Exploring the impact of new medical technology on workforce planning

Exploring the impact of new medical technology on workforce planning Maynou, Laia 1,3, McGuire, Alistair 1 and Serra-Sastre, Victoria 2,1 1 LSE Health and Social Care, London School of Economics 2 Department of Economics; City, University of London 3 Center for Research in Health and Economics (CRES), Universitat Pompeu Fabra May, 2017 Abstract This paper contributes to the existing literature on the diffusion of medical technologies. The objective of this research is twofold. First of all, we examine the substitution or complementarity effects across different types of new technologies introduced into the NHS. In particular, the analysis is looking at the introduction of PTCA for the treatment of cardiovascular disease. Secondly, we compute estimates of the degree to which the workforce reacts to the introduction of new technology, based on elasticity of supply measures. Data is combined from different sources to analyse these relationships: mainly, the Hospital Episodes Statistics (HES) and the NHS Electronic Staff Records (ESR). We apply panel data techniques to determine the manner in which technology is diffused across the NHS, with a particular emphasis on the impact that technology has on the workforce composition. Our analysis is at the provider level and the empirical specification explores the relationship between volume and workforce also controlling for provider and at risk population characteristics. Given the lack of quantitative evidence on the degree of substitution or complementarity across different forms of input in treating surgical cases, such analysis gives indicative estimates of productivity gains attributable to flexible workforce planning and technology uptake. Keywords: Technology, workforce, substitution, production function JEL Classification: O33, I12, C41, C33, J2 Contact Author. Email: l.maynou-pujolras@lse.ac.uk 1

1 Introduction The role of technology in health care has been identified as the main driver of health care expenditure for most developed countries (Newhouse 1992, Smith et al. 2009) but has also been accompanied by improvements in health outcomes (Skinner & Staiger 2015). Although technological change boosts economic growth through increases in productivity, this is unproven in the health care sector. Many medical innovations introduced into the health care market have a lower initial per unit cost, but may also increase overall expenditure through increasing demand. Even for those technologies that have a higher initial unit cost than that of existing technologies, the higher marginal cost may be offset by a higher marginal benefit to the patient in the long-run. Regardless of the technology s unit cost, when a medical innovation is introduced in the healthcare market, two effects are likely to arise: an expansion and/or substitution effect (Cutler & Huckman 2003). An expansion effect occurs when the technology opens up new treatment possibilities that were previously unavailable. In addition, there may also be a substitution effect that leads to a replacement (partial or total) of an existing technology. These two effects have been studied in the context of the introduction of Percutaneous Transluminal Coronary Angioplasty (PTCA) for the treatment of cardiovascular disease. PTCA was introduced as a less invasive and cheaper procedure than the existing technology to treat cardiovascular disease, Coronary Artery Bypass Graft (CABG). Cutler & Huckman (2003) estimated that the substitution effect accounted for 25-35% of new PTCAs that followed an initial expansion effect in the USA. McGuire et al. (2010) found similar results for the UK, however, the substitution effect was lower and the expansion effect was higher in the UK than in the USA. Similarly, other technologies have shown an expansion effect. Serra-Sastre & McGuire (2012) estimated a complementarity effect between laparoscopic and endoscopic prostatectomy, in addition to a substitution effect between medical surgery and prescription drugs targeting the treatment of the same condition. The current evidence on the impact following the introduction of technology into the health care sector is however extremely limited both in volume and nature. Generally studies have considered the diffusion and substitution aspects of new technologies in the health care sector (Serra-Sastre & McGuire 2009, 2012, 2013). The overall impact on the health care production has not been considered. Given that the workforce accounts for roughly 70% of the expenditure for NHS health care providers (Imison & Bohmer 2013, House of Commons Health Committee 2007), there remain unanswered questions over the impact that medical innovations have on staffing levels. Is the up-take of medical innovation accompanied by a reduction in labour-intensive inputs? For instance, many new surgical interventions are less invasive than previous technologies. If there is a reduction in surgical time required per procedure, should we observe a reduction in the workforce associated? There has been scant analysis of these major substitution effects within the NHS. The predicted 2

