Estimating the costs of Hospital Acquired Infection

Nicholas Graves University of Auckland and Diana Weinhold London School of Economics

February 2002 preliminary draft, comments welcome Please do not quote

Abstract: Hospital acquired infections (HAI) are a largely avoidable and costly burden on public health systems. Accurate estimates of the costs of HAI are necessary to gauge the cost efficiency of alternative infection control measures, but are traditionally very difficult to compute. In particular, existing studies of the costs of HAI do not take into account the possible endogeneity bias between the risk of HAI and the length of stay in the hospital, and thus may over-estimate costs. We propose a multi-stage instrumental variables estimation strategy that explictly controls for the endogeneity between risk and length of stay and find that a ten point reduction in ex ante risk of HAI results in an expected savings of $1,800. JEL Code: I10, I18 Keywords: Hospital acquired infection, costs

Development Studies Institute, London School of Economics, Houghton Street, London UK. Email: [Link]@[Link] The authors thank J.S. Butler and Richard Plowman for helpful comments. Remaining errors are the responsibility of the authors.

1. Introduction Approximately one in ten hospitalised patients will acquire an infection after admission to hospital (Plowman et al., 1997) and many have reported that substantial economic costs result (Coello et al., 1993; Haley et al., 1981b; Rubinstein et al., 1982; Scheckler, 1978; Poulson et al., 1994; Pittet et al., 1994; Kappstein et al., 1992; Girard et al., 1983; Plowman et al., 1999). Haley (1991) argues that the primary cost is that patients with hospital-acquired infection (HAI) have their stay prolonged during which time they occupy scarce bed days and are supplied additional diagnostic and therapeutic interventions. The costs from these infections is substantial; the annual economic burden is estimated to be approximately $4 billion per year in the US (Haley, 1985) and 1 billion ($1.41 billion) in the UK (Plowman et al., 1999) and evidence suggests up to 20,000 deaths in the US and 5,000 in the UK result directly from HAI (Plowman et al., 1997). Many of these costs could be avoided as 15-30% of all HAI are considered preventable (Haley, 1985; National Audit Office, 2000). However, to judge the cost effectiveness of any given infection control program, policy makers need data on their incremental costs and benefits. While estimates of the cost of a particular program can be calculated in the usual way, estimating costs of HAI has proven much more complicated and has attracted much methodological debate. There are a myriad of problems facing researchers attempting to affix a cost to HAI. Patients with HAI are often very different to a control group regardless of their HAI status. For example, they may be older, sicker and have a different socio-demographic profile and so quite different cost outcomes regardless of their HAI status. The challenge is to control for these differences in any comparison. The current state of the art technique is statistical regression modelling using data on a cohort of patients with the cost of hospital stay as the outcome variable 1 , and HAI, alongside a number of other terms, as independent variables. However, to date all studies of the costs of HAI have suffered from one serious shortcoming. On one hand there is good evidence that a primary channel through which HAI effects total costs is through increasing the length of stay in hospital. On the other, significant clinical evidence suggests that increased length of stay is itself a primary cause of increased risk of HAI. Thus length of stay and risk of HAI are endogenously determined within any model of the costs of HAI. In failing to account for this bias, the estimates of all previous studies should be interpreted with extreme caution. The purpose of this paper is to formulate a statistical model of the effect HAI on cost outcomes while specifically addressing the endogeniety between length of stay and the likelihood of getting an infection. We use a data set, collected from one UK hospital, of medical and costs data for 899 patients and an instrumental variables estimation strategy to predict the expected cost outcomes resulting from an ex-ante change in the risk of infection. As the potential benefits of competing infection-control interventions are themselves often measured as changes in ex-ante risk of infection, our estimates of the cost savings, also from ex-ante changes in the risk of infection, are ideally suited for analyzing the cost-effectiveness of any proposed risk-reduction measure. Contrary to
1

