Convenient primary care and emergency hospital utilisation

Participation and utilisation decisions lie at the heart of many public policy questions. I contribute new evidence by using hospital records to examine how access to primary care services affects utilisation of hospital Emergency Departments in England. Using a natural experiment in the roll out of services, I first show that access to primary care reduces Emergency Department visits. Additional strategies then allow me to separate descriptively four aspects of primary care access: proximity, opening hours, need to make an appointment, and eligibility. Convenience-oriented services divert three times as many patients from emergency visits, largely because patients can attend without appointments. © 2019 The Author. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). For some services discrepancies between social and individual benefits warrant government action on efficiency grounds. In other cases, society may intercede to ensure individuals can access some hitherto unattainable level of service. Interventions to improve the accessibility of services conceivably come in many guises, for instance improving affordability or widening eligibility; providing more, closer, or better services; shorter waiting times; or more convenient opening hours (e.g. Millman et al., 1993; Hiscock et al., 2008). The ways in which interventions are designed and structured may have consequences for utilisation, service costs, and the attainment of policy objectives. This paper investigates how dimensions of access to primary care affect the demand for unplanned use of hospital Emergency This paper was part completed while I was working at the Centre for Economic Performance which is funded by the ESRC under grant number ES/M010341/1. My thanks to the editor and two anonymous referees for valuable suggestions, and to Gwyn Bevan, Philippe Bracke, Peter Dolton, Steve Gibbons, Christian Hilber, Alex Jaax, Alistair McGuire, Henry Overman, Sefi Roth, and Olmo Silva as well as participants at the LSE/Spatial Economics Research Centre seminar, the Royal Economic Society conference, and staff at NHS Improvement for helpful comments, or other input to earlier versions of this work. I am grateful to Professor Peter Griffiths and Trevor Murrells for generously providing ancillary data. This paper was produced using Hospital Episode Statistics provided by NHS Digital under Data Sharing Agreement NIC-354497-V2J9P. This paper has been screened to ensure no confidential information is revealed. E-mail address: edward.pinchbeck@city.ac.uk Departments (EDs). I draw on Equitable Access to Primary Medical Care (EAPMC), a policy reform in the English National Health Service (NHS) designed to make primary care more convenient across the country, and to address geographical imbalances in access. Under EAPMC, around 250 new primary care services were deployed between 2008 and 2012. More than half were “walk-in clinics”: practices with evening and weekend opening hours, and offering walk-in services with no need to register or make an appointment. The remainder, targeted to administrative districts with the lowest concentration of primary care physicians, were “extended hours practices”: regular services requiring registration but open at least 5 hours per week more than conventional practices. The comprehensive nature of the English NHS, where all patients have access to free primary care, allows me to abstract from insurance issues, and to focus on physical proximity and other less well-understood, but potentially important, convenience dimensions of access. To contrive a quasi-experimental research design from the EAPMC policy reform I use hospital records to capture the evolution of hospital utilisation in small neighbourhoods, then generate a measure of primary care access as a non-parametric function of distance to EAPMC services.1 Restricting regression samples to places receiving new facilities under the policy, specifications estimate 1 More concretely access intensity is computed by counts of open services in a series of distance buffers centred on the neighbourhood centroid, where distance https://doi.org/10.1016/j.jhealeco.2019.102242 0167-6296/© 2019 The Author. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). 2 E.W. Pinchbeck / Journal of Health Economics 68 (2019) 102242 an average treatment effect on the treated (ATT) from changes in hospital outcomes when an EAPMC service opens or closes, with a control group composed of areas suitable for similar services but not experiencing access changes at that particular time. Using timing differences for identification is underpinned by evidence that: (i) service roll-out is unrelated to pre-reform primary care access measures; and (ii) trends in ED visits are broadly parallel across cohorts. This aligns with policy documents that indicate service deployment timetables were driven by administrative factors that are plausibly unrelated to the determinants of hospital utilisation. This research design is leveraged to generate three sets of findings. The first documents policy-relevant estimates of the impact of walk-in clinics on neighbourhood wide ED use. Conditional on fixed neighbourhood factors, labour-market trends, and demographic changes, proximity to these convenience-oriented services results in strongly significant reductions in unplanned ED visits. Reductions in ED visits are in the order of 1.5–4%; implying that each facility reduces annual ED throughput by approximately 1000–2000 visits. The robustness of these estimates is bolstered by auxiliary analyses, the use of alternative sources of variation, as well as numerous robustness and falsification tests. Parameter estimates imply that some 5–20% of walk-in clinic visits substituted for a trip to an ED. Despite the lower costs of primary relative to ED care, this implies a net increase in health care spending (in the region of £ 10–20 per walk in visitor), but says nothing about possible patient benefits or diversion from regular primary care services. The former, which include anxious patients being able to consult with primary care physicians promptly when faced with an uncertain need for care, may be considerable given that many services proved extremely popular. A full welfare analysis, which lies beyond the scope of the current paper, would need to account for these benefits. Regarding the latter, NHS primary care physicians are paid a capitation fee per registered patient so the analysis undertaken is a helpful guide to budgetary implications. A second suite of results exploits the richness in my data to unearth further patterns. First, using an event study approach I trace out the time dynamics of ED diversion from time of first exposure. This is inconclusive: a test rejects equality of the post exposure event time indicators, yet there is no clear pattern nor a significant linear trend in effects. The spatial dimension in the data, however, reveals a much sharper relationship: diversion from EDs is subject to a strong, near-linear, decay with distance to a walk-in clinic. Turning next to characteristics of ED visits, subsequent findings indicate that diversion from EDs is largely driven by patients whose visit does not result in a hospital admission, and by patients that were neither referred to the ED nor conveyed there in an ambulance. This points towards a conclusion that the effects of walk-in clinics mainly arise from influencing care utilisation decisions of