Initial Policy Evaluation

In evaluating whether the policy of allowing free early education places for all 3-year olds has had an effect on results, the first basic model to look at is a simple linear panel regression model with fixed effects at the level of the LEA, and with just a simple [0,1] dummy to indicate the years in which the policy is in operation:

Rs_jt+3=α0+α1policyjt+T

0

ϕ+cj +jt+3 (2.1)

where Rs

jt+3 are the results outcomes: the percentage of children in LEA j at time t+3 who attain a specified level in subjects, wheres∈ {reading, writing, maths}. The levels that I look at are L2B or higher, which is the level that children are expected to attain at KS1, and L3 or higher which represents very high achievement at KS1;

policy_jtis the [0,1] dummy to indicate that the policy is in operation in LEA j at timet;

T is a vector of year dummies; cj is the fixed effect for LEA j; and jt+3 is the idiosyncratic error term for LEAj at timet+3.

It is necessary to estimate the model using the fixed effects estimator due to the nature of selection into the group of LEAs that have the policy implemented in the first phase. As outlined above, it was the ‘poorer’ LEAs, the 65 deemed to be in greatest deprivation, that were given the Nursery Education Grant to pay for early education places for all 3-year olds in 1999-2000, while the ‘better off’ LEAs did not receive the funding until 2000-2001. These first 65 LEAs to receive the policy funding are the ‘pathfinder’ LEAs, with the remaining 85 LEAs the ‘non-pathfinder’ LEAs. As well as being the first to have the policy implemented, these ‘poorer’ LEAs also have a results distribution which is lower than the distribution for the ‘better off’ LEAs. Consequently, in a cross sectional estimate of the effect of policy on results,

we will pick up some of the negative effect of being a poorer LEA through the coefficient on the policy dummy – there is a correlation between the fixed unobserved component of the error term and the policy dummy, biasing the coefficient downwards.

Implementing a fixed effects regression controls for all of the time invariant characteristics of the LEA, treating the unobserved component of the error term for each LEA as a parameter to be estimated and therefore allowing a clean estimate of the policy effect. So the selection issue is dealt with since selection into the early treatment group is on the basis of fixed unobservable characteristics that are subsumed in the fixed effect and thus controlled for. The policy effect is identified through differences between LEAs in their within variation in policy status and results.

Included in the model are a set of year dummies with the first results year the omitted comparison year in each case. It is necessary to include year dummies to take account of any common trends in results due to the year of the test – though the assessments are standard across the country, there may be countrywide cohort effects or marking leniency changes that equally effect all LEAs in the different years. As the policy is implemented at a different time for two different groups of LEAs there is not a problem of the year dummies and the policy dummy being collinear. It is necessary to assume that the year effects are common across the two groups of LEAs in order for this identification strategy to be successful. However I believe that this is a reasonable assumption given that there should be no reason why the year effects would not be common to all LEAs if they are driven by variations in marking standards, since the assessments are marked to an externally implemented national criteria that is standard across the country. Moreover there is no reasona priori to assume that there are differing cohort effects depending on whether an LEA is in the pathfinder group or not.

As I also wish to look at the extent to which the policy has affected the areas of most concern i.e. the LEAs that were deemed most in need of the policy and therefore had the policy implemented first, I also estimate the policy evaluation regressions allowing for a different policy effect depending on whether the LEAs were the poorer ‘pathfinder’ or the better off ‘non- pathfinder’ LEAs:

Rs_jt+3=γ0+γ1policyjt∗pfj+ξ1policyjt∗(1-pfj) +T

0_φ+c

j +jt+3 (2.2) where all variables are as per their definitions above, with pfj a [0,1] dummy for the path-finder

for the better off LEAs. Clearly this weakens the identification strategy as I can no longer fully exploit the difference in timing of the policy’s introduction between the two groups of LEAs. The assumption of common year effects means however that I can identify separate policy effects for the two groups of LEAs without them being collinear with the year dummies.

Whether estimated separately or in single policy dummy variable, it is also necessary to assume that the policy effect is an intercept shift that is constant in each year that the policy is in operation – otherwise it is clear that some of the policy effect in the later years, when all LEAs have the policy in operation, could be subsumed in the year dummies thus we could not delineate the separate policy effect were we not to assume it to be constant and identified through the years when the policy dummies are not equal to each other.