The Effect of Extra Funding for Disadvantaged Pupils on Achievement

Computer-subsidy discontinuity samples: DS+/-5 = 150 schools and 3,954 pupils; DS+/-10 = 328 schools and 8,263 pupils (2002 figures), against a national population of 5,938 schools and 150,821 pupils. Estimation pools four cohorts of end-of-primary test-takers - 1999 and 2000 (pre) against 2002 and 2003 (post). Of 7,028 primary schools in 1999, 564 (8 percent) were eligible for the computer subsidy and 551 received it.

Dutch primary-school pupils in their final (8th) year, roughly age 12, in schools where at least 70 percent of pupils came from a disadvantaged group (ethnic minority background or low-educated parents) as counted on 1 October 1999 - compared with schools just below that threshold. The discontinuity samples are considerably more disadvantaged and more urban than the national average and score about 0.3-0.45 SD below the national mean.

THE INTERVENTION: a one-time Dutch government payment of about $90 per pupil - roughly $1,250 per classroom - paid in December 2000 and EARMARKED for computers, software and language materials. Eligibility: at least 70 percent of pupils from any disadvantaged group. (The same paper evaluates a second, separate scheme paying about $2,225 + $585 + $2,440 per teacher to schools with at least 70 percent ETHNIC MINORITY pupils; the two samples barely overlap, because within the discontinuity windows almost every personnel-subsidy school also qualifies for the computer subsidy but almost no computer-subsidy school qualifies for the personnel subsidy.) FUNDING AND INDEPENDENCE: academic authors at the University of Amsterdam, Stockholm University, Tinbergen Institute and CPB Netherlands Bureau for Economic Policy Analysis; the Dutch Ministry of Education and CITO supplied the data. No vendor involvement of any kind - `independence: independent`. IDENTIFICATION, and it is strong: a SHARP regression discontinuity at the 70 percent threshold, exploited inside a local DIFFERENCE-IN-DIFFERENCES framework rather than as a bare RD, to gain precision. The estimating equation is a fixed-effects regression of pupil test score on (eligibility x post), with school fixed effects, a school-cohort (class) random effect and year dummies, run on "discontinuity samples" - all eligible schools within x percentage points above the cutoff and all ineligible schools within x points below (DS+/-5 and DS+/-10). Standard errors are clustered at the school-by-year (class) level and heteroscedasticity-robust. MANIPULATION OF THE RUNNING VARIABLE IS RULED OUT: schools would have had to anticipate the scheme a year in advance, and the plotted density of schools around the 70 percent cutoff shows no bunching. IMPERFECT COMPLIANCE is small and handled two ways - 551 of 564 eligible schools took the money and 16 ineligible schools received it by error; the DiD absorbs time-invariant school effects, and a 2SLS specification instrumenting receipt with eligibility gave very similar point estimates that could not be rejected as equal. TIMING: the pre-intervention tests (February 1999 and February 2000) both precede the December 2000 payment; 2001 is dropped as ambiguous; the outcome tests are February 2002 and February 2003, one and two years after payment. THE OUTCOME IS UNUSUALLY GOOD: the nationwide CITO end-of-primary test, taken by 8th graders in over 80 percent of Dutch primary schools, more than 200 multiple-choice items over three days, scored in language, arithmetic and information processing (world orientation optional). It is genuinely high stakes - the score determines placement into secondary-school tracks with strict admission thresholds, school averages are published and used by parents to choose schools, and the national press covers it every year. So this is not a low-stakes measure that pupils or schools could plausibly be indifferent to. Scores are standardised to population mean 0 and SD 1. WHAT THE CONTROL SCHOOLS DID WITH THE SAME TIME, AND WHAT THE MONEY ACTUALLY BOUGHT: this is a resources study, not a curriculum study - control schools simply carried on. The critical context is that BOTH groups already had ample hardware, roughly one computer per five pupils, against a government target of one per ten. A questionnaire to 171 DS+/-5 schools (153 responses, ~90 percent response rate; 63 eligible, 90 not) shows the subsidy bought NO extra hardware: the computer-pupil ratio difference is +0.017 (SE 0.016) raw and -0.018 (SE 0.033) controlling for the disadvantage share, and computer age is likewise unaffected. What it did buy was USE: pupils in eligible schools spent just over FIFTY MINUTES MORE PER WEEK at a school computer (+0.851 hours, SE 0.418, significant at 5 percent), of which about 20 minutes went to language and 10 to arithmetic (both imprecise). The authors infer the money went on software and internet connections. So the treatment that produced the negative estimates is: an hour a week more of computer-based work, using software the extra money bought, displacing whatever else those minutes would have held. GRADE B, squarely - the methodology's "regression discontinuity / difference-in-differences with clean shocks" row, with a verified no-manipulation check, a documented first stage on actual usage, and a compliance-robustness check. It is not A because the identification is not randomised, the estimates are local to schools at the 70 percent threshold, and the significant negatives appear only at the wider DS+/-10 bandwidth where the common-trend assumption is more strained. VERSION READ AND WHAT IS AND IS NOT VERIFIED: every numeric estimate below was checked against the full text of IZA Discussion Paper 1122 (April 2004), the openly available version. The published REStat 2007 abstract was verified verbatim and matches in substance. THE GIRLS RESULT IS NOT IN THE 2004 WORKING PAPER - it was added for publication, and is recorded here ONLY from the published abstract's exact wording ("Extra funding for computers and software seems especially detrimental for girls' achievement"). No point estimate, standard error or sample size for the gender split has been verified against the published tables, and none is claimed.