Kulcher and Golan (1999) argue that the total economic cost of HAI should include employment and welfare loss of the patient as well. While we agree this would be ideal, in the absence of such data it seems that the hospital costs are a reasonable, and more readily available, substitute.

conventional wisdom on the direction of the bias in the traditional literature we find that the costs of HAI are substantially larger than existing estimates suggest. We show that our results are not consistent with a simple feedback mechanism between length of stay and risk of infection, as is generally thought. Rather, the relationship between these two variables appears to be much more complex. The paper proceeds as follows: in section 2 we review the existing literature and methods used to attribute cost to HAI. In section 3 we describe our data and choice of variables. In section 4 we outline the econometric estimation strategy and present results and in section 5 we discuss our findings and draw some conclusions. 2. Background The attribution of cost to HAI has long challenged researchers. Some early attempts involved the crude weighting of a patients length of stay and the multiplication of the extra days with a value of the cost per bed day. These back of the envelope estimates were used to create political urgency and raise awareness of the issue (Haley, 1992). Two further approaches were applied in an attempt to introduce rigour to the exercise; the concurrent and the comparative methods. The concurrent method requires professionally qualified staff to estimate the additional resources that should be attributed to HAI. A study by Scheckler (1978) used a nurse/epidemiologist to estimate by how much patient length of stay had been extended by HAI, the researchers found additional costs of US$1,414 for infections of surgical wounds. Haley et al. (1981b) studied three hospitals and found cost ranges of between $1,300 to $1,450 for lower respiratory tract infections, $500 to $925 for surgical wound infections, and a huge range of $50 to $2,700 for blood stream infections. However McGowan (1982) argued the concurrent approach would produce judgements not reproducible over time or from person to person. Thus in an attempt to make the method more reliable Wakefield et al. (1987) used specially trained staff who worked to carefully prepared protocols that assessed each day of the patient's hospital stay according to whether it was: (a) attributable to the reason for admission; (b) jointly attributable to the reason for admission and the HAI; or (c) attributable to the HAI alone. He found that 20.9% of total length of stay for the sample could be attributed to HAI, equivalent to an extra 7.1 days per infection. The comparative method requires that data are collected on resources used by patients with and without infection, and the level of resource use is compared between the two groups. As these patient groups may have quite different characteristics, and some of these might impact on resource use, there is a need to control for confounding factors. Accordingly, infected patients are matched with uninfected controls on key characteristics such as sex, age, diagnosis, treatment procedures and co- morbidities. Using this method Rubinstein et al. (1982) and Coello et al. (1993) found the costs of a surgical wound infection to be $1,290 and 1,295 ($1,850), respectively, and the cost of multiple sites of infection to be $1,800 and 3,237 ($4,600). Davies and Cottingham (1979) and Green and Wenzel (1977) found the cost of a surgical wound infection to be 775 ($1,100) and $689, respectively. For the comparative method, however, a large sample of controls is required if all infected patients are to be matched, and still there may well be a residue of infected patients for whom no match can be found. The results could be biased if these patients

