So I have a binary model where $y_1^*$ is the latent unobserved variable and $y_1 \in \{0,1\}$ the observed. $y_2$ determines $y_1$ and $z_2$ is thus my instrument. So in short the model is. \begin{eqnarray} y_1^*&=& \delta_1 z_1 + \alpha_1 y_2 + u_1 \\ y_2 &=& \delta_{21} z_1 + \delta_{22}z_2 + v_2 = \textbf{z}\delta + v_2 \\ y_1 &=& \text{1}[y^*>0] \end{eqnarray} Since the error terms aren't independent but, \begin{eqnarray} \begin{pmatrix} u_1 \\ v_2 \end{pmatrix} \sim \mathcal{N} \left(\textbf{0} \; , \begin{bmatrix} 1 &\eta \\ \eta &\tau^2 \end{bmatrix} \right). \nonumber \end{eqnarray} I make use of an IV-probit model.

I'm having trouble deriving the likelihood function. I get that I can write one of the error terms as a linear function of the other so, \begin{eqnarray} u_{1} = \frac{\eta}{\tau^2}v_{2} + \xi, \qquad \text{where} \quad \xi \sim \mathcal{N}(0, 1-\eta^2). \end{eqnarray}
and that $\xi$ should be used in order to impose a normal CDF.

I've looked in the Stata manual (http://www.stata.com/manuals13/rivprobit.pdf) for IV-probit and they suggest using the definition of the conditional density \begin{eqnarray} f(y_1, y_2 \mid \textbf{z}) = f(y_1 \mid y_2, \textbf{z}) f(y_2 \mid \textbf{z}) \end{eqnarray}
in order to derive the likelihood function but I really don't use it (and yea I end up with the wrong result...). My attempt so far is,

\begin{eqnarray} \mathcal{L}(y_1) &=& \prod_{i=1}^n \Pr(y_1=0 \mid y_2, \textbf{z} )^{1-y_1} \Pr(y_1=1 \mid y_2, \textbf{z} )^{y_1} \nonumber \\ &=& \prod_{i=1}^n \Pr(y_1^* \leq 0)^{1-y_1} \Big(\Pr(y_1^* > 0) f(y_2 \mid \textbf{z}) \Big)^{y_1} \nonumber \\ \text{[standardizing]} &=& \prod_{i=1}^n \Pr \Big( \frac{\xi}{\sqrt{1-\eta^2}} \leq - \frac{\delta_1 z_1 + \alpha_1 y_2 + \frac{\eta}{\tau^2}(y_2 - \textbf{z})}{\sqrt{1-\eta^2}}\Big)^{1-y_1} \\ &\cdot& \Big(\Pr \Big( \frac{\xi}{\sqrt{1-\eta^2}} < \frac{\delta_1 z_1 + \alpha_1 y_2 + \frac{\eta}{\tau^2}(y_2 - \textbf{z})}{\sqrt{1-\eta^2}}\Big) f(y_2 \mid \textbf{z}) \Big)^{y_1} \nonumber \\ &=& [1-\Phi(w)]^{1-y_i} \left[ \Phi(w)f(y_2 \mid \textbf{x}) \right]^{y_1} \end{eqnarray} As I said, I haven't used the definition for the joint density function as stated above. Moreover, I end up with also $f(y_2 \mid \textbf{z})$ being raised to $y_1$ which appears to be wrong. Can somebody give me a hint on how to derive the correct (log-) likelihood function or where I went wrong?


1 Answer 1


Remember that for a bivariate normal variable $$\begin{pmatrix}X \\ Y\end{pmatrix}\sim\mathcal{N}\left(\begin{bmatrix}\mu_X\\\mu_Y\end{bmatrix}, \begin{bmatrix}\sigma_X^2 & \rho\sigma_X\sigma_Y\\\rho\sigma_X\sigma_Y & \sigma_Y^2\end{bmatrix}\right),$$ the conditional distribution of $Y$ given $X$ is $$Y\mid X \sim \mathcal{N}\left(\mu_Y+\rho\sigma_Y\frac{X-\mu_X}{\sigma_X},\sigma_Y\left[1-\rho^2\right]\right).$$

In the present case, we have \begin{align} u_1 \mid v_2 &\sim \mathcal{N}\left(0+\frac{\eta}{1\cdot\tau}\cdot1\frac{v_2-0}{\tau}, 1\cdot\left[1-\left(\frac{\eta}{1\cdot\tau}\right)^2\right] \right) \\ &= \mathcal{N}\left(\frac{\eta}{\tau^2}v_2, 1-\frac{\eta^2}{\tau^2} \right), \end{align} which means that $$u_1=\frac{\eta}{\tau^2}v_2+\xi$$ where (and this was your first mistake) $$\xi\sim\mathcal{N}\left(0,1-\frac{\eta^2}{\tau^2}\right).$$

We can thus rewrite the first equation \begin{align} y_1^* &= \delta_1 z_1 + \alpha_1 y_2 + u_1 \\ &= \delta_1 z_1 + \alpha_1 y_2 + \frac{\eta}{\tau^2}v_2+\xi \\ &= \delta_1 z_1 + \alpha_1 y_2 + \frac{\eta}{\tau^2}(y_2-\textbf{z}\delta)+\xi. \end{align}

Now, remember that the conditional probability density function of $X=x$ given $Y=y$ is $$f_{X}(x \mid y)=\frac{f_{XY}(x,y)}{f_{Y}(y)}.$$

