Suppose that $S_i$ is continuously distributed, not necessarily non-negative. The conditional expectation function of interest is $h(t):=E[Y_i|S_i=t]$ has derivative $h'(t)$.

Equation 3.3.8 of Mostly Harmless Econometrics is:

$$\frac{E[Y_i(S_i- E[S_i])]}{E[S_i(S_i-E[S_i])]} = \frac{\int h'(t)\mu_t dt}{\int \mu_t dt} $$ where $\mu_t :=[E[S_i|S_i\ge t]-E[S_i |S_i<t]][P(S_i\ge t)[1-P(S_i\ge t)]]$ and the integrals run over the support of $S_i$.

That equation is not obviously true to me and I am looking for a proof.

  • 2
    $\begingroup$ You can find a discrete version of the derivation in Appendix A2 of Angrist and Krueger (1999): Empirical strategies in labor economics in the Handbook of Labor economics (Ashenfelter and Card) volume 3. link $\endgroup$
    – tdm
    Apr 21, 2022 at 12:33
  • $\begingroup$ The "+" sign should be "-" in the inner integral $\endgroup$
    – laozhang
    Oct 25, 2023 at 3:58

1 Answer 1


The appendix to that section in Mostly Harmless, section 3.5, has a derivation.

$$Cov(Y_i,S_i) = E[h(S_i)(S_i-E[S_i])]$$

Let $k_{-\infty}=\lim_{t\rightarrow -\infty} h(t)$. By the fundamental theorem of calculus,

$$h(S_i)=k_{-\infty} +\int_{-\infty}^{S_i} h'(t)dt $$

Thus, $$ E[h(S_i)(S_i-E[S_i])] = \int_{-\infty}^\infty\int_{-\infty}^{S_i} h'(t) (s-E[S_i])g(s)dtds$$

where $g(s)$ is the density of $S_i$ at $s$. Apply Fubini's theorem to switch the order of integration, $$ E[h(S_i)(S_i-E[S_i])] = \int_{-\infty}^\infty h'(t)\int_{t}^{\infty} (s-E[S_i])g(s)dtds$$

The inner integral is $E[S_i|S_i \ge t]Pr(S_i\ge t)+E[S_i]Pr(S_i \ge t) $

$= E[S_i|S_i \ge t]Pr(S_i\ge t)+(E[S_i|S_i \ge t]Pr(S_i \ge t) + E[S_i|S_i < t]Pr(S_i < t))Pr(S_i \ge t) $

$=\mu_t :=(E[S_i |S_i \ge t]-E[S_i |S_i <t])(Pr(S_i \ge t)(1-Pr(S_i \ge t)$

Then, setting $S_i =Y_i$, the denominator can similarly be derived.


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.