The impact of risk management standards on the frequency of MRSA infections in NHS hospitals 1

Transcription

1 The impact of risk management standards on the frequency of MRSA infections in NHS hospitals 1 Paul Fenn Nottingham University Business School Alastair Gray Health Economics Research Centre, University of Oxford Neil Rickman University of Surrey and CEPR Dev Vencappa Nottingham University Business School Oliver Rivero Health Economics Research Centre, University of Oxford Emanuela Lotti University of Surrey 19th February 2007 Abstract We test for the effects of risk management standards on the frequency of MRSA infections in UK hospitals. ML and GMM estimators are used to test for the relationship between risk management standards and the number of infections observed. Our findings are consistent with the exercise of greater infection control measures by hospitals with higher assessed risk management standards, after controlling for activity levels, bed utilisation rates, and casemix. In the context of wider debates about medical malpractice reform, these results are also consistent with hospitals reacting to the incentives provided by discounts applied to their liability risk-pooling contributions. KEYWORDS: Medical malpractice, insurance, litigation JEL Classification: I18, K13 1 Corresponding author Paul Fenn, Nottingham University Business School, Jubilee Campus, Wollaton Rd, Nottingham NG8 1BB, UK. paul.fenn@nottingham.ac.uk. We thank the National Health Service Litigation Authority for kindly permitting access to data. This research forms part of a project entitled Public Services: liability, risk pooling and health care quality, which is funded by the ESRC as part of the Public Services Programme, and we are grateful to other participants in that programme for comments on our work.

2 1 Introduction The search for mechanisms to encourage patient safety has been active in numerous healthcare systems for many years (see Fenn et al, 2004, for a survey). 2 In principle, one important mechanism for encouraging patient safety is through the provision of financial incentives to hospitals. Whilst tort liability seeks to achieve this (by making suppliers of demonstrably substandard care liable for the damage caused), its effects may be diluted by hospitals use of liability insurance to pool such potentially large risks. Liability insurers can seek to overcome this by setting deductibles, by experience rating their premiums, and (the focus of this paper) by the use of merit rating where the insurer specifies a direct relationship between the premium charged and verifiable procedures put in place by hospitals to minimise risk. 3 The purpose of merit rating is to constrain the hospital to take more care and, as such, to mitigate the moral hazard problem. 4 In this paper, we seek to assess the extent to which the discounts offered to hospitals for reaching higher risk management standards have incentivised hospitals to improve their safety procedures, and the extent to which this has effected a reduction in risks to patients, measured here by the incidence of hospital-acquired infections. We are able to control for various observable measures of hospital risk type, including hospital throughput, bed utilisation, 2 In the National Health Service (NHS), this search has recently culminated in a public report by the Chief Medical Officer (DoH, 2003) and a subsequent Act of Parliament (the NHS Redress Act 2006, which received Royal Assent on 8th November 2006); it has also attracted the attention of the National Audit Office (2000). 3 Such mechanisms are not, of course, observed only in health care. For example, see XXXX for a discussion of their use in employers liability insurance. 4 Rickman et al (2007) demonstrate that such policies can also be useful in a setting where the insurer faces hospital moral hazard and adverse selection (in terms of risk type). In this case, incentive compatible risk management standards (with premium discounts) can separate the risk types and address the moral hazard problem. 2

3 and casemix. To our knowledge this is the first study to examine the link between risk management standards and patient safety. 5 The hospital-acquired infections that we examine are methicillin resistant Staphylococcus aureas (MRSA) infections. MRSA (and similar infections such as clostridium difficile) have increased markedly in the NHS over the last fifteen years (deaths from MRSA alone from 51 in 1993 to 955 in ), leading to the introduction of a mandatory surveillance system in England in 2001 (Health Protection Agency, 2006), vapour treatments in some hospitals, the specific employment of a hygiene company by the NHS and several ward closures. These costs, along with well-documented compensation payouts to a number of MRSA sufferers, additional treatment costs of approximately 1bn per year (National Audit Office, 2000) and reputation costs to the NHS itself make MRSA a high profile and important focus for the patient safety debate in the NHS. There are consequently clear public policy benefits from understanding the role of risk management standards in controlling MRSA. Moreover, such an understanding helps to evaluate whether this particular mechanism is effective in controlling an insured s behaviour, which is important from theoretical and institutional perspectives that extend beyond the health care setting. The paper is structured as follows. Section 2 presents an overview of the risk pooling measures available to NHS hospitals with respect to their liability to patients, and outlines the basis for the calculation of each hospital s contribution to the risk pool. Section 3 reviews the panel dataset available to us from UK hospitals, and the following section outlines the estimation methodology by which we use this panel dataset to test hypotheses concerning 5 In a companion paper, we explore the relationship between hospitals policy excesses and their care decisions: Fenn et al (2007). 6 This pace of increase has continues to the present. Deaths involving MRSA have increased by 39% (to 1,629) since 2004/05. Recent Department of Health analysis suggests that the NHS will miss targets set for MRSA outbreaks by April

