The geography of job automation in Ireland: what urban areas are most at risk?

Future automation and artificial intelligence technologies are expected to have a major impact on labour markets. There is a lack of analysis which considers the sub-national geographical implications of automation risk posed to employment. In this paper, we identify the proportion of jobs at risk of automation across all Irish towns, using the occupational methodology of Frey and Osborne (2017) and compare these results with those of the task-based methodology of Nedelkoska and Quintini (2018). The job risk of automation varies significantly across towns, and while there is a substantial difference in the magnitude of risk identified by the occupational and task-based approaches, the correlation between them is approximately 95% in our analysis. The proportion of jobs at high risk (> 70% probability of automation) across towns using the occupational based methodology varies from a high of 58% to a low of 25%. In comparison, the proportion of jobs at high risk using the task-based methodology varies from 26 to 11%. Factors such as education levels, age demographics, urban size, and industry structure are important in explaining job risk across towns. Our results have significant implications for local and regional urban policy development in the Irish case.


Introduction
Technological change is identified as the critical driver of long run growth (Schumpeter 1942;Romer 1990).Yet, the process of technological change, and in particular the change associated with job replacement, creates anxiety and angst within societies (McClure 2017).72% of adults in the USA are worried of a future where robots and computers can do many human jobs (Smith and Anderson 2017).Already, the spatial disparities following technological disruptions are increasingly suggested as the underlying drivers of a 'geography of discontent' (Rodríguez-Pose 2018; Dijkstra et al. 2018;Ballas et al. 2017;Frey et al. 2017;Hudson 2017;McCann 2016).
Existing estimates of how automation may impact workers are wide ranging and some predict extreme disruption.The most widely cited work in the examination of automation on employment by Frey and Osborne (2017) estimated that 47% of US employment is at 'high risk' of being automated in the future.There exists some understanding of the impact of automation at a national level (Frey and Osborne 2017;White et al. 2019;Arntz et al. 2016), a regional level (Crowley et al. 2021;OECD 2018) and at a county and city level (Rijnks et al. 2022;Muro 2019;Frank et al. 2018).But given the significance of the political and social upheaval that may be caused by future technological disruptions, there is an absence of an accurate understanding of the sub-national automation job risk across cities and smaller urbanised areas in regions and countries.Policymakers are consequently making policy responses at the regional and local level in a knowledge vacuum (Leigh and Kraft 2018).
This paper makes two significant contributions to the existing nascent field of regional differences in the spatial distribution of the job risk of automation.Firstly, we use the same automation risk methodology developed by Frey and Osborne (2017) and Nedelkoska and Quintini (2018) at a national level and develop a regionalisation disaggregation method to analyse sub-national variations in regional job risk to automation.There is disagreement in the literature as to whether the occupational approach used by Frey and Osborne (2017) or the task-based approach used by Nedelkoska and Quintini (2018) and Arntz et al. (2016) is the most appropriate way to measure future employment disruptions from automation (White et al. 2019;Frey and Osborne 2018).We overcome this contention in the literature and make a significant contribution to the discussion by applying both an occupational and task-based methodology in this paper.We utilise Irish data to highlight the regional variations in job risk, and specifically, we identify the proportion of jobs at risk of automation across the 200 towns in Ireland, which have a population of 1500 or more using data from the 2016 Irish census.
In our second contribution, we go beyond identifying job risk across towns by delving deeper into what factors explain the variation in job exposure across Irish towns to automation.In doing so, we use an economic geography framework to examine what types of local place characteristics are most likely to be associated with high risk towns.To our knowledge, this is the first paper to assess automation job risk at such a disaggregated sub-national town level and to attempt to explain what factors may explain the exposure of jobs to automation at town level.
The next section provides a review of the existing literature on the job risk of automation and how this can be linked to on-going discussions in economic geography.The precise data used, the automation methodology debate, the regionalisation imputation implemented and the methodologies employed are described in Sect.3. Section 4 reports the results, and the paper concludes with a conclusion and discussion section.

