Analysing further education recruitment trends using a large language model

28th November 2024 by Timo Hannay [link]

Update 19th November 2024: See also this summary from The Gatsby Foundation and this coverage in FE News.

Since 2017, we have been reporting in detail on recruitment trends among schools and colleges in England, including not only vacancies for teachers, but also those for support and auxiliary staff. The analysis presented here takes a more detailed look than ever at recruitment trends in the further education (FE) sector. This has been enabled in part because we have used a large language model (LLM) – specifically, OpenAI's ChatGPT 4o – to extract not just basic information such as job title and institution, but also unprecedented levels of further detail from the semi-structured and unstructured text in the body of recruitment adverts.

To summarise:

• We analysed over 36,000 adverts posted between 1st May 2022 and 31st August 2024 by 134 different FE colleges (representing nearly 70% of all those in England). There was some seasonality, with peaks in the summer, but much less so than we have previously observed for schools.
• About half of the adverts were for teaching postions and most of the rest were for support positions. The most common subject areas for teaching positions were in construction, engineering and health. The most common specific subject was maths.
• The vast majority of postions were permanent and most of them were full time. Among adverts for part-time vacancies, teaching postions were most likely to specify 0.5, 0.6 or 0.8 FTE, while support positions were more likely to specify higher time commitments. Annual leave entitlements varied widely, but were most commonly in the range 30-40 days.
• Annual FTE salaries for support positions were most commonly in the £20-25k range, those for teaching positions peaked in the £30-35k range and those for leadership positions in the £40-45k range (but with wider variation).
• An analysis of putative repeat advertising rates, which made use of the ability of LLMs to analyse the semantic similarity of adverts, suggested that around 40% of adverts may be followed by repeats for the same vacancy.

To find out more, read on. Our sincere thanks to the Association of Colleges (AoC), who kindly gave us permission to use data from their popular recruitment platform AoCJobs, and to The Gatsby Foundation for their generous ongoing support of this work.

 

Recruitment and recruitability

As in other parts of the public sector, hiring in education has been particularly challenging in recent years and it has become a prominent political issue. At the same time, enhancement of the nation's collective technical skills is also high up the agenda, as demonstrated by the recent launch of Skills England. These two prominent national challenges combine in the often underappreciated area of recruitment by FE colleges. If we cannot hire the staff to teach young people new skills then what hope is there that they will find productive employment?

Recruitment is of course a complex process that is influenced by a wide variety of factors, including salary competitiveness, job satisfaction, workload and overall perceptions of the sector. Despite efforts to address these, anecdotal evidence suggests that FE colleges are struggling to attract and retain qualified staff, with particular challenges in certain subject areas, such as STEM (science, technology engineering and maths), where there is strong competition from industry. The research presented here seeks to understanding recent levels and patterns of recruitment at FE colleges in England. It also examines rates of apparent repeat advertising to provide an indicator of the difficulty that these institutions are experiencing in filling their vacancies.

We used traditional data-analytical and algorithmic approaches to determine basic information such as employer identity and job title. But in addition, we passed the full content of each job advert to an LLM via an API (application programming interface) along with a prompt from us instructing it to extract certain specific information and return this in a structured form that was amenable to further analysis. In this sense, the LLM made working with the free text of adverts much more like analysing a structured database. It was not always reliable, but following optimisation of the prompt using a subset of adverts, spot checks indicated that it was generally correct at a similar level to that of a diligent but relatively naive human. Of the few detectable errors that remained in the final output, many were of a kind that we might expect a human to make too, especially given the occasionally ambiguous or erroneous information in the adverts themselves.

 

Up (and down)

We analysed a total of 36,603 FE college adverts published by FE colleges on FEJobs between 1st May 2022 and 31st August 2024. These were posted by a total of 134 distinct FE employers (about 68% of all FE colleges in England), with the top 40 of these collectively accounting for 81% of adverts. As shown in Figure 1, there was some apparent seasonality, with peaks in the summer, but this was far less regular than that we have previously seen in schools.

