Complex places for complex times: An analysis of the complexity of UK local economies, and their future evolution
In the third data pilot in Arloesiadur, our innovation analytics project for Welsh Government, we draw on cutting edge economic geography research to identify related industries for clustering analysis, measure the complexity of local economies in the UK, explore the links between complexity and better economic outcomes such as productivity and wealth, and generate predictions about the future specialisations of UK local economies based on their current profiles.
The world isn’t flat
One of the key economic lessons of the last 20 years is that, in spite of globalisation and the Internet, place still matters: industries tend to cluster in particular places; the industrial composition of those places shape future opportunities and constraints . Too much specialisation creates fragile local economies reliant on the fortunes of a small number of sectors. Too much variety could reduce the scope for spillovers between them.
The questions for policymakers are:
- What is the current profile of a local economy?
- What are the potential and desired future paths of development?
- How to intervene (if at all) to achieve desired outcomes?
These are very topical issues: many analyses have linked the discontent behind the results of the European Referendum to regional economic divides, and the same seems to apply to the advance of Donald Trump in the USA. National and local industrial strategies to drive industrial growth are back on the agenda.
The challenge is how to design and implement these strategies in a way that creates new markets, strengthens industrial ecosystems and drives growth, rather than prop up inefficient industries or satisfy vested interests. Ideas such as ‘smart specialisation’ or ‘entrepreneurial discovery’, originated in academia and adopted by the European Commission are an important contribution to these debates. In a nutshell, they ask regions to focus on those industries where they have established strengths rather than trying to build new clusters from scratch, and to identify what the new opportunities are in a process driven by entrepreneurs instead of officials.
This blog post reports the findings of an innovation analytics pilot where we apply new approaches developed by economic geographers and complexity scientists with the goal of generating data that can inform policy decisions in this important area. The pilot is part of Arloesiadur, our project to develop an innovation dashboard for Welsh Government.
We discuss some of the findings in the rest of the blog. Although we introduce key concepts and intuitions for the methods used in the relevant sections, we don’t go into their detail. If you want further information, check the GitHub repository with the scripts we have used to process and analyse the data. We've made a visualisation of the data available here.
Before starting, we should point out this is an exploratory pilot with limitations that we highlight where relevant, so the (suggestive) results should be taken with caution. Our goal is to take this analysis further when we start building the Arloesiadur platform in the coming months, incorporating any feedback or suggestions you might have.
First step: segmenting industries
The UK Interdepartmental Business Register contains information about the geographical distribution of 543 4-digit SIC codes – however, many of these SIC codes are interrelated (e.g. PR, Marketing and Advertising). In order to identify higher-level aggregates that simplify analysis, we have drawn on the approach that US Economists Mercedes Delgado, Michael Porter and Scott Stern recently applied in the US. The basic idea is to calculate different measures of ‘similarity’ between industries – based on their propensity to co-locate, employ people from the same occupations, and trade – and use these measures to identify segments of highly interrelated industries we might expect to see clustering in the same place. This approach allows us to generate a mutually exclusive, completely exclusive allocation of SIC-codes to a manageable (64) set of industries we can map in a comparable basis across UK Local Authority Districts .
It is worth noting that these 64 segments don’t just replicate the SIC structure: 71.8% of the clusters contain 4-digit SIC codes from an assortment of different SIC divisions. Having said this, the top division in each segment accounts for 71.6% of the 4-digit SICs codes, which is what one would expect if SIC structure captured, to a strong degree, meaningful similarities between sectors. One example that illustrates how our segmentation method brings together interrelated sectors from across the SIC structure is the case of rural services, a segment where our clustering analysis places those Wholesale services related to the sale of live animals (46.23), agricultural machinery (46.61), and veterinary activities (75.00) amongst other things.