4 the effect of risk management standards on the frequency of MRSA infections. In particular, we hypothesise that the risk management discounts available to hospitals led to lower patient risk proxied by fewer hospital-borne infections, via the incentives these discounts provided to improve externally assessed risk management standards. Section 4 outlines our estimation methodology. Section 5 presents our results and a final section concludes. 2 Liability risk-pooling in the NHS The main body responsible for administering schemes allowing NHS trusts in England to pool the costs of liabilities to patients arising from the carrying out of their functions is the NHS Litigation Authority (NHSLA), a Special Health Authority established in November The NHSLA is responsible for administering the Clinical Negligence Scheme for Trusts (CNST), a voluntary scheme to which all English NHS Trusts and PCTs currently belong, and covers clinical incidents occurring on or after the date when the Trust joined the scheme. The current CNST scheme is a pay as you go scheme, in which contributions for each year are calculated so as to cover expected costs, including claims and administrative costs. The pay as you go approach minimises the cash reserves that are required to operate the scheme. The basis of the scheme is an actuarial model that forecasts the expenditure based on the predicted number and value of claims arising in each period. If contributions are set at a level that exceeds the outturn expenditure, these are passed back to members in the form of rebates on contributions in future years. Having estimated the aggregate contributions required to finance the scheme, individual contributions from each scheme member are calculated. Contributions are based on the number of staff employed in different risk 4

5 categories by each Trust member. Whole time equivalent staff are counted and then divided into five risk categories, reflecting the national experience of claims rates and values by specialty: low risk, medium risk, high risk, very high risk, and obstetrics & gynaecology. Weightings are attached to each of these, on the basis of the predicted number and value of claims in each category, and the weights are given a cash value that will equate the total contributions required from all members times the total staff employed by all members. It is then possible to calculate the annual contribution required from each member. Two adjustments are then made to this estimated contribution. 7 First, each Trust is experience rated on the basis of actual compared with expected claims based on risk exposure measured using the WTE staff data. The experience rating takes one of five values: +10%, +5%, 0%, -5% and -10%. Second, and most importantly for this paper, the scheme also gives a role to the risk management processes that members have in place when determining contributions. Assessments are currently routinely conducted every 2 years, but organisations failing to attain the first level of standard are assessed annually. The assessment is based on seven core standards 8. Trusts which are assessed as complying with the standards will be entitled to a discount from their scheme contribution for the following two financial years. The discounts on CNST contributions are 10% (level 1 compliance), 20% (level 2 compliance) and 30% (level 3 compliance). The discount earned by members is applied to contributions in the financial year following a successful assessment and is valid for 2 years. 7 In addition, until 2002, hospitals were permitted to choose deductibles or excess levels below which they were responsible for the patient s claim, and which in turn affected the contribution paid. This facility no longer exists. 8 In addition, there is a separate standard for adult mental health services, which is applicable only to Trusts that provide such services. 5

6 3 Data 3.1 MRSA infections Data on MRSA infection rates were extracted from the Health Protection Agency Communicable Disease Surveillance Centre. These cover the period from April 2001 (when mandatory surveillance began) to September Data were available for each NHS Trust by number of MRSA bacteraemia reports and MRSA rates per 1000 bed days. Figure 1 presents the mean MRSA bacteraemia reports per year and the annual MRSA rates per 1000 bed days. There appears to be no strongly increasing or decreasing trend over these years. Figure 2 shows how the number of MRSA infections per hospital were distributed across NHS hospitals over each of the four years of observation. Clearly, there are considerable differences in the number of infections each year, which will be influenced by hospital throughput and casemix, but which may, of course, also be a function of the infection control measures introduced in these hospitals. *** Figure 1 here *** *** Figure 2 here *** 3.2 Risk management discounts The payment for CNST cover varies depending on the hospital s casemix, the claims experience, and the risk management standards applied. Hospitals with low risk management standards face a higher CNST contribution than those 6

