How to explain that showing X affects Y doesn't mean we can answer how Y affects Z? I'm working on a paper with a senior colleague who provided data. To make it general, we are studying whether something (measured by X) affects an outcome (measured by Y). Since unobserved factors may affect both X and Y, I argue that we can consistently estimate the effect of X on Y with an instrumental variable W for X.
My co-author's comment is, "OK, so you're saying that X increases Y. Good. But how does this increase in Y affect Z (another outcome)?"
My question is how do I explain to my co-author that we can't answer this question with the data we currently have? My understanding is based on the following two directed acyclic graphs:

The one on the left shows the idea behind using W as an instrumental variable for X, and the one on the right shows what I think the problem is: Y is a collider variable, so we shouldn't treat it as an independent variable. 
Without these graphs, I casually mentioned to my co-author that under his reasoning, Y is an endogenous variable, so OLS won't consistently estimate the effect of Y on Z, but I'm not sure if the point got across.
So to reiterate, how do I explain that showing X affects Y doesn't mean we can answer how Y affects Z?
 A: If you believe the graph you draw in the second case is correct (just including again the arrow $W \rightarrow X$ as in the first graph) and if you are assuming a linear model, actually the answer is yes, you can identify all the effects using W as an instrument both for X and Y.
According to your graph, W can be used an instrument for: (i) the direct effect of X on Y; (ii) the direct effect of Y and X on Z (conditioning on X) and (iii) the total effect of X on Z.
So all the effects of X and Y on Z are identified in the model below:

To make things clear, let's see a simulated example.
rm(list = ls())
set.seed(10)
n <- 1e6
w <- rnorm(n)
a <- rnorm(n)
b <- rnorm(n)
x <- 1*w + a + rnorm(n)
y <- 2*x + a + b + rnorm(n)
z <- 3*x + 4*y + b + rnorm(n)
df <- data.frame(w, x, y, z)

Here we have that the direct effect of $X$ on $Y$ is 2. The direct effect of $Y$ on $Z$ is 4. The direct effect of $X$ on $Z$ is 3. And the total effect of $X$ on $Z$ is $3 + 2\times4 = 11$. 
Let's estimate these with instrumental variables. I will use the package AER in R with the function ivreg. The syntax of the function is formula|instruments.
Direct effect of $X$ on $Y$:
library(AER)
ivreg(y ~ x | w, data = df)
Call:
ivreg(formula = y ~ x | w, data = df)

Coefficients:
(Intercept)            x  
   0.000973     1.996818  

Direct effect of $Y$ and $X$ on $Z$:
ivreg(z ~ y + x| x + w, data = df)
Call:
ivreg(formula = z ~ y + x | x + w, data = df)

Coefficients:
(Intercept)            y            x  
    0.00196      4.00469      2.98934  

Total effect of $X$ on $Z$:
ivreg(z ~ x | w, data = df)
Call:
ivreg(formula = z ~ x | w, data = df)

Coefficients:
(Intercept)            x  
   0.005856    10.985980  

Notice that for all the cases if you run a simple OLS you will get biased results.
