Using PISA data for Ireland I show a kernel regression of the students (15 year olds) against class-size. The results show that as you lower class size from 40 to about 35 there is a higher score- as you might expect. After that however, its down-hill all the way: smaller classes are associated with worse scores. This is actually robust to inclusion of school effects and various other controls. In work in progress with a colleague it even seems to be robust to controling for endogeneity (using two distinct identification strategies).
4 comments:
Very interesting/exciting stuff. Look forward to hopefully seeing this in the future. What does the PDF of class size look like? Could the non-monotonic relationship in the graph here be driven by outliers?
1) Its very spiky, most of the density is at points like 25,30 etc. That may be "digital preference" i.e. reporting bias.
2) If it is, it would be the downward sloping part on the right I think.
Is it possible to do this for primary school classes Kevin? Is there any reason to believe that class size is more important (and beneficial with regard to reading scores) for 7 and 8 year-olds, compared to 15 and 16 year-olds?
I am also intrigued that reducing class-size below 35 is associated with worse scores. Without knowing anything about your identification strategies, could something else be at play? Are class-sizes lower than 35 more likely to be "lower-streamed" classes in "streamed" schools. I don't know anything about the extent of streaming in second-level, except that it happened in my school.
There is no publicly available data for primary schools that I am aware of. It might be possible to get at this with the Growing Up in Ireland data. It may well differ for primary schools.
Streaming is one possibility. Remedial classes also tend to be small I think. The graph was not intended (& I hope this was clear) as a statement about the causal relationship so while it doesn't prove that big classes are good it certainly does not prove the opposite.
Moreover much popular discussion in these matters relies either on no evidence at all or anecdotes or simple correlations. So I think this is an improvement on that and I think the onus is on others to show differently.
As I said, adding various controls doesn't really change the picture. Hopefully the paper (with Veruska Oppedisano) will be publicly available soon-ish.
Post a Comment