7 with high standards. 9 Moreover, these standards have been assessed and reassessed by the NHSLA over time. Hospitals have a clear financial incentive to improve their procedures via the risk management discounts which are applied. This creates an opportunity to test whether this incentive led to improvements which resulted in better control over hospital-borne infections. Table 1 shows the variation in assessed risk management standards across all English hospitals in the scheme from It is evident that the trend over time has been for an overall progressive improvement in the assessed standard of risk management procedures, although considerable variations between hospitals have existed, and remain throughout the period of observation. *** Table 1 here *** One fundamental issue for the subsequent analysis of these data is the potential endogeneity of the risk management standard. It is quite plausible that the perception of liability risk from infections was an important consideration when hospital management decided upon the extent to which risk management procedures should be targeted and improved, with a consequent risk of simultaneity bias in the estimates. We therefore test for this possibility in the estimates reported below. 3.3 Size and casemix variables The most important factors potentially influencing inter-hospital variations in MRSA infections relate to the size, capacity utilisation and casemix of the hospital direct measures of exposure to infection. Clearly, raw activity level measures such as the number of admissions or treatment episodes at a particular hospital will be a factor determining the number of infections. In 9 10% discount for attaining level 1, 20% for level 2, 30% for level 3. 7

8 addition, the nature of the treatment episodes will influence the frequency of infection: hospitals with a large proportion of acute beds or surgical patients may be more open to infection than others, for instance. Table 2 summarises the panel of data we have in relation to hospital size (measured by the total number of bed days), bed utilisation rates (i.e. bed days relative to bed capacity), and casemix variables (i.e. the proportion of bed days allocated to the main treatment specialities). ***Table 2 here*** In what follows we use these broad utilisation and casemix variables as proxies for the risk type of individual hospitals that is, their exposure to risk of infection. 4 Estimation 4.1 Maximum likelihood estimation Following Fenn et al (2007), for a given hospital, the process over time by which observed data are generated on the numbers of MRSA infections could be characterised as a Poisson process with a constant rate of occurrence, µ. The observed number of infections would clearly depend on the population exposed to risk that is, the number of treatment episodes for the hospital in a given year, N. Consequently, the expected number of infections in a given year at a given hospital would be the product Nπ, where π represents the mean probability of an infection being required for a given treatment episode. Given an assumed Poisson process, this would imply that the observed number of infections (x) in a hospital in a given year is distributed with density 8

9 Nπ x e ( Nπ ) f( x; N, π ) = (1) x! The parameter π can be modelled as a function of observed covariates and unobserved random variables. With a conventional loglinear specification of this function, we have π = exp( β ln N + β d + βρ + ε) (2) where ln N is included as a regressor to capture the possibility that the daily likelihood of an infection is sensitive to the activity level of the hospital. If β 1 = 1 the number of infections increase proportionately with the number of bed days; if β 1 < 1 the number of infections in a hospital increases less than proportionately with the number of bed days, and if β 1 > 1 the number of infections increases more than proportionately with the number of bed days. The variable d measures the risk management standard achieved by the hospital and ρ is a vector of measures capturing the hospital s casemix and bed utilization rates; β 1, β 2 and β 3 are the associated coefficients. The error term ε measures the impact of unobserved heterogeneity in the underlying risk across hospitals. Incorporating (2) into (1) leads to overdispersion of the Poisson distribution. A mixed distribution can be obtained once an assumption is made about the distribution of exp(ε). A common assumption for the heterogeneity is the gamma distribution, and the resulting Poisson/gamma mixture can be shown to generate a negative binomial distribution for x: 1 1 Γ ( α + x) α Nπ f( x; N, πα, ) = Γ( α ) Γ ( x + 1) α + Nπ Nπ + α α 1 x (3) 9

10 where α defines the one-parameter gamma distribution for the heterogeneity variable exp(ε). If the distributional assumptions explicit in the above are correct, estimation of the parameters α, β 1, β 2 and β 3 by maximum likelihood is straightforward. Moreover, given the panel data structure of our dataset, we can allow for the heterogeneity parameter itself to vary across hospitals. A random effects overdispersion model can be estimated using STATA s xtnbreg,re command, which allows this parameter to vary randomly across hospitals such that it follows a two parameter Beta distribution. As pointed out above, we believe that d is a potentially endogenous regressor due to the discretion by which hospital management decides what level of expenditure is required on risk management activities (including infection control measures). While it is feasible, though difficult, to test and allow for this endogeneity in a two-step estimation process (see Wooldridge, 1997; p384), we have a computationally less complex alternative available to us which is described in the following section. 4.2 General method of moments estimation If we maintain the loglinear specification for the relationship between the incidence of infections and observed covariates (equation (7)), a more general specification of the count data model with multiplicative unobservables ω would be 10 x = exp( β ln N + β d +βρ ) ω (10) Windemeijer and Santos Silva (1997) point out that if the error term was additive rather than multiplicative, then an instrument set which is orthogonal to one may not be orthogonal to the other. We abstract from this issue here by assuming that the error enters multiplicatively. 10

