25 – 1 = 24 . Step 3: Calculate the F-statistic The F-statistic is computed using the RSS_UR and RSS_R. The formula is: F = [ (RSS_R – RSS_UR) / m ] / [ RSS_UR / (n – k) ] Where:

• RSS_R = Residual Sum of Squares from the restricted model.
• RSS_UR = Residual Sum of Squares from the unrestricted model.
• m = Number of restrictions imposed in the null hypothesis. In this case, m = 2 (since β₂=0 and β₃=0).
• n = Sample size = 25.
• k = Number of parameters in the unrestricted model = 3.

Step 4: Make a Decision

We compare the calculated F-statistic (F_calc) with a critical value (F_crit) from the F-distribution at a chosen level of significance, α (e.g., 0.05). The degrees of freedom for the critical value are

m

for the numerator and

(n – k)

for the denominator.

In this case, the degrees of freedom would be (2, 22).

The decision rule is:

• If F_calc > F_crit(α, m, n-k) , we reject the null hypothesis (H₀). This means that β₂ and β₃ are jointly different from zero, and the variables X₂ and X₃ are jointly significant in explaining Y.
• If F_calc ≤ F_crit(α, m, n-k) , we fail to reject H₀. We cannot conclude that X₂ and X₃ are jointly significant.

Q4. When do you encounter the problem of heteroscedasticity in data ? Explain how you would detect the presence of heteroscedasticity.

Ans. Introduction to Heteroscedasticity: Heteroscedasticity means “unequal variance.” In econometrics, it refers to a situation where the variance of the error term (uᵢ) in a regression model is not constant across all observations. This is a violation of the homoscedasticity assumption of the Classical Linear Regression Model (CLRM), which assumes that Var(uᵢ) = σ² (a constant). Under heteroscedasticity, Var(uᵢ) = σᵢ², meaning the variance changes with each observation i. When Heteroscedasticity is Encountered: The problem is particularly common in the following types of data and situations:

• Cross-sectional data: This is the most common setting for heteroscedasticity. For example, when analyzing data on household income and consumption, higher-income households have more discretionary income and choice in their spending, leading to greater variability in their consumption patterns compared to low-income households. Similarly, the profits of large firms might show more variability than those of small firms.
• Error-learning models: As time passes, people learn from their mistakes, and the magnitude of errors may decrease. For instance, in learning to type, the number and variance of errors tend to decrease over time.
• Presence of Outliers: Extreme values (outliers) in the data can cause large deviations in the residuals, increasing the error variance for those observations.
• Model Specification Error: If an important explanatory variable is omitted from the model or if the wrong functional form is used, its effect may be captured by the error term, causing the error variance to change systematically and appear as heteroscedasticity.

Detection of Heteroscedasticity:

There are several informal and formal methods to detect heteroscedasticity:

1. Graphical Method (Informal): This is one of the simplest methods. Plot the squared residuals (ûᵢ²) from the OLS regression against the predicted Y (Ŷᵢ) or against one of the explanatory variables (Xᵢ).
   • If the spread of the points is random and shows no obvious pattern, homoscedasticity is likely.
   • If a systematic pattern emerges—such as a cone shape where the spread increases (or decreases) as Ŷᵢ or Xᵢ increases—it is a strong indication of heteroscedasticity.
2. Park Test: This is a formal test that involves regressing the logarithm of the squared residuals on the logarithm of an explanatory variable: ln(ûᵢ²) = α + β ln(Xᵢ) + vᵢ If the slope coefficient β is statistically significant (using a t-test), it supports the presence of heteroscedasticity related to Xᵢ.
3. Glejser Test: Similar to the Park test, this regresses the absolute values of the residuals (|ûᵢ|) on some function of an explanatory variable Xᵢ, such as: |ûᵢ| = α + βXᵢ + vᵢ or |ûᵢ| = α + β√Xᵢ + vᵢ If β is statistically significant, the hypothesis of heteroscedasticity is supported.
4. White’s General Heteroscedasticity Test: This is a very general and popular test because it does not assume a specific form of heteroscedasticity. It runs an auxiliary regression of the squared residuals (ûᵢ²) on the original explanatory variables, their squares, and their cross-products. The overall significance of this auxiliary regression is evaluated via an F-test or an LM-test (the nR² statistic). If the test is significant, we reject the null hypothesis of homoscedasticity.

