Reception Baseline Assessment, algorithmic bias and the reification of ‘ability’

Credit: Phil Meech for UCL IOE.

Guy Roberts-Holmes and Lucy Kaufmann.

The Department for Education (DfE) had attempted to introduce its contested and controversial Reception Baseline Assessment (RBA) for four-year-olds since 2015. Reflecting a wider realisation of the COVID-19 pandemic as a powerful catalyst for ‘re-imagining’ education with digital education technologies, the DfE implemented RBA as a statutory assessment in September 2021.

RBA is an automated national standardised numeracy and literacy test designed to facilitate the measurement of progression across seven years of primary school, and thereby, school accountability. Teachers receive a series of short, narrative statements about how their pupils performed in the assessment, but otherwise, the school’s results are black boxed until Year 6, when their progression rates are measured and compared across those seven years. Beginning in the summer of 2028, the DfE will use the data to create comparable school-level progression tables showing the progress pupils make. The DfE will publish these comparable datasets, adding to the data available to parents to calculate from when choosing a primary school.

The test is carried out in the first six weeks of a child starting Reception, with each four-year-old being individually tested on a tablet device. RBA functions via an algorithmic line of code which responds to the pre-determined trigger of the teacher input of ‘yes’ or ‘no’ responses. The algorithm sends the child on a pre-programmed differential pathway. This routing algorithm is presented positively by the DfE, in terms of preventing children from answering consecutive questions incorrectly and experiencing ‘failure’. What is obscured is that the algorithm facilitates higher assessment scores in children it identifies as ‘more able’ by rerouting their test pathway through additional and increasingly advanced questions, notably in the maths aspect. In practice, then, the algorithm separates children based on their responses, creating differential assessment pathways and resultant scores determined by an algorithmic interpretation of ‘ability’. Whilst some children have the opportunity to answer all of the questions and score the maximum mark of 39, other children will only be able to score, at most, 22, because that’s the maximum possible with the questions the algorithm presents to them.

Thus, the algorithm functions as a classificatory mechanism sorting children by ‘ability’. On the one hand, the algorithm favours and accelerates the progress of children who enter Reception equipped with the required maths and literacy measurable skills. On the other hand, it creates limitations for others, disproportionately facilitating lower scores in those children whom it does not recognise as successful learners. There is a possibility here of ‘ability’ automation bias with raced and classed implications, particularly as the RBA can only be conducted in English.

This is made more problematic by the fact that a child’s RBA performance can feed into and inform broader assessments of his or her ‘ability’ – on occasion actively so and with very real consequences. For example, recent research has shown instances where teachers were attempting to make sense of the RBA for organisational teaching purposes. They were noting each child’s test responses and subsequently completing class grids and checklists. This allowed them to identify and compare each child’s maths and literacy test scores. They then looked across these RBA datasets and combined them with their own in-house school Baseline judgements. Here, algorithmically generated machine data were being combined with a teacher’s professional decision making, in a process known as synthetic governance, that is, the combination of both human and machine knowledge. In this case, an unintended effect of this synthetic data governance could be that it further solidified notions of ability as teachers classified and sorted children into groups in an attempt to optimise progression.

The statutory implementation of the RBA demonstrates the government’s belief in technological solutions as an appropriate, perhaps even optimal, approach to highly complex societal problems. There is a risk, however, that the RBA algorithm precludes already disadvantaged children from the opportunity of obtaining a high score and since the RBA data are black boxed, this algorithmic bias remains invisible and incontestable. There is a danger that the RBA algorithm may have the inadvertent effect of reproducing disadvantage at the start of children’s schooling.

Guy Roberts-Holmes is Professor of Early Childhood Education at IOE, UCL’s Faculty of Education and Society, and Lucy Kaufmann is an alumna of IOE’s Early Years Education MA programme. Her MA dissertation is titled ‘The Effects of the 2021 Reception Baseline Assessment on Classroom Practice’.