11 This is a standard exponential regression model for non-negative dependent variables and one which requires fewer distributional restrictions than the ML estimators discussed above. Furthermore, if x is strictly positive (as with our data), equation (10) can be transformed for estimation as ln( x) = β ln N + β d + βρ + ln( ω) (11) If we can further assume that E[ln(ω) z] is a constant independent of the instrumental variable vector z, then even where d is shown to be endogenous it is possible to estimate the parameters consistently using a linear IV estimator (see Mullahy, 1997; p590). However, because the underlying count data process has an intrinsic heteroskedasticity, and because the standard linear IV estimator is inconsistent in the presence of endogeneity together with heteroskedasticity of uncertain origin, we prefer to estimate the model in equation (11) using the GMM estimator which is fully robust and asymptotically efficient (Baum et al, 2003). Given the panel structure of our dataset, efficient GMM estimation requires a consistent estimator of the covariance matrix. STATA s ivreg2 estimator provides robust estimation in the presence of clustering (Baum et al, 2003; p9). In this context the estimator allows for observations on specific hospitals to be correlated over time. 5 Results The first column of results in Table 3 shows the random effects negative binomial ML estimates for the number of MRSA infections observed over the period The coefficient on the logarithm of the hospital s total number of bed days was strongly significant and greater than one, indicating 11

12 that the number of infections increases more than proportionately with the number of bed days provided. Larger hospitals appear to be more at risk. Moreover, the coefficient on the bed utilisation rate is also positive, implying that hospitals of a given size are more at risk of MRSA the closer they are to full capacity. 11 The casemix variables have plausible impacts; the likelihood of infections is highest for hospitals with more bed days devoted to surgery of all kinds and gynaecology in particular. ***Table 3 here*** The key variable of interest, the attainment of risk management standards 2 or 3, 12 is negative and significant. The potential endogeneity of this variable is tested in the second column of results, which report the coefficients from an efficient GMM estimator, robust to clustering at the hospital level. The instruments used for the endogenous risk management variable are its own first and second period lagged values. The test statistics show, respectively, the F-test for the overall fit of the regressions; the Pagan-Hall χ 2 test for heteroskedasticity; the Wu-Hausman F-test for the endogeneity of the risk management standard; and Hansen s J-test of overidentifying restrictions (a test of instrument exogeneity). The null of exogeneity of the risk management standard under the Wu-Hausman test cannot be rejected. The J-test of overidentifying restrictions does not allow us to reject the null of instrument orthogonality, which means we cannot reject the possibility that all of the instruments are exogenous and uncorrelated with the error. Moreover, the null of homoskedasticity cannot be rejected, indicating that heteroskedasticity does not appear to be problematic. Overall, therefore, our diagnostic tests imply that the prior concern over both endogeneity and heteroskedasticity may not 11 This result is consistent with evidence of a link between hospital capacity pressures and infection rates in Northern Ireland. Whilst our objective is not to consider the effects of capacity pressures directly, it is notable that capacity pressures have been linked to dilutions of care quality in other studies: e.g. XXXX. 12 We pool together levels 2 and 3 to produce a single categorical variable, given the low numbers of hospitals to date which have been assessed at level 3. 12

13 be justified, and not surprisingly, therefore, both sets of results are qualitatively similar: risk management standards are negatively related to the MRSA infection rate, after controlling for the effect of observed variations in bed utilisation, hospital size and casemix, as well as the effect of unobserved heterogeneity. While the impact of risk management standards is significant in both equations, the predicted scale of the effect differs; the ML estimator shows a reduction of 11% in infections for hospitals with higher risk management standards, whereas the GMM estimator shows a reduction of 20% Conclusion Ultimately, whether hospital care levels respond to the financial incentives explicitly incorporated into risk pooling contributions is an empirical matter, and we believe that the data available from the NHSLA has opened up a unique opportunity for research on this topic. A combination of financial autonomy at hospital level and risk management assessments in the years from 2001 to 2004 means that data exist on the extent to which hospitals responded differently to these incentives by implementing improved risk management procedures. Because of the progressive implementation and assessment of standards by the NHSLA, it has been possible to construct a panel of data in which the variations in these standards across hospitals and over time can be captured and related to the measurements in reported MRSA infections. The results reported in this paper are indeed consistent with the implementation of improved risk management procedures having a positive effect on patient safety (i.e. a reduction in the number of infections for given throughput and casemix). Moreover, our subsidiary results indicate that inter-hospital 13 There are similar differences in the predicted impact of other covariates: the predicted elasticity of MRSA infections with respect to bed utilisation is 0.68 in the negative binomial regression, and 1.33 in the GMM regression. 13