are simply omitted from the study if, as a group, they are particularly expensive (or cheap) to treat. Thus in attempt to minimise this bias researchers have used statistical regression techniques to control for differences in cost outcomes for those with and without HAI. This approach is more rigorous as all patients are included in the comparisons and it has the added advantage of providing confidence intervals around the cost estimates. Using regression analysis, Hyryla and Sintonen (1994) found that patients with a wound infection stayed four times longer than those without and Plowman et al. (1999) found that patients with one or more hospital acquired infections had their stay prolonged by on average 2.8 times at a cost of 2,917 ($4,170) per case. Nevertheless, while the application of these statistical methods represents the current state of the art, all the studies potentially suffer from a severe endogeneity bias that was discussed in the introduction. In particular, a primary source of increased cost of HAI is through length of hospital stay, but in turn length of hospital stay has been cited as a primary risk factor for HAI (see below). Thus the direction of causality between incidence of HAI and length of stay is not well identified. Plowman et al. (1999) acknowledged the potential for this bias and made an attempt to demonstrate its effect, but in the end could not control for it and concluded, as a result, that care should be exercised when interpreting their results. In what follows we suggest a instrumental variables approach that explicitly takes into account the endogeneity between length of stay and HAI to provide consistent estimates of the resulting costs. 3. Data Our data was collected for a study of the Socio-economic burden of Hospital Acquired Infection conducted by (Plowman et al., 1999). The design was a prospective survey of 4,000 consecutive adult admissions to the non-tertiary medical and surgical services of a district general hospital in the UK. The aim of the study was to determine the overall economic burden imposed by the most common sites of HAI occurring in the in-patient population. Data collection was undertaken by research nurses employed specifically for the project that adhered to strict guidelines for data collection and entry. Information was collected for demographic and clinical variables, whether or not HAI presented during the hospital stay or after discharge, all resources used during the inpatient stay and after discharge, private costs incurred by patients and families and productivity losses. The model developed here will use a subset of the data collected for the Plowman et al. (1999) study. All patients over 65 years, admitted to the general medicine and care of the elderly specialities with or without a hospital acquired infections of the lower respiratory tract (LRTI) were selected. Patients with other sites of HAI were excluded. The reason for this was that a relatively simple relationship between LRTI and cost might exist in this patient group. They demonstrate many risk factors for LRTI and rates of LRTI were highest among them as compared to the rest of the sample. In all, we end up with a sample of 899 observations to be used in the estimation and 29 of these patients acquired LRTI during their hospital stay. The dependent variable we seek to explain is total hospital costs (TOTCOST). The methods used to estimate values for this variable are reported in the second volume of the final report of the Socio-economic burden of Hospital Acquired Infection (Plowman et al., 1999). Briefly, only those resources that could be shown to have a positive opportunity cost in their existing use were included in the cost assessment

exercise 2 . These include capital infrastructure and the time of healthcare professionals, as well as items of diagnosis or therapy, such as x-ray investigations or drug therapy. Thus the aggregate cost vectors for each patient reflected the time of doctors, nurses, other professions allied to medicine, all diagnostic and therapeutic procedures, all drugs, dressings and other consumables used, the overheads of the organization and a value was included for the depreciation of capital assets. The set of control variables was selected by reviewing the literature and identifying important determinants of hospital costs from other studies. From the literature on cohort studies with matched infected and uninfected patient groups (Davies and Cottingham, 1979; Girard et al., 1983; Green and Wenzel, 1977; Rubinstein et al., 1982; Coello et al., 1993) a number of suitable control (i.e. matching) variables were identified. In particular, these were age of the patients, sex of the patient, speciality to which they were admitted, primary surgical procedure, social class, year of discharge, indices of obesity, and the smoking status of the patient. In addition, Graves (2001) finds that admission type and number of co- morbidities significantly influence cost outcomes and postulates that the number of co- morbidities could act as an indicator of the severity of the patients morbidity so affect cost outcomes. Finally, Scitovsky (1984), Scitovsky (1994) Mendleson and Schwartz (1993), Emmanual (1996) and McGrail et al. (2000) all found that proximity to death was an important predictor of cost outcomes. Data were not available from the Socio-economic burden of Hospital Acquired Infection for obesity and smoking status, and primary surgical procedure is not relevant for medical and care of the elderly patients. The complete set of variables that is available and were used in the analysis, along with their definitions, is presented in Table 1. Table 2 presents summary statistics on a subset of important variables. In particular we see that the average age of patients is 79, the average stay was 12 days for an average totcost of 2131. About 55% of the sample is female and just over 9% of the sample died while in hospital. 3.2% of the patients acquired an LRT infection.

Jointly used resources were analysed by the activities for which they were employed and the costs of the activities allocated by an appropriate cost driver, using the method of Johnson and Kaplan (1987). Unit costs were developed for the single items of diagnosis or therapy using standard cost accounting practice (Kaplan and Atkinson, 1989).

