I am trying to simulate a selection model for a variable $Y$ dependent on covariate vector $X$, so that two groups/sub-sets $S=(0,1)$ of observations on $Y$ are created, which have a fixed difference in expectations (means) $\theta$ (i.e., selection bias) $$E(Y|S=1)-E(Y|S=0)= \theta$$

I am looking for solutions to approach this problem, where in principle I could also simulate two variables $Y_1$ and $Y_2$ and overlay these as a mixture distribution $Y$. However, note that the selection mechanism $S$ should be conditionally independent of $Y$ given $X$ (ignorable selection) $$P(S|Y,X) = P(S|X),$$ but not be marginally independent $$P(S|Y) \ne P(S).$$

Whereas it is easy to parameterize a model for $P(S|X)$, and also relate $E(Y|X)$, it seems very difficult to me to specify this model to achieve a fixed $\theta$ and keep variances of $Y$ across sub-sets approximately constant (the latter is not absolutely necessary, but desirable). If the mixture approach is chosen, I do not know how to overlay distributions in dependence of $X$. Thanks for suggestions!

  • $\begingroup$ From a mixture perspective, a model like $Y(x,s) = Y_1(x)\cdot\pi(s) + Y_2\cdot\left(1-\pi(s)\right)$ could easily be written down. But here the choice probabililty $\pi(s)$ does not marginally depend upon $x$. However if we would set $\pi(x,s)$ the model cannot be identfied. Hence some further restrictions need to be applied. A possibility would be to only look at the case where the impact of $x$ is the same for both groups, i.e., the same for $Y_1$ and $Y_2$. $\endgroup$
    – Druss2k
    Aug 4 '14 at 12:53

If I understood your question correctly, you want observations $(X_i, Y_i, treat_i)$ where $Y$ is different across treatment groups, but where, conditionally on $X$, $Y$ is independent of treatment. If that is the case, this R script works:

n.obs <- 1e3
prop.1 <- .5
n1 <- ceiling(prop.1 * n.obs)
sub1 <- sample(1:n.obs, n1)
treatment <- (1:n.obs) %in% sub1

x <- rnorm(n.obs, mean= -1)
x[sub1] <- rnorm(n1, mean= 1)
y <- 1 + 2 * x + rnorm(n.obs)

plot(x,y, col= 1 + treatment)
by(y, treatment, mean)
summary(glm(treatment~x+y, family = binomial(logit)))
summary(glm(treatment~y, family = binomial(logit)))

Here $X$ follows a mixture distribution: a different Gaussian in each treatment group (means $\bar x_1 \neq \bar x_2$), so that $P(S|X)\neq P(S)$. Then $Y$ is defined as an arbitrary (noisy) function $f$ of only $X$. Then we have the three equalities you wanted:

$E(Y|S=1)−E(Y|S=0)= E(f(X_1)) - E(f(X_2))= \theta$

$P(S|Y,X)=P(S|X)$ (because $X$ explains away $Y$)

$P(S|Y)≠P(S)$ (because distribution of $Y$ is different across treatment groups).


Maybe I am oversimplifying this but why don't you draw for the unselected group $Y_0 \sim N(\gamma,\sigma_0)$ and for the selected group $Y_1 \sim N(\gamma+\theta,\sigma_1)$ with

$$Y = \left\{ \begin{array}{ll} Y_0 & \mbox{if $S=0 $};\\ Y_1 & \mbox{if $S=1 $}.\end{array} \right.$$

You have $$E(Y|S=1)-E(Y|S=0)= \theta$$

Now either define $X$ as a function of selection $X=f(S)$ or as a scalar transformation $X=cY$. Either of these will give you ignorability $P(S|Y,X) = P(S|X)$ since knowing $X$ either makes knowing $Y$ uninformative about selection.

That said, I doubt either of these options is what you had in mind since it renders your selection model somewhat silly. Beyond these options I find it difficult to imagine an alternative schema in which $Y$ does not provide information with regards selection conditional upon $X$.

Perhaps there are some other transformations of $Y$ that will get you there. Perhaps something like $X$ being a continuous function defined as $Y$ is above and $Y$ being a function which contains some of the information from $X$ though not all of it such as $Y=Round(X)+c$. Thus knowing $X$ gives you all of the information you could have gained by knowing $Y$.


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.