Q5. Explain why measurement error in the explanatory variable will lead to inconsistent estimators.

Ans. Measurement error in an explanatory variable is a serious problem in econometrics because it violates a key assumption of the Classical Linear Regression Model (CLRM), causing the OLS (Ordinary Least Squares) estimators to be biased and inconsistent . Inconsistency means that even as the sample size increases, the estimator does not converge to the true parameter value. Let’s demonstrate this with a simple regression model: True Model: Yᵢ = β₁ + β₂Xᵢ* + uᵢ Here, Yᵢ is the dependent variable, Xᵢ is the true, but unobservable, explanatory variable, and uᵢ is an error term satisfying the classical assumptions (notably, E(uᵢ)=0 and Cov(Xᵢ , uᵢ)=0). Introduction of Measurement Error: We cannot observe Xᵢ* directly. Instead, we observe a proxy variable, Xᵢ, which contains measurement error: Xᵢ = Xᵢ* + eᵢ Here, Xᵢ is the observed variable and eᵢ is the measurement error. We typically assume that eᵢ has a zero mean (E(eᵢ)=0), is uncorrelated with the true value Xᵢ (Cov(Xᵢ , eᵢ)=0), and is also uncorrelated with the model’s error term uᵢ (Cov(uᵢ, eᵢ)=0). Effect on the Estimated Model: When we run the regression using the observed data, we are actually estimating the following model: Yᵢ = β₁ + β₂Xᵢ + vᵢ To see what the error term vᵢ is, we substitute Xᵢ* = Xᵢ – eᵢ into the true model: Yᵢ = β₁ + β₂(Xᵢ – eᵢ) + uᵢ Yᵢ = β₁ + β₂Xᵢ – β₂eᵢ + uᵢ So, the composite error term in the estimated model is: vᵢ = uᵢ – β₂eᵢ The Cause of Inconsistency: A key condition for OLS estimators to be consistent is that the explanatory variable must be uncorrelated with the error term, i.e., Cov(Xᵢ, vᵢ) = 0. Let’s check this covariance: Cov(Xᵢ, vᵢ) = Cov(Xᵢ* + eᵢ, uᵢ – β₂eᵢ) Expanding this using the properties of covariance: Cov(Xᵢ, vᵢ) = Cov(Xᵢ , uᵢ) – β₂Cov(Xᵢ , eᵢ) + Cov(eᵢ, uᵢ) – β₂Cov(eᵢ, eᵢ) Based on our assumptions:

• Cov(Xᵢ*, uᵢ) = 0 (assumption of the true model)
• Cov(Xᵢ*, eᵢ) = 0 (measurement error is unrelated to true value)
• Cov(eᵢ, uᵢ) = 0 (measurement error is unrelated to regression error)
• Cov(eᵢ, eᵢ) = Var(eᵢ) = σₑ² (the variance of the measurement error)

Substituting these in, we get:

Cov(Xᵢ, vᵢ) = 0 – β₂(0) + 0 – β₂σₑ² =

-β₂σₑ²

Conclusion:

Since measurement error exists (σₑ² > 0) and we assume β₂ ≠ 0, the covariance

Cov(Xᵢ, vᵢ) ≠ 0

. The observed explanatory variable Xᵢ is correlated with the new error term vᵢ. This violates a fundamental assumption of OLS.

This correlation renders the OLS estimator β̂₂ biased and inconsistent. It can be shown that the probability limit of the estimator is:

plim(β̂₂) = β₂

[ Var(X

) / (Var(X

) + Var(e)) ] = β₂

[ σₓ

² / (σₓ

² + σₑ²) ]

Since the term in the brackets is less than 1, the OLS estimator will be biased towards zero. This is known as

attenuation bias

. The estimator underestimates the magnitude of the true coefficient, and this bias does not disappear even as the sample size grows, hence the estimator is inconsistent.

Q6. Explain the concept of Best Linear Unbiased Estimators (BLUE). Prove that OLS estimators are BLUE.