Among other things, the LLM was prompted to distinguish between different types of role: teaching (student-facing roles associated with a particular subject, including coaches, instructors and curriculum leads), leadership (at the institutional level and not associated with a particular subject, including heads of college, admissions and operations, among other roles), and support (neither of the above, including assistants, administrators and auxiliary staff such as groundspeople). Broadly similar temporal patterns were apparent across all roles, teaching roles, support roles and leadership roles.

(Use the menu below to select a role type to view. Hover over the lines to see corresponding data values.)

Over the entire period covered, teaching roles accounted for 50% of adverts, support roles for 45% and leadership roles for 5%, as shown in Figure 2.

(Hover over the bars to see corresponding data values.)

 

Desperately seeking teachers

This section examines in more detail the 50% of adverts that were for teaching positions. For these, the LLM extracted information about the subject area and this was compared against a taxonomy that we have developed previously (see see Footnote 1 for details). Advert numbers by top-level subject group are shown in Figure 3, with the economically and politically prominent areas of construction, engineering and health featuring most often. These general proportions did not vary much over time. Note that any given advert might occasionally fall into more than one group, particularly if the role encompasses multiple subjects.

These provide indicators of the subjects for which teachers are in most demand at FE colleges. The demand itself is driven by a combination of at least three factors: (a) how many students are studying the subject, (b) the turnover of teachers in the subject, and (c) the difficulty of filling any vacancies that arise. It is hard to tease these apart, though later on we will look at a potentially useful indicator of difficulty in hiring.

(Hover over the bars to see corresponding data values.)

At a higher level of detail, Figure 4 shows the 24 most commonly occurring subjects, with maths coming top.

(Hover over the bars to see corresponding data values.)

 

Terms of employment

This section looks at the kinds of contracts offered. As shown in Figure 5, around 85% of all positions advertised were permanent. This proportion fluctuated from month to month by a few percentage points, but there were no obvious seasonal or long-term trends.

(Hover over the bars to see corresponding data values.)

As shown in Figure 6, most postions were full-time. Generally speaking, part-time positions have declined from 20-25% in 2022 to around 15% in more recent months. We also asked the LLM to identify trainee and apprentice positions (2.4% of all adverts, including 2.2% of teaching positions and 2.9% of support positions), explicit mentions of job sharing (3.5% of all adverts) and vacancies intended only for internal candidates (0.7% of all adverts).

(Hover over the bars to see corresponding data values.)

We instructed the LLM, where possible, to extract information about the numerical FTE (full-time equivalent) commitment of the role as a decimal fraction. Most adverts did not mention this explicitly, but the distributions for those that that did are shown in Figure 7. Across all positions, 0.5, 0.6 and 0.8 FTE were the most common. This was also true of teaching positions. Support positions were more likely to be for over 0.8 FTE.

(Use the menu below to select a role type to view. Hover over the columns to see corresponding data values.)

About a third of adverts mentioned leave allowance in terms of number of days. For part-time positions, we instructed the LLM to convert these into FTE values where possible. The largest sample sizes were for teaching positions, which are shown in Figure 8. Leave allowances of 30-40 days a year were most common, but with some lower and higher – occasionally much higher – values. Note that extreme outliers can indicate mistakes or ambiguities in the text of the adverts, or errors made by the LLM, so should not be taken too seriously.

(Hover over the columns to see corresponding data values.)

Across all years, around 31% of adverts specifically mentioned 'flexible' working, but this seems to have increased in recent years, from less than a quarter in 2022 to over a third in 2023 and so far in 2024. This is shown in Figure 9. Only time will tell if this is constitutes a longer-term trend. Also, it is not always clear exactly what is meant by this term and to who the flexibility is conferred, the employer or the employee. Understanding this will require further analysis.

(Hover over the columns to see corresponding data values.)

For adverts that mentioned specific annual salary amounts or ranges, we asked the LLM to summarise them. For part-time positions, we additionally instructed it to convert these into FTE values where possible. Results are are shown in Figure 10. For teaching positions the mode annual salary was in the £30-35k range, for support positions it was in the £20-25k range, and for leadership positions it was in the £40-45k range – though with a much wider spread. As for the leave entitlements shown in Figure 8 above, extreme outliers may sometimes represent errors in the advert text and/or in the LLM's interpretation, so any columns representing implausible values that consist of 1% or less of adverts are probably best ignored.