changes in medical workforce are mixed (depending on the care setting); however, there is an overall expected hospital doctor oversupply with some specialties facing recruitment challenges (Imison & Bohmer 2013). There has been an overall increase in consultants since 2004 with an estimated 22% increase between 2010 and 2016 only and an average annual decrease in productivity of 2.3% (Lafond et al. 2017, Bojke et al. 2016). Given the trends in hospital workforce and complexities of workforce planning, understanding the relationship between workforce and innovation is key to achieve the efficient deployment of the workforce. There is also a related question of the impact of new innovation up-take on volume and outcome relationships. In particular, is the impact of technological change constant with respect to any substitution effect on labour? This paper builds upon, and extends, existing research that focused on the diffusion of medical technologies. The aim is to focus empirically on estimating substitution or complementarity effects across different types of labour as new technologies are introduced into the NHS. There has been very little work on input substitution or complementarity generally within the NHS (see Gray & McGuire (1989) on aggregate relationships and Elliott et al. (2010) on the impact of local labour market forces on workforce deployment). On the substitution between technology and labour in a context of regulated industry, Acemoglu & Finkelstein (2008) find that moving to a fixed-price system such as the Medicare Prospective Payment System (PPS) induces a reduction in the capital-labour ratio through a decline in labour and a push for the adoption of new technologies. Their model predicts a substitution between technology and labour. In their analysis, Acemoglu & Finkelstein (2008) use input levels and technologies available at the aggregate hospital level whereas our paper looks at a particular set of inputs for a single technology. However, one objective is to single out whether their aggregate predictions also hold when examining the technologylabour elasticity of substitution in the NHS secondary care setting. There have been different lines of research relating to technology uptake focusing on different technology types. Evidence on drug diffusion (Coscelli & Shum 2004, Crawford & Shum 2005), health technology information (Lammers 2013, Dranove et al. 2015) or surgical procedures (Cutler & Huckman 2003, McGuire et al. 2010) have shed light on different aspects of diffusion such as informational spillovers, organisational structure and the role of insurance. Our proposed research will initially focus on surgical procedures, given that we are fundamentally interested in technology-labour relationships, and these are easier to identify explicitly for these type of health technologies. Our case-study focuses on the introduction of PTCA as a less invasive and cheaper treatment than CABG. This case is an example of a technology that was introduced as a replacement of an existing one. For each of the chosen case-studies hospital level data will be used to assess the degree of substitutability/complementarity with the workforce. Estimates of the degree to which the workforce reacts to the introduction of new technology, based on 3

elasticity of supply measures, is quantified. 2 Empirical Strategy For the purpose of this paper, the diffusion process and its effect on workforce planning is examined at the provider level through the following specifications. Equation 1 shows the diffusion equation that captures the elasticity of substitution or complementarity between CABG and PTCA. The diffusion model is based on Cutler & Huckman (2003) and McGuire et al. (2010). However, for the purpose of our research, we reverse the position of the main variables in the equation. In this case, the dependent variable is the PTCA volume, instead of the CABG volume as specified in Cutler & Huckman (2003) and McGuire et al. (2010), as we are also interested in the impact of the new technology on labour levels. We use a panel data model using two approaches: 1) levels, with fixed-effects, and 2) changes, with first-differences. The specification of Equation 1 is the following: P T CA pop45 it = α + β 1 CABG pop45 it + (β s β 1 )[ CABG pop45 it (tɛs)] + γ Z it + d t + c i + u it (1) where the dependent variable, number of PTCA procedures performed by provider i and year t over population aged 45 and over, is regressed against CABG volume over population aged 45 and over, an interaction term of the population-adjusted CABG volume with tɛs, a vector of indicators of 4-5 year period (1999-2002, 2003-2007, 2008-2012), a set of control variables, Z it, defined in Table 1 in the Appendix, and finally, c i is a fixed-effect for provider i, d t is a fixed-effect for year t and u it is the disturbance term. PTCA and CABG are procedures used mostly to treat coronary heart disease prevalent in patients aged 45 and above and therefore volume is controlled for population at risk at the Primary Care Trust (PCT) where the provider is located. The variable of interest is the volume of CABG (adjusted by population at risk) and its interactions with the time periods. The coefficients of greatest interest in Equation 1 are represented by β 1 and the vector (β s -β 1 ) (Cutler & Huckman 2003, McGuire et al. 2010). By using time-varying coefficients the degree of substitution between the procedures is allowed to change over time as PTCA matures. Cutler & Huckman (2003) note that unobservable factors u it will be correlated with both CABG and PTCA and therefore the OLS estimator for β 1 and the β s will be biased and consequently the value of substitution by any given period. By assuming that the bias in any given period is constant, they argue that (β s -β 1 ) can be estimated without bias. By then assuming that β 1 =0, Cutler & Huckman (2003) obtain the substitution rates over time. Other non-surgical technologies may influence the volume trajectory for PTCA/CABG. For instance, statins (cholesterol-lowering drugs) are prescribed for primary and secondary 4