Ans. The Concept of Best Linear Unbiased Estimator (BLUE): BLUE is an acronym for an estimator that possesses three crucial properties: it is the Best , Linear , and Unbiased estimator. It is a central concept in econometrics for describing the properties of OLS (Ordinary Least Squares) estimators, as stated in the Gauss-Markov Theorem. Let’s break down each term:

• Linear: An estimator is linear if it is a linear function of the values of the dependent variable, Y. For example, the slope estimator β̂₂ in a simple regression can be written as β̂₂ = ΣkᵢYᵢ, where the kᵢ are weights that depend only on the values of the explanatory variable X.
• Unbiased: An estimator is unbiased if its expected value or average value (if we were to take repeated samples) is equal to the true population parameter. Mathematically, E(β̂) = β. This means that, on average, the estimator gives the correct value.
• Best: ‘Best’ means having the minimum variance. A BLUE estimator has the smallest variance in the class of all linear unbiased estimators. This means it is the most efficient and precise, as its values are less spread out around the true parameter value.

In short, a BLUE estimator is the most efficient one out of all estimators that are both linear and unbiased.

Proof that OLS Estimators are BLUE (Gauss-Markov Theorem):

The Gauss-Markov theorem states that under the assumptions of the CLRM (zero mean of errors, homoscedasticity, no autocorrelation, and zero covariance between explanatory variables and the error term), the OLS estimators are BLUE.

Let’s prove this for the slope coefficient β₂ in the simple linear regression model Yᵢ = β₁ + β₂Xᵢ + uᵢ.

The OLS estimator β̂₂ = Σxᵢyᵢ / Σxᵢ² = Σxᵢ(Yᵢ – Ȳ) / Σxᵢ². This can be rearranged in terms of Yᵢ:

β̂₂ = Σ(xᵢ/Σxᵢ²)Yᵢ = ΣkᵢYᵢ, where kᵢ = xᵢ/Σxᵢ².

1. Linearity:

Since β̂₂ can be written as ΣkᵢYᵢ, it is a linear function of the dependent variable Yᵢ. Hence, it is

Linear

.

2. Unbiasedness:

We substitute Yᵢ = β₁ + β₂Xᵢ + uᵢ into the formula for β̂₂:

β̂₂ = Σkᵢ(β₁ + β₂Xᵢ + uᵢ) = β₁Σkᵢ + β₂ΣkᵢXᵢ + Σkᵢuᵢ

Using properties of the weights kᵢ, it can be shown that Σkᵢ=0 and ΣkᵢXᵢ=1.

So, the equation becomes:

β̂₂ = β₂(1) + Σkᵢuᵢ = β₂ + Σkᵢuᵢ

Now, take the expected value:

E(β̂₂) = E(β₂ + Σkᵢuᵢ) = β₂ + E(Σkᵢuᵢ) = β₂ + ΣkᵢE(uᵢ)

Since E(uᵢ) = 0, we get E(β̂₂) = β₂.

Thus, the OLS estimator is

Unbiased

.

3. Best (Minimum Variance):

First, find the variance of β̂₂:

Var(β̂₂) = Var(β₂ + Σkᵢuᵢ) = Var(Σkᵢuᵢ) = Σkᵢ²Var(uᵢ)

Using CLRM assumptions (homoscedasticity, Var(uᵢ)=σ²):

Var(β̂₂) = σ²Σkᵢ² = σ²Σ(xᵢ/Σxᵢ²)² = σ² (Σxᵢ² / (Σxᵢ²)²) =

σ²/Σxᵢ²

Now, consider any other linear unbiased estimator of β₂, say β₂*.

β₂* = ΣcᵢYᵢ, where cᵢ are some other weights.

For β₂* to be unbiased, it must satisfy Σcᵢ=0 and ΣcᵢXᵢ=1.

The variance of β₂

is Var(β₂

) = Var(ΣcᵢYᵢ) = σ²Σcᵢ².

We want to show that Var(β₂*) ≥ Var(β̂₂).

Let cᵢ = kᵢ + dᵢ. Using the unbiasedness conditions, one can show that Σdᵢ=0 and Σdᵢxᵢ=0.

Var(β₂*) = σ²Σ(kᵢ + dᵢ)² = σ²(Σkᵢ² + Σdᵢ² + 2Σkᵢdᵢ)

The term Σkᵢdᵢ can be shown to be zero: Σkᵢdᵢ = Σ(xᵢ/Σxᵢ²)dᵢ = (1/Σxᵢ²)Σxᵢdᵢ = 0.

So, Var(β₂*) = σ²(Σkᵢ² + Σdᵢ²) = Var(β̂₂) + σ²Σdᵢ².

Since σ²Σdᵢ² ≥ 0, we have:

Var(β₂*) ≥ Var(β̂₂)

This means the variance of any other linear unbiased estimator is greater than or equal to the variance of the OLS estimator. Therefore, the OLS estimator is

Best

.