3
The geography of job automation in Ireland: what urban areas… 2 Background and literature review

Automation and job risk
Humans have always developed new and superior products and production technologies to satisfy new wants and needs and to produce greater economic output with less human effort.This is what is meant broadly by technological change, and it is a significant catalyst of long run economic growth (Romer 1990;Schumpeter 1942).In this context, the idea of computerisation, where job automation is by means of computer-controlled equipment (Frey and Osborne, 2017), is not new.New technologies (like computerisation) and new markets have replaced and generated new and more productive jobs and will continue to do so in the future (OECD 2018).Automation technologies have a long history, with inventions like the steam engine, electricity, mechanised weaving looms, industrial robots, or automated teller machines and more recently with information technologies (Acemoglu and Restrepo 2018).From this perspective, robots have been around for a long time.
As far back as Simon (1965), the decline of routine jobs was predicted, where computers were argued to hold the comparative advantage in 'routine' rule-based activities which are easy to specify in computer code.Autor et al. (2003) presented a simple model, referred to as the ALM model, of how the rapid adoption of computerisation by firms is changing the tasks performed by workers at their jobs, which in turn changes the market demand for human skills.For them, robots (machines) are substituting 'routine' tasks but will not substitute non-routine tasks that involve problem-solving, complex social and emotional communication activities and tacit knowledge.Non-routine tasks that cannot be substituted by automation are generally complemented by it (Autor 2015).Routine tasks are tasks that can follow explicit computerised code rules, and by way of simple contrast, non-routine tasks cannot be specified in a computer code (Frey and Osborne 2017).Autor (2015) argues that the tasks that can be automatable are bounded by the limits of human knowledge and what he refers to as Polanyi's paradox: Since humans 'know more than they know they know' (i.e.know things that are difficult to explain as a matter of codified, programmable steps), there are limits to the substitutability of human tasks.Consequently, there will always be a division of 'human tasks' and 'machine tasks'.Susskind and Susskind (2015) have identified key weaknesses of the theoretical ALM model, where tasks that were considered 'non-routine' in the ALM task-based literature have already become 'routine', due to technological changes leading to the likelihood of a more dramatic job displacement reality, from current modern computerised technologies.Big data and artificial intelligence are progressing at incredible speed.Susskind and Susskind (2015) observations pre-empted the findings of Frey and Osborne (2017), who identified the scope of jobs that could be automatable in the future.Frey and Osborne (2017) identified that 47% of jobs in the USA are susceptible to automation, at some unspecified time in the future.They argue their estimate also aligns closely to the extent of job displacement that has happened in the past.For instance, in 1900, 40% of the US workforce was employed in agriculture.Now it is less than 2%.
In general, the grave concern of technological unemployment from theorists such as Keynes (2010) has been unfounded.The process of creative destruction has led to as many new jobs being created in new sectors, as destroyed in traditional sectors.Hence, jobs lost were replaced, with often better paid, more highly skilled jobs in their place.Despite the unfounded unemployment concerns of the past, automation still created widespread disruption with a greater polarisation in jobs and incomes, as technologies (robots and algorithms) replaced, middle-skilled (in particular) people based-routines (OECD 2018).And future disruption is predicted.Schwab (2016) reflecting on the fourth industrial revolution, at the economic summit in Davos stated 'We stand on the brink of a technological revolution that will fundamentally alter the way we live, work, and relate to one another.In its scale, scope, and complexity, the transformation will be unlike anything humankind has experienced before'.