Second step: measuring the economic complexity of UK local authority districts
After generating a list of industry segments, we have looked at their geographical distribution across UK Local Authority Districts. We use the resulting sectoral specialisation patterns to generate measures of complexity. In doing this, we have attempted to translate the analysis of economic complexity that Cesar Hidalgo and Ricardo Hausman pioneered using export data at the country level, to the analysis of local specialisation patterns in UK local authorities.
In a nutshell, this approach, called ‘the method of reflections’ is a recursive algorithm that takes the specialisation profile of a local authority and weights it by the extent to which the sectors in which it specialises tend to appear in diversified areas. After several iterations, the ranking of areas becomes stable. We have taken those scores and normalised them to generate an indicator of complexity.
The violin plot below shows the distribution of economic complexities in local authority districts in different Regions of Britain.
It shows that London and the South tend to have, on average, local authority districts with higher levels of economic complexity, consistent with the idea that these are more diversified economies. At the same time, there is a great degree of variation in the economic complexity within regions and nations of the UK – we find more complex, generally urban, and less complex, often rural economies inside all of then. The shape of each of the violins shows that some regions such as London or the North West tend to have a ‘normal’ distribution of economic complexity, with most of their local authorities in the middle, while Wales, Scotland and Northern Ireland have a few highly complex areas, and many less complex areas. The South East and the South West are closer to a bimodal distribution, with many economically complex areas, and many non-complex ones.
Why would anyone care about any of this? Hidalgo and Hausman answer the ‘so what’ question by showing that more economically complex countries tend to perform better along a wide range of measures: they tend to have higher GDP per capita, faster growth and lower inequality.
What about the measures of complexity that we have generated for UK local authority districts?
Although replicating the econometric analyses undertaken by Hidalgo, Hausman and their collaborators goes beyond the scope of this paper, we have carried out some preliminary analyses of the link between economic complexity and measures of local economic performance, and the results are suggestive. We find a strong correlation (p=0.67) between our indicator of complexity and an area’s average annual earnings based on the Annual Survey of Hours of Earnings, which fits with the idea that those local authorities with more complex economies tend to be more productive.
Is this result simply driven by the fact that economically complex areas tend to be predominantly urban, and wealthier? This doesn’t seem to be the case. In the chart below, we have plotted economic complexity against normalised annual average salaries for English Local Authority Districts, using different colours for ‘urban’ (blue) and ‘rural’ (orange) areas based on ONS definitions. The slope of the relationship seems to be positive for both types of areas (although the urban areas tend to be, in general, more economically complex, as shown by the displacement of the blue dots towards the right of the chart relative to the orange ones).
Interestingly, there also appears to be a link (p=0.47) between economic complexity and economic satisfaction, which we proxy through the voting decisions of British Local Authority Districts in the European Referendum (the assumption here, in line with a growing body of literature on the geography of Brexit, is that areas that were more economically fragile and unhappy with the free trade/free movement of people status quo voted to leave). Local authorities in the top quartile of economic complexity were 4 times more likely to vote to Remain than those that didn’t (the actual percentages were 64% and 16%).
What industries tend to appear in more complex areas?
he method of reflections can also be used to quantify the complexity of different sectors – the extent to which they appear in highly complex local economies or not.
In the bar-chart below, we represent industries ranked by their propensity to be present in more economically complex areas, and their share of activity in different local authorities, shaded by the economic complexity of those areas. It shows that creativity and knowledge-intensive services activities tend to be a stronger feature of economically complex areas, although some manufacturing activities such as Optics, Advanced Manufacturing and – interestingly – Manufacture of Leather and Footwear, and Paper also appear in more complex areas. Primary activities such as fishing, mining or forestry are, perhaps unsurprisingly, those that characterise less economically complex areas.
Third step: from describing the present to exploring what might happen
Finally, we have carried out an experimental analysis of the future path of developments of UK LADs using predictive analytics methods: the idea has been to train a machine-learning model on the patterns of change in specialisation in British LADs between 2010 and 2015, and then use that model to predict what might happen to them in the future, based on levels of specialisation in 2015.