Table 1: List of Variables TOTCOST STAY LRTI AGE SEX DIED ELDERLY ONECOM TWOCOM MORECOM JOB0 JOB1 JOB2 JOB3 JOB4 JOB5 JOB6 AD1 AD2 AD3 AD4 NG ET OX total hospital cost of patients stay length of stay in hospital dummy=1 if patient acquired LRTI during hospital stay3 age of patient sex of patient ; 1=female, 0=male dummy =1 if patient died while in the hospital dummy for admission under elderly care dummy for one co- morbidity present dummy for two co-morbidities present dummy for three or more co- morbidities present dummy for professional job dummy for managerial and technical job dummy for skilled non- manual job dummy for skilled manual job dummy for partly skilled job dummy for unskilled job dummy for retired, student or armed forces job dummy=1 if elective admission dummy =1 if urgent admission dummy=1 if emergency admission dummy=1 if other admission dummy =1 if patient has nasogastric tube dummy=1 if endo-tracheal suction was required dummy=1 if patient had oxygen therapy

Table 2: Summary Statistics for Selected Variables, N=899 Variable TOTCOST STAY AGE SEX DIED INFECT2 NG OX Mean Standard Deviation Minimum Maximum 2131.270 2545.290 252.180 37087.430 12.528 12.806 1.000 125.000 79.112 7.367 65.030 98.000 0.554 0.497 0.000 1.000 0.093 0.291 0.000 1.000 0.032 0.177 0.000 1.000 0.020 0.140 0.000 1.000 0.187 0.390 0.000 1.000

For prescise definition of LRTI see Plowman, R. P., Graves, N., Griffin, M., Roberts, J. A., Swan, A. V., Cookson, B. C. and Taylor, L. (1999) The Socioecomic Burden of Hospital Acquired Infection, Public Health Laboratory Service, London.

4. Empirical specification The estimation strategy is designed to control for the possibility that LRTI and TOTCOST are endogenous in that LRTI is associated with TOTCOST through the effect of length of stay; and, length of stay is a recognised risk factor for LRTI. In particular, in the first stage we instrument for LRTI using variables that are correlated with the incidence of infection (LRTI) but are not themselves causally related to length of stay. Since LRTI is a dichotomous variable taking the value 1 or 0, this instrumenting first stage analysis takes the form of a probit regression and the result is an estimate of the (exogenously determined) probability of getting an infection which is not dependent on length of stay, this variable is called PROBI. In the second stage we then estimate the expected increase in total hospital costs from an increase in the probability of getting an infection. However, as we use constructed variables in the second stage of this procedure there is the danger that our estimated standard errors, and hence tests of hypotheses, will be biased. Thus we bootstrap the entire two-stage process to provide consistent, robust standard errors. The results give us an estimate of the expected cost savings from reducing the chance of a hospital infection by a given percentage. Sections 4.1-4.4 outline this estimation strategy in detail and present the results. 4.1 Risk of infection and hospital length of stay The proposed estimation strategy is designed to circumvent a serious problem that has plagued all previous attempts at measuring the cost of acquiring a hospital infection; namely that the primary source of additional cost of an infection is through increasing the length of stay in the hospital, but that the longer a patient remains in the hospital the greater are their chances of getting an infection. As discussed above, the literature supports this view as Celis et al. (1988), Glynn (1997) and Rebollo et al. (1996) all identify length of stay as an important risk factor for LRTI. To show that this is the case in our data, in Table 3 we present an analysis of the determinants of length of stay (STAY) and show that the incidence of infection (LRTI) is a highly statistically significant contributor. This is consistent with the literature and suggests our sample is by no means extraordinary in this regard; however, the regression suffers from endogeneity bias between LRT1 and STAY and thus the estimates are expected to be biased and inconsistent. In Table 4 we show that the contribution of LRTI to total hospital costs is through this effect on length of stay. In particular, table 4 presents two regressions of total cost (TOTCOST) on our set of controls and LRTI. In regression (i) we omit length of stay (STAY) and show, as expected, that LRTI is a positive and significant contributor to total hospital costs. In regression (ii), however, we additionally control for length of hospital stay. In this specification STAY is highly significant but LRTI is not. Thus we conclude that the mechanism through which hospital infections increase total cost is through its effect on increased length of hospital stay. As before, however due to the endogeneity between STAY and LRT1 we would expect the estimates of regression 4(i) to be biased.