In the present case, we have $$f_{1}(y_1 \mid y_2, \mathbf{z})=\frac{f_{12}(y_1,y_2 \mid \mathbf{z})}{f_{2}(y_2 \mid \mathbf{z})},$$ which can be rearranged to your expression $$f_{12}(y_1, y_2 \mid \mathbf{z})= f_{1}(y_1 \mid y_2, \mathbf{z})f_{2}(y_2 \mid \mathbf{z}).$$

Then, we can write the likelihood as a function of the densities of the two independent shocks $v_1,\xi_1$: \begin{align} \mathcal{L}(y_1,y_2\mid \mathbf{z}) &= \prod_i^n f_{1}(y_{1i} \mid y_{2i}, \mathbf{z}_i)f_{2}(y_{2i} \mid \mathbf{z}_i) \\ &= \prod_i^n \Pr\left(y_{1i}=1\right)^{y_{1i}}\Pr\left(y_{1i}=0\right)^{1-y_{1i}}f_{2}(y_{2i} \mid \mathbf{z}_i) \\ &= \prod_i^n \Pr\left(y_{1i}^*>0\right)^{y_{1i}}\Pr\left(y_{1i}^*\leq0\right)^{1-y_{1i}}f_{2}(y_{2i} \mid \mathbf{z}_i) \\ &= \prod_i^n \Pr\left(\delta_1 z_{1i} + \alpha_1 y_{2i} + \frac{\eta}{\tau^2}(y_{2i}-\textbf{z}_{i}\delta)+\xi_i>0\right)^{y_{1i}}\\ &\qquad\quad \Pr\left(\delta_1 z_{1i} + \alpha_1 y_{2i} + \frac{\eta}{\tau^2}(y_{2i}-\textbf{z}_i\delta)+\xi_i\leq0\right)^{1-y_{1i}}\\ &\qquad\quad f_{2}(y_{2i} \mid \mathbf{z}_i) \\ &= \prod_i^n \Pr\left(\xi_i>-\left[\delta_1 z_{1i} + \alpha_1 y_{2i} + \frac{\eta}{\tau^2}(y_{2i}-\textbf{z}_i\delta)\right]\right)^{y_{1i}}\\ &\qquad\quad \Pr\left(\xi_i\leq-\left[\delta_1 z_{1i} + \alpha_1 y_{2i} + \frac{\eta}{\tau^2}(y_{2i}-\textbf{z}_i\delta)\right]\right)^{1-y_{1i}}\\ &\qquad\quad f_{2}(y_{2i} \mid \mathbf{z}_i) \\ &= \prod_i^n \Pr\left(\frac{\xi_i-0}{\sqrt{1-\frac{\eta^2}{\tau^2}}}>-\frac{\delta_1 z_{1i} + \alpha_1 y_{2i} + \frac{\eta}{\tau^2}(y_{2i}-\textbf{z}_i\delta)+0}{\sqrt{1-\frac{\eta^2}{\tau^2}}}\right)^{y_{1i}}\\ &\qquad\quad \Pr\left(\frac{\xi_i-0}{\sqrt{1-\frac{\eta^2}{\tau^2}}}\leq-\frac{\delta_1 z_{1i} + \alpha_1 y_{2i} + \frac{\eta}{\tau^2}(y_{2i}-\textbf{z}_i\delta)+0}{\sqrt{1-\frac{\eta^2}{\tau^2}}}\right)^{1-y_{1i}}\\ &\qquad\quad f_{2}(y_{2i} \mid \mathbf{z}_i) \\ &= \prod_i^n \Pr\left(\frac{\xi_i}{\sqrt{1-\frac{\eta^2}{\tau^2}}}>-w_i\right)^{y_{1i}} \Pr\left(\frac{\xi_i}{\sqrt{1-\frac{\eta^2}{\tau^2}}}\leq-w_i\right)^{1-y_{1i}} f_{2}(y_{2i} \mid \mathbf{z}_i) \\ &= \prod_i^n \left[1-\Pr\left(\frac{\xi_i}{\sqrt{1-\frac{\eta^2}{\tau^2}}}\leq-w_i\right)\right]^{y_{1i}} \Pr\left(\frac{\xi_i}{\sqrt{1-\frac{\eta^2}{\tau^2}}}\leq-w_i\right)^{1-y_{1i}} f_{2}(y_{2i} \mid \mathbf{z}_i) \\ &= \prod_i \left[1-\Phi(-w_i)\right]^{y_{1i}} \Phi(-w_i)^{1-y_{1i}} \varphi\left(\frac{y_{2i}-\mathbf{z}_i\delta}{\tau}\right) \\ &= \prod_i^n \Phi(w_i)^{y_{1i}} \left[1-\Phi(w_i)\right]^{1-y_{1i}} \varphi\left(\frac{y_{2i}-\mathbf{z}_i\delta}{\tau}\right) \\ &= \Phi(w)^{y_{1}} \left[1-\Phi(w)\right]^{1-y_{1}} \varphi\left(\frac{y_{2}-\mathbf{z}\delta}{\tau}\right) \\ \end{align} where \begin{align} w_i = \frac{\delta_1 z_{1i} + \alpha_1 y_{2i} + \frac{\eta}{\tau^2}(y_{2i}-\textbf{z}_i\delta)}{\sqrt{1-\frac{\eta^2}{\tau^2}}}. \end{align} $\Phi(z)$ and $\varphi(z)$ are the cumulative density function and probability density function of the standard normal distribution.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.