Since the OLS estimator is Linear, Unbiased, and Best, it is

BLUE

.

Q7. Consider the following Keynesian model of income determination : Consumption function : Cₜ = β₀ + β₁Yₜ + uₜ Income identity : Yₜ = Cₜ + Iₜ Assume that 0 < β₁ < 1. The error term follows the classical assumptions : E(uₜ) = 0, E(uₜ)² = σ², E(uₜ, uⱼ) = 0. Show that the model has an endogeneity problem.

Ans. Introduction to the Endogeneity Problem: In econometrics, an endogeneity problem arises when an explanatory variable in a regression model is correlated with the error term. This violates a key assumption of OLS (Ordinary Least Squares), which requires Cov(X, u) = 0. When endogeneity is present, OLS estimators become biased and inconsistent. In the given Keynesian model, we have two equations: 1. Consumption function: Cₜ = β₀ + β₁Yₜ + uₜ 2. Income identity: Yₜ = Cₜ + Iₜ In the consumption function, the explanatory variable is income (Yₜ). To show that the model has an endogeneity problem, we need to demonstrate that Yₜ is correlated with the error term uₜ, i.e., that Cov(Yₜ, uₜ) ≠ 0 . Demonstrating Endogeneity: We start by deriving a ‘reduced form’ equation for income (Yₜ), which expresses Yₜ as a function of only exogenous variables and the error term. In this model, we assume investment (Iₜ) is exogenous. 1. Substitution: Substitute the consumption function (equation 1) into the income identity (equation 2): Yₜ = (β₀ + β₁Yₜ + uₜ) + Iₜ 2. Solve for Yₜ: Move the terms involving Yₜ to one side of the equation: Yₜ – β₁Yₜ = β₀ + Iₜ + uₜ Yₜ(1 – β₁) = β₀ + Iₜ + uₜ 3. Obtain the Reduced Form: Divide by (1 – β₁) to isolate Yₜ: Yₜ = [ β₀ / (1 – β₁) ] + [ 1 / (1 – β₁) ]Iₜ + [ 1 / (1 – β₁) ]uₜ This equation expresses Yₜ in terms of the exogenous variable Iₜ and the structural error term uₜ. 4. Calculate the Covariance: Now, we can calculate the covariance between the explanatory variable Yₜ and the error term uₜ: Cov(Yₜ, uₜ) = Cov ( [ β₀ / (1 – β₁) ] + [ 1 / (1 – β₁) ]Iₜ + [ 1 / (1 – β₁) ]uₜ , uₜ ) Using the properties of covariance (covariance of a constant is zero, and we assume the exogenous investment is uncorrelated with the error term, i.e., Cov(Iₜ, uₜ)=0): Cov(Yₜ, uₜ) = Cov([β₀/(1-β₁)], uₜ) + Cov([(1/(1-β₁))Iₜ], uₜ) + Cov([(1/(1-β₁))uₜ], uₜ) Cov(Yₜ, uₜ) = 0 + (1 / (1 – β₁)) Cov(Iₜ, uₜ) + (1 / (1 – β₁)) Cov(uₜ, uₜ) Cov(Yₜ, uₜ) = 0 + 0 + (1 / (1 – β₁)) Var(uₜ) 5. Final Result: Since the variance of the error term is Var(uₜ) = σ² (which is greater than zero), we get: Cov(Yₜ, uₜ) = σ² / (1 – β₁) Given that 0 < β₁ < 1, the denominator (1 – β₁) is not zero. Also, σ² > 0. Therefore, Cov(Yₜ, uₜ) ≠ 0 . Conclusion: Because the explanatory variable income (Yₜ) is correlated with the error term (uₜ), there is an endogeneity problem in the consumption function. Income and consumption are determined simultaneously, making Yₜ an endogenous variable. A direct application of OLS to this equation will yield biased and inconsistent estimates of β₀ and β₁. To solve this problem, methods such as Indirect Least Squares or Two-Stage Least Squares would be required.

Q8. Suppose you have to build an econometric model for determinants of household consumption. What model selection criteria will you use ?

Ans. When building an econometric model for the determinants of household consumption, a combination of several criteria should be used to select the ‘best’ model. No single criterion is sufficient, and the selection process should involve both statistical rigor and economic theory. The key criteria I would use are as follows: 1. Theoretical Plausibility: This is the most important criterion. The model should be grounded in established economic theories, such as the Keynesian consumption function, the life-cycle hypothesis, or the permanent income hypothesis.