14 variation in throughput, casemix and capacity utilization can help explain the incidence of MRSA infections, and consequently that there is a degree of predictability to the geographical distribution of these events. Our results may have useful implications for policy makers. At a time when MRSA infection rates appear difficult to control, the results indicate that financial incentives could be given an important role in this area. More broadly, it appears that an insurer can provide some degree of financial incentives to hospitals so that the presence of insurance itself need not completely dilute the incentives for care provided by tort liability. Such insights are relevant to the wider evaluation of alternative mechanisms for compensating medical injuries as well as those directed at improving specific aspects of patient safety. 14

15 References Baum, C, Shaffer M, Stillman S (2003), Instrumental variables and GMM: estimation and testing Boston College Dept. of Economics, Working Paper No. 545 Department of Health (2003): Making Amends: A consultation paper setting out proposals for reforming the approach to clinical negligence in the NHS. A report by the Chief Medical Officer, Department of Health: London Fenn, P.; Rickman, N.; Gray, A.(2004)., "The Economics of Clinical Negligence Reform in England", Economic Journal, Vol.114 (496), pp. F Fenn, P., Gray, A. and Rickman, N. (2007) Liability, insurance and medical practice, Journal of Health Economics, in press. Health Protection Agency (2006). Mandatory Surveillance of Healthcare Associated Infections Report London: Health Protection Agency. Mullahy, J (1997), Instrumental-Variable Estimation of Count Data Models: Applications to Models of Cigarette Smoking Behavior, Review of Economics and Statistics, Vol. 79(4): National Audit Office (2000) The Management and Control of Hospital Acquired Infection in NHS Acute Trusts in England, HC 230 Session London, The Stationery Office. Rickman, N., Lotti, E, Fenn P, Gray A, Vencappa D, Rivero O, (2007) Ex ante regulation in a dynamic model with asymmetric information and uncertainty, mimeo, Department of Economics, University of Surrey. Windmeijer, F A G & Silva, J M C Santos, (1997) "Endogeneity in Count Data Models: An Application to Demand for Health Care," Journal of Applied Econometrics, vol. 12(3), pages Wooldridge, J. (1997). Quasi-maximum likelihood methods for count data. In Pesaran, H. and Schmidt, P. (eds.) Handbook of Applied Econometrics, Vol II, Oxford; Blackwells. 15

16 Tables and Figures Figure 1: MRSA bacteraemia reports and MRSA rates per 1000 bed days from 2001 to MRSA Bacteraemia reports MRSA rates per 1000 bed days MRSA bacteraemia reports Period 0 16

17 Figure 2: Distributions of MRSA infection numbers, NHS hospitals Number of hospitals Number of infections Graphs by Year 17

18 Table 1: Risk management standards, NHS hospitals, Year Risk management standards:

19 Table 2: Bed days, bed utilisation rates and casemix variables, NHS hospitals Total bed days (mean per Trust) Proportion general medicine Proportion general surgery Proportion gynaecology Proportion obstetrics Proportion paediatrics Proportion trauma & orthopaedics Proportion urology Proportion other surgery Proportion other medicine Bed utilisation rate

20 Table 3: Negative binomial ML and GMM estimates for MRSA infections, NHS hospitals Random effects GMM NB Risk management standard 2 or (2.35) (2.07) Ln[Total bed days] (13.07) (10.19) Proportion general medicine (2.23) (1.61) Proportion general surgery (4.05) (1.63) Proportion gynaecology (2.71) (2.49) Proportion obstetrics (0.57) (1.07) Proportion paediatrics (1.42) (0.51) Proportion trauma & orthopaedics (1.94) (1.42) Proportion urology (1.73) (0.50) Proportion other surgery (3.50) (2.70) Proportion other medicine (2.28) (1.72) Bed utilisation rate (1.94) (3.66) year= (0.13) year= (0.01) (0.63) year= (0.17) (0.26) Observations Number of hospitals Wald test P-value Hansen s J test P-value Pagan-Hall test statistic P-value Wu-Hausman F test