Table 3: Dependent Variable: STAY Variable Coefficient t-stat. INTERCEP 10.12 1.41 DIED 7.18 4.73 * AGE -0.01 -0.1 SEX -0.01 -0.01 ELDERLY 2.09 1.44 ONECOM 2.19 1.77 TWOCOM 1.01 0.77 MORECOM 0.44 0.34 NG -0.18 -0.05 ET 7.59 1.64 OX 1.10 0.99 JOB1 -3.99 -1.5 JOB2 -1.67 -1.05 JOB3 -1.03 -0.7 JOB4 -3.32 -2.37 * JOB5 -0.05 -0.03 JOB6 -0.25 -0.15 AD2 1.78 0.61 AD3 0.96 0.45 AD4 8.67 2.64 * LRTI 7.38 3.05 * R-square 0.0966 Adj R-sq 0.0760 * Significant at the 5% level

Table 4: Dependent Variable: TOTCOST (i) Variable Coefficient INTERCEPT 3970.26 DIED 2769.76 AGE -32.41 SEX 167.76 ELDERLY -43.59 JOB1 -721.08 JOB2 -166.90 JOB3 -119.94 JOB4 -502.23 JOB5 -71.83 JOB6 -135.67 AD2 127.83 AD3 118.91 AD4 866.57 ONECOM 711.87 TWOCOM 555.30 MORECOM 422.86 STAY LRTI 1276.12 R-sq AR-sq (ii) Coefficient T-stat. 2215.93 2.56 1541.24 8.60* -28.14 -2.49* 171.14 1.38 -341.41 -1.95 -117.57 -0.37 84.33 0.44 47.09 0.27 -5.03 -0.03 -55.33 -0.30 -83.47 -0.41 -174.83 -0.49 -56.95 -0.22 -415.23 -1.04 370.32 2.47* 397.29 2.5* 331.22 2.10* 151.15 37.06* 82.13 0.28 R-sq 0.6634 AR-sq 0.6565

T-stat. 2.867 9.835* -1.792 0.845 -0.156 -1.403 -0.540 -0.422 -1.854 -0.244 -0.412 0.224 0.286 1.359 2.979* 2.186* 1.679 2.743* 0.1380 0.1213

* Significant at the 5% level

4.2 First stage instrumenting probit regression Having thus established that the possib le endogeneity between length of stay and infection risk discussed in the literature is in fact a problem in our dataset, we now turn to the instrumenting procedure designed to account for this bias. In the first stage we instrument for LRTI using variables indicating whether the patient required a nasogastric tube (NG), and whether or not the patient required oxygen therapy (OX). There are several reasons to believe that NG and OX are good instruments for LRTI. First, NG and LRTI are recognized in the literature to be correlated. Glynn (1997) found if a patient required a nasogastric (NG) tube for one day, the risk of LRTI increased 1.48 times, and for two days the risk increased 1.53 times. Rebollo et al. (1996) also found that the presence of NG tube was also found to increase the risk of any site of infection, in that case over eight- fold. There is also evidence to support the use of oxygen therapy (OX) as an instrument; Celis et al. (1988) found the use of respiratory devices, such as oxygen therapy, significantly increased the risk of LRTI. To be valid instruments NG and OX must not only be correlated with LRTI but must also be causally unrelated to length of stay, except via the mechanisms of increasing the chances of an infection. None of the literature surveyed identifies NG or OX as determinants of length of stay, and in Table 2 we have shown that NG and OX are

statistically insignificant in a regression of length of stay (STAY) on NG, OX, LRTI and the basic set of control variables discussed above 4 . The first stage regression thus takes the form:

LRTI = f ( AGE , NG , OX , u )

(1)

In Table 5 we present the results of our first stage probit regression. As the age of patient has also been found in the literature to be a major contributor to LRTI, this variable was also included as a control in the first stage regression, although it is not an identifying instrument. From the results in Table 5 and a calculation involving the functional form of the probit equation (see, for example, Green (1997), p.874) the estimated conditional probability of LRTI (PROBI) was calculated for each observation, and the cumulative distribution of this variable is presented in Table 6. Most individuals had a low risk of LRTI, with 95.44% of individuals demonstrating a risk of LRTI of between 0-10%. The remaining individuals had a risk of infection between 10% and 40%. No one had a risk of infection over 40%. The variation in these estimated probabilities across patients is interpreted as a source of variation (in the chances of getting an infection) exogenous to length of stay.

Table 5: First Stage Probit Regression, Dependent Variable = LRTI Variable Coefficient Chi-Square p-value Intercept* -6.21 15.56 <.0001 AGE* 0.05 7.36 0.01 NG* 0.88 4.62 0.03 OX* 0.50 5.14 0.02 DIED* 0.57 6.00 0.01 SEX -0.06 0.11 0.73 AD2 0.03 0.00 0.96 AD3 -0.28 0.35 0.56 AD4 -0.04 0.00 0.96 ONECOM 0.11 0.15 0.70 TWOCOM -0.10 0.10 0.75 MORECOM -0.52 2.42 0.12 ELDERLY 0.00 0.00 0.99 Number of Observations: 899 Log Likelihood: -113.0622467 * Significant at the 5% level

We also considered a variable on whether the use of endo-tracheal suction was required to help the patient breathe (ET) as an instrument for LRTI, but rejected this variable on the grounds that it was not sufficiently statistically insignificant in the regression in Table 2.

Table 6 - Frequency distribution of PROBI Estimated probability Count Cumulative of infection (PROBI) % Frequency 0-5 739 82.4 5-10 99 93.4 10-15 36 97.4 15-20 13 98.9 20-25 3 99.2 25-30 2 99.4 30-35 5 100 35-40 0 100

4.3 Second stage estimation and bootstrap of standard errors In the second stage regression we observe how total hospital costs vary according to variation in (exogenous) probabilities of infection in patients (PROBI). The interpretation of just what such an estimate means deserves additional consideration, as it is not immediately transparent. In particular, we seek to estimate the expected total savings in hospital costs that would result from some given decrease in the chances of infection in any given patient due to some investment in infection control activities. We think this is exactly the definition in which health service policy makers should be interested. For example, while ex-post costs for a given individual will vary by whether they do or do not get an infection, ex-ante expected costs for a patient or group of patients would depend on the probability of getting an infection. Whether a proposed measure is cost effective or not will depend on the cost of the improvement and on what the savings are for the associated decrease in the (ex-ante) probability of future patients getting an infection. This latter amount is exactly what our model estimates. Furthermore, our estimates are free from the heretofore-ubiquitous endogeniety bias between the probability of getting and infection and the length of hospital stay. After a series of robustness checks (see section 4.4 below) and checks for the appropriate functional form and outlyers, the final model takes the form:

Tot cos t = + i Control i + PROBI + u

(2)

As described, our procedure uses constructed variables in the second stage so we would expect the standard errors from this regression to be biased. Model (2) is thus estimated using a bootstrapping procedure (for a good review see Hall (1992)) in which the data were repeatedly re-sampled with replacement. For each bootstrapped sample, the first stage probit model (1) was estimated and the predicted values of the associated probabilities (calculated as described above) used in the estimation of the final model (2), which was estimated using a Huber-White covariance estimate to make the standard errors heteroscedasticity-consistent. After 1000 samples are subjected to this procedure, the bootstrapped distribution of the t-test is constructed as follows:

 * t* = *

where * , and * are the bootstrapped coefficient estimates and the

bootstrapped Huber-White standard errors, respectively, and is the least squares coefficient estimate from the original data set. Thus t* is the bootstrapped t-statistic. From the distribution of bootstrapped t*s the critical values for the top and bottom 2.5th percentiles, which correspond to the critical values for a two sided test at 5%, were calculated. The Huber-White standard errors and t-statistics were calculated from the data itself and compared with the critical values calculated from the bootstrapped tdistribution to determine statistical significance. In Table 7 we present the results of our second stage regression as well as the top and bottom critical t- values, for tests at 5% level of significance, calculated from the bootstrapping procedure (Tb1 and Tb2). Interpretation of the results is discussed fully in section 5 below. Table 7: Second stage regression with bootstrapped critical t-values, Dependent Variable = Total Cost Variable Coefficient t-stat Tb1 Tb2 INTERCEP 3523.34 2.81 -3.64 1.50 DIED 1419.14 2.96 -2.44 2.50 AGE -28.56 -1.58 -1.51 3.86 SEX 243.34 1.79 -2.11 1.87 ELDERLY 89.07 0.42 -2.16 1.77 AD2 18.22 0.05 -2.09 2.54 AD3 192.86 0.85 -2.25 2.00 AD4 1113.53 1.63 -3.56 1.74 TWOCOM 59.67 0.37 -2.51 1.79 THREECOM -100.87 -0.39 -1.74 2.42 MORECOM 191.00 0.67 -3.65 1.43 PROBI 6936.36 2.15 -5.33 1.83 R-square 0.2448 Adj R-sq 0.2328

4.4 Checks of Robustness and Model Selection The specifications and functional forms of the models presented in sections 4.1-4.3 were chosen after testing for alternatives, and this model selection process is discussed here. In particular, among all possible control variables available to us in the data set and suggested by theory, a comprehensive analysis of the extent of multicollinearity was undertaken5 . Our variables of interest remained statistically significant in every specification, but the analysis permitted a relatively more parsimonious set of control variables to be utilized and facilitated the interpretation of the results for readers interested in other variables.

A full discussion of these variables and the results of our correlational, functional form and specification analyses is available from the authors upon request.

In addition, the functional form of the model was explored by adding exponential and logarithmic terms, but no improvement to the adjusted R2 were achieved and so the model was assumed to be linear. Another concern regards the inclusion of the variable indicating whether a patient died or not (DIED) in the analysis. In particular, some may wonder whether or not DIED is itself endogenous in that costly hospital procedures may be more dangerous and increase the likelihood of death. However, we believe this to be unlikely and counter- intuitive. It is much more likely that the causality runs from DIED to total cost through the costliness of end of life procedures. Nevertheless the analysis was repeated dropping DIED from the list of control variables in model (2) but these are not presented here to save space but are available from the authors upon request. The main effect of omitting DIED was that the R2 was reduced and the magnitude of the coefficient on PROBI increased. This suggests that PROBI and DIED are correlated, either through PROBI increasing the chances of death, or through conditions close to the end of life (which are costly) increasing PROBI. The current analysis cannot discriminate between these two possibilities, but fa iling to control for DIED will mistakenly attribute some of the cost of end-of- life treatment to LRTI, overestimating its impact on total cost. By including DIED we generate the more conservative estimate of the cost of infection. Finally two outlyers in terms of total cost were identified that fall well over 8 standard deviations from the mean. Thus our final dataset contains 897 individuals. 5. Discussion of Results and a Simple Simulation Exercise The results of model (2) presented in Table 7 provide our estimate of the total hospital costs of LRTI. The model can explain almost one quarter of all the variation in hospital costs among our sample, