3.1. Data Sources and Setup
Our article relies on patient-level Hospital Episodes Statistics (HES) data from 2002 to 2008. In addition to observations for patients with an emergency AMI, our analysis contains data on patients undergoing elective hip replacement, knee replacement, knee arthroscopy, cataract repair and hernia repair, which we use in the construction of our competition indices. At the hospital level, we know hospital site postcodes, the NHS trust to which the site belongs, and we have indicators of the hospital type (teaching hospitals, FTs status) and hospital size. Our work improves on previous research by using the hospital site-specific locations, rather than the trust headquarters. There are typically multiple treatment sites for each trust, separated by distances of up to 50 km, so using Trust locations provides only a very approximate indicator of the location at which treatment is carried out.8
We use GP and hospital site postcodes to calculate distances between patients’ GPs and the hospital where care was delivered. This distance is a key component in our analysis and is used as an input into most of our competition measures. For our main analysis, we use matrices of straight-line distances between GPs and NHS sites. For some of our supplementary results, we calculate origin-destination matrices from minimum road travel times along the primary road network.9
3.2. Measures of Health Care Quality
Our measure of hospital quality is the 30-day mortality rate for patients with an AMI.10 In our analysis, we include every patient who had a main International Classification of Disease 10 code of I21 or I22 and only include emergency AMI admissions and admissions where the patients’ length of stay was three days or more (unless the patient died within the first three days of being admitted) (World Health Organization, 2009).11
We chose to use AMI mortality as our measure of performance for four primary reasons. First, AMIs are a relatively frequent, easily observable medical occurrences that are clinically identifiable and have a substantial mortality rate. For example, in 2008 the overall, 30-day mortality rate for emergency AMI was 11.7% compared to a mortality rate of 0.20% for elective hip replacement. Second, with AMIs, there is a clear link between timely and high-quality medical intervention and patients’ survival (Bradley et al., 2006; Jha et al., 2007). Contrast this with a quality indicator such as readmissions for elective hip replacements, where a patient failing to stick to a rehabilitation programme after they were discharged could produce poor outcomes or lead to a readmission. Third, unlike other measures of performance, like hospital waiting times, AMI mortality (and death rates in general) are not subject to gaming or manipulation by hospitals. Fourth, AMIs are an emergency procedure where patients are generally taken directly to their nearest provider for care with little discretion over which hospital they attend, which mitigates hospitals ability to risk-select healthier patients for care. The fact that AMI is a non-elective procedure also mitigates biases due to the endogeneity of market structure to elective quality. This point is illustrated in Appendix A.
A further impetus for using AMI mortality is that it is frequently used by governments and private organisations to rank and compare hospital performance (including by the UK government).12,13 Consequently, 30-day AMI mortality is also often used in the academic literature as a measure of overall hospital performance in the UK and the US (Kessler and McClellan, 2000; Volpp et al., 2003; Propper et al., 2004, 2008; Kessler and Geppert, 2005; Bloom et al., 2010; Gaynor et al., 2010; Propper and van Reenen, 2010). Consistent with its use as a measure of hospital performance, a recent study assessing the relationship between hospitals’ management quality and their overall performance found a statistically significant relationship between overall hospital management performance and hospital level 30-day AMI mortality (Bloom et al., 2010). Likewise, according to data made publicly available by Dr. Foster Health, despite accounting for less than 3% of total hospital deaths, standardised AMI mortality in English hospitals was positively correlated (r = 0.33) with overall hospital mortality for the financial year beginning in 2009.14 Likewise, in our administrative data, we have found that raw AMI mortality is positively correlated with elective hip and knee replacement waiting times (r = 0.33) and positively correlated with length of stay for elective hip and knee replacement (r = 0.11 and r = 0.22, respectively).
While 30-day AMI mortality is a frequently used measure of hospital quality, there are several issues with its use. First, as with all quality measures, despite being correlated with other dimensions of performance, there is a question of whether or not a single measure can capture the multi-dimensional nature of health care quality (McClellan and Staiger, 1999). A second issue with 30-day mortality is the noise inherent with this type of measure. This noise is particularly acute when researchers use hospital level data, where it is difficult to suitably risk adjust and hospital performance can vary from year to year. Our use of patient-level data, which allows for controls for patients’ socioeconomic status, age and co-morbidities, mitigates this problem. In our estimation, we control for co-morbidities using the Charlson co-morbidity index (Charlson et al., 1978) and control for patients’ socio-economic status using the income vector of the 2007 Index of Multiple Deprivation, which we include at the Census Output Area level (Communities and Local Government Department, 2009).15,16
3.3. Market Measures and Estimates of Market Structure
Identifying the impact of competition in the wake of NHS reforms requires accurately measuring market structure. In this article, we estimate market structure in the English NHS using both counts of providers and Herfindahl-Hirschman Indexes (HHIs) calculated using actual and predicted patient flows. Our aim in developing a range of measures of market structure is to illustrate that our results are robust across a number of measures of market structure, since there is not a single, agreed upon measure that is immune to each and every form bias.
