Audio presentations

School league tables: what can they really tell us?

Libby Purves interviews Harvey Goldstein.

School league tables are eagerly scanned by parents hoping to find the best school for their child, and by teachers hoping to see their school rise in the rankings. But do they tell parents what they need to know? Harvey Goldstein and George Leckie argue that league tables are not fit for that purpose, and say it is time that their publication should cease.

Residuals - An Introduction

Residuals can be important if we want to rank our units after controlling for a set of covariates (for example when drawing up league tables of schools), or if we are interested in the effect of a particular level 2 unit (for example if we want to see what effect a particular school is having on its students' performance). In this presentation, we show how both level 1 and level 2 residuals are calculated for multilevel models (though in practice this calculation will usually be performed by the software), and explain why residuals in multilevel models are shrunk in towards the overall regression line.

Measuring Dependency

 

We use multilevel modelling when we have dependent data, i.e. there is similarity between observations from the same group (for example, heights of children from the same family). This presentation explains how to measure the dependency using the variance partitioning coefficient (VPC). We explore the interpretation of the VPC through example graphs. We also see that the VPC shows how much of the variance is due to each level of the model, and thus gives some insight into what extent the response is determined at each level.

Covariance and Correlation Matrices

We use multilevel modelling when we have dependent data, i.e. there is similarity between observations from the same group (for example, heights of children from the same family). An obvious question is: just how does the multilevel model take this dependency into account? In this presentation, we examine the structure of the model: we see what the correlation is between each pair of level 1 units in our dataset. This allows us to see how the relation between different observations from the same group is specified by the model. We contrast this to a single level model, for which we see there is no correlation between different observations from the same group.

Significance Testing

Tests for coefficients of individual variables: eyeballing standard errors, Wald tests, and calculating p-values using the tail areas screen in MLwiN. Tests for comparing models: the Likelihood Ratio Test and the Deviance Information Criterion.

Top of the Class

by Harvey Goldstein

Problems? In order to hear the show you may need to Download Realplayer

BBC's Radio 4 programme, More or Less told the story of a measurement that has become a defining feature of the way we now run Britain

For nearly 15 years, the league table, and its raw material, performance assessment, has been the driving force behind policy; and nowhere more so than in education.

In the beginning, it seemed simple: tell parents how good the local school is. But, as Andrew Dilnot discovered, that ambition began an enduring clash of political will with statistical headache.

We have steadily learnt over the years, and through repeated reform, how hard to achieve that simple objective is.

First with raw tables of 5 passes at A* to C grades at GCSE, then with value added tables, then tables that required the inclusion of Maths and English, and now, from next year, 15 years on and still trying, with contextualised value added tables which attempt to take account of the social and economic factors that influence how successful we can expect a child to be.

So it is not over yet.

But as the government introduces this next major and complex reform of school performance assessment next year, we ask if the ambition of simplicity and accountability will finally be realised, or if it will prove to be a 15 year march up a blind alley, or perhaps a third possibility: whether it will change the way we look at and use data about pupil performance.

Presenter: Andrew Dilnot
Producer: Michael Blastland

BBC Radio 4's More or Less was broadcast on Monday, 11 December, 2006 at 1630 GMT.

Edit this page