Modelling school assignment with administrative data

1. Modelling school assignment with administrative data Simon Burgess Based on “School Assignment, School Choiceand Social Mobility” by Simon Burgess and Adam Briggs, CMPO DP 06/157

2. Introduction • Not all schools are good schools • Which pupils go to the good schools? • To the extent that children from poor families are allocated to worse schools, this perpetuates disadvantage, reducing social mobility • Questions: • What is the extent (if any) of a differential chance of going to a good school? • How does it happen? www.bris.ac.uk/Depts/CMPO

3. What we do: • We estimate the chances of poor and of non-poor children getting places in good schools • One of the key factors is location – distance between school and home. • Our dataset allows us to measure distance precisely and characterise the pupil’s very local area. • We compare pupils living in the same place. Exploit within-street variation and also control for other personal characteristics including prior test scores. www.bris.ac.uk/Depts/CMPO

4. Results • Poor children half as likely to go to good schools. • Much of that, but not all, comes through location. That is, accounting fully for location, the gap is much smaller, but not zero. • Children from poor children tend not to go to a good school, even if it is their nearest. • (more …. See paper) www.bris.ac.uk/Depts/CMPO

5. Plan • Modelling Framework • Data • Results • Conclusions www.bris.ac.uk/Depts/CMPO

6. Modelling Framework • We model the assignment of children to schools, as a function of the characteristics of the school and of the children. It’s a matching problem. • The observed data on the outcome of this assignment are realisations of an underlying process, composed of two decisions: • applications by parents and children for places in particular schools (demand), • the administrative procedures that allocate children to schools given their choices (assignment rule) www.bris.ac.uk/Depts/CMPO

7. Allocation • Write a general model of the outcome of the allocation as: • where www.bris.ac.uk/Depts/CMPO

8. Reverse causation? • We interpret the estimated relationship between the school’s quality score qa(i, t), t-6 and a student’s personal characteristic, fit, as representing the outcome of the assignment process. • Alternative: from student characteristics to the outcome score. www.bris.ac.uk/Depts/CMPO

9. Timing: the quality score derives from the performance of a group of children 6 years older than the current intake. • But: persistence in school attendance. Two interpretations: • “Islands story”: Schools located on “islands”, with no mobility between them. All students from succeeding generations therefore go to the school on their island. • Correlation from one generation’s poverty to the next. • But: this is not what England’s schools look like • Half of children do not go to local school • See map of Birmingham www.bris.ac.uk/Depts/CMPO

10. Figure 1: School Distance Contours in Birmingham www.bris.ac.uk/Depts/CMPO

11. “Dynasties”: pupils living in particular locations always go to the same school. And with persistence in area poverty, particular locations always house poor families. • poverty of succeeding generations is correlated, score of one generation of pupils drawn from that area is correlated with the poverty of the next. • Econometrically, estimating: • Will be biased because omitted variable of the nature of i’s location is correlated with fi, and with the nature of the previous cohort of pupils who generated the school quality score. • Response: control for location to remove omitted variable bias; within postcode variation. www.bris.ac.uk/Depts/CMPO

12. Data • Data on pupils • Data on schools • Data on location • Our sample www.bris.ac.uk/Depts/CMPO

13. Pupils • PLASC/NPD: Census of all children in state schools in England, taken each year in January. • First in 2002; use three PLASCs • Gender, within-year age, ethnicity, SEN,.. • Key-stage 2 test taken at age 11 as the pupils finish primary school. This is a nationally set group of tests (in English, Maths and Science), marked outside the school • Key variable for our purposes is an indicator of family poverty, the eligibility for Free School Meals (FSM). www.bris.ac.uk/Depts/CMPO

14. Schools • “Quality” of the secondary school that each child attends. • Use the publicly available and widely quoted measure of the proportion of a school’s pupils which passes at least 5 GCSE exams at age 16. • These exams are important, are nationally set and come at the end of compulsory schooling. • Define a “good school” as a school in the top third nationally of the distribution of %5A-C scores. • Dating – we use the score for each school from the time that the cohorts made their decisions on school applications, so deriving from the results of a cohort of pupils 6 years older. www.bris.ac.uk/Depts/CMPO

15. Location • We have access to each pupil’s full postcode. This locates them quite precisely. • Also the coordinates of the school, which locates it exactly. • We rely on the postal geography of the UK for this analysis. Overall, there are about 1.78m unit postcodes covering 27.5m addresses. On average, there are 15 addresses in a unit postcode, but this varies. • Using pupils’ postcodes, we match in data on neighbourhoods, on two scales: postcode, and area (ward). www.bris.ac.uk/Depts/CMPO

16. Location (2) • Mosaic classification, a postcode level dataset that categorises each postcode in the UK into one of 61 different types. • Over 400 variables used. This is commercial geo-demographic data, kindly provided to us by Experian. • Index of Multiple Deprivation (IMD) produced by the Office of the Deputy Prime Minister. Ward level dataset that ranks every ward in England on a range of criteria. • Distance can be measured in a number of different ways. It is only computationally feasible to use straight-line distances. www.bris.ac.uk/Depts/CMPO

17. Sample • We take the cohort of new entrants into secondary school from each PLASC, so pupils in their first year of secondary school. Roughly 0.5m pupils in each cohort, so our full sample is 1.57m pupils. • State schools in England; non-selective LEAs (this cuts out 13.4% of the pupil total); omit pupils from some special schools, and pupils are omitted if they have missing values for data. • Sample for the overall regressions in Table 2 is 1.24m – some 91% of the available total in non-selective LEAs. www.bris.ac.uk/Depts/CMPO

18. Table 2: Probit of whether pupil goes to a good school www.bris.ac.uk/Depts/CMPO

19. Results • Controlling fully for Location • How much of the difference in probability of attending a good school is due to location? • Need to control completely for location. • Interpretation: location not exogenous – estimating how important choice of location is for parents’ strategy. www.bris.ac.uk/Depts/CMPO

20. www.bris.ac.uk/Depts/CMPO

21. www.bris.ac.uk/Depts/CMPO

22. www.bris.ac.uk/Depts/CMPO

23. Table 5: Statistics on numbers of pupils per postcode www.bris.ac.uk/Depts/CMPO

24. Figure 5: Differences in school quality by differences in FSM status www.bris.ac.uk/Depts/CMPO

25. Table 6: Postcode-cohort FE regressions of school quality www.bris.ac.uk/Depts/CMPO

26. Figure 6: Probability of pupils attending their nearest school www.bris.ac.uk/Depts/CMPO

27. Conclusions • Poor children half as likely to go to good schools. • Much of that, but not all, comes through location. That is, accounting fully for location, the gap is much smaller. • Controlling for location, this gap doesn’t vary much by degree of choice. • Children from poor children tend not to go to a good school, even if it is their nearest. • Our econometric strategy is not to identify causal relationships in this paper (future work). www.bris.ac.uk/Depts/CMPO