The economic geography of job risk and automation
Job polarisation and technological unemployment are often the most discussed adjustment concerns raised around automation and the rise of artificial intelligence.But an often-overlooked implication and, possibly a greater problem, is the displacement to places created by technological change.The past century has resulted in a large migration of people from rural to urban areas.By 2050, Eurostat (2016) predicts that about 80 per cent of the European population will be living in urban areas.
Urban theorists argue that agglomerations are intellectual breeding grounds for new ideas and innovations (Jacobs 1969;Marshall 1890).The geographical proximity between economic actors helps foster, facilitate and nurture face to face interaction, the flow of local knowledge, ideas and innovations (McCann and Shefer 2005: 302).In theoretical approaches to knowledge flows and innovation, it is assumed that it is easier for knowledge to transfer over shorter distances than longer distances, or as Glaeser et al. (1992Glaeser et al. ( : 1126) ) put it, 'intellectual breakthroughs must cross hallways and streets more easily than oceans and continents'.Given the spikey distribution of prosperity around the world, between countries and within countries, the importance of 'place' in contributing to economic growth has been a central research topic (Florida 2005).
From the literature, it is apparent that urban areas play a critical role as creative and intellectual spaces (Frank et al. 2018).The 'tolerant' and 'bohemian' nature (Florida 2005); the creative spill-overs that are 'in the air' (Marshall 1890); dense diversity (Jacobs 1969); and the creative 'buzz' (Storper and Venables 2004) in urban areas drive innovation resulting in 'swarms of creative clusters' to emerge in global and capital cities (Chapain et al. 2010).Florida's thesis outlines that the disparity in the creative class across cities is driving the disparity in innovation and productivity outcomes across space.Educated, creative, more highly skilled people are migrating to urban areas to exploit amenities, higher wages and employment opportunities in favourable business environments (Moretti 2012;Florida et al. 2011;Marlet and van Woerkens 2007).Patterns of human capital accumulation are following a self-propelling process, where well-educated individuals are attracted to areas that already are concentrated with well-educated individuals (Waldorf 2009).
The geography of job automation in Ireland: what urban areas… Consequently, 'superstar' agglomerations are providing an enhanced ecosystem to drive novel combinations of knowledge and ideas which are further compounding previous geographical divides between regions (Iammarino et al. 2018) and urban and rural areas (Moretti 2012;Rosenthal and Strange 2004) through time.
The shift from manufacturing to predominantly service driven economies has made geographical proximity matter more, resulting in increased divergence between places (Moretti 2012).There is also strong evidence that digital technologies are resulting in greater divisions between large and small urban areas (Muro 2019;Frank et al. 2018).In general, urban areas with a more critical mass are more productive (Melo et al. 2017;Coulibaly et al. 2008), more innovative (Carlino et al. 2007), more accessible to industries and anchor institutions (Hamidi and Zandiatashbar 2019) offer more job opportunities (Hendrickson et al. 2018) and higher wages (Wang 2016; Echeverri-Carroll and Ayala 2011).
But it is also not just a story of 'large' versus 'small' and 'leaders' versus 'laggards'-a much more complex geographical landscape is emerging with brain hubs, declining manufacturing cities and a number of cities that could thrive or decline (Moretti 2012).The reality of shrinking urban areas is not a recent phenomenon (O'Driscoll et al. 2022;Wolff and Wiechmann 2018).It is suggested that economic stagnation resulting from a decline of economic and social opportunities in smaller cities, declining regions and rural areas, was the force behind recent populist voting patterns in the USA (Hendrickson et al. 2018;Frey et al. 2017;Shearer 2016), in France, Germany and the UK (Rodríguez-Pose 2018; Dijkstra et al. 2018;Ballas et al. 2017).In response, many are calling for new frameworks, strategies and policies to tackle the concerns of lagging areas (Rodríguez-Pose 2018; Shambaugh and Nunn 2018;Storper 2018;Iammarino et al. 2018).
This regional disparity may be exacerbated and intensified by an uneven spread of automation risk.Indeed, the aggregate number of jobs at high risk of automation at a national level hides the sub-national regional variation of job exposure.Automation is predicted to have a different impact across regions (OECD 2019).The share of jobs at high risk of automation across OECD defined territorial level 2 regions varies considerably from a low of 4% to a high of 40% (OECD 2018).It is also predicted that automation will result in improved welfare (particularly for better educated workers) in leading regions, and losses in welfare (particularly for lower educated workers) in lagging regions, as production will shift to locations with higher wages and skills, due to more efficient capital over time (Lecca et al. 2018).Indeed, Okamoto (2019) identified that individuals relocate to states with more automation in skilled occupations and less automation in unskilled occupations, demonstrating that automation has a complementary effect on skilled occupations and a substitution effect on unskilled occupations.In effect, if the adoption of present and future automated technologies is rapid, and new firms in new technological industries follow the similar past trends of high knowledge-based firms, by spatial clustering, by locating around successful cities, talent and centres of learning and research (Florida et al. 2011;Venables 2010;Berry and Glaeser 2005;Florida 2005), then it is inevitable that future automation disruptions will further exacerbate regional and local disparities.Present trends in the area of artificial intelligence suggest that the industry is already following a clustering and localised co-location pattern (Klinger et al. 2018).Barzotto and De Propris (2021) identified that firms which collaborate with local suppliers of enabling technologies linked to industry 4.0 have improved levels of performance.
So far, the extent of the automation job risk literature has focused on the level of national exposure (Frey and Osborne 2017), European regional territorial level 2 exposure (Crowley et al. 2021;OECD 2018;Arntz et al. 2016) or US county and city-level exposure (Muro 2019; Frank et al. 2018).While the OECD (2018) and Crowley et al. (2021) papers are significant steps in assessing the job exposure to automation across different regions, such (regional territorial level 2) defined boundaries are unlikely to identify deeper local labour market areas in the European case (Coombes 1995).It is common across countries for development plans and policies to transcend across many different levels of government, and it is typical to have a combination of local, regional and/or national intervention programmes for economic development (OECD 2016a(OECD , 2018)).Consequently, it is important to have a competent empirical understanding of local labour market dynamics (Leigh and Kraft 2018).This paper focuses specifically on ensuring a deeper disaggregated regionalisation of automation job exposure to urban settlements in Ireland to get a better understanding of the localised automated job risk exposure across Ireland.