The debate over measuring market structure centres around thwarting potential endogeneity between hospital quality and market structure, avoiding measures of market structure that simply reflect urban population density and defining a market size that accurately reflects the choice sets available to NHS users. Concerns over the endogeneity between measures of market structure and firm performance have been frequently cited in the literature and stem from three aspects of the construction of the market structure measures.
These different forms of bias could positively or negatively affect our estimates. First, the physical market size itself could be associated with hospital performance, which would bias our estimates upwards. For example, a high-quality provider might attract patients from a larger area, and hence appear to be operating in a less concentrated market. Second, the actual patient flows that are used to estimate market shares (and form the key component of the HHI) could be associated with quality because high quality providers could attract all the local business and as a result appear to be operating in more concentrated markets and bias our estimates downwards. Third, the actual location of hospitals and of new market entrants may be associated with performance. For example, if new hospitals were reluctant to locate near high quality providers, this would artificially show high quality providers to be operating within concentrated markets and bias our estimates of the treatment effect downwards.
In addition to concerns about endogeneity, there are also fears that the various measures of market structure will be spuriously correlated with urban population density, which stem from two causes. First, densely populated cities have more hospitals within smaller geographic areas, and as a result, urban areas will likely appear more competitive. Second, measures of market structure that are calculated within fixed geographic markets may be biased because the time it takes to travel 30 km in an urban area will differ significantly from the time it takes to travel 30 km in a rural area.
In our estimates of market structure, we calculate competition within the market for elective secondary care for NHS funded patients. We focused on competition for elective care because this was the only hospital market where competition occurred during the time period we are studying. We study five high volume procedures – hip replacement, knee replacement, arthroscopy, hernia repair and cataract repair – and develop composite measures of market structure, which are weighted averages of the competition measures that we calculated for each of the individual procedures. The bulk of our measures of market structure are based on actual patient flows. However, Kessler and McClellan (2000) have suggested that any measures of market structure based on actual patient flows could be endogenous to hospital quality because they may be correlated with various unobserved characteristics of either patients or providers. As a result, in addition to using an instrumental variable strategy, we also estimate a measure of market structure, similar to the measure used in Kessler and McClellan (2000), which is based on predicted patient flows generated from models of patient choice.
We centre all of our markets on GP practices, rather than on hospitals, because this mirrors the post-2005 NHS market structure, where patients select their hospital in conjunction with their GP (Dixon et al., 2010). In addition, were we to centre our measures of market structure on hospitals, then there is the risk that if unobserved determinants of hospital choice are correlated with patient characteristics, there could be spurious and problematic associations between health status and market structure.
To measure market concentration using actual patient flows, we calculate the negative natural logarithm of an HHI (nlhhi) based on hospitals’ market shares. This transformation is convenient because the nlhhi increases with competition, with zero corresponding to monopoly and infinity to perfect competition. In addition, this measure is equivalent to the natural log of the number of equal size firms in the market, which makes interpreting the index more intuitive. Thus, for given market area j, our concentration index is:
Here, nk is the number of procedures carried out at hospital k within market j and Nj is the total number of procedures carried out in market j. Note that nk includes procedures performed at hospital k that were not referred from market j.
