Confounds alternative- additional variable Let's say that we would like to predict y from x, and we are aware that there are many confounds we wouldn't be able to deal with.
Would it be a reasonable idea to find another variable, z, which has less confounds with x, but may be predicted by x, and has less confounds with y, but may predict y.
Does this sounds reasonable? 
If yes- is this a common strategy, in cases of unknown confounds, or confounds which may not be eliminated for some reason?
Thanks!
 A: Confounders have to deal with causal inference not prediction. With prediction, the ends justify the means, so given $x$ and a list of what you call confound(er)s, I will simply pick whatever subset of them predicts $y$ accurately and reliably. In that case, all that matters is that those variables (possibly) predict $y$. 
Confounders are more subtle. See this paper for a discussion on the rationale for a causal declaration of variable(s) as possible confounders.
With causal inference, however, you specifically want to study the $x$ relationship to $y$ controlling for other variables.  If confounders are omitted from analyses, the usual estimates of effects are biased (bias from confounding).
Without directly measuring the confounding of interest there are two ways of improving inference. 
The first is an instrumental variables analysis. Identify a set of variables $Z$ which predict the exposure of interest $x$ and which are uncorrelated with the residual variance of the biased $x$ $y$ analysis. Then, predicted values of $X$ obtained from the unconfounded $Z$ are conditionally independent of the confounders. These models are estimated with two stage least squares.
The second approach is a similar application of structural equation modeling. Unlike the first case, you can identify a set of mediators $M$ that summarize the effect of $x$ on $y$ and calculate a mediated effect. Fit the model with $x$ to $m$ and to $y$ and again from $m$ to $y$ and calculate the pathwise coefficient and significance of the product effect $a$ from $x$ to $m$ and $b$ from $m$ to $y$. Note you partial out the direct effect of $x$ on $y$.  An example illustrating this problem: labatilol is a short acting treatment for blood pressure. Patients who receive labatilol in the inpatient setting may in fact experience a sensation of relief just to get a placebo believing it will treat their blood pressure, which in turns reduces the blood pressure. We could identify the bioavailability of labatilol by assaying the patients blood and treat the concentration as the mediator $m$ which predicts blood pressure controlling for the mere administration of the drug. This effect is consistent with what is called the method efficacy of drug administration.