A regression discontinuity on Dutch school funding earmarked for computers and software finds NEGATIVE point estimates on every outcome, in every year, at every bandwidth - and significant ones on language. Pooling 2002 and 2003 at the wider DS+/-10 bandwidth, language is -0.079 SD (SE 0.030, significant at 1 percent) and arithmetic -0.061 SD (SE 0.033, significant at 10 percent); information processing is -0.032 SD (SE 0.030), not significant. At the tighter DS+/-5 bandwidth the point estimates are the same sign but imprecise (language -0.039, arithmetic -0.072, information -0.011). The arithmetic effect is MORE negative in 2003 than 2002, which the authors read as evidence the harm is not a short-run transition cost. The confidence intervals rule out anything but tiny positive effects: the DS+/-5 arithmetic estimate excludes positive effects larger than 0.02 SD, and the DS+/-10 information-processing estimate excludes positives larger than 0.03 SD. The mechanism evidence is what makes this bite: the money did not buy hardware - both groups already had a computer for every five pupils, better than the national target - it bought software and internet, and pupils in funded schools spent just over 50 minutes more per week on a computer. That is the treatment, and it moved achievement slightly the wrong way. The companion personnel subsidy in the same paper was also negative in point estimate and mostly insignificant, so this is not a general "money does not matter" result being read into the computer arm - both arms of a well-funded compensatory system hit the same wall. The published version additionally reports that the computer funding was "especially detrimental for girls' achievement" (see design_notes for the verification status of that claim).

Low. Not randomised, but identification comes from a sharp administrative threshold with verified absence of manipulation, combined with school fixed effects and pre-intervention cohorts - pupil composition on either side of the 70 percent cutoff is near-identical by construction and any time-invariant compositional difference is differenced out.

The Dutch computer-subsidy discontinuity has not itself been re-estimated by another team. The class of finding it belongs to - large-scale public funding earmarked for school computers has no positive effect on achievement - has been tested repeatedly with a split record: nulls in Goolsbee & Guryan (2006, E-Rate California) and Dynarski et al. (2007, NCEE), a negative in Angrist & Lavy (2002, Israel), and a positive in Machin, McNally & Silva (2007, UK). The authors themselves note their result "concurs with findings of other recent studies" across countries, education levels and identification strategies, and read that convergence as suggesting "computer aided instruction may after all be an inferior mode of teaching."