(Use the menu below to select a role type to view. Hover over the columns to see corresponding data values.)

Figure 11 shows the top 10 qualifications required by applicants, where mentioned.

(Hover over the bars to see corresponding data values.)

 

Repeaters

Another aspect of school and college recruitment that we have examined in the past is putative repeat advertising. When a vacancy has not been successfully filled, an institution may re-advertise, so detecting repeat adverts can provide an indicator of difficulty in hiring. We use the term 'putative' because, short of asking the employer, it is usually impossible to be certain that any given advert is truly a repeat of another one that has run before it – sometimes institutions really do hire multiple similar positions in quick succession. But these analyses can be informative, particularly when looked at in aggregate and over time.

However, they are not easy to do and require somewhat subjective rules about (for example) how similar a job title should be for it to correspond to the same position, as well as the time window within which to search for repeats. Fortunately, LLMs provide an alternative, arguably more objective approach. Specifically, they enable the creation of semantic 'fingerprints' for each advert in the form of so-called embeddings: lists of numbers that represent the collective meanings of the words and phrases contained in an advert or other document. In this case, we used OpenAI's 'text-embedding-ada-002' model. By comparing the embeddings of one advert with another, we calculated their similarity score (technically, the dot product of the two vectors represented by the embeddings of each advert). This gives a value between 0 (completely dissimilar) and 1 (semantically identical).

Figure 12 shows the distribution of these scores – in terms of percentages – when we compare every advert with all the other adverts issued over this period by the same employer. (Adverts from different employers are clearly not potential repeats.) Most adverts from the same employer are 80-95% similar. This is to be expected: adverts from the same employer, even if for different positions, are very much alike in the grand scheme of things. But look closely and you can see another much smaller peak at the far right-hand side of the figure, close to 100%. These adverts, with similarity scores of around 98-100%, are almost or completely identical, at least in terms of their semantic content. In other words, they are putative repeats.

(Hover over the bars to see corresponding data values.)

But similarity is not the only criterion, timing is important too. Figure 13 shows the numbers of putative repeat adverts (ie, >=98% similarity to a preceding advert from the same employer) by number of days since the preceding advert was posted. There are large peaks within a few days of the original, which we take to be short-term corrections or updates rather than true repeats. There are also peaks at around one month and a clear weekly cycle, which suggests that it is common to wait a month or a whole number of weeks before reposting. The existence of these regular temporal patterns also helps to confirm that at least some significant proportion of these similar-look adverts are genuine repeats.

For the current analyses we used a definition of a repeat advert as one posted 14-65 days after another advert of >=98% similarity. Figure 14 shows the resulting repeat rate over time. There are a few technical aspects to note. First, some employers (especially large colleges or college groups) advertise certain positions (eg, teaching assistants) on an almost continual basis. We have excluded these advertisements, defined as those with 20 or more appearances over the whole period covered (ie, appearing on average roughly every month-and-a-half or more). Also, the apparent drop in putative repeat advertising rate is at least partly an artefact caused by insufficient time having passed for repeat ads to have appeared.

The upshot is that the putative re-advertising rate is high, at around 40%. In other words, about 40% of all adverts are followed by a putative repeat within 14-65 days, though this proportion has at least been broadly stable over the period analysed. An obvious conclusion is that it reflects the difficulty institutions face in filling vacancies, with potential consequences for persistent understaffing, as well as increased recruitment workload and expense.

 

The Sense of an Ending

We hope that this analysis has provided a useful overview of the current state of recruiting activity at FE colleges in England. In addition, it provides a glimpse of the potential for using LLMs and similar software systems in analysing data from unstructured and semi-structured sources.

These are still early days, and much work remains to be done in both improving the underlying technologies and optimising their effectiveness in use cases such as this one. But they are already helpful and cost-effective (although we used one of the more expensive LLM models available, the work described here incurred API charges of less than £0.01 per advert). Especially for underappreciated domains like FE recruitment, this holds out the hope of better, more timely information to inform sectorial activities and government policies. The information is out there if we know where to look and how to analyse it.

We hope to continue monitoring this space in order to uncover longer-term trends and more detailed patterns (eg, by subject or geographical location), as well as building on and progressing the emerging research methods described.