3
The geography of job automation in Ireland: what urban areas… abilities of different occupations in the workforce.FO created a 'training dataset' using machine learning experts to assess the automatability of 70 occupations using detailed task descriptions.Specifically, they asked the experts to assess whether each task for these occupations was likely to be automated given current knowledge on computerisation capabilities and possibilities.Nine properties were used to assess occupations such as the level of manual dexterity or social perceptiveness.1They then used AI algorithms from the training dataset to examine and create an automation probability for a total of 702 occupations that correspond to the 2010 version of the US Standard Occupational Codes (SOC), making up 97% of the U.S. workforce.We base our analysis on transposing these 702 occupations to the Irish context.
The FO approach to measuring automation has proved contentious.This contention has primarily arisen due to the underlying assumption in their method that entire occupations are at risk of substitution, despite the possibility that tasks may vary within an occupation, at the individual level, across different workplaces and locations (White et al. 2019;Frey and Osborne 2018;Nedelkoska and Quintini 2018;Arntz et al. 2016).To overcome these concerns, the task-based approach adopted by Arntz et al. (2016) and Nedelkoska and Quintini (2018) (hereafter referred to as NQ) uses the OECD's Survey of Adults Skills (PIACC) which contains detailed worker data on how individuals in the same occupations may have different tasks across different workplaces, as well as demographic variables such as gender, education, age and income of workers.For example, construction labourers may have different tasks depending on the type of work being conducted and their individual capabilities, which may vary across different workplaces.The Arntz et al. (2016) study exploits the demographic data where for example an individual's income or education level may increase or reduce their automation risk.Frey and Osborne (2018) criticise this approach asking the question 'why should automation discriminate on the basis of worker characteristics?'The NQ approach did not use worker characteristics and instead applied the FO automation probabilities to the task make-up of occupations at the individual level.NQ created a training data set using a logistic regression and data from the Survey of Adults Skills (PIACC) for Canada, as Canada had a substantially larger sample size than other countries.Firstly, they manually compared the 70 O*NET occupations used by FO with ISCO data.Secondly, they identified occupation task variables in PIACC that best corresponded to the nine engineering bottlenecks identified in FO. 2 Lastly, PIACC has individual level data on occupation tasks allowing NQ to control for within occupation differences that may occur at an individual level across different workplaces, which is the key differentiating factor between the FO and NQ approach.The estimated coefficients from the Canadian training set were then applied to all other individuals across 32 countries in PIACC.Frey and Osborne (2018) criticise the NQ paper for not providing greater clarity on the degree of variability of tasks within occupations, across different workplaces, and question why tasks performed by different workers within occupations would vary that greatly in the first place.
Widely contrasting predictions emerge when the two methods are applied to the entire labour market to predict the proportion of jobs at risk to automation.FO estimated that 47% of jobs in the USA were at high risk of being automated, whereas NQ estimated that 14% of OECD jobs are at high risk.However, NQ also estimated that a further 32% of OECD jobs have a 50-70% likelihood of significant change and the median job is estimated to have a 48% probability of being automated.The key distinction between the occupational versus task-based methodologies lies in the distribution of jobs from low to high risk of automatability.The FO method produces a U-Shaped distribution with extremes at both ends, whereas the Arntz et al. (2016) method produces an inverted U-shaped distribution.The NQ method produces similar inverted U-shaped distributions for countries across the OECD which can be seen in Nedelkoska and Quintini (2018).Despite the differential distributions patterns, what is consistent with both methodologies is the large proportion of jobs that is expected to experience significant change.

