There has been empirical studies done in the USA that seem to support the finding you discuss. You might go to the MDRC.ORG website and the urban.org website to find more information. One explanation it that work helps to contextualize education. It also creates a demand for discipline that may be lacking for many young people.

In general, the differentail effect for different ability groups is not uncommon. I wrote a theory paper about (http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1142859): I argue that fraternities, who might not contribute directly to skills and abilities of participants, but select on ability, on average might show a job market premium for members. However, this benefit is likely to be distributed unevenly: best students might be even hurt by associating with the fraternity. Therefore, there's nothing surprising that a certain screening device (work, or even apply to work, in the town council: I am actually stupefied why would they randomize!) benefits those who have low grades, but does not benefit those who have high grades: those with high grades have already a convincing enough way to show off their worth to the employing company. 

In your study, I'd worry more about applications. Who would apply to work for the city council? Maybe more able students apply to work in investment banks for that summer,  or go study abroad, or have some other options; so able student with an experience in working for the town council raises brows of the employers. Referees might complain about endogeneity here. Also, it would be interesting to distinguish those who had an offer from those who actually worked: if people with an offer who did not worked for the council get the same wage premium as those who did worked, you get the story about self-assessment (especially if employment for the council is indeed random), but if the offer does not matter whereas actual working does, it is about experience obtained. You can see how this has different university policy implications.

John: Thanks I will check it out.

Sergey: I was also anticipating differential effects, it is the magnitude of it that is mind-boggling to me.

Overall, there was little difference between the applicants and non-applicants to the Council's offer as regard their incomes as adults. Recall that the girls were 15-16 years old at the time of application and would stand a slim chance of being offered a investment-banker position for that summer. As for offer vs working the offer was almost always accepted.

Thanks for your interesting comments.

I also would have anticipated that the more disadvantaged group may have experienced a slightly greater benefit compared to the more advantaged (based on grades). However, the size of the effect is surprisingly large. I wonder a bit about the randomization process. The fact that access to employment was randomized does not mean that all individual characteristics were controlled for in the randomization process.The low and high grade variable may not be best way to group girls. I might start by grouping them on the outcome variable (wages) and looking to find factors that differentiate between the groups.

I agree with the comments above.  I think that any positive signal of having this employment opportunity during high school will be disproportionately beneficial to students who otherwise do not have strong signals (the low GPA group).  However, I do think that there is some concerns about what "low GPA" is really telling us.  You demonstrate that GPA is correlated with applications, not surprising, so the question becomes: is there something about the low GPA students who did apply?  For example, they may be particularly motivated and driven, compared to the other low GPA students, but they are low GPA because of a lack of interest in formal schooling or because they are involved in other activities that lower their GPA (extracurricular activities).  In this case, it is not that all low-GPA students would benefit from working during school but some of them would benefit a lot (and they are the ones that did apply).

Curtis good points! Because there was no randomization process for selecting the sample, unobserved factors, such as self-selection bias, may be playing a powerful role in determining the observed findings.

@John and Curtis: Yes, the randomization of the girls was not a block design where the girls first were divided by grades and thereafter randomized, so the analysis I refer to is a post-stratification and therefore selection bias might be an explanation. Of course, we should remember that the intention never was to estimate differential effects (due to the design of this (natural) experiment) and the number of observations is modest and this form of fishing-expedition is likely to provide some spurious results. But I still wonder if that is the whole story.

As the comments above, I agree that self-selection is probably driving the results. I recommend to check my paper on Economics of Education Review where we find opposite selection results for boys and girls.

Self-selection is the driving factor behind the effect. However, ability-enhancing effect of work-while-in-school increases the returns to education and hence causes wage differentials. Thus, the effect differs as those with work skill enhances education potential. More to come.

@Pedro. It could be mentioned that we have checked boys in another report. There was no difference in work experience during high school between the offered and the non-offered and consequently it was not possible to estimate the causal effect of high school work on the ensuing incomes for boys.

The non finding for boys is very interesting. In USA research often find impacts on range of intervention aimed at promoting positive transition to adulthood  are less robust for boys compared to girls. I'm surprised to find that in your population there was no difference in work experience during high school for boys.

@John. The report that examined the boys is available by the attached link. It seemed that the boys worked more or less equally much in high school irresepctive of getting the random offer or not. Consequently, we did not try to further estimate the causal effect.

Data The effect of summer jobs on post-schooling incomes