individuals with less urgent health problems, and with more discretion over the location of their treatment. The third and final part of the analysis unpicks further channels though which primary care access determines ED utilisation. Here I rely on a descriptive approach that compares walk-in clinics and extended hours practices in under-doctored administrative districts which received both types of service under EAPMC. When estimated simultaneously, walk-ins divert three times as many patients from EDs. Although both types of service have meaningful effects on ED visits outside of standard practice hours, the greater bite of walk-in services predominantly occurs during these standard hours. To the extent that services are well matched on unobserved features, this suggests that being able to attend withbuffers vary across space based on the distribution of distances travelled to access emergency care locally. out registering or pre-booking strongly influences where patients seek treatment. This paper’s overarching contribution is to provide new evidence on the extent to which convenient primary care reduces visits to hospital EDs. Shifting care from EDs to primary care is likely to be socially beneficial because some 15–40% of ED visits are for health problems that could be safely treated in less costly settings outside hospitals (Mehrotra et al., 2009; Weinick et al., 2010; Lippi Bruni et al., 2016). Moreover, rapid growth in ED use in many OECD countries has resulted in well-documented congestion in EDs (Berchet, 2015).2 The possible adverse effects of crowding has made reducing pressure at EDs an increasing priority (e.g Pines et al., 2011; Morley et al., 2018). In the English NHS, various initiatives have been adopted, or are proposed. Primary care access features prominently. For example, prior to the 2015 election both major political parties put forward access policies in expectation that reduced pressure at EDs would follow (Cowling et al., 2015). Since that time the government has introduced a 7-day primary care policy, and is currently rolling out Urgent Treatment Centres (UTCs) across the country.3 Despite this, the evidence available to inform policy decisions is far from clear-cut. A review by Ismail et al. (2013) cautions against using evidence from studies prior to 2011, and findings from more recen

* City, University of London and Centre for Economic Performance. Address: City, University of London, Northampton Square, London, EC1V 0HB. Email: edward.pinchbeck@city.ac.uk. My thanks to Gwyn Bevan, Philippe Bracke, Peter Dolton, Steve Gibbons, Christian Hilber, Alex Jaax, Alistair McGuire, Henry Overman, Sefi Roth, and Olmo Silva as well as participants at the LSE/Spatial Economics Research Centre seminar and the Royal Economic Society conference for helpful comments, suggestions, or other input to earlier versions of this work.
For some services discrepancies between social and individual benefits warrant government action on efficiency grounds. In other cases, society may intercede to ensure individuals can access some hitherto unattainable level of service. Interventions to improve the accessibility of services conceivably come in many guises, for instance improving affordability or widening eligibility; providing more, closer, or better services; shorter waiting times; or more convenient opening hours (e.g. Millman et al., 1993;Hiscock et al., 2008).
The ways in which interventions are designed and structured may have consequences for utilization, service costs, and the attainment of policy objectives, particularly given mounting evidence that psychological factors can lead seemingly minor service features to play a part in utilization and participation decisions (e.g. Duflo et al., 2006;Bertrand et al., 2006;Baicker et al., 2015). This paper investigates how dimensions of access to primary care affect the demand for unplanned use of hospital Emergency Departments (EDs). I draw on Equitable Access to Primary Medical Care (EAPMC ), a policy reform in the English National Health Service (NHS) designed to make primary care more convenient across the country, and to address geographical imbalances in access. Under EAPMC, around 250 new primary care services were deployed between 2008 and 2012. More than half were "walk-in clinics": practices with evening and weekend opening hours, and offering walk-in services with no need to register or make an appointment. The remainder, targeted to administrative districts with the lowest concentration of primary care physicians, were "extended hours practices": regular services requiring registration but open at least 5 hours per week more than conventional practices. The comprehensive nature of the English NHS (where all patients have access to free primary care), allows me to abstract from insurance issues, and to focus on physical proximity and other less well-understood, but potentially important, convenience dimensions of access.
To contrive a quasi-experimental research design from the EAPMC policy reform I use hospital records to capture the evolution of hospital utilization in small neighborhoods, then generate a measure of primary care access as a non-parametric function of distance to EAPMC services. 1 Restricting regression samples to places receiving new facilities under 1 More concretely access intensity is computed by counts of open services in a series of distance buffers centered on the neighborhood centroid, where distance buffers vary across space based on the distribution of distances traveled to access emergency care locally. the policy, specifications estimate an average treatment effect on the treated (ATT) from changes in hospital outcomes when an EAPMC service opens or closes, with a control group composed of areas suitable for similar services but not experiencing access changes at that particular time. Using timing differences for identification is underpinned by evidence that: (i) service roll-out is unrelated to pre-reform primary care access measures; and (ii) trends in ED visits are broadly parallel across cohorts. This aligns with policy documents that indicate service deployment timetables were driven by administrative factors that are plausibly unrelated to the determinants of hospital utilization.
This research design is leveraged to generate three sets of findings. The first documents policy relevant estimates of the impact of walk-in clinics on neighborhood wide ED utilization. Conditional on fixed neighborhood factors, labor-market trends, and demographic changes, proximity to these convenience-oriented services results in strongly significant reductions in unplanned ED visits. Reductions in ED visits are in the order of 1.5 -4%; implying that each facility reduces annual ED throughput by approximately 1,000 -2,000 visits. The robustness of these estimates is bolstered by auxiliary analyses, robustness and falsification tests.
A second set of results exploits data richness to show that diversion from EDs is subject to a strong, near-linear, decay with distance to a walk-in clinic. Further findings indicate that diversion from EDs is largely driven by patients whose visit does not result in a hospital admission, and by patients that were neither referred to the ED nor conveyed there in an ambulance. This points towards a conclusion that the effects of walk-in clinics mainly arise from influencing care utilization decisions of agents with less urgent health problems, and with more discretion over the location of their treatment.
