I am wondering about the effect of an event $Z$ on an outcome $Y$. I assume $Z$ does not affect $Y$ except through the channel of an intermediate outcome $D$, that is endogenous to $Y$.

To understand the effect of $Z$ on $D$, I run the following regression:

$$ D = X'\beta + \sum_{\tau \neq -1} \delta_\tau Z_\tau + \epsilon $$ Where $X$ is a vector of controls, and $Z_\tau$ is a set of dummy variables equal to 1 if the time period is $\tau$ periods since the event.

How can I use the event, $Z$, as an instrument for the intermediate outcome $D$, and show a period-by-period effect of $D$ on $Y$?

Are there papers with examples using something similar?


Without modeling the influence of $D$ on $Y$ you won't be able to identify the influence of $Z$ on $Y$.

For example, assuming linear relation $Y=\gamma_0+\gamma_1 D + \eta$ and that $\eta$ and $\epsilon$ are independent, then the effect of $Z$ on $Y$ is given by $\gamma_1 \cdot \delta_\tau$. The general rules of causal inference are provided in Pearl (2009). The example above is a simple instance of a linear structural equation model with a causal structure of chain.

Pearl, J. (2009). Causality: models, reasoning and inference. Second edition. Cambridge: MIT press.

  • $\begingroup$ So, given the linear relation you suggested, I could get an effect for each period ($\gamma_1 \cdot \delta_1$,$\gamma_1 \cdot \delta_2$,$\gamma_1 \cdot \delta_3$, $\gamma_1 \cdot \delta_4$, ...)? $\endgroup$ – gannawag Apr 12 '17 at 20:51
  • 1
    $\begingroup$ Yes, that's the idea. Your comment makes me however wonder, what the meaning of the state in which $\sum_{\tau \neq -1}Z_t =0$ is? Because the effect describes the change in $Y$ due to change from$Z_\tau=0$ to $Z_\tau=1$. $\endgroup$ – matus Apr 12 '17 at 21:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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