prevention of cardiovascular disease. If primary prevention via medical treatment influences the volume of surgical procedures the estimates in the results section may be upward biased and the specification may suffer from potential omitted variable bias. To alleviate this potential source of bias we also examine the role that non-surgical technology may have in the diffusion of PTCA. Data availability is restricted to the period 2008-2012 for which we have the total number of statins prescribed at the PCT level. The next step is to quantify the elasticity between the workforce and the PTCA/CABG using a dynamic panel model: P T CA W F it = α + β 1 W F it 1 + β 2 + (β s β 2 )[ P T CA (tɛs)] + γ Z it + d t + c i + u it (2) CABG it CABG it where the dependent variable W F it is the ratio between two counts of different FTE in the workforce. We distinguish between cardiologists and cardiothoracic surgeons as they specialise in the provision of cardiovascular surgery. Cardiologists are the consultants that perform PTCA in patients in a catheterisation laboratory whereas cardiothoracic surgeons perform CABG in an operating theatre (NHS UK, Gray et al. (2000) and Molina & Heng (2009)). The underlying assumption is that if PTCA volume has been increasing over time this would theoretically be followed with an increase in the skill mix of the workforce, that is, an increase in cardiologists and relative to cardiothoracic surgeons. We therefore examine the ratio of the following set of skilled labour related to the provision of PTCA/CABG: Ratio of cardiologists (C) over cardiothoracic surgeons (CS) per provider i and year t, each adjusted by the total number of cardiologist and cardiothoracic surgeons per year t in England, respectively. This adjustment is intended to capture any systematic differences in workforce growth between the provider and the national trend. In addition to this ratio, the empirical specification also includes the same ratio differentiating by the seniority level of the physician: consultants, specialists (Associate Specialist, Specialty Doctor, Staff Grade and Specialty Registrar), trainees (Core Medical Training, Foundation Doctor Year 1 and Year 2). C/CScountry it = Cardiologist it /T otalcardiologist t CardioSurgeons it /T otalcardiosurgeons t (3) Ratio of cardio surgeons (CS) over total nurses and the ratio of cardiologists (C) over total nurses per provider i and year t. The consolidation of PTCA may also lead to a change in the physician/nurse ratios. PTCA is a minimally invasive procedure that requires only an overnight stay, and therefore this can change the composition of medical and nurse staff. However, this will depend on the degree of expansion and substitution (or complementarity) effect between PTCA and CABG. 5

At the time of writing this paper we only have total number of nurses per provider and not specialty-related number of nurses, therefore we anticipate the estimates for these ratios will not be precisely estimated. As part of future research we are working on refining the variable for nurses to FTEs that are directly involved in the provision of these surgical procedures. CS/nurses it = CardioSurgeons it Nurses it (4) C/nurses it = Cardiologists it Nurses it (5) Equation 2 also accounts for the lag of the dependent variable to control for any adjustment cost associated to workforce training. Any changes in workforce and medical specialisation are not immediate and by including the lagged workforce we aim to account for the constraints introduced by for medical training. While Equation 1 can be estimated as a standard panel data model, Equation 2 is estimated using dynamic panel data methods, which control for individual heterogeneity among providers. The inclusion of a lag of the dependent variable introduces correlation between the error term and the regressors. Under this specification standard panel data methods give biased results. To control for this correlation, the system GMM estimator by Blundell & Bond (1998) and Bond (2002) are used to estimate the coefficients of interest. There is also a further problem of endogeneity given by the simultaneity of the workforce and the PTCA/CABG relationship. To account for this we used lagged values of the PTCA/CABG relationship as instruments for the current PTCA/CABG ratio. In order to obtain consistent estimators, first-differences are taken to eliminate c i and remove any bias caused by the correlation of c i and the lagged dependent variable. As the first difference of the lagged dependent variable is now correlated with the first-differenced error component, Instrumental Variable (IV) estimators must be used in order to obtain consistent estimates (Bond 2002). The instruments required to control for the correlation between the lagged dependent variable W F it 1 and the error term are W F it 2, W F it 3 and W F it 4 for the equations in first-differences and W F it 1, W F it 2 and W F it 3 for the equations in levels, where W F it 1 = W F it 1 W F it 2. To account for the simultaneity problem between the dependent variable and the PTCA/CABG ratio we use the same lag structure of instruments as for the workforce variable, and also for the interactions between the PTCA/CABG ratio and the time dummies. 6

