Tutorial 4

CREST - Institut Polytechnique de Paris - IPP

March 11, 2026

Roadmap

In this tutorial, we are going to

  1. Do a recap on failures of the Gauss-Markov theorem assumptions
  2. Learn how to implement an instrumental variable strategy in R
  3. Replicate the key results from Acemoglu, Johnson & Robinson (2001)

Recap on OLS failures

Gauss-Markov

Under the following assumptions:

    1. Linearity of the “true” model \(y = x'\beta + \epsilon\)
    1. Observed sample \((y_i,x_i)_{i= 1...n}\) is a random sample with the same distribution as \((y,x)\)
    1. No perfect collinearity between the covariates in the sample (and in the population)
    1. Zero conditional mean \(E(\epsilon|x) =0\)
    1. Homoskedasticity \(Var(\epsilon|x_1,...,x_n) = \sigma^2\)

The OLS estimator is BLUE: best linear unbiased estimator

  • Linear estimator: \(\hat{\beta}_j = \sum\limits_1^n w_{ij} y_i\), \(w_{ij}\) functions of \(x_1,...x_n\)
  • Unbiased: \(E(\hat{\beta}|x_1,...,x_n) = \beta\)
  • “Best”: smallest variance of all linear unbiased estimators, conditional on observed \(x_1,...x_n\)

Endogeneity: measurement error

  • True model: \(y_i = \alpha + \beta x_i + \epsilon_i\)
  • Observed variable: \(\tilde{x}_i = x_i + \mu_i\)
  • Estimated model: \(y_i = \alpha + \beta \tilde{x}_i + \nu_i, \quad \nu_i = \epsilon_i - \beta \mu_i\)
  • Further assumptions: \(E(\mu) = 0\), \(Cov(\mu,x) = 0\), \(Cov(\mu,\epsilon) = 0\)
  • Estimated coefficient: \(E(\hat{\beta}|\tilde{x}) = \frac{Cov(\tilde{x},y)}{Var(\tilde{x})} = \frac{Cov(x+\mu,\alpha + \beta x + \epsilon)}{Var(\tilde{x})} = \beta \frac{Var(x)}{Var(\tilde{x})}\)
  • Key insight: \(\tilde{x}\) and \(\nu\) are correlated by construction, so the exogeneity assumption is violated
  • Sign of the bias: attenuation bias, \(\hat{\beta}\) is biased towards zero

Endogeneity: reverse causality / simultaneity

  • Demand model of interest: \(y_{di} = \alpha_d + \beta_d p_i + \epsilon_i\)
  • Simultaneous supply equation: \(y_{si} = \alpha_s + \beta_s p_i + \mu_i\)
  • We observe equilibrium price: \(p_i = \frac{\alpha_d - \alpha_s + \epsilon_i - \mu_i}{\beta_s - \beta_d}\) which is a priori correlated with \(\epsilon_i\)
  • Estimated coefficient: \[E( \hat{\beta_d} | P) = \frac{Cov(p,y)}{Var(p)} = \beta_d + \frac{Cov( \frac{ \epsilon - \mu}{\beta_s - \beta_d} , \epsilon )}{Var(p)} \]
  • Key insight: An increase in \(P\) decreases \(Y\) through the demand equation, but an increase in \(Y\) increases \(P\) through the supply equation
  • Sign of the bias: Typically, \(\beta_d <0 < \beta_s\) so the bias is positive in that case

Omitted variable bias

  • True model: \(Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + \epsilon\), \(E[\epsilon | X_1, X_2] = 0\).
  • The estimated model is: \(Y = \alpha_0 + \alpha_1 X_1 + u\), where \(u = \beta_2 X_2 + \epsilon\) and \(X_2\) can be expressed as \(X_2 = \gamma_0 + \gamma_1 X_1 + v\) with \(E[v | X_1] = 0\)
  • Estimated coefficient: \(E(\hat{\alpha}_1|X)= \frac{Cov(\beta_0 + \beta_1 X_1 + \beta_2 X_2 + \epsilon, X_1)}{Var(X_1)} = \beta_1 + \beta_2 \gamma_1\)
  • Key insight: \(X_1\) is correlated with the error term of the estimated model: \(u = \beta_2 \gamma_0 + \beta_2 \gamma_1 X_1 + \beta_2 v + \epsilon\) as long as \(\beta_2 \gamma_1 \neq 0\)
  • Sign of the bias depends on the sign of \(\beta_2 \gamma_1\)