The third and final part of the analysis unpicks further channels though which primary care access determines ED utilization. Here I rely on a descriptive approach that compares walk-in clinics and extended hours practices in under-doctored administrative districts which received both types of service under EAPMC. When estimated simultaneously, walk-ins divert three times as many patients from EDs. Although both types of service have meaningful effects on ED visits outside of standard practice hours, the greater bite of walk-in services predominantly occurs during these standard hours. To the extent that services are well matched on unobserved features, being able to attend without registering or pre-booking strongly influences where agents seek treatment.
This paper provides new evidence on the extent to which convenient primary care reduces visits to hospital EDs, and by using a national policy reform and the universe of hospital visits, I build on a small literature that obtains plausibly causal estimates in narrower empirical settings (e.g. Dolton and Pathania, 2016). Shifting care from EDs to primary care is likely to represent a social benefit because some 15 to 40% of ED visits are for health problems that could be safely treated in less costly settings outside hospitals (Mehrotra et al., 2009;Weinick et al., 2010;Lippi Bruni et al., 2016). Moreover, strong recent growth in ED use in many OECD countries (Berchet, 2015) has found many hospitals operating above capacity and resulted in well-documented congestion in EDs, such that reducing pressure at EDs has become an increasing priority (Rust et al., 2008;Pines et al., 2011). 2 Findings indicate that EAPMC walk-in clinics significantly reduced ED activity in areas of England with and without primary care shortages. The ATT estimates are directly relevant to any ex post evaluation of the EAPMC policy, and may generalize to other settings in which policy-makers are seeking to reconfigure existing services or expand primary care in suitable locations. Coefficients imply that 5-20% of walk-in clinic visits substituted for a trip to an ED. Despite the lower costs of primary relative to ED care, by itself this implies a net increase in health care spending, but says nothing about patient benefits or diversion from regular primary care services. A full welfare analysis would need to take account of these factors. However, NHS primary care physicians are paid a capitation fee for each registered patient so this analysis is a relevant metric in understanding the effects of the policy on the NHS budget. This paper's second more general contribution is to adopt a multi-dimensional view of health care access, and to demonstrate that several dimensions simultaneously drive health care utilization patterns. Related research has typically focused on individual components of access in isolation -for example affordability (Selby et al., 1996); service opening hours (Dolton and Pathania, 2016;Whittaker et al., 2016); or physical proximity (Van Dort and Moos, 1976;Currie and Reagan, 2003;Buchmueller et al., 2006). The results documented here suggest that a wider view is warranted, and could help to resolve puzzles and anomalies in care utilization patterns. For example, as Chen et al. (2011) 2 Between 1995 and 2010 visits to US EDs increased by 34% (National Center for Health Statistics, 2013), while visits to Accident & Emergency departments in the England rose by 40% (Appleby, 2013). acknowledge, the availability of primary care physicians may be behind heterogeneity in how reforms that alter affordability lead to changes in ED utilization and substitution across care settings. 3 Finally, the paper is related to a growing literature that allows for behavioral factors in models of individual participation and utilization decisions (e.g. Mullainathan et al., 2012). Baicker et al. (2015) document numerous examples where psychological factors plausibly result in underutilization or overutilization of health care. In this paper, I make a start in applying these insights to agents' choice of treatment setting. Proximity and the ability to attend without appointment are important factors in determining the extent to which primary care diverts agents from EDs, which chimes with evidence from other settings that inconvenience and hassle can be powerful barriers to participation (e.g Bertrand et al., 2006;Kahn and Luce, 2006 Consultant-led 24 hour services with full resuscitation facilities catering for all kinds of emergency (equivalent to Emergency Departments, or EDs); or consult with a family physician -know locally as a General Practitioner (GP) -at a primary care practice. It is widely acknowledged that providing care in EDs is considerably more costly than in settings outside hospitals such as physicians' offices (e.g Mehrotra et al., 2009). This is also true in the NHS context: for example, recent figures indicate that a visit to A&E costs £124 while a GP practice consultation costs £32.
Despite universal coverage and no demand-side cost sharing, patients using NHS emergency care incur time and travel expenses. In addition, EDs and primary care are subject to access frictions. In this regard EDs are arguably more convenient than conventional primary care: patients can visit any ED whenever they wish, and due to closely monitored performance targets, can normally expect to wait less than 2 hours for treatment.
Conversely, access to specific primary care services requires registration, and is usually only available to patients living within a practice's catchment boundary. Access is via an appointment, an emergency appointment, or -where available -by using a primary care Out of Hours service on evenings and weekends. Although almost all individuals are registered at a primary care practice, they may have to wait a week or longer to obtain a regular appointment with a family doctor; and, although often available, same day emergency appointments can be difficult to book. Even then, appointments may not be convenient. 4 From the late-1990s, alternative ways to access unplanned care emerged in the shape of new urgent care services designed for patients with minor medical problems. These included a telephone advice service and facilities offering easy access to face-to-face advice and treatment. NHS Walk-in Centres are one type of urgent care service that were introduced in this period. 5 These facilities provide routine and emergency primary care for minor ailments and injuries with no requirement for patients to pre-book an appointment or to register (Monitor, 2014). Most are located away from hospitals although some are co-located with hospital EDs, so that on arrival patients are directed (triaged) to the appropriate service. In total approximately 230 Walk-in Centres have opened in England since 2000. Some 150 (or 65%) of this total number were commissioned following a report in 2007 that led to the creation of the Equitable Access to Primary Medical Care (EAPMC ) policy reform.
EAPMC was set-up with the twin objectives of delivering more personalized and responsive primary care across England, and improving access in the most under-doctored areas. To meet these objectives EAPMC comprised two discrete initiatives. The first funded 100 new primary care practices in the 38 Primary Care Trusts (PCTs) with the 4 For example, surveys indicate the average wait to get a GP appointment is around 13 days Pulse (2016). Average waiting time for GP appointment increases 30% in a year, June 10. In the July 2017 GP Patient survey, 32% of patients did not find it easy to get through to their practice by phone; 29% were not able to see or speak to someone at the time they wanted; 31% who wanted a same day appointment could not get one; 24% say that their practice is not open at times that are convenient for them; and only two thirds of patients rate their overall experience of out-of-hours NHS services as good.
5 Others include Urgent Care Centres and Minor Injury Units, both of which usually do not provide primary care services. See Monitor (2014) for a review. lowest provision of family doctors. 6 These practices were similar to conventional primary care services already available, but had to meet certain core criteria such as having at least 6,000 patients and being accredited training practices. They were also required to facilitate access opportunities through extended opening hours, with a minimum of 5 hours per week beyond Monday to Friday 8.30am-6.30pm, and by setting large catchment boundaries (Department of Health, 2007) (see Appendix A for the full list of criteria). I refer to these services henceforth as "extended hours practices". The second strand of EAPMC compelled each of the 152 PCTs to establish a "GP-led Health Centre", a new service type designed to offer more convenient access to primary care. These facilitieswhich I refer to throughout as "walk-in clinics" -had to offer both a regular registered primary care service with bookable appointments, as well as a walk-in service for any member of the public from 8am-8pm, 365 days a year. Core criteria required the centers to be located in areas maximizing convenient access and opportunities to integrate with other local services (Department of Health, 2007).    had opened before May 2009, more than two thirds by the end of 2009, and all but two before 2011. What drove this pattern of deployment? Guidance issued by the Department of Health highlights that local administrators were under pressure to establish the new services quickly, with an expectation that all procurements should be finished in financial year /9 (Department of Health, 2007. Although some did meet this timetable, many others did not, with sources suggesting that deployment timings were mainly driven by local administrative factors, for example readiness on the part of administrators to specify services and identify suitable premises, the speed of procurement processes, and the time needed to prepare sites.

Primary care and ED utilization
In standard formulations, individuals seek care when the private costs of obtaining treatment p are lower than the perceived private treatment benefits b(σ), which are increasing in illness severity σ. From a social perspective, treatment is warranted when private benefits are higher than the social costs of treatment c; such that when p < c there is moral hazard and some over-treatment. When EDs and primary care are substitutes, patients seek treatment in primary care when they perceive a net private benefit from primary care (b P C − p P C > 0); and when primary care offers a higher perceived net benefit than an ED (b P C − p P C > b ED − p ED ). A more general case allows for behavioral biases and is set out in Appendix B.
Primary care access interventions reduce private costs of primary care, either through lowering co-pays or, as in the English NHS, by reducing time costs and travel expenses.
Any such intervention can have an effect at the extensive margin by inducing marginal agents to utilize primary care instead of not seeking any kind of care. Additionally, through the second condition an intervention can divert patients from EDs to primary care. The stylized facts presented in Section 1.1 suggest the following. First, EDs can treat all patients, but those with illness severities above some pointb(σ) P C are not treatable in primary care. Second, ED care is available at any time but primary care can only be accessed during practice opening hours. Third, treatment costs in EDs are strictly higher than primary care (c ED < c P C ). These stylized facts predict that following an increase in primary care access: (i) agents with less severe medical problems should be Later analysis uses micro-data to estimate the extent to which EAPMC services divert patients from EDs. This is warranted because the aggregate utilization data depicted in Figure 3 is inconclusive. In the period when EAPMC services were being deployed (2008/9-2010/11), visits to walk in clinics and other urgent care services for minor problems (denoted "Type 3 units") rose steadily while ED visits (denoted "Type 1 Departments") remained flat. These trends could be consistent with effects purely at the extensive margin i.e. walk in clinics meeting previously unsatisfied demand. 7 However, the same outcome can also arise when walk-ins substitute for ED care. In the limit aggregate demand for emergency care is perfectly inelastic, such that all else equal every clinic visit is offset by one less visit to an ED. Under such conditions, the aggregate trends in Figure   3 could reflect unrelated shifts in emergency care demand, for example from, say, an aging population or increased patient expectations. Hospital activity data is drawn from two main sources: the Quarterly Monitoring of Accident and Emergency (QMAE) dataset published by NHS England, and Hospital Episode Statistics (HES) records provided by HSCIC. QMAE was the official source of information on ED activity in the period 2009 to 2012, and is generally considered to be the most comprehensive and reliable source of aggregate information on emergency care activity. It captures aggregate ED visit counts at the NHS Trust, rather than the site, level. For most NHS Trusts this is inconsequential as there is only one ED, but some NHS Trusts have multiple emergency care sites, in which case the split of attendances across sites cannot be observed. To account for mergers, I group together earlier data for NHS Trusts which will eventually merge in order to create a consistent panel.
For the neighborhood level analysis, three hospital utilization variables are derived from two distinct HES data resources. Both contain anonymized patient records, and include the patient's residential location (LSOA) as well as details of care received. The first and principal utilization measure records unplanned visits to hospitals: (1) the total number of visits to hospital EDs. Two further measures relate to admissions to hospital: (2) the total number of admissions, and (3) the proportion of unplanned admissions that could potentially have been avoided with appropriate primary care. 8 The source for the second and third variables is the HES Admitted Patient Care dataset while the first is by necessity drawn from the (separate) HES A&E dataset. This distinction is important because Admitted Patient Care data contain detailed diagnosis information but A&E data do not. This omission precludes analysis of ED visits by the categories used in Taubman et al. (2014) i.e. "Non-urgent," "Urgent, primary-care treatable," etc.
The HES A&E dataset is a rich source of data on ED activity but was published as experimental statistics until 2012/13. The use of these data to compute ED visits is challenging because in early iterations of the data collection health care service providers were not strictly required to record the type of emergency unit that a patient attended (for example an ED or another type of emergency care facility, such as an eye hospital or Minor Injury Unit). Completing this field in the data then subsequently became mandatory. As a result emergency unit type codes are missing for close to 30% of patient records for NHS Trusts in 2009/10. The share of missing codes then falls to around 11% in 2010/11, 3.5% in 2011/12, and 1.5% of records by 2012/13, a trend depicted in the series of bars labeled 1 in Figure 4.