3 Data In order to test empirically the impact of new technology on workforce, we use two main datasets: Hospital Episode Statistics (HES) and the NHS Electronic Staff Records (ESR) from NHS Digital. HES data contains all episodes for patients admitted into hospitals in England. The sample we retrieved from HES data includes all records from the financial year 1999/2000 to financial year 2012/13 for each patient admitted into hospital for CABG or PTCA. Each patient record contains clinical information on the admission date, date of operation, discharge date, main operation and all other operations the patient might have had as well as the main diagnosis. Additionally, the dataset includes organisational and geographical information. For patient hospital admission, we consider both, elective and emergency admissions. Finished consultant episodes are retrieved for CABG and PTCA based on the procedure codes (OPCS-4) either for the main operation or for any secondary operation. The graph on the left in Figure 1 shows the increase over time in PTCA volume compared to a decreasing trend in CABG. Figure 1: Volume and Workforce (I) (II) The ESR provides counts of FTE medical professionals by specialty and provider in each financial year for the same financial years as the HES data. The data also includes information on nurses, health visitors and midwifes but it is not disaggregated by specialty. The graph on the right in Figure 1 shows the overall increase in total number of doctors in England. The graph also shows the increase by seniority level and the highest increase are for consultants and specialists. Figure 2 shows the increase in number of cardiologists and cardiothoracic surgeons. The increasing numbers for cardiothoracic surgeons is specially striking considering the decreasing trend in CABG volumes to the lowest levels in the last years of our study period. Regarding the whole period of analysis (1999-2012), the average annual rate of growth was 6.8% for the doctors (all specialties), while it was 8.8% for cardiologists and 6.2% for cardiothoracic surgeons. Figure 3 shows the evolution on the number of total nurses 7

in England. The rise in the last part of the study period is a response to the safety concerns specified in the Lord Darzi s 2008 report of the NHS Next Stage Review, High Quality Care for All (Department of Health 2008). The HES data was aggregated at the provider level to construct a longitudinal dataset with the total volume for PTCA and CABG by provider by year. The volume data was then merged with the ESR data to include the workforce information. The result of the merge gave an unbalanced panel data of 199 providers from 1999/2000 to 2012/13. However, after removing providers with volumes of PTCA and CABG below 50 procedures per year 1 (less than one intervention per week) and taking into account the providers mergers during the study period, the final dataset accounts for 79 providers from 1999/2000 to 2012/13. Figure 2: Cardiothoracic Surgeons and Cardiologists (I) (II) Figure 3: Nurses (I) Patient information extracted from HES such as the % of male patients, % of emergency procedures, the average Charlson morbidity index and the average Index of Multiple Depri- 1 Providers that have a high volume of PTCA, but few for CABG (or reverse) are kept. Moreover, providers that have low volume levels at the beginning of the period, but then the volumes go above 50, are kept for the whole period. 8

vation of the area where the patient resides are included to control for the average case-mix of patients being treated by each individual provider. We also include a number of provider characteristics such as whether they have foundation trust status, they are a teaching trust, bed occupancy rate, total number of sites which the trust occupies for health care services delivery and total annual cost for estate services. 2 We also account for the population rate of individuals aged 45 plus at the PCT level and some specifications also account for the prescription of statins in each PCT to control for the use of non-surgical technologies that may reduce the need for PTCA/CABG as these are prescribed also for primary prevention of cardiovascular disease. Table 1 in the Appendix shows a description of the variables used and some descriptive statistics. 4 Results Tables 1 and Table 2 present the results for Equation 1. The dependent variable is the volume of PTCA over the population aged 45 and above (in 1000s) and the main explanatory variable is the corresponding measure for CABG. Table 1 shows the estimates of a fixed-effects panel data model, while Table 2 shows a first-differenced panel data model. In both tables, the estimates in Column 1 account for the whole sample 79 providers and Column 2 restricts the sample to those providers with volume above 50 procedures per year (for both PTCA and CABG) for the period 1999-2012. Column 3 and 4 include the log of statins to control for other non-surgical competing technologies that may affect the volume of PTCA. The sample now only covers the period 2008-2012, as this is the period for which we have data on prescription of statins. Regarding the number of providers, Column 3 has all available providers for the period, while Column 4 only accounts for the one with high volumes of PTCA and CABG. The results in Tables 1 and Table 2 show the degree of substitution or complementarity between both technologies. Each interaction term represents a change in the CABG rate coefficient relative to 1999-2002. As Cutler & Huckman (2003) stated, the value of the substitution for any given period the sum of the base CABG coefficient and the relevant interaction term is potentially biased. Nevertheless, the change in this coefficient over time captured by the interaction terms will be unbiased (based on Cutler & Huckman (2003) assumptions). The main result of Table 1 is that there exist complementarity between these two technologies (positive relationship). The first coefficient of the model (expansion effect) shows an increase in output of 69% (Column 1) and 75% (Column 2). The interaction terms (marginal complementarity) are positive and significant for Column 1 and 2. Moreover, they increase from 2003-2007 to 2008-2012 compared to the baseline period 1999-2002. Another important 2 Teaching trust is not included in the models due to multicollinearity. 9

