From this online document, it is assumed that we have a treatment and control group, with $N_1$ and $N_0$ being the number of treated and control units, respectively. $D$ is the treatment variable while $Y^1$ and $Y^0$ are the treated and control potential outcomes. Then, they define a "Native" Average Treatment Effect and decompose it into the ATE, Selection Bias, and Differential Effect Bias.

\begin{align} & E\left[\frac{1}{N_1} \sum_{i:d_i=1}{Y_i} - \frac{1}{N_0} \sum_{i:d_i=0}{Y_i}\right] \\ & \\ & = \underbrace{E[Y^1 | D=1] - E[Y^0 | D=0]}_{\text{NATE}} \\ &= \underbrace{E[Y^1 - Y^0]}_{\text{ATE}} + \underbrace{E[Y^0 | D=1] - E[Y^0 | D=0]}_{\text{Selection bias}} + \underbrace{E[D]*(E[\Delta | D=1] - E[\Delta | D=0])}_{\text{Differential effect bias}} \end{align}

I am uncertain if the third line holds without assumptions, as I cannot figure out how it is derived. Are there certain assumptions invoked above?


1 Answer 1


I believe that there is a small mistake in the third line. $E[D]$ should be $1-E[D]$, apart from that we only really need counterfactual consistency. Starting from the ATE

\begin{align} & E[Y^1-Y^0]=E[\Delta] \\ & \\ & = E[\Delta|D=1]P(D=1)+E[\Delta|D=0]P(D=0)\\ &\\ &= E[\Delta|D=1]\left(1-P(D=0)\right)+E[\Delta|D=0]P(D=0)\\ &\\ &=\left(E[\Delta|D=0]-E[\Delta|D=1]\right)P(D=0)+E[\Delta|D=1]\\ &\\ &=\left(E[\Delta|D=0]-E[\Delta|D=1]\right)P(D=0)+E[Y^1|D=1]-E[Y^0|D=1]\\ &\\ &=\left(E[\Delta|D=0]-E[\Delta|D=1]\right)P(D=0)+E[Y^1|D=1]-E[Y^0|D=1]\pm E[Y^0|D=0]\\ &\\ &=\left(E[\Delta|D=0]-E[\Delta|D=1]\right)P(D=0)+(\underbrace{E[Y^1|D=1]-E[Y^0|D=0]}_{\text{NATE}})- \underbrace{(E[Y^0|D=1] -E[Y^0|D=0])}_{\text{Selection Bias}}\\ \end{align}

Where all we used above was counterfactual consistency. Remember that since D is binary $P(D=1)=E[D]$ and take things to the left side and you obtain: \begin{align} &\underbrace{E[Y^1|D=1]-E[Y^0|D=0]}_{\text{NATE}}\\ &\\ &=\underbrace{E[Y^1-Y^0]}_{\text{ATE}}+\underbrace{(E[Y^0|D=1] -E[Y^0|D=0])}_{\text{Selection Bias}}+\left(E[\Delta|D=1]-E[\Delta|D=0]\right)P(D=0)\\ &\\ &=\underbrace{E[Y^1-Y^0]}_{\text{ATE}}+\underbrace{(E[Y^0|D=1] -E[Y^0|D=0])}_{\text{Selection Bias}}+\left(E[\Delta|D=1]-E[\Delta|D=0]\right)(1-E[D]) \\ \end{align}

  • $\begingroup$ What is the meaning of the selection bias here? I understand the bias arises if the treated and untreated group differ systematically in ways that affect their outcomes, even without treatment. But what is being "selected"? thanks. $\endgroup$
    – user321627
    Commented May 10 at 3:23

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.