The implication of this is that changes in ED visits observed in the raw data between 2009 and later years will in part reflect better coding practices rather than genuine ED activity changes. This is problematic as better coding coincides strongly with the introduction of EAPMC services. I circumvent this problem in two steps, which are visually illustrated in Figure 4. First, I exploit that the QMAE data described above indicates that some hospital-quarter cells only contain ED attends whereas others contain only non-ED attends. Cross-referencing to QMAE thus allows me to impute true type codes for more than half of the the uncoded NHS Trust attendances in the HES A&E dataset in my sample window. Nevertheless, as depicted in the second series of bars in Figure 4, a substantial number of missing codes remain. (2) after imputing missing codes using QMAE; (3) after dropping quarter-neighborhood cells that contain fewer than 50 ED visits; (4) as (3) but retaining a balanced panel of neighborhoods with non-missing data in each quarter.
Second, after removing duplicate records and collapsing the data to quarter-neighborhood cells, I then exclude any quarter-neighborhood cells that contain fewer than 50 ED visits from the final estimation sample. As shown in the third (the resulting unbalanced neighborhood panel) and fourth (the associated balanced neighborhood panel) series of bars in Figure 4, this strategy is effective in eliminating missing codes problem because it reduces the number of uncoded emergency care visits to inconsequential levels. However, and while this strategy is unlikely to be a source of bias, it potentially raises generalizability concerns as results may be specific to places with high ED use. Later findings that indicate a close correspondence between the neighborhood and NHS Trust level results, as well as an alternative strategy detailed in full in section 3.6, give reassurance that this is not the case.

Empirical Approach
My data constitutes two quarterly panels of hospital utilization measures outcomes at the NHS Trust and the LSOA administrative geography and a database of EAPMC services including opening and closing dates. The general estimation framework common across both panels is: where observation units indexed by subscript i ∈(LSOAs, NHS Trusts). The dependent variable is a hospital utilization outcome in quarter t. EAPMC is a primary care access intensity measure that captures EAPMC services within distance buffer b from unit i at time t. Time varying controls variables are contained in the vector x, and f (i, t) are fixed effects which allow for unobserved time and place variation.
The majority of estimates that follow are generated from neighborhood-level (i =LSOA) regressions that take the form: Here x captures time varying counts of population in five age bands (aged less than 10, aged 10-19, aged 20-49, aged 50-69, aged 70+) and their squared values to control flexibly for changes in neighborhood population and demographics. To account for unobserved variation, specifications include LSOA fixed effects (φ i ), quarter indicators interacted with labor-market area (indexed by m) dummies (φ tm ), and separate year (indexed by T ) indicators for all neighborhoods that obtain exposure to services in distance buffer b at any time in the panel (φ T (t)b ). These fixed effects are intended to eliminate factors that could bias results, including any time invariant neighborhood characteristics such as access to a walk-in clinic that existed prior to the EAPMC policy, as well as general labor-market wide changes, for example in the supply of hospital or community care. Ancillary specifications at the hospital level (i =NHS Trust) are useful as they require no sample restrictions to deal with data coding issues. Regressions take the form: Besides the different unit of observation, in contrast to equation 2 these regressions omit demographics given there is no simple way to assign population to NHS Trusts, and account for area trends at the level of 9 regions (London, South East, South West, West Midlands, North West, North East, Yorkshire and the Humber, East Midlands, and East of England, indexed by g), reflecting that in many cases a labor-market area contains only a single ED.
In both panels, the principal object of interest is EAMPC, a vector that captures timevarying primary care access intensity as a non-parametric function of proximity to services. These measures are generated from counts of the number of open walk-in clinics (or extended hours practices) within concentric distance buffers surrounding the centroid of each neighborhood or NHS Trust. As shown in Appendix Figure A2, the median travel distance to access emergency care in England differs considerably over space, so I allow distance buffers to vary according to the distribution of observed distances in the data.
In practice, this means buffers are computed for each of the 149 labor-markets in my data then assigned to all neighborhoods/NHS Trusts with centroids falling in that area. 9 9 The Office for National Statistics calculates labor-market areas, known locally as Travel to Work Areas (TTWAs), using commuting data. They each contain one or more cities and they nest LSOAs. The labor-market distance buffers are computed using distances traveled to attend EDs in the HES data between 2008/9 to 2012/13. I approximate patient starting location as registered primary care practice and ED visit location as the closest ED (relevant if an NHS Trust has more than one ED). Using patient trips to EDs is driven by practical considerations (walk-in clinic attendances are not well recorded in HES) but also has the benefit of ameliorating concerns about the endogeneity of resulting buffers.
Results in an earlier working paper (Pinchbeck, 2014) show that computing buffers across alternative Buffers are constructed in a discrete way such that each service falls into only one buffer for each neighborhood or NHS Trust. In most cases effects in three distance buffers are estimated: the lower quartile distance traveled (p25), the median (p50), and the upper quartile (p75). Around 15 of the walk-in clinics in my data are co-located at hospital EDs. To allow for different effects for these services I create a separate treatment for all such services within the median travel distance i.e. within the first two buffers. This yields four buffers in total, and the following estimated equation for walk-in clinics:

Endogenous Primary Care Availability
When identification rests on policy-induced variation a key methodological challenge is endogenous policy targeting. Here, primary care access is determined by a series of decisions by health administrators, for example where and when to open a new facility. This decision-making process is a black-box, and the suspicion must be that local service availability is correlated with unobserved underlying drivers of hospital outcomes. It might be reasonable to assume that EAPMC services were targeted to places experiencing increasing ED attendances, or to places expected to experience future ED attendance growth. If true, any estimate of associations between primary care access and ED attendances that ignore policy targeting would be biased towards finding that better access to primary care leads to more ED visits i.e. results would be underestimated.