result of this model is that the statins variable is negative and significant (Column 3 and 4), showing a substitution effect between this type of drugs and the PTCA procedure. Table 2 follows the same structure as Table 1, but instead of levels, it considers differences. In other words, Equation 1 is re-specified as a first-differences panel data model. Comparing both tables, the main difference is that in this last model the interactions are negative, but not significant. In this table, statins are not significant for Column 3 and 4. Table 1: Results Equation 1: PTCA/CABG volume - regression in levels Dependent: PTCA/1000pop45+ (1) (2) (3) (4) CABG/1000pop45+ 0.690*** 0.750*** 0.473 0.318 (0.202) (0.211) (0.370) (0.445) CABG/1000pop45+ * (2003-2007) 0.336*** 0.320*** (0.0586) (0.0806) CABG/1000pop45+ * (2008-2012) 0.391*** 0.564*** (0.0951) (0.128) log Statins -10.44** -15.73** (4.041) (7.633) Constant 11.45** 23.99* 162.5*** 240.6** (4.833) (14.01) (56.30) (92.76) N 868 357 363 144 R 2 0.442 0.523 0.416 0.583 Providers 79 29 76 29 Year fixed-effects Yes Yes Yes Yes Provider fixed-effects Yes Yes Yes Yes Controls Patients Yes Yes Yes Yes Controls Providers Yes Yes Yes Yes Years 1999-2012 1999-2012 2008-2012 2008-2012 Notes: Robust standard errors in parentheses. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. A dynamic panel model was also considered for Equation 1. But, basically, the result of the variables of interest were very similar. The coefficient for overall output was the same and the interactions too, so for brevity we merely report the static specifications. With regards to Equation 2, as explained in the Empirical Strategy section, the model is specified as a dynamic panel data model, following the one-step robust system GMM. Table 3 and 4 present the results of Equation 2. The difference between Column 1 to 6 is the dependent variable used in each model. Column 1 to 4 uses as dependent variable the ratio of cardiologist over cardiothoracic surgeons per provider and year (adjusted by total number of cardiologist/cardiothoracic surgeons in England per year). However, while Column 1 is the sum of consultants, specialists and trainees, Column 2 to 4 disaggregate the ratio by the seniority level of the physician. Finally, Column 5 and 6 have as dependent the ratio of cardiologist over nurses or cardiothoracic surgeons over nurses, respectively. The results of Table 3 and 4 show that the coefficient relating to the effect of workforce 10

Table 2: Results Equation 1: PTCA/ CABG - regression in differences Dependent: PTCA/1000pop45+ (1) (2) (3) (4) CABG/1000pop45+ 1.089*** 1.183*** 0.700 0.770 (0.0916) (0.117) (0.503) (0.527) CABG/1000pop45 * (2003-2007) -0.130-0.149 (0.311) (0.296) CABG/1000pop45 * (2008-2012) -0.352-0.421 (0.532) (0.575) Log Statins -0.104-0.250 (0.150) (0.368) Constant -0.0734 4.095-1.702-3.286 (1.250) (5.331) (1.879) (8.986) N 784 347 353 143 R 2 0.484 0.559 0.168 0.241 Year Dummies Yes Yes Yes Yes Controls Patients Yes Yes Yes Yes Controls Providers Yes Yes Yes Yes Years 1999-2012 1999-2012 2008-2012 2008-2012 Notes: Robust standard errors in parentheses. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. training (lag of the dependent) is positive and highly significant, confirming strong learning effects associated with the previous workforce training. This supports our dynamic specification. The main result of Table 3 is that there is a positive and significant effect of the ratio of PTCA over CABG on the ratio of cardiologist over cardiothoracic surgeons for Column 1 (for the overall and the interaction with the last time period). In other words, the relationship among these two technologies (complementary, as seen from the results of Equation 1) affects the workforce planing by increasing the number of cardiologist over cardiothoracic surgeons. This result was expected due to the increase of PTCA over CABG in the study period. However, when disaggregating it by the seniority level of the physician only consultants seem to be driving this result, the rest are not significant. Regarding the ratio over nurses, Column 6 shows a negative and significant effect of the ratio of PTCA over CABG on the ratio of cardiothoracic surgeons over nurses. Table 4 follows the same structure of Table 3 but, in this case, the PTCA/CABG ratio is now defined as the PTCA rate/cabg rate as adjusted for population at risk. The results does not show any significant effect of these rate on the workforce ratio (except from the cardiothoracic surgeons over nurses ratio). Equation 2 model is supported by several specification tests. The Sargan test of the null 11