We are always pleased to hear from anyone with questions, comments or suggestions. Please write to: [email protected].

 

 

Footnotes:

1. The taxonomy for subjects was similar to the one used in our previous analysis of college website adverts. 583 terms were organised into a hierarchical structure that encompassed 104 subjects organised into 58 subject areas and ultimately aggregated into 30 subject groups: Agriculture, environmental and animal care; Business and administration; Care services; Catering and hospitality; Computing and ICT; Construction and the built environment; Crafts, creative arts and design; Creative and design; Digital; Education and childcare; Engineering and manufacturing; Foundation learning; Geographical and environmental studies; Hair and beauty; Health and science; Historical, philosophical and religious studies; Languages, literature and culture of British Isles; Learning support; Legal, finance and accounting; Mathematics; Media and communication; Other languages, literature and culture; Performing arts; Protective services; Psychology; Sales, marketing and procurement; Sciences; Social sciences; Sport, leisure and recreation; Transport and logistics. This taxonomy was developed in collaboration with The Gatsby Foundation.

Grouping schools to tackle disadvantage

8th November 2024 by Timo Hannay [link]

Update 8th November 2024: See also this summary from The Gatsby Foundation and this further coverage from Schools Week.

This is the second of two posts about disadvantage in England's schools. If you have't done so already, we suggest reading Part One first.

 

Dimensions and degrees of disadvantage

Our previous post outlined a wide variety of often underappreciated headwinds faced by schools serving more disadvantaged communities, including difficulties in recruiting, greater use of supply teachers, lower levels of staff experience, higher spend on training, higher levels of sickness leave and (perhaps unsurprisingly given all the above) different areas of focus in Ofsted reports. Thus disadvantage manifests itself in many different ways, not just in rates of eligibility for free school meals and academic outcomes.

As we explored in some detail last year, social deprivation is not fully captured by income indicators alone: aspects like health, crime and the environment also have their effects. Furthermore, the main income metrics used for schools – pupil eligibility for free school meal (FSM) and the Pupil Premium (PP) – use a binary threshold to divide pupils into two somewhat arbitrary groups deemed 'disadvantaged' and 'not disadvantaged', and therefore fail to reflect the fact that deprivation exists on a continuum.

The upshot is that even schools with identical PP measures often experience very different local conditions with respect to levels of income deprivation, as well as with other aspects of deprivation such as crime, education, the environment, health and housing. In important ways, each school is different when it comes to deprivation, a fact that tends to get lost in the FSM and PP statistics.

One solution might be to prioritise by place. Education Investment Areas (EIAs) were proposed by Britain's previous Conservative government in 2023 as a way to target educational support to parts of England that had fallen behind. But as we have previously argued these also do a bad job of grouping together truly similar schools – except in the trivial sense that they happen to be located in the same local authority areas.

Are there better ways of grouping schools that retain important nuances without having to treat each one as its own special case? We believe there are. The work presented here provides a preliminary segmentation of schools in England based on Index of Multiple Deprivation (IMD) characteristics and POLAR4 (higher-education participation rates) of their local areas. For now, we'll look only at mainstream state secondary schools, but similar approaches could be applied to primary schools, as well as to other phases and forms of education.

The tl;dr version is that we end up with six clusters, roughly characterised as follows:

• Cluster 1, Suburban: This represents 'middle England' outside the major cities. Socioeconomic and educational indicators are mostly unexceptional.
• Cluster 2, Affluent Suburban: Richer suburban and rural neighbourhoods. The incidence of income deprivation is very low, but educational outcomes are not as good as you might expect.
• Cluster 3, Affluent Urban: Richer city areas, especially in London. Much greater levels of income deprivation than Cluster 2, but also higher levels of educational engagement and better outcomes.
• Cluster 4, Poor Urban: Especially in the North and the Midlands, but also to the east of London and elsewhere. Lots of adverse socioeconomic indicators, coupled with relatively weak educational outcomes – though not as bad as you might think given the levels of poverty.
• Cluster 5, Poor Suburban: Again, mainly in the North and Midlands. IMD indicators are mixed, but income deprivation is high and educational outcomes are poor. These are the areas that have fallen furthest behind.
• Cluster 6, Urban: Middling city areas in London, Birmingham and Manchester, among other places. Moderately high levels of income deprivation, but relatively good educational outcomes.