To mitigate these issues all specifications use difference-in-difference strategies on samples composed solely of neighborhoods or NHS Trusts that already have close access to an EAPMC primary care service, did so in the past, or will do so in the future. With this sample restriction in place I estimate an average treatment effect on the treated (ATT) based on localized changes in hospital outcomes in places close to centers when the availability of services changes, against a control group provided by other places that have (or had, or will have) a similar facility close by, but where the availability of services administrative geographies other than TTWAs leaves results materially unchanged, but that setting buffer distances universally based on national averages introduces substantial noise. Later results are unaffected when the sample is restricted to places with median buffers that lie between the 25th (3km) and 75th (5km) percentiles of the buffer distribution, in which case buffers distances are very similar.
does not change at that particular time. This strategy is particularly helpful because EAPMC prescribed criteria for facility location and service specification (see Appendix A), suggesting services should be similar on observable and unobservable dimensions.

Results
Table 1 provides descriptive statistics for hospital utilization and control variables. The "NHS Trust sample" refers to the 118 NHS Trusts that were exposed to walk-in clinics under the EAPMC policy. The "Walk-in clinic sample" refers to the sample of neighborhoods that were exposed to walk-in clinics under the EAPMC policy. The "Under-  vs. national average £196,000).
By consequence of the research design all neighborhoods in the LSOA regression samples were exposed to at least one type of new EAPMC service. To illustrate how access to walk-in clinics varies by neighborhood, I create a variable capturing "maximum exposure" to walk-in clinics -i.e. the highest number of ED and other walk-in clinics that each neighborhood becomes exposed to at any point in the period April 2009 to September 2012, and cross-tabulate results in Appendix Figure A3. Neighborhoods in the main sample were exposed to between 0 and 9 non-ED walk-in clinics and either 0 or 1 EDbased clinics. However, the vast majority of neighborhoods gained access to only one or two clinics at any time: around 60% were exposed to one WiC in the panel period, whereas some 80% were exposed to no more than two.  The remaining columns in Table 2  The last two specifications in Table 2  Similarly, the last column signals no evidence of effects on the proportion of admissions that may have been prevented with appropriate primary care. 12 10 Appendix Table A2 demonstrates robustness to using a binary 1/0 WiC exposure variables instead of the count-based treatment intensity variables, specifying the dependent variable in levels, and removing the population and buffer control variables. This Table also shows that effects on ED visits for children and elderly people are slightly smaller than the baseline effects, and that impacts are significantly larger in the most deprived neighborhoods. Appendix Table A3 finds that standard errors are larger when clustering at the MSOA level than standard errors that follow Conley (1999) to allow for continuous forms of spatial autocorrelation up to a distance cut-off of 2km. Allowing for spatial autocorrelation is computationally demanding so here I specify the dependent variable as log ED visits per 1000 residents and drop population controls. Coefficients are robust to this specification change.

Walk-in Clinics and Hospital Utilization
11 For example, applying the first coefficient (-0.0376) to the sample mean ED visits (140) implies 5.2 fewer ED visits whereas the corresponding calculation for the second column implies 4.8 fewer visits.
12 Earlier versions of this work used mean ED waiting times as another outcome. However, identification is complicated by possible endogenous responses in hospital resourcing and operating decisions, as well as the likely violation of Stable Unit Treatment Value Assumption (SUTVA) i.e. because to the extent that walk in service affect waiting times, they will do so for all patients using an ED regardless of whether they themselves gain better primary care access. I therefore leave this analysis for future work. Note that results for ED visits are unchanged when controlling for ED waiting times, which should rule out SUTVA-type spillover concerns on my main results.

Substitution and Health Care Spending
Previous results indicate that walk-in services reduce visits to EDs but do not indicate the degree to which they substitute for ED activity. The mean number of ED visits for neighborhood-quarter cells in my main sample is 140 (Table 1), and the average walk-in clinic in my data has slightly under 50 neighborhoods in the first distance buffer, 50 more in the (thinner) second buffer, and a further 100 in the third. The point estimates in column one of Table 2 thus imply that an average ED walk-in clinic reduces annual ED visits by 2106 (=0.0376*140*100*4) whereas the average walk-in clinic located elsewhere reduces visits by 1145 (=0.0265*140*50*4 + 0.0144*140*50*4). Based on auxiliary information I assume each walk-in clinic is visited 18,000 times annually, suggesting that around 12% of patients visiting an ED walk-in clinic and around 6% of those visiting a clinic elsewhere were diverted from an ED. 13 The (unreported) 95% confidence intervals indicate that between 5 to 20% of patients attending ED based walk-in clinics and 5 to 10% of patients attending other clinics were diverted from an ED.
These rough calculations imply the lion's share of walk-in visits do not substitute for a visit to an ED. Clearly this is an incomplete analysis of the full possible effects of walk-in services, because the clinics may also substitute for care in regular primary care practices.
Such an analysis lies beyond the scope of this paper. Nevertheless, it is possible to make some assessment of the resource implications of the clinics because regular primary care services are funded through capitated budgets in the NHS setting. Available figures indicate that the average unit cost of a visit to an ED is three times the cost of a visit to a walk-in clinic. 14 Based on these direct costs, diversion rates of under 20% imply that walk-in services in the NHS lead to a net increase in health care spending.

Balancing Tests and Trends
Difference-in-difference applications assume that trends in treatment and control groups are parallel absent treatment. The research design outlined in Section 2.3 means that places that host EAPMC services act as both a treatment and control group, with identification coming off the timing of service deployment. The identifying assumption is that service deployment time should be unrelated to the determinants of ED visits, conditional on general labor-market trends. If in fact new services are deployed to places at times when ED visits are rising or falling more quickly than the general trend, then the control group of past and future locations for services will not provide a valid counterfactual.