hypothesis that the overidentifying restrictions are valid is accepted 3. The t-statistics for the null of no first-order autocorrelation fail to reject the null hypothesis, while the null of no second-order autocorrelation is not rejected at any significance level. As showed by Arellano & Bond (1991) the presence of first-order autocorrelation does not affect the consistency of the specification of the model as long as there is no second-order autocorrelation, as it is the latter the required assumption for the correct specification of GMM methods. So, our models fulfills this requirement. 3 Except from model 3 and 6 of Table 3 and 1, 3 and 6 of Table 4. 12

Table 3: Results Equation 2: Workforce - PTCA/CABG ratio C/CS C/CS consultant C/CS specialist C/CS trainees C/nurses CS/nurses (1) (2) (3) (4) (5) (6) Lag dependent 0.595*** 0.611*** 0.324*** 0.468*** 0.683*** 0.680*** (0.0683) (0.0672) (0.0611) (0.0768) (0.0652) (0.0890) vol PTCA/vol CABG 0.0901*** 0.0554* 0.0644-0.00396-0.0122-0.106*** (0.0325) (0.0306) (0.0447) (0.0351) (0.0195) (0.0338) vol PTCA/vol CABG* (2003-2007) 0.0204 0.00916 0.0125-0.00586 0.00587-0.0175 (0.0200) (0.0189) (0.0322) (0.0406) (0.0130) (0.0178) vol PTCA/vol CABG* (2008-2012) 0.0479* 0.0354 0.0271 0.00990 0.00390-0.0390* (0.0253) (0.0267) (0.0354) (0.0467) (0.0168) (0.0228) Constant -0.928-1.598* -1.023-1.590-1.348** -1.491** (0.862) (0.865) (1.116) (1.081) (0.677) (0.708) N 789 789 789 789 789 789 Providers 79 79 79 79 79 79 Year Dummies Yes Yes Yes Yes Yes Yes Controls patients Yes Yes Yes Yes Yes Yes Controls provider Yes Yes Yes Yes Yes Yes Years 1999-2012 1999-2012 1999-2012 1999-2012 1999-2012 1999-2012 Notes: Robust standard errors in parentheses. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. Sargan test is not fulfilled for models of Column 3 and 6. The Abond test for autocorrelation is fulfilled for all models. 13

Table 4: Results Equation 2: Workforce - PTCArate/CABGrate ratio C/CS C/CS consultant C/CS specialist C/CS trainees C/nurses CS/nurses (1) (2) (3) (4) (5) (6) Lag Dependent 0.718*** 0.743*** 0.383*** 0.560*** 0.759*** 0.821*** (0.0827) (0.0826) (0.0703) (0.0641) (0.0640) (0.0582) PTCArate/CABGrate -0.125-0.0913-0.175-0.0863-0.0197 0.0186 (0.0809) (0.0651) (0.166) (0.128) (0.0904) (0.0438) PTCArate/CABGrate*(2003-2007) 0.116 0.0256 0.0336 0.0798-0.00496-0.0246 (0.102) (0.0720) (0.175) (0.142) (0.117) (0.0825) PTCArate/CABGrate*(2008-2012) -0.0275-0.0176 0.148-0.118 0.00636 0.140* (0.0847) (0.0719) (0.188) (0.154) (0.120) (0.0803) Constant -1.608-1.608* -2.118* -0.632-0.311 1.137 (1.011) (0.921) (1.187) (0.897) (0.743) (0.724) N 681 681 681 681 681 681 Providers 79 79 79 79 79 79 Year Dummies Yes Yes Yes Yes Yes Yes Controls Patients Yes Yes Yes Yes Yes Yes Controls Providers Yes Yes Yes Yes Yes Yes Years 1999-2012 1999-2012 1999-2012 1999-2012 1999-2012 1999-2012 Notes: Robust standard errors in parentheses. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. These models have just one lag of PTCArate/CABGrate ratio as instrument. Sargan test is not fulfilled for models of Column 1,3 and 6. The Abond test for autocorrelation is fulfilled for all models. 14