Converting US SOC codes to the detailed Irish occupational classifications
In our analysis, as a starting point, we use the probabilities of automation estimated by Frey and Osborne (2017), for 702 occupations in the US case, to identify the risk probabilities associated with occupational classifications in the Irish case.FO's occupational classifications are based on the 2010 version of the US Standard Occupational Codes (SOC).The Irish Central Statistics Office (CSO) bases their occupational classifications on the UK SOC.The US and UK SOC are not directly comparable, and there is no direct conversion available.Therefore, in order to convert the US codes to their UK counterparts (which are approximately identical to the Irish codes used by the CSO) we transform these data using a series of established international classifications.This allows us to perform an analysis of the risk of automation to occupations in Ireland.Once this process has been completed our analysis begins with, what the CSO define as, the detailed occupational classifications for Ireland of which we have FO risk probabilities associated with 273 detailed occupations across 24 intermediate occupation aggregates, in total covering approximately 86% of all jobs in Ireland.Appendix 2 provides a more detailed discussion of the matching procedure and the occupations that cannot be considered in the Irish case.

Description of the Irish data
The most disaggregated Irish data in terms of the scope of occupations are available at a national level.This is referred to by the CSO as the detailed occupations classifications and covers, in our analysis, 273 occupational classifications.This detailed classification is also available at Local Authority level of which there are 31 in Ireland.At these levels, data are available from the Irish censuses of 2016 on the number of people employed in each detailed occupational classification.Therefore, The geography of job automation in Ireland: what urban areas… the exposure of Ireland, and the Local Authorities to automation, can be obtained by identifying the number of people employed across different 'probability of an occupation being automated' groupings.Maintaining alignment with the literature (OECD 2018;Frank et al. 2018;Frey and Osborne 2017;Arntz et al. 2016), we identify jobs at high risk of automation as those with a probability of automation greater than 70%.We further define medium risk as being 50 to 70% probability of automation and low risk as less than 50% probability of automation to align with the methodology employed in Nedelkoska and Quintini (2018). 3t a national level, using FO, we observe a general U-shaped pattern where jobs in Ireland are mainly susceptible to high (44.6%)or low (46.4%)risk of automation with relatively fewer jobs falling into the medium risk category (9%).This U-shaped pattern is largely in line with the findings of FO for the USA.At a national level, using NQ, 18.4% of jobs are at high risk, 8% are at medium risk, and the remainder are at low risk.When our analysis is compared with the NQ analysis of the OECD, the proportion of jobs at high risk is greater in Ireland, but a considerably lower proportion is at medium risk (32% of jobs in OECD are at medium risk).