Each of these has distinct characteristics, not only in terms of the socioeconomic factors by which they are defined, but also in terms of school characteristics and educational outcomes. The journey to creating these clusters is as informative as the destination, so we hope you'll join us and read on.

Our sincere thanks once again to the Gatsby Foundation for collaborating in, and generously supporting, this work.

 

Counting clusters

Our main approach here will be to apply a clustering algorithm to the IMD and POLAR4 metrics of each school's local area (defined as postcodes within a 4km radius). Specifically, we will use k-means, which is a relatively simple unsupervised machine-learning method that can divide entities (in this case schools) into arbitrary numbers of groups containing statistically similar members. In this case, it will cluster together schools in local areas with similar IMD and POLAR4 characteristics.

IMD is composed of several different measures: Crime, Education, Employment, Environment, Health, Housing and Income¹), to which we have added POLAR4. Some of these correlate strongly with each other, so it makes sense to examine whether these eight parameters can be reduced in order to eliminate those that provide little or no extra information. This process, known as principal component analysis (PCA), is a common prelude to conducting any kind of clustering analysis.

By applying PCA, we find that about 58% of the statistical variation between schools can be accounted for by a single component (Component 1) and a further 20% by a second component (Component 2). Adding a third component accounts for another 8% of the variation, but was ultimately found to have only minor effects on the clustering results, so for the sake of simplicity the rest of this analysis uses clustering based only on Components 1 and 2 – though we will briefly revisit below the implications of adding a third component.

What real-world characteristics do these components represent? Formally, they are just statistical measures, but it is fairly clear when looking at the clustering results that they roughly correspond to poverty and urbanisation, respectively. This shouldn't be too surprising: life is very different between rich and poor communities, and also for residents of urban versus suburban or rural locations. (The real-world correlate of Component 3 is less clear, but inspection of the clusters suggests that, like Component 2, it may be related to population density.)

Having chosen to use two statistical components, we also need to decide how many clusters or segments to apply to the population of schools. Figure 1 is a so-called 'elbow plot' that shows how well the clusters represent their constituent schools as we increase their number. This uses a statistical measure known as 'inertia' (only loosely related to the physical sense of that word), for which lower values indicate better representation. Sometimes the inertia reduces rapidly up to a certain number of clusters, with little improvement above this. In those cases, it is common to choose the number at which the gradient of the line flattens since that provides the best fit to the data with the lowest number of clusters. However, in this case there is no very clear inflection point, or 'elbow', in the line, so we could plausibly choose almost any number of clusters within the range shown, In the analysis that follows we will therefore explore the effects of varying the number of clusters between 2 and 6.

(Hover over the graph to see corresponding data values.)

 

Two by two

To give a sense of the effects of clustering, let's first see what two clusters look like. Figure 2 shows how these divide up based on the two principal statistical components, with each dot representing a school. They are mostly separated by Principal Component 1 (horizontal axis) rather than Principal Component 2 (vertical axis). The two clusters are not completely distinct, and there is some intermingling. In that sense, grouping schools in this way is a bit more like slicing pizza (you could choose to cut in a range of different places) rather than breaking off dough balls (which are naturally lumpy). This doesn't make the clusters arbitrary – far from it – but it does mean that we shouldn't read too much significance into which side of the boundary individual schools fall.

(Click on the legend to turn individual clusters on or off; double-click to show one cluster on its own. Hover over the dots to see corresponding school information.)

Figure 3 shows the locations of schools in Cluster 1. In general, these are in poorer areas, including cities in the North and Midlands, but excluding most schools in London. On average, these localities have worse IMD measures, especially higher crime rates, but also lower levels of education, employment, health and income. They also have very low POLAR4 scores (32% versus 52% for Cluster 2). Conversely, the housing indicator is good, probably at least in part because homes are relatively cheap in these areas. Within schools, Pupil Premium rates are high (an average of 30% versus 23%) and almost all educational indicators are below average (eg, Attainment 8 is 44 versus 51 and Progress 8 is -0.16 versus +0.20), with exclusion rates particularly high (fixed-term exclusions 26% versus 12%). Cluster 2 – which of course is composed of all those schools not in Cluster 1 – has precisely complementary characteristics. Note that in both cases the schools are scattered all over the country, albeit unevenly.