5 Discussion The objective of this research was twofold. First of all, we examined the substitution or complementarity effects across two technologies. In particular, the analysis looked at the introduction of PTCA for the treatment of cardiovascular disease compared to the existing technology, CABG. Secondly, we computed estimates of the degree to which the workforce reacts to the introduction of new technology, based on elasticity of supply measures. These objectives were achieved through panel data models that determine the manner in which technology is diffused across the NHS, with a particular emphasis on the impact that technology has on the workforce composition. Moreover, we made use of two main databases to make these analyses: the Hospital Episodes Statistics (HES) and the NHS Electronic Staff Records (ESR). The results relating to the first objective show a complementarity effect between CABG and PTCA (both adjusted by population at risk). This is not in line with previous research on the topic that found a substitution effect across these technologies (Cutler & Huckman 2003) and (McGuire et al. 2010). The reason for this different finding could be the time period of analysis. In our research, from 1999 to 2012, throughout the period, PTCA started to be considered as a more mature technology, while in Cutler & Huckman (2003) (1982 to 2000) and McGuire et al. (2010) (1989-2003), it was considered a recently introduced technology. Another interesting finding is that the prescription of statins for the use of non-surgical technologies show a negative and significant effect on the need for PTCA. As a result, the model is showing a substitution effect between statins and PTCA. With regards to the second objective, the estimates show a positive and significant effect of the PTCA/CABG ratio on the workforce planning (cardiologist over cardiothoracic surgeons ratio). As a result, it seems that the two technologies relationship (complementarity), is relevant to explain the number of FTE cardiologist and cardiothoracic surgeons. This result was expected because as the data showed, there has been an increase of PTCA interventions in the study period and the cardiologists are the ones performing this type of procedure. However, it is interesting to see that it is mainly driven by the increase on the number of consultant rather than specialists or trainees. It would appear that new technology up-take and diffusion does affect the skill mix of the medical workforce. This is important, especially at a time when funding constraints appear to be having a impact on staffing levels. While our results are preliminary it would appear that the complex regulation of staffing and specialty mix is even further complicated once account is taken of the impact of new technology on the hospital production process. This paper presents the preliminary results of our research. We are still waiting for more disaggregated data for the workforce variables. We believe that controlling for wages and physicians characteristics (such as, age or gender) will improve our estimates. Moreover, an updated version of the paper will include another type of patient admissions into NHS 15

hospitals for PTCA and CABG. At the moment, there are elective and emergency procedures, but, we also want to include the procedures coded as transfer to another provider. Further research will not only incorporate more specific data on workforce and PTCA/CABG, but it will also analyse other technologies. For instance, the introduction of laparoscopic procedures in bariatric surgery or in prostatectomy interventions. Acknowledgement This paper is part of the research project entitled Health Care Technology Diffusion in the NHS and workforce impact funded by The Health Foundation. 16

References Acemoglu, D. & Finkelstein, A. (2008), Input and technology choices in regulated industries: Evidence from the health care sector, Journal of Political Economy 116(5), 837 880. Arellano, M. & Bond, S. (1991), Some tests of specification for panel data: Monte carlo evidence and an application to employment equations, The Review of Economic Studies 58(2), 277 297. Blundell, R. & Bond, S. (1998), Initial conditions and moment restrictions in dynamic panel data models, Journal of Econometrics 87, 115 143. Bojke, C., Castelli, A., Grasic, K., Howdon, D. & Street, A. (2016), Productivity of the english nhs:: 2013/14 update, Workingpaper importmodel: Workingpaperimportmodel, Centre for Health Economics, University of York. Bond, S. (2002), Dynamic panel data models: a guide to micro data methods and practice, Portuguese Economic Journal 1(2), 141 162. Coscelli, A. & Shum, M. (2004), An empirical model of learning and patient spillovers in new drug entry, Journal of Econometrics 122(2), 213 246. Crawford, G. S. & Shum, M. (2005), Uncertainty and learning in pharmaceutical demand, Econometrica 73(4), 1137 1173. Cutler, D. M. & Huckman, R. S. (2003), Technological development and medical productivity: the diffusion of angioplasty in new york state, Journal of Health Economics 22(2), 187 217. Department of Health, U. (2008), High quality care for all. nhs next stage review final report, London, The Stationery Office CM7432. Dranove, D., Garthwaite, C., Li, B. & Ody, C. (2015), Investment subsidies and the adoption of electronic medical records in hospitals, Journal of health economics 44, 309 319. Elliott, R., Ma, A., Sutton, M., Skatun, D., Rice, N., Morris, S. & McConnachie, A. (2010), The role of the staff mff in distributing nhs funding: taking account of differences in local labour market conditions, Health economics 19(5), 532 548. Gray, A. & McGuire, A. (1989), Factor input in nhs hospitals, Applied economics 21(3), 397 411. Gray, H., Swanton, R., Schofield, P., Murray, R., Brooksby, I., Venn, G., Perrins, J., debelder, M., Smith, L., Hall, R. & Cumberland, D. (2000), Coronary angioplasty: guidelines for good practice and training?, Heart 83, 224 235. 17