Regionalising the analysis
In the Irish case, occupational classifications for towns with a population of more than 1500 people are only available at the intermediate occupational classification, of which there are 24 occupational classifications (see Appendix 3 for a detailed discussion of them).The lack of availability of detailed occupational classification is not limited to Ireland as it is common across most countries.
The regionalisation methodology applied in this paper of the FO approach is based on firstly identifying the automation risk for each detailed occupational classification nationally.The method of identifying the automation risk at town level can be summarised as follows (with a detailed example provided in Appendix 4): 1. Use Frey and Osbornes (2017)  The process for regionalising the NQ approach is more straightforward.The authors were provided by NQ with the proportion of the workforce at high and significant risk of automation at ISOC 2 digit level.This was mapped to the Irish intermediate SOC classifications, and then the number of people in each town employed in each intermediate occupational class was multiplied by the NQ proportions at high and significant risk of automation.
Therefore, at the Irish town level, all analysis is conducted based on aggregate intermediate occupational classifications regardless of whether we consider the FO or NQ approach.More routine-based occupations are at greater risk and for a breakdown of risk by occupation across the intermediate occupational classifications, see Appendix 5. We conduct checks to ensure our regionalisation approach is robust (which is outlined in Appendix 6 for interested readers).

The impact of automation on towns in Ireland
In total, we examine 200 Irish towns (including Ireland's five cities: Dublin, Cork, Limerick, Galway and Waterford) with a population over 1500 people.A map showing the distribution of these towns across Ireland and the spatial autocorrelation of automation risk at town level is presented in Appendix 7. The average percentage of jobs at high risk of automation across Irish towns using the FO approach is 44%, compared to 18% using the NQ approach, highlighting a large gap in average exposures between the different methodologies, which is consistent with this literature (Frey and Osborne 2018).However, the correlation between the FO and NQ proportion of jobs at high risk across towns is 0.9524 (with an associated p-value of 0.0000).The key observed difference is the downward shift in risk exposure associated with the NQ approach, relative to the FO approach.
Separately, a spatial dependence Moran's I plot is estimated to identify significant town risk clusters of the proportion of jobs at high risk of automation, using the FO method.The results of the local spatial autocorrelation plot are presented in Map A and B in Fig. 1.Significant clusters (Map B) of at lower risk towns can be mainly found in the greater Dublin city region (dark blue circles), and separately, significant clusters of at higher risk towns (light green circles), can be found in the West and South-West of Ireland.More detail on this analysis is provided in Appendix 7.

3
The geography of job automation in Ireland: what urban areas…

Modelling the factors which impact exposure of Irish towns to risk of automation
We deepen our analysis by considering some factors which may determine the proportion of jobs at high risk of automation.We estimate Eq. ( 1) for the 200 Irish towns for which we have generated job risk data.
where AutomationRisk i is the proportion of jobs in town i at high risk of automa- tion.The IndependentVariables include: urban population size dummy variables as per Table 1; the age profile of town i dummy variables; the proportion of the population in town i aged 24 or over with a third level qualification; the proportion of the population employed in a given sector in town i; the proportion of the population in town i that are Irish; and the employment rate calculated as the number of people who indicate that they are employed, divided by the number of people who indicate that they are employed, plus the number of people unemployed, plus the number of people seeking their first job.The data are derived from the Irish Census 2016.Table A8.1 in Appendix 8 provides descriptive statistics of all variables utilised.Appendix 9 presents a variance inflation factor (VIF) table from the estimation of Eq. (1) to test for multicollinearity.The mean VIF is below 5 indicating that multicollinearity isn't an issue.5 Appendix 10 presents a correlation matrix (1)   The geography of job automation in Ireland: what urban areas… of our variables.We estimated a number of models as part of a sensitivity analysis including variations of estimation techniques and variations of the dependent and independent variables.These sensitivity analyses are outlined in Appendix 11 for interested readers, and the results are consistent and robust across different model specifications.Moving forward, we report the results from the ordinary least squares regression.

Results
Table 1 presents a comparison of results for Eq. ( 1) for the FO occupational approach and the NQ task-based approach.As can be identified, with respect to significance levels, the results are similar across both models, with differences occurring in relation to some industry dummies, the diversity measure and a population size dummy.A further distinctive difference in the two outputs can be identified with respect to the magnitude of the coefficients, where the coefficients in the NQ model are generally smaller than the coefficients for the FO model.This is not surprising since the key difference identified earlier in the outcome of the two approaches is the proportion of jobs at high risk, where the NQ approach results in lower proportions, relative to the FO approach.Despite these minor differences, it can be ascertained that the factors associated with automation exposure at town level are largely not dependent on what type of measurement approach is applied.Consequently, we proceed with reporting the main results from the FO output.Firstly, we find that a higher proportion of third level graduates in a town reduces the proportion of jobs at high risk of automation.It would be expected that places endowed with a greater proportion of educated workers are more likely to create and attract more highly skilled businesses and jobs, which would be less exposed to automation.
Towns with a higher proportion of the population in the 25 to 44 age bracket have the greatest exposure to automation risk.This is followed by the 45 to 64 age category and then the 19 to 24 age category.We conclude that towns with younger populations are generally less exposed to automation risk than towns with older populations. 6The most dramatic pattern often described in more peripheral areas is the hallowing out and flight of young talent from more remote areas (OECD 2016b;Storper 2018;Hendrickson et al. 2018;Henning 2019).
It can be identified from the results that industrial structure is a key determinant of automation risk, which corresponds to similar findings in the literature (Crowley et al. 2021;OECD 2018;Frey and Osborne 2017).In the FO model, towns with a greater proportion of employment in mining and quarrying (relative to manufacturing) are the least exposed to automation.This is followed by the arts, entertainment and recreation sector (relative to manufacturing).Areas with significant employment in transport and storage, health and support services, public administration and defence, accommodation and food related services, construction and wholesale and retail trade related jobs are also less exposed to automation, relative to areas more exposed to a manufacturing industrial base.Towns with a greater proportion of employment in agriculture are more susceptible to automation, relative to areas more exposed to manufacturing.Surprisingly, towns with higher levels of unemployment are less exposed.This may be a signal that some towns are already struggling with previous automation disruptions and are in a transition phase and are in effect less exposed to future automation as they are already dealing with past automation disruption (we note that this would be worthy of more detailed analysis in the future, using time series data).
We observe in the FO output that towns with a population greater than 5000 people and less than 10,000 people are most exposed to automation, relative to our base category (1500-4999 population).We note that there is no other significant difference associated with urban size.This implies that larger towns (> 10,000) and smaller towns (< 5000) have similar exposures to automation risk.Mid-sized towns, which may not be able to benefit from advantages of critical mass accruing to larger 1 3 The geography of job automation in Ireland: what urban areas… areas or may not be able to benefit from specific niche industries as smaller towns, are most at risk.7

