IV
Instituto de Matemática, Estatística e Computação Científica (IMECC),
Universidade Estadual de Campinas (UNICAMP).
Seja \(Y\) uma v.a e \(D\) uma v.a (ou vetor aleatório) com segundo momento finito tais que \[Y = D^T \beta + e,\] em que \(\mathbb{E}(D\epsilon) = 0.\)
O coeficiente populacional por MQO é dado por \[\beta = \arg \min_b \mathbb{E}(Y - D^T b)^2 = \mathbb{E}(DD^T)^{-1}\mathbb{E}(DY).\]
O estimador de MQO é dado por \[\hat{\beta} = \Big (\displaystyle \sum_{i = 1}^n D_i D_i^T \Big)^{-1} \sum_{i = 1}^n D_i Y_i\]
\[\begin{align} \hat{\beta} & = \Big (\displaystyle \sum_{i = 1}^n D_i D_i^T \Big)^{-1} \sum_{i = 1}^n D_i Y_i, \\ & = \Big (\displaystyle \sum_{i = 1}^n D_i D_i^T \Big)^{-1} \sum_{i = 1}^n D_i (D_i^T \beta + e_i), \\ & = \beta + \Big ( \displaystyle \sum_{i = 1}^n D_i D_i^T \Big)^{-1} \Big (\sum_{i = 1}^n D_i e_i\Big), \\ & \xrightarrow{p} \beta + \mathbb{E}(DD^T)^{-1} \underbrace{\mathbb{E}(D\epsilon)}_{0}, \\ & = \beta \end{align}\]
E se o modelo for da forma \[Y = D^T \beta + e, \quad \text{mas } \quad \mathbb{E}(D\epsilon) \neq 0?\]
Neste caso, teremos que
\[\hat{\beta} \xrightarrow{p} \beta + \mathbb{E}(DD^T)^{-1} \mathbb{E}(D\epsilon) \neq \beta\]
Isto muda completamente o modelo anterior e o estimador deixa de ter boas propriedades.
Quando \(D\) é endógena, \(\hat{\beta}\) é inconsistente para \(\beta\). Assim, se quisermos um estimador consistente precisamos utilizar informação adicional, surgindo assim o modelo de variáveis instrumentais (IV).
Consideremos o caso em que \(Z\) e \(D\) tem a mesma dimensão e \(\mathbb{E}(Z D^T)\) tem posto completo.
\[\begin{align} \mathbb{E}(Ze) & = \mathbb{E}(Z[Y - D^T\beta]), \\ & = \mathbb{E}(ZY) - \mathbb{E}(ZD^T)\beta, \\ \beta & = \mathbb{E}(ZD^T)^{-1}\mathbb{E}(ZY). \end{align}\]
O estimador MQO é dado por
\[\hat{\beta}_{IV} = \Big (\displaystyle \sum_{i = 1}^n Z_i D_i^T \Big)^{-1} \sum_{i = 1}^n Z_i Y_i\]
Neste modelo,
\[\mathbb{C}ov(Z, Y) = \beta \mathbb{C}ov(Z, D)\]
\[\beta = \dfrac{\mathbb{C}ov(Z, D)}{\mathbb{C}ov(Z, Y)} = \dfrac{\mathbb{C}ov(Z, D) / \mathbb{V}(Z)}{\mathbb{C}ov(Z, Y) / \mathbb{V}(Z)}\]
Que é equivalente ao coeficiente associado a \(Z\) na regressão por MQO de \(Y\) e \(D\) sob \(Z\).