(Use the menu to select a cluster. Use the map controls to pan and zoom. Hover over the dots to see corresponding school names.)

 

Fab four?

But using two clusters achieves little more than recapitulating the traditional disadvantaged / not disadvantaged distinction by other means, so how about dialling it up to four? As we shall see, this starts to separate out different geographical groups too, especially across the urban / suburban divide. Figure 4 shows the resulting cluster memberships plotted against the two principal components.

(Click on the legend to turn individual clusters on or off; double-click to show one cluster on its own. Hover over the dots to see corresponding school information.)

Figure 5 shows the locations of schools in these four clusters.

Cluster 1 schools are overwhelmingly in or around London and tend to have good scores across most IMD measures, though environment is moderately poor and housing deprivation is high (in large part because homes are expensive). POLAR4 scores are very high (average of 52%). Academic attainment and progress rates are high (Attainment 8 is 52, Progress 8 is +0.29) and exclusions are low (fixed-term exclusions are 11%), even though Pupil Premium rates are also moderately high (28%). In other words, poverty and wealth tend to coexist in these areas, and educational engagement and outcomes are generally good, especially given the high incidence of income deprivtion.

Cluster 2 schools tend to be in cities in the Midlands and the North, or in other disadvantaged areas around the coast. Average IMD scores are uniformly bad and POLAR4 scores are low (33%). Pupil Premium rates are high (35%), as are exclusion rates (26%). Academic attainment (44) and progress (-0.14) are both low.

(Use the menu to select a cluster. Use the map controls to pan and zoom. Hover over the dots to see corresponding school names.)

Schools in Cluster 3 are located in towns and around major conurbations such as London, Birmingham and Manchester. They have very low Pupil Premium rates (19%), but only moderately high POLAR4 rates (43%) and only moderately good academic outcomes (Attainment 8 is 49, Progress 8 is +0.08). IMD metrics are generally good with the exception of housing (again, because it is expensive).

Cluster 4, schools are also mostly outside cities, but grouped in particular locations in the North, the Midlands and to the east of London. They have moderate Pupil Premium levels (29%) and IMD scores are mostly mediocre rather than low, though environment and housing are good. However, POLAR4 is very low (29%) and school exclusions are very high (also 29%), while academic attainment (44) and progress (-0.21) are both very low.

What we are starting to see here is that the relationships between poverty and education are more complicated than is often appreciated. True, richer areas tend to do better. But some urban areas – notably London – do well despite having high levels of income deprivation (see Clusters 1 and 2), while suburban and rural areas tend to underperform by comparison, especially after allowing for their relative poverty levels (see Clusters 3 and 4).

 

V6

Finally, let's crank the cluster number up to six. Among other things, this provides a bit more nuance on the poverty-affluence axis. Figure 6 shows the resulting cluster memberships plotted against the two principal components.

(Click on the legend to turn individual clusters on or off; double-click to show one cluster on its own. Hover over the dots to see corresponding school information.)

Just for fun, Figure 7 shows what six clusters look like when generated using three principal components. The extra dimension aside, the resulting clusters are actually very similar to the corresponding clusters created using just two principle components: overlaps in school membership are all over 85% and most of them (four out of six clusters) are well over 90%. For this reason, we will continue to use the clusters created using two principal components.

(Click on the legend to turn individual clusters on or off. Hover over the dots to see corresponding school information. Click and drag to rotate the figure. Scroll to zoom in or out.)

Figure 8 shows the locations of schools in these six clusters. At the risk of oversimplifying, we will assign each one with a descriptive name in order to distinguish them more easily. It is important to emphasise that these names are not definitions, just labels. The clusters are defined only by the IMD and POLAR4 characteristics of their constituent of the local areas around each of their schools.

We will call Cluster 1 "Suburban". (It includes rural areas too, but these account for only around 7% of the total school population compared to over 40% for the suburbs, so we will use "suburban" in all of our shorthand descriptions.) Broadly speaking, these represent middle England outside cities. IMD indicators and POLAR4 (35%) are mostly unexceptional. The same goes for school measures: the average Pupil Premium rate is 24%, average Attainment 8 is 46 and average Progress 8 is -0.09. The mean fixed-term exclusion rate is 21% and the absence rate is 9% of sessions.