The discussion in Section 1.1 suggests the actual timetable for the new centers was driven by administrative factors (e.g. availability of suitable premises and speed of procurements etc.) which are plausibly unrelated to ED visits. To test this premise, pre-reform primary care access variables (measured in both levels and changes) are regressed on the number of quarters between the policy announcement and the neighborhood's first exposure to a service, as well as the time-invariant analogues of the control variables listed above. As data for pre-reform access is not available at the neighborhood level, values are assigned to neighborhoods from the nearest primary care practice. The upper panel of Table 3 shows no significant correlation between EAPMC treatment timing and pre-reform primary assesses the extent to which groups of neighborhoods that were first exposed to new services at different times follow similar trends in ED visits. Reassuringly Figure 5, which is displayed in actual time rather than event time due to the seasonal pattern, reveals the unconditional trends in ED visits are broadly similar across all groups throughout the sample period.

Distance Decay
Section 3.1 reported spatial patterns in the impacts of walk-in clinics on ED attendances.
Distance decay is more precisely teased out in Table 4 which drops neighborhoods close to walk-in facilities at EDs and expands the number of distance buffers to seven. Column (a) of Figure 6 summarizes the first two columns of this   The conceptual framework in Section 1.2 suggests walk in clinics should divert patients with less serious medical problems from EDs. Most patients that are not admitted during an ED attendance likely fall into this category. These outpatient visits make up around three quarters of all ED visits in this setting. Column (b) of Figure 6 plots the effect of walk-in access on these patients, corresponding to the third and fourth columns of Table   4. The pattern of effects is highly similar to the first column. During walk in hours effects in the first buffers are slightly larger and the distance decay is a little steeper, although these differences are not statistically distinguishable. The corollary is that around three quarters of the overall effect of walk in services arises through diverting patients who would not be admitted through an ED, with the remainder of the effect coming through patients who would be admitted.
Hospital records also indicate how patients came to be at the ED. Column (c) of Figure 6, corresponding to columns five and six of Table 4, tracks impacts on patients recorded as self-referring to the ED. This group represents around 60% all visits to EDs. The remain- Figure A4 shows that results are highly similar but that effects are slightly larger at very short and very long distances for neighborhoods closer to walk-ins than EDs.  Table 4 (black lines) and the bounds of the associated 95% confidence intervals. Notes: Sample contains quarter-LSOA cells with 50 or more ED visits but drops neighborhood exposed to walk-in facilities co-located at EDs. Dependent variables are in logs. WiC hrs are between 8am-8pm Monday through Sunday. Standard errors in parentheses clustered at the MSOA level der are patients that ostensibly had less discretion in the location of their treatment. This is because they were referred to the ED from another source (most commonly a family doctor), or conveyed to the ED in an ambulance. As with the non-admitted group of patients the effects are qualitatively similar to the overall patterns shown in column (a), but here coefficients during clinic open times are roughly one to one and a half times as large. One possible explanation is that self-referred patients have less severe health needs which can be treated in lower acuity facilities like walk-ins more readily. This finds support in the data: only 12% of the self-referred group are admitted following their attendance compared to more than 40% of the other group.

Dimensions of Access
What further dimensions of access drive diversion from EDs? This section aims to shed light on this question through a descriptive comparison of the impacts of walk-in services and extended hours practices (denoted PCPs in this section) opened under the EAPMC policy. As noted previously, PCPs are conventional primary care services that require patients to be registered to receive services. They offer extended opening relative to core primary care hours but operating hours fall short of the 7 day services at walk-in clinics.
Making comparisons across these service-types can help to ascertain how opening hours and the need to make an appointment condition the extent to which patients are diverted from EDs.
The PCPs opened under the EAPMC policy were located in areas of the country with the lowest concentration of family doctors. To ensure a like-for-like comparison samples underpinning all regressions in this section are restricted to neighborhoods in administrative areas eligible for both types of EAPMC service. Regressions, reported in Table 5, first estimate the impacts of walk-in services and extended hours practices separately, then simultaneously in the third and fourth columns. Because of the narrower geographical sample these regressions differ in two ways to earlier specifications. First, they include region rather than labor-market trends here as there is insufficient variation to separately identify the latter from changes in primary care access driven by the policy reform. Second, because there are very few walk-in facilities co-located at EDs in this sample, all neighborhoods in close proximity to such services are dropped throughout this section.
Before comparing service types I first use this sample to assess the robustness of prior results for walk-in services. The first column of Table 2 estimated the impact of walk-in clinics across the country as a whole. In Table 5 walk-in impacts are estimated in underdoctored areas of the country on their own (in column 1), and conditional on changes in regular primary care access (in column 3). The coefficients on the walk-in service variables across these three specifications are remarkably similar, giving reassurance that the omission of variables capturing access to regular primary care in earlier regressions is unlikely to be a major source of bias.
The third and fourth columns of Table 5 estimate the effect of both types of primary care service concurrently. The third column uses a dependent variable constructed from  The final column of Table 5 compares effects on ED visits in core primary care hours: 8.30am and 6.30pm on Monday-Friday. During these hours, both types of service are open so any differences in diversion cannot be driven by variation in opening hours of services. The mean number of ED visits taking place during these times is 60 (see Table   1), so that results imply that walk-in clinics divert roughly 755 ED trips whereas PCPs divert 172, or 583 fewer visits. Both service types thus appear to divert a significant proportion of patients outside core primary care opening hours, but some 80% of the difference in the overall effects arise when both types of service are open. If EAPMC walk-in clinics and extended hours practices are similar on unobserved dimensions, these findings signal that the ability for patients to attend without registering or making an appointment may have a large bearing on the ED diversion.
Appendix Table A4 finally reports the impacts of access to primary care services inside and outside practice catchment boundaries during core primary care practice hours. In theory a patient living outside a practice's boundary cannot register for regular primary care services but can attend a walk-in clinics (where these exist) as a non-registered patient. In line with this prior, extended hours practices have zero impacts in neighborhoods outside catchment boundaries. For walk-in clinics ED diversion occurs inside and outside boundaries but is systematically larger in neighborhoods falling inside boundaries.