\[\begin{align} \hat{\beta}_{MQ2E} & = \Big ( \displaystyle \sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big)^{-1} \sum_{i = 1}^n \hat{D}_i Y_i, \\ & = \Big ( \displaystyle \sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big)^{-1} \sum_{i = 1}^n \hat{D}_i (D_i^T \beta + \epsilon_i), \\ & = \Big ( \displaystyle \sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big)^{-1} \sum_{i = 1}^n \hat{D}_i D_i^T \beta + \Big ( \displaystyle \sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big)^{-1} \sum_{i = 1}^n \hat{D}_i \epsilon_i), \\ \end{align}\]
Da primeira regressão temos que \(D_i = \hat{D}_i + \tilde{D}_i\), tal que \[\displaystyle \sum_{i = 1}^n \hat{D}_i \tilde{D}_i^T = 0\]
Então, \[\sum_{i = 1}^n \hat{D}_i D_i^T = \sum_{i = 1}^n \hat{D}_i \hat{D}_i^T.\]
\[\hat{\beta}_{MQ2E} = \beta + \Big ( \displaystyle \sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big)^{-1} \sum_{i = 1}^n \hat{D}_i \epsilon_i\]
Da primeira regressão temos que \[\hat{D}_i = \hat{\Gamma}^T Z_i.\]
\[\begin{align} \hat{\beta}_{MQ2E} & = \beta + \Big (\hat{\Gamma}^T \big ( n^{-1}\displaystyle \sum_{i = 1}^n Z_i Z_i^T\big) \hat{\Gamma} \Big)^{-1} \hat{\Gamma}^T \underbrace{n^{-1}\sum_{i = 1}^n Z_i \epsilon_i}_{\xrightarrow{p} \mathbb{E}(Z\epsilon)} \xrightarrow{p} \beta \\ \end{align}\]
Cuja variância é dada por
\[\hat{V}_{MQ2E} = \Big ( \displaystyle \sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big )^{-1} \Big (\sum_{i = 1}^n \hat{\varepsilon}_i^2 \hat{D}_i \hat{D}_i^T \Big ) \Big (\sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big )^{-1},\] em que \(\hat{\varepsilon}_i = Y_i - \hat{\beta}_{MQ2E}D_i\)
Ou seja,
\[\hat{V}_{MQ2E} = \Big ( \displaystyle \sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big )^{-1} \Big (\sum_{i = 1}^n \hat{\varepsilon}_i^2 \hat{D}_i \hat{D}_i^T \Big ) \Big (\sum_{i = 1}^n \hat{D}_i \hat{D}_i^T \Big )^{-1},\]
é igual ao estimador EHW da variância mas substituindo \(\hat{\varepsilon}_i = Y_i - \hat{\beta}_{MQ2E}\hat{D}_i\) por \(\hat{\varepsilon}_i = Y_i - \hat{\beta}_{MQ2E}D_i\).
Consideremos o seguinte modelo:
\[\begin{align} Y_i & = \beta_0 + \beta_1 D_i + \beta_2^T X_i + \epsilon_i, \\ D_i & = \gamma_0 + \gamma_1 Z_1 + \gamma_2^T X_i + \varepsilon_i, \end{align}\] em que \(D\) é endógena, \(Z\) é a IV e \(X\) são outras variáveis (todas exógenas).
Este é um caso especial do visto anteriormente com:
Rows: 3,010
Columns: 36
$ X <int> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18…
$ id <int> 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 19, 2…
$ nearc2 <int> 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1…
$ nearc4 <int> 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1…
$ educ <int> 7, 12, 12, 11, 12, 12, 18, 14, 12, 12, 9, 12, 11, 11, 16, 14,…
$ age <int> 29, 27, 34, 27, 34, 26, 33, 29, 28, 29, 28, 26, 24, 30, 31, 2…
$ fatheduc <int> NA, 8, 14, 11, 8, 9, 14, 14, 12, 12, 11, 11, 11, 11, NA, 15, …
$ motheduc <int> NA, 8, 12, 12, 7, 12, 14, 14, 12, 12, 12, 6, 6, 6, 8, 12, 8, …
$ weight <int> 158413, 380166, 367470, 380166, 367470, 380166, 367470, 49663…
$ momdad14 <int> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 0, 1, 1, 1…
$ sinmom14 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0…
$ step14 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ reg661 <int> 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1…
$ reg662 <int> 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0…
$ reg663 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ reg664 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ reg665 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ reg666 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ reg667 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ reg668 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ reg669 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ south66 <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ black <int> 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0…
$ smsa <int> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1…
$ south <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1…
$ smsa66 <int> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1…
$ wage <int> 548, 481, 721, 250, 729, 500, 565, 608, 425, 515, 225, 400, 4…
$ enroll <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…
$ KWW <int> 15, 35, 42, 25, 34, 38, 41, 46, 32, 34, 29, 34, 22, 27, 43, 3…
$ IQ <int> NA, 93, 103, 88, 108, 85, 119, 108, 96, 97, 84, 89, 93, 74, 1…
$ married <int> 1, 1, 1, 1, 1, 1, 1, 1, 4, 1, 1, 1, 1, 1, 1, 6, 1, 1, 6, 1, 1…
$ libcrd14 <int> 0, 1, 1, 1, 0, 1, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 0, 0, 1…
$ exper <int> 16, 9, 16, 10, 16, 8, 9, 9, 10, 11, 13, 8, 7, 13, 9, 4, 16, 9…
$ lwage <dbl> 6.306275, 6.175867, 6.580639, 5.521461, 6.591674, 6.214608, 6…
$ expersq <int> 256, 81, 256, 100, 256, 64, 81, 81, 100, 121, 169, 64, 49, 16…
$ region <int> 661, 661, 661, 662, 662, 662, 662, 662, 662, 662, 669, 662, 6…
Carlos Trucíos (IMECC/UNICAMP) | ME920/MI628 - Inferência Causal | ctruciosm.github.io