• Choice of variables: Theory suggests key determinants should include current disposable income, wealth, interest rates, and expectations of future income. Demographic factors like family size, age, and education are also relevant.
• Expected signs of coefficients: Theory guides the expected signs of the coefficients. For instance, income and wealth should have a positive effect on consumption (β_income > 0), while the effect of interest rates might be ambiguous (higher rates could encourage saving, reducing consumption).

2. Goodness-of-Fit:

This criterion measures how well the model fits the data.

• R-squared (R²): Indicates the proportion of variation in consumption that is explained by the model. A high R² is desirable, but it should not be the sole goal.
• Adjusted R-squared (Adjusted R²): This is superior to R² as it penalizes the inclusion of additional variables that do not significantly improve the explanatory power. When comparing different models, I would prefer the one with a higher adjusted R², provided other criteria are also met.

3. Statistical Significance of Coefficients:

• t-tests: Each individual coefficient in the model should be statistically significant (typically at the 5% level). An insignificant coefficient may suggest that the corresponding variable can be dropped from the model, unless theory strongly suggests its inclusion.
• F-test: The model as a whole must be significant, as measured by the F-test. This tests the joint hypothesis that all slope coefficients are zero.

4. Information Criteria:

When comparing non-nested models (i.e., one model is not a subset of the other), AIC and BIC are useful.

• Akaike Information Criterion (AIC)
• Bayesian Information Criterion (BIC)

Both criteria balance the model’s goodness-of-fit with the number of parameters (complexity). The model with the

lowest AIC or BIC value

is preferred. BIC tends to impose a larger penalty for more parameters than AIC, and thus it favors more parsimonious models.

5. Diagnostic Tests:

The final chosen model should be checked for violations of the CLRM assumptions.

• Heteroscedasticity: Should be checked using White’s test, as it is common in consumption data. If present, robust standard errors should be used.
• Multicollinearity: Check for high correlation between variables by calculating the VIF (Variance Inflation Factor).
• Specification Error: The RESET test can be used to check if there are omitted relevant variables or if the functional form is incorrect.

In summary, my approach would be to select a model that is theoretically coherent, has a high adjusted R², significant coefficients, and passes all relevant diagnostic tests. Among different specifications, I would favor the one with the lowest AIC/BIC.

Q9. Explain how dummy variable method can be used for testing of structural stability in an econometric model.

Ans. Structural stability refers to whether the coefficients of an econometric model are constant or stable across different subsets of the data. A structural break occurs when these coefficients change at a certain point in time or between different groups (e.g., male/female, pre-war/post-war). While the Chow test is a traditional way to test for structural stability, the dummy variable method provides a more flexible and informative alternative. This method not only tests if a break has occurred but can also indicate where the break is—in the intercept, the slope, or both. Steps for Testing using the Dummy Variable Method: Let’s assume we want to test the stability of a simple regression model: Yᵢ = β₁ + β₂Xᵢ + uᵢ And we suspect there might be a structural break between two periods (Period 1 and Period 2). Let there be n₁ observations in Period 1 and n₂ in Period 2. Step 1: Define a Dummy Variable We create a dummy variable, D, that distinguishes between the two periods: Dᵢ = 0 if observation i is from Period 1 Dᵢ = 1 if observation i is from Period 2 Step 2: Create an Unrestricted Model Instead of running separate regressions, we estimate a single, unrestricted model on the entire dataset (n = n₁ + n₂) that allows both the intercept and the slope to change. This is done by including the dummy variable as an intercept modifier and, multiplied with an explanatory variable, as a slope modifier (an interaction term): Yᵢ = α₁ + α₂Dᵢ + β₁Xᵢ + β₂(Dᵢ * Xᵢ) + vᵢ The coefficients in this model have the following interpretations:

• α₁: The intercept for Period 1.
• α₁ + α₂: The intercept for Period 2. ( α₂ is the differential intercept for Period 2).
• β₁: The slope for Period 1.
• β₁ + β₂: The slope for Period 2. ( β₂ is the differential slope for Period 2).

Step 3: Formulate the Hypothesis

The null hypothesis (H₀) of structural stability is that the coefficients are the same in both periods. This means the differential intercept and differential slope are both zero.