Although I provide no direct tests, I speculate these patterns could reflect benefits from continuity of care or competition from walk-in clinics driving improvements in practices outside my sample. 16

Further Robustness Checks and Placebos
In all preceding estimations fixed effects partial out time-invariant unobservables at the neighborhood level and region-or labor market-wide trends, while population counts control for demographic changes. Besides these controls, previous sections reported some natural robustness checks, for example by examining the impacts of services during service open or closed hours and inside or outside practice catchment boundaries. A number of further placebo and robustness checks that lend further support to these results. In all cases I report graphical evidence, relegating associated regression outputs to Appendix 16 The difference in overall effects between the service types could also be driven by differences in boundary sizes. In Table A5 I   Tables A6 and A7.
A first robustness check evaluates the sample restriction under which neighborhood quarter cells with few ED visits were dropped. This restriction was adopted to avoid conflating changes in data reporting practices with genuine changes in ED volumes and to circumvent problems inherent in using count data. To test this strategy I re-estimate walk-in impacts under different samples but now using the difference between the logarithm of self-referred ED visits and the logarithm of ambulance or referred ED visits as the dependent variable. Changes in reporting should affect both of these patient groups symmetrically. The three plots in Figure 7 demonstrate that distance decay of walk-in clinics on this measure are qualitatively similar when no cells are dropped (left-most plot), when cells with less than 10 ED visits are dropped (middle plot), and with the full sample restriction (right-most plot). Given earlier findings these estimates are driven largely by the self-referred patient group so it is reassuring that the patterns in all plots are broadly consistent with those in Figure 6. More generally, differencing between these types of attendances partials out any unobserved time varying neighborhood factors that affect both groups so provides a powerful check on earlier results. Figure 8 presents two more general checks on walk-in access effects. In the first, I re-run the first regression in Table 2  control only for less granular regional trends. The coefficients for walk-in services are similar across these specifications yet it remains possible that the latter estimates are partially driven by common shocks within labor markets. Figure 9 follows the approach in Busso et al. (2013) by estimating the effect of primary care access on the neighborhoods (log) rank position on ED visits within the labor-market distribution, where rank 1 is assigned to the neighborhood with the lowest count of ED visits in the labor-market that quarter. The estimated pattern of effects is qualitatively similar to those in Table 5 albeit stronger for walk-in services relative to the extended hours primary care practices.  Table 2. LHS plot refers to the same regression in column 1 of   confound estimates should they correlate with factors driving hospital utilization and directly coincide with EAPMC service changes. Figure 10 indicates that changes in access are uncorrelated with average house prices which goes some way to alleviating this concern. 18 4 Conclusion This paper examines a policy reform that introduced a substantial change in primary care access across England within a short time-frame. The reform is helpful because its implementation provides a source of plausibly exogenous variation, and of particular interest because it created new primary care services which differ along several organizational dimensions. The first part of the analysis finds that access to convenient primary care services significantly reduces visits to hospital Emergency Departments, and documents a range of further findings that support the robustness of this result.
Parameter estimates imply that somewhere between 5 and 20% of patient visits to a 18 In the last Column of Table A7, I also show that EAPMC services have no significant impact on ED visitors arriving by ambulance.
walk-in facility substitute for a visit to an ED. The lower unit costs of care in the clinics relative to EDs is insufficient to offset the costs of the new utilization, so that walk-in clinics imply a small net increase in health care spending. A full assessment of the welfare implications of walk-in services lies outside the scope of this work. Shifting care outside of EDs is likely to be socially beneficial because of the lower costs of care in primary care settings. However, further work would be needed to evaluate whether the social benefits of the substantial new utilization of walk in clinics outweigh the social costs of providing the services.
Subsequent sections of this article then distinguish empirically between four aspects of primary care access: proximity to services, convenience of opening hours, the need to make an appointment, and eligibility to receive care. Estimates indicate that two convenience dimensions of access -proximity and the ability to attend without appointment -are paramount in determining the extent to which primary care services divert patients from hospitals. Given that the private costs of distance and making appointments are small, these results suggest that psychological factors influence how agents' choose to obtain treatment. This tallies with recent evidence showing that hassle factors can prove to be an important barrier to participation decisions.

B Behavioral factors and ED use
Given evidence of non-standard decision-making in a health context, Baicker et al. (2015) modify the standard treatment decision to p < b(σ) + (σ) where is a general behavioral bias factor that can encompass several phenomena such as hyperbolic discounting, nonstandard beliefs, or inattention. Some behavioral hazards lead to under-utilization of care, for example < 0 when treatment benefits are delayed and agents are present biased, when agents underestimate the benefits of treatment, when obtaining care is subject to "hassle" factors, or when symptoms are not salient. Conversely, > 0 implies behavioral hazards that lead to over-utilization even when p = c.
Agents then seek treatment in primary care when two conditions hold: patients perceive a net private benefit from primary care (b P C + P C − p P C > 0); and primary care offers a higher perceived net benefit than ED care (b P C + P C − p P C > b ED + ED − p ED ). Primary care access intervention then work through two channels. First, they may reduce the private costs of obtaining treatment in primary care, either through lowering co-pays or -as in the NHS case -by reducing the time and travel expenses incurred to access services. Second, for behavioral agents, policies that ease access can also work through an additional channel, for example by mitigating or eliminating inconvenience and hassle factors associated with obtaining treatment (Bertrand et al., 2006).   Table 4 (using ED vists during walk-in open hours) but interacts the count of walk-in clinics in each buffer with a variable that takes the value of 1 when the average distance to walk-in clinics is smaller than the distance to the nearest Emergency Department (LHS) or 0 when the reverse is true (RHS). Figure A5: One quarter lead and lag effects for walk-in clinics   Notes: Table reports coefficients from a regression of log ED visits per 1000 population on the full set of fixed effects. In column (1) standard errors are clustered on LSOAs, in column (2) standard errors are clustered on MSOAs, in column (3) standard errors follow Conley (1999) and are robust to continuous forms of spatial autocorrelation.  X Table A6: Regression outputs for Figure 7 (1) (2)