Conclusions and discussion
This paper has presented an analysis of the job risk of automation at town level in Ireland using Irish Census data.In doing so, it is the first study to provide a detailed sub-national town analysis of automation risk, comparing the Frey and Osborne (2017) occupational methodology with the Nedelkoska and Quintini (2018) taskbased methodology in the literature.We identify that the occupational approach estimates Irish towns to be harder hit from automation, relative to the task-based approach.Despite the large differences in the anticipated disruption between the two methodologies, we get an accurate indication of the relative risk of automation disruption to different towns, since the methodologies have almost a perfect correlation in terms of town rankings.Furthermore, the associated factors explaining the level of exposure using both approaches are similar.We find that an educated workforce, age demographics and sectoral differences will be essential in explaining a town's resilience and adaptation to future technological change.
Ireland's weaker regions are more vulnerable to the labour substitution implications of automation and artificial intelligence.A greater proportion of the towns at higher risk and associated high risk clusters are predominantly located in the lagging provincial regions around the Border area, South-West and West of the country.These regions are dwarfed by the relative size of the capital city region and contain much weaker agglomeration, concentration and diffusion effects and are suffering from significant comparative disadvantages.As identified here, these mainly stem from industrial structure, education, and age demographic differences which are driving the geography of job automation disparities on the island.This is consistent with brain drain or skills trap patterns identified in the U.S. (Okamoto 2019) and across OECD countries (OECD 2016b) with upwards-directed labour filtering effects evident in the regional hierarchy of other developed economies (Henning 2019).Future automation and artificial technologies are predicted to hit already lagging areas harder.Present and past national and local institutional settings and policies in Ireland have failed to dampen significant regional inequalities from occurring.We, therefore, support the recent international calls for new frameworks, strategies and policies to tackle the concerns of lagging areas (Crowley et al. 2021;Rodríguez-Pose 2018;Shambaugh and Nunn 2018;Storper 2018;Iammarino et al. 2018) to be considered carefully by policymakers in the Irish case.
While this study provides a rare insight into the regional employment disparities associated with automation, it is worth highlighting several factors which could not be considered.Firstly, detailed occupational data are not available at a low level of spatial disaggregation.It should be noted that this is common for occupational data.Therefore, we rely on intermediate data when we consider the town level analysis.
We have shown this to be highly correlated (in excess of 0.95 correlation coefficient) with the detailed occupational classification analysis, but it remains an imperfect limitation (which is not surmountable).However, our approach highlights a mechanism of disaggregating national data to regional level to provide rich insights into spatial disparities.And, given the importance of these spatial disparities in explaining voting patterns, evolving income inequality, etc., a robust approach to analysing regional disparities, along the lines of those conducted in this paper, is essential.
In this paper, the FO and NQ approaches produce wide ranging predictions of the proportions of jobs at risk to automation in the labour force at town level.The concerns previously highlighted in the literature between these contrasting approaches can also be extended to the regional analysis presented here.FO and NQ make valid points in defence of their methodologies and the true picture of absolute automation risk probably lies somewhere between the predictions produced by these methodologies.Furthermore, the rapid advancement of technology capabilities will not solely explain future long-run employment patterns as job risk exposure will also be affected by local and international dynamics stemming from the availability of cheap labour; migration flows; the social, ethical, political and institutional regulatory environments; and the market demand for robot substitution (Frank et al. 2018).Nevertheless, if we take the absolute risk of automation predicted with a degree of caution, applying either the FO or NQ approach to a regional analysis will garner a reliable indicator of relative risk of automation disruption.
A further consideration is that there are other ways of measuring automation risk (Dengler and Matthes 2018).By using the methodologies of Frey and Osborne (2017) and Nedelkoska and Quintini (2018), we examine automation exposure from the perspective of technology experts.Recently, Dengler and Matthes (2018) also used a task-based approach using occupational experts to assess automation risk for tasks specific in the German case.Replicating the Dengler and Matthes (2018) methodology to a disaggregated regional analysis would have merit.

Fig. 1
Fig. 1 Maps of local spatial autocorrelation.Map A: spatial dependence of towns (significant and insignificant).Map B: spatial dependence of towns (significant) At the national level calculate the proportion of the individuals in each of the 24 intermediate occupational classification at high, medium and low risk of automation by aggregating the risk of the 273 detailed occupational classification, into the 24 intermediate occupational classifications; 4. At the town level multiply the national values obtained in (3) of the proportion of the individuals in each of the 24 intermediate occupational classification at high, medium and low risk of automation by the number of people employed in each occupation in that town; 5. Based on the number of people identified at high risk of automation in (4), we calculate the proportion of the workforce employed in high risk occupations for each town. 4

Table 1
Results for FO and NQ approach to estimation of Eq. (1)