H₀: α₂ = 0 and β₂ = 0

The alternative hypothesis (H₁) is that at least one is not equal to zero, indicating a structural break.

Step 4: Test the Hypothesis

This is a joint hypothesis test on the coefficients of the dummy variable (Dᵢ) and the interaction term (Dᵢ * Xᵢ). We use an

F-test

:

1. Estimate the unrestricted model from Step 2 and get its Residual Sum of Squares (RSS_UR).
2. Estimate the restricted model, which is what results when H₀ is true (α₂ = 0, β₂ = 0). This is just the original pooled regression on the whole dataset: Yᵢ = α₁ + β₁Xᵢ + vᵢ. Get its Residual Sum of Squares (RSS_R).
3. Calculate the F-statistic: F = [ (RSS_R – RSS_UR) / m ] / [ RSS_UR / (n – k) ] where m = number of restrictions (2 here), n = total number of observations, and k = number of parameters in the unrestricted model (4 here).

Step 5: Conclusion

Compare the calculated F-value to a critical F-value with (m, n-k) degrees of freedom.

• If F_calculated > F_critical , we reject H₀. We conclude there is a statistically significant structural break.
• If F_calculated ≤ F_critical , we fail to reject H₀, suggesting the relationship is structurally stable.

An advantage of the dummy variable approach is that we can also perform individual t-tests on α₂ or β₂ to see if the break is only in the intercept, only in the slope, or in both.

Q10. Write short notes on any two of the following : (a) Dummy variable trap (b) RESET test (c) Test for normality of the error variable

Ans. (a) Dummy Variable Trap The dummy variable trap is a situation that arises in regression analysis when a dummy variable is included for every category of a qualitative variable, in addition to an intercept term in the model. This leads to perfect multicollinearity , making it impossible to compute the OLS estimators. Example: Suppose we have a qualitative variable ‘Gender’ with two categories: ‘Male’ and ‘Female’. If we create two dummy variables: D_male = 1 if the person is male, 0 otherwise D_female = 1 if the person is female, 0 otherwise And then attempt to estimate the model: Yᵢ = β₀ + β₁D_maleᵢ + β₂D_femaleᵢ + … + uᵢ The problem is that for every single observation, D_maleᵢ + D_femaleᵢ = 1 . This is the same as the regressor associated with the intercept term (β₀), which is a column of 1s for all observations. This exact linear relationship among the regressors causes perfect multicollinearity. As a result, the data matrix (X’X) becomes singular and cannot be inverted, which is necessary for calculating the OLS estimates. Solution: To avoid the trap, the rule is to always include one fewer dummy variable than the number of categories. If a qualitative variable has ‘m’ categories, include only ‘m-1’ dummy variables in the model. The omitted category becomes the base or reference category . The coefficients on the included dummy variables are then interpreted as the difference relative to this base category. (c) Test for Normality of the Error Variable A key assumption of the Classical Linear Regression Model (CLRM) is that the error terms (uᵢ) are normally distributed. This assumption is particularly crucial for the validity of hypothesis tests (t-tests and F-tests) in small samples. In large samples, due to the Central Limit Theorem, OLS estimators are approximately normally distributed even if the errors are not, provided other CLRM assumptions hold. The most widely used test for error normality is the Jarque-Bera (JB) test . Jarque-Bera (JB) Test: This test checks whether the residuals from an OLS regression have skewness and kurtosis consistent with a normal distribution.

• Null Hypothesis (H₀): The errors are normally distributed.
• In a normal distribution, Skewness (S) = 0 and Kurtosis (K) = 3.

Procedure:

1. Run the OLS regression and obtain the residuals (ûᵢ).

2. Calculate the sample skewness (S) and sample kurtosis (K) of the residuals.

3. The JB statistic is calculated using the formula:

JB = (n/6) * [ S² + ( (K-3)² / 4 ) ]

where ‘n’ is the sample size.

Interpretation:

Under the null hypothesis, the JB statistic has a chi-squared (χ²) distribution with 2 degrees of freedom.

• If the calculated JB statistic’s value is greater than the critical χ² value at a chosen significance level (e.g., 5%), or if the p-value of the JB statistic is less than 0.05, we reject the null hypothesis.
• Rejecting H₀ means there is evidence that the errors are not normally distributed.

If normality is rejected, in large samples we can still rely on the estimates, but in small samples, the results of hypothesis tests may be unreliable.