Differential baseline bias vs Heterogenous treatment effect Where it says 'Differential baseline bias only', I see both a differential baseline bias and a differential treatment effect bias. From my understanding, to have no differential treatment effect bias, $y_i^{1}$ -  $y_i^{0}$ has to be the same for both the treatment group and the control group. However it is (20-10 = 10) and (20-0 = 20) in the first panel.
Also please refer to the comment on the first answer by Dimitry.

Stephen L. Morgan and Christopher Winship ”Counterfactuals and Causal Inference. Methods and Principles for Social Research”. Cambridge University Press.
 A: In case anyone is paying attention, or comes across this page, the resolution is simple:  it is a mistake!  
As explained on the Errata page for the second edition here:
http://socweb.soc.jhu.edu/faculty/morgan/papers/Errata_2nd_Edition.pdf
the first cell of the second row should be 10.  The other cells then change in that panel in response to this change, which results in an example that actually follows what is claimed in the panel label:  an example where bias is restricted only to baseline bias.  
Please do not hesitate to contact me directly if you find other things that appear to you to be mistakes.  You may be correct!  (In this case, thanks to Dave Harding and his students for noticing this one.)
A: I also found this baseline vs. differential treatments effect distinction confusing.
The $v$s are the potential outcomes centered around their population-level expectations. We are worried about a relationship between $d$ (being selected for treatment) and the $v$s since they are in the error term. When $v$s are large in absolute value, the individual is fairly different in terms of his potential outcome. When the $y$s are the same for both individuals, the $v$s for that state of the world will be zero.  
The OLS regression error term is equal to $v^0$ for those in the control and $v^1$ for those in the treatment. In the first panel, there's no relationship between $d$ and $v^1$ and a positive relationship between $d$ and $v^0$, which is the baseline, no-treatment potential outcome. Hence this is called baseline bias. People who have a lower outcome in the no-treatment world are more likely to be in the control group.
In the second panel, you have a relationship between $d$ and $v^1$, which is the treated potential outcome. That is the differential treatment bias: people who really benefit from treatment are likely to receive it. In the third panel, you have both kinds of biases simultaneously. 
These are different from heterogeneous treatment effects (people respond differently to the treatment), which you have in all 3 panels. Here are two examples that may make the distinction clearer. 
Suppose you have a job training program for 2 bicycle messengers that was given on a first-come, first-served basis with limited availability: only one training slot was available. Once it is announced, the fast biker shows up first since he can pedal quickly (at 10 vs 0). But the program does not really help him more (relative to the slow biker): everyone treated cycles at 20.
Contrast that with a job training program for 2 math students (aka math camp). Only the student who enjoys math signs up because she is going to learn a lot. For her the program has a huge payoff in terms of getting better at taking derivatives. After the camp, she can differentiate at 20. The student who didn't sign up would spend the class time napping and he would not learn all that much (only up to 15). If the cost of the program was foregone income from a job selling hot dogs on the boardwalk, the student with the higher return would sing up, while the other may not. Without the program, everyone would have the same level of math skills (10), so the baseline is the same.
You can have selection into treatment based on two types of potential outcomes: the first is the untreated outcome, the second is the treated outcome. In either case, however, if you were to pluck a random biker or student from a population to train, you would get a pretty different effect from the naively estimated one.  