The variable we are predicting is whether a local authority will increase its specialisation in its industry, and the predictors are its current specialisations in all industries, as well as the average specialisations of its neighbours. Again, we want to emphasise that this analysis is preliminary – we have not had time, within our pilot, to experiment with many different predictive algorithms, tune them, or explore different types of classification errors across classes. These are all things we plan to do as we progress our work in Arloesiadur in the coming months, hopefully with your feedback.
The initial, preliminary results at the Local Authority level can be explored in the data dashboard we are publishing with this blog (more on that below). The chart below shows the proportion of local authorities in a region or nation that are, according to our model, likely to grow their specialisation in a given sector (again, we have ranked these in decreasing level of complexity).
The chart suggests a stronger propensity for most regions to gain specialisation in more highly complex services industries. Other sectors, such as metallurgy and agrochemical manufacturing are, according to the model, are unlikely to gain much specialisation across the UK, although this is partly linked to the fact that they tend to be based in a small number of locations in the first place (see the thickness of bars in the previous chart on the distribution of industry segments over LADs).
Bringing it all together: An interactive dashboard with information about local specialisation profiles, past growth and future prospects
The results of the analysis we have started exploring in this pilot are granular at the level of UK local authority districts and sectors, and include many variables of interest –past performance, current situation, and future potential. This ‘multidimensionality’ makes our results hard to communicate using static reports and blogs. Yet, reducing this complexity means removing detail that might be useful for local stakeholders who, as we pointed out before, are making decisions and taking actions to spur local economic development that could benefit from this detailed information.
Interactive web technologies can help us manage this trade-off, which is what we have sought to achieve with the companion local industrial profile experimental dashboard you can explore here: the dashboard includes a map with information about the economic complexity of different local authority districts in Britain, and for each of them, a bubble chart representing the recent growth performance of its constituent industries, potential prospects (based on the probabilities of future development generated by our predictive model) and current weight in the local economy. We think that tools like this can help their users to more quickly understand the industrial composition of their economy, help them to explore different future scenarios, and consider interventions to drive economic growth (for example, by encouraging collaboration between related sectors, or trying to develop more diversified economic profiles).
As with everything else in this blog, the dashboard should be considered as a demo or ‘beta’ that we publish, in part, to obtain feedback that helps us to improve it in the run-up to the creation of Arloesiadur from this Autumn (leave comments below or get in touch with us directly at [email protected] or [email protected] ).
Bearing this in mind, we think that the preliminary results we have presented already highlight the potential benefits of using new ways of analysing and visualising official data to inform policies that make UK local economies more diverse, resilient and able to grow.
 There are some differences between the approach adopted by Delgado et al and what we have done in our analysis: within the timeframe of our pilot, we only had time to access public IDBR business-count data from ONS open labor market data NOMIS repository. We therefore didn’t include employment co-location in our analysis. Our reliance on publicly available data also dictated our need to focus on Local Authority Districts as our focus of analysis, instead of using economically functional geographies such as Travel to Work Areas. Finally, Delgado et al include a final phase on their cluster detection process, where they quality assure the composition of their clusters, looking for abnormally allocated sectors and reallocating them by hand. Since we are planning to update our data sources and unit of analysis further down the line, we have not performed this labour and domain-knowledge intensive task. The takeaway is that this analysis is experimental, and the results should be taken with caution, as a ‘demo’ of what is possible to do with this data and methods rather than anything more conclusive.
 Average salary is a rough proxy for productivity.
 We focus on English LADs because there are readily available lookups for their urban-rural classification.
 We could not do this for Northern Ireland because the shapefiles for Northern Ireland (which we would have needed to generate measures of activity in an area’s neighbors) are not available from the ONS Open Geography Portal.
 We did however search for the parameters that generated more accurate predictions with the model that we used, a logistic regression model with L2 regularisation and balanced classes.
 All of this is based on business counts, which as we mentioned, are the main measure of economic activity we have used in this pilot.