House of Commons Health Committee (2007), Workforce planning: Fourth report of session 2006 07, London: The Stationery Office. Imison, C. & Bohmer, R. (2013), Nhs and social care workforce: meeting our needs now and in the future, London: The Kings Fund. Lafond, S., Charlesworth, A. & Roberts, A. (2017), A year of plenty?, The Health Foundation. Lammers, E. (2013), The effect of hospital physician integration on health information technology adoption, Health economics 22(10), 1215 1229. McGuire, A., Raikou, M., Windmeijer, F. & Serra-Sastre, V. (2010), Technology diffusion and health care productivity: Angioplasty in the uk. Molina, J. A. & Heng, B. H. (2009), Global trends in cardiology and cardiothoracic surgery an opportunity or a threat?, Ann Acad Med Singapore 38(6), 541 545. Newhouse, J. P. (1992), Medical care costs: how much welfare loss?, The Journal of Economic Perspectives 6(3), 3 21. Serra-Sastre, V. & McGuire, A. (2009), Diffusion of health technologies: evidence from the pharmaceutical sector, The economics of New health technologies pp. 53 69. Serra-Sastre, V. & McGuire, A. (2012), Technology diffusion and substitution of medical innovations, in The Economics of Medical Technology, Emerald Group Publishing Limited, pp. 149 175. Serra-Sastre, V. & McGuire, A. (2013), Information and diffusion of new prescription drugs, Applied Economics 45(15), 2049 2057. Skinner, J. & Staiger, D. (2015), Technology diffusion and productivity growth in health care, Review of Economics and Statistics 97(5), 951 964. Smith, S., Newhouse, J. P. & Freeland, M. S. (2009), Income, insurance, and technology: why does health spending outpace economic growth?, Health Affairs 28(5), 1276 1284. 18

Appendix Table 1: Variables Definition and Descriptive Statistics Variable Definition Source N Mean St.Dev v ptca PTCA volume HES 927 486.6 489.2 v cabg CABG volume HES 927 220.7 298.5 p males % male patients HES 927 0.733 0.168 imdi Index of Multiple Deprivation HES 927 0.138 0.0523 (IMD) income charlson Charlson Comorbidity Index HES 927 4.283 2.424 p emergency % emergency operations HES 927 0.336 0.286 ftrust =1 if provider has FT status HES 927 0.300 0.458 teaching =1 if teaching status ERIC 894 0.320 0.467 occuprate Overnight bed occupancy rate NHS England 888 85.06 5.490 #sites # sites the trust occupies for services delivery ERIC 886 6.949 8.848 estatescost Annual revenue cost ERIC 912 16.257 18.47 ( 000,000) to provide the whole of the Estate services Pop45 % population aged 45 and ONS 916 38.37 6.640 above 45 by PCT dptcacabg =1 if provider has high volumes HES 927 0.425 0.495 (above 50) of PTCA and CABG dptca =1 if provider has high volumes HES 927 0.575 0.495 (above 50) of PTCA and low CABG volumes Total Doctors # FTE doctors ESR totconsultant Consultants 927 223.8 118.2 totspecialist Specialist 927 226.3 154.5 tottrainess Trainees 927 91.73 74.73 Cardiologists # FTE cardiologists ESR c consultant Consultants 927 7.192 4.665 c specialist Specialists 927 7.864 7.265 c trainees Trainees 927 2.699 3.734 Continued on Next Page... 19

Table 1: Variables definition and Descriptive Statistics Continued Variable Definition Source N Mean St.Dev Cardio surgeons # of FTE cardio surgeons ESR cs consultant Consultants 927 3.444 4.551 cs specialist Specialists 927 4.388 6.455 cs trainees Trainees 927 0.849 1.960 nurses # FTE nurses ESR 927 1311 619.9 statins Number of items prescribed (2008-2013) NHS Digital 364 507979 291682 Notes: p males, imdhd, imdi, imdle, charlson and p emergency are averages over patients treated by trust i each year. ESR (Electronic Staff Records), NHS Digital. ERIC (Estates Return Information Collection) from Hospital Estates and Facilities Statistics. 20