Cluster 2, "Affluent Suburban", represents richer suburban and rural neighbourhoods. IMD measures are uniformly good, with the exception of housing, which is expensive. POLAR4 is quite high (50%). Pupil Premium rates are very low (17%), while academic attainment (51) and progress (+0.14) are reasonably good. Rates of exclusion (13%) and absence (8%) are low. Cluster 3, "Affluent Urban", represents richer areas in cities, especially London. These schools tend to do better than those in Cluster 2 despite having higher levels of income deprivation. With the exception of environment and housing, IMD measures are good, and POLAR4 is exceptionally high (71%). Academic attainment (54) and progress (+0.38) are both very good even though Pupil Premium rates are quite high (27%). Rates of exclusion (11%) and absence (8%) are both low.

(Use the menu to select a cluster. Use the map controls to pan and zoom. Hover over the dots to see corresponding school names.)

Locations in Cluster 4, "Poor Urban" have uniformly poor IMD metrics, with the exception of housing, which is cheap. POLAR4 rates are rather low (33%). Pupil Premium rates are high (35%), while academic attainment (44) and progress (-0.15) are both moderately low. Absence (10%) and exclusion (24%) rates are quite high. Cluster 5, "Poor Suburban", has good environment and housing scores, but poor values for other IMD metrics and very low POLAR4 scores (27%). Pupil Premium rates are high (34%). The same goes for rates of absence (10%) and exclusion (35%), while academic attainment (42) and progress (-0.26) are both low. Taking these two clusters together, we once again see urban areas outperforming suburban and rural ones with similar levels of income deprivation.

Cluster 6, "Urban", has moderate IMD scores, with the exception of housing, which is poor (ie, expensive). POLAR4 (48%) is somewhat high. Pupil Premium rates (32%) are also quite high, but rates of absence (8%) and exclusion (13%) are both low, while academic attainment (49) and progress (+0.21) are both quite good. This is another example of an urban group that appears to outperform its socioeconomic fundamentals in terms of educational outcomes.

 

What k-means might mean

It is important to emphasise that that the educational metrics mentioned here – attainment, progress, exclusions and absences – were not used to create the clusters, yet they display clear disparities across them. This is a reminder that educational effectiveness and outcomes do indeed correlate with social factors. Furthermore, this way of grouping schools illustrates the importance not just of income but also of place – in terms of the level of urbanisation rather than the name of the region or local authority area. Indeed, there appears to be an interaction between the two, with some urban areas performing exceptionally well despite high levels of income deprivation. From this analysis alone we cannot say exactly why, but perhaps we should be considering the educational impact of the cultural and social capital often associated with more densely populated places, not just the single dimension of affluence and poverty that currently gets all the attention.

Either way, as we have hopefully shown, it makes little sense to segment schools based solely on their FSM or PP rates. As well as being crude threshold measures, they hide important information about different forms of deprivation. Neither is it sensible to simply rely on the part of the country, rather than the type of place, in which the school is located. The approach presented here in effect combines income, geography and other factors. In an age of plentiful data, this more nuanced view is what we should mean by 'similar' schools, at least when it comes to the local socioeconomic environments they experience.

The work presented here is just a start and we hope to continue developing and applying it. In particular, it would be interesting to see how robust the clustering is to changes in the underlying data and algorithmic parameters, as well as to repeated rounds of clustering (given that k-means makes use of random searches to create clusters, which might therefore represent a local optimum rather than a global one). We would also like to further explore the educational and other characterics of these and similar clusters, as well as their possible policy implications. Of course, we can also increase the number of clusters further to cater for situations in which we want to discriminate more at the cost of additional complexity.

In the meantime, subscribers to SchoolDash Insights can explore these trends further, especially in the Schools section. Non-subscribers can request a trial account or demo. To keep up to date with more analyses like this one, sign up for our free monthly-ish newsletter. We also welcome questions or comments; please write to: [email protected].

 

Footnotes:

1. For Income, we actually used the IDACI subcomponent, which specifically covers income deprivation among families with children.