We construct our preferred market definition as follows: consider an elective procedure, e.g. hip replacements, in one year, e.g. 2002. We use matrices of patient flows from GP practices to hospitals for hip replacement in 2002 to deduce GP-centred markets. Specifically, we find the radius that represents the 95th percentile of distance travelled from a GP practice to hospitals for hip replacements in 2002. This defines the feasible choice set for patients at this GP practice in 2002. We then compute the HHI based on all hospitals providing hip replacements within this GP’s market, regardless of whether this GP actually refers patients to all of these hospitals. This process is repeated for all GPs, for all years 2002–8 and for all five key elective procedures. A single elective HHI is calculated for each GP per year as a weighted average of the procedure-specific HHIs with weights proportional to the volume of patients in each procedure category.17
In addition to calculating this HHI within a variable radius market, we also compute a number of alternative HHIs using other market definitions. These include an HHI measured within a fixed radius market, which is derived in a similar way to the variable radius HHI described above, except that we use a fixed 30 km radius drawn around each GP practice in the country to delineate the market boundaries. The second alternative index is an HHI based on travel times along the primary road network from each GP. Here, we include hospitals in our relevant markets if they fall within a 30-min car ride from a referring GP. A third alternative is based on our 95% variable radius market but it does not treat sites within the same trust as competitors and only views sites from a different Trust as viable alternatives in the calculation of our HHI. We have also calculated our preferred purchaser-perspective measures using the count of hospitals within each market in lieu of using HHIs. In addition, we calculate one measure of market concentration from the provider’s perspective, where the market is centred on hospitals and the market is defined a fixed radius of 20 km drawn around each site.
Alongside the HHIs we generated using actual patient flows, we also created an HHI derived from predicted patient flows that is based on the strategy used in Kessler and McClellan (2000). Building our predicted patient flow HHI is a two-step procedure. The first step involves estimating a patient choice model based on hospital and GP locations, and hospital and patient characteristics.18 From this step we predict the number of patients each GP refers to their local hospitals, controlling for patient and provider characteristics and GP-hospital differential distances. We then use these numbers to generate the HHIs.19 However, whereas Kessler and McClellan (2000) used a conditional logit to model patient choices, we use a Poisson regression on aggregate GP-hospital flows, which is equivalent but is simpler to compute (Guimaraes et al., 2003).
As Table 1 illustrates, all these measures of GP-centred competition are moderately correlated. The indices from fixed radius and time-based market definitions are highly correlated. Indices based on market definitions using GP hospital flows are quite highly correlated with each other and only moderately correlated with the fixed distance and time-based indices. We favour the variable radius methods that infer markets from de-facto patient choices over hospitals, not least because this is less correlated with urban density.20
Correlations Between Different Measures of Market Structure
|−log(HHI)-95%||1.00|| || || ||0.7483||0.5639|
|−log(HHI)-30 km||0.48||1.00|| || ||1.4860||0.9053|
|−log(HHI)-30 min||0.43||0.92||1.00|| ||1.2686||0.8081|
As a further check of robustness, we estimate (1) substituting an indicator variable for our competition measure, which is equal to one if a patient’s GP practice is located in an urban area.21 For further robustness, we also reconstruct the competition index using the shares of secondary school pupils in schools within our GP-centred markets (defined by the 95% referral radius during the pre-policy period) for use in a placebo test. These tests are designed to confirm that our results are driven by competition, rather than spurious associations with urban density.
3.4. Instrumental Variable Estimation
In an effort to thwart the endogeneity that we described earlier, in addition to creating HHIs from predicted patient flows, we have also developed an instrument for competition. Our preferred instrument takes advantage of the historically determined hospital locations in England and is based on the variation in distance to a patient’s nearest four hospitals. Specifically, our instrument for market structure is the standard deviation of distances from GPs to their nearest four hospitals, conditional the on the distance to the patient's nearest hospital (a control which we introduce in order to control for potential urban/rural differences in GP location). This IV strategy rests on the fact that NHS hospital and GP relative positions are unrelated to hospital quality, which is supported by the fact that hospital locations in England are largely a historical artefact which have not changed substantially since the NHS was founded in 1948 (Klein, 2006).
To illustrate our IV strategy, imagine two hospital markets centred on two individual GP practices (A and B). The nearest provider in the area of GPA is located at 5 km, and the remaining three at 15, 20 and 30 km. The nearest provider to GPB is also at 5 km, but with the remaining three all within 10 km (in different directions). In this situation, while the distance to the nearest provider is the same in both cases, the alternatives available to patients of GPB are much more substitutable than the alternatives available to patients of GPA because they are all within a similar travel distance, so patients of GPA are much more likely to attend the nearest provider. We therefore assume that GP-centred markets characterised by a high dispersion in distances to local providers are low choice and therefore low competition markets.
In practice, we have three instrumented variables, which include the baseline measure of market structure, the pre-policy time trend interacted with market structure and the post-policy time trend interacted with market structure. We perform our IV with a 2SLS estimator and include GP and hospital fixed effects.