Instrumental variable in R

Recap on the strategy

  • True model: \(y_i = \alpha + \beta x_i + \epsilon_i\)
  • OLS estimator is biased because \(E(\epsilon|x) \neq 0\)
  • Consider an instrument \(z\) such that
    • \(E(\epsilon|z) = 0\) (exogeneity)
    • \(Cov(z,x) \neq 0\) (relevance)
  • First stage: \(x_i = \pi_0 + \pi_1 z_i + v_i\)
  • Predicted values: \(\hat{x}_i = \hat{\pi}_0 + \hat{\pi}_1 z_i\)
  • Second stage: \[y_i = \alpha + \beta x_i + \epsilon_i = \alpha + \beta \hat{x}_i + \epsilon_i + \beta \hat{v}_i = \alpha + \beta \hat{x}_i + \mu_i \]
  • Estimated coefficient: \(E(\hat{\beta}|\hat{x}) = \frac{Cov(\hat{x},y)}{Var(\hat{x})} = \beta + \frac{Cov(\hat{x},\mu)}{Var(\hat{x})}\), and \(Cov(\hat{x},\mu) = 0\) by construction

the ivreg function

The function ivreg from the package AER allows to estimate an instrumental variable regression in one step. The syntax is as follows:

ivreg(y ~ x1 + x2 | z1 + z2, data = dat)

where y is the dependent variable, x1 and x2 are the true and (potentially) endogenous regressors, and z1 and z2 are the instruments (including exogenous regressors in x).

The function will automatically perform the two stages of the regression and return the estimated coefficients with the appropriate standard errors

Replication of Acemoglu et al (2001)

Summary of the paper

  • Daron Acemoglu, Simon Johnson and James Robinson received the 2024 Nobel Prize for their work on understanding the differences in prosperity between nations
  • Their key contribution is to study the role of institutions in economic development, and to show that good institutions are a key driver of economic growth. You can find their Nobel Prize lecture here
  • Institutions include formal rules (e.g. property rights, rule of law) and informal constraints (e.g. social norms, culture)
  • In 2001, they published a very influential paper: “The Colonial Origins of Comparative Development” in the American Economic Review
  • They show that the heterogeneous way former European colonizers set up institutions in their colonies explains a large share of the observed disparity in their current economic performance

OLS regression

In a first step, the authors study the relationship between economic development (log GDP per capita 1995) and institution strength (expropriation risk in 1985-1995) using ordinary least square on a set of 64 former colonies. \[ \log y_i = \mu + \alpha R_i + \mathbf{X}_i'\gamma + \epsilon_i \]

    1. Discuss the Gauss-Markov assumptions in that context
    1. Download the dataset and describe the data
    1. Replication of Figure 2: create a scatter plot of log GDP per capita in 1995 against average expropriation risk in 1985-1995
    1. The result of the main OLS specification is given in columns 2, 5 and 6 of Table 2. Interpret the results
    1. Replicate these three regressions and export the results into a latex table using stargazer

IV strategy

To address endogeneity of institution levels, the authors instrument it by the mortality level of settlers at the start of colonization. The full model they have in mind is:

\[log(y_i) = \mu + \alpha R_i + \mathbf{X}_i'\gamma + \epsilon_i\] \[R_i = \lambda_R + \beta_R C_i + \mathbf{X}_i'\gamma_R + \nu_{Ri}\]

\[C_i = \lambda_C + \beta_C S_i + \mathbf{X}_i'\gamma_C + \nu_{Ci}\]

\[S_i = \lambda_S + \beta_S \log(M_i) + \mathbf{X}_i'\gamma_S + \nu_{Si}\]

where \(y_i\) is the GDP per capita in 1995, \(R_i\) is the average expropriation risk in 1985-1995, \(C_i\) is a a measure of early institutional development, \(S_i\) is a measure of European settlement, and \(M_i\) is the mortality rate of settlers.

    1. The instrument for \(R\) chosen is \(\log(M)\). Why not \(C\) or \(S\)?
    1. What are the assumptions needed for \(\log(M)\) to be a valid instrument?
    1. Replicate Figure 3. Do you think settler mortality is a valid instrument?
    1. Table 4 panel B (column 2) presents more formally this first stage. How do you interpret it?

The result of the full IV is presented in panel A (column 2).

    1. The authors state: “measurement error is likely to be more important than reverse causality and omitted variable bias”. Do you agree with this statement? Why?
    1. Interpret the coefficient in terms of causal effect.
    1. Replicate the first and second stage regressions using the lm and predict functions
    1. Run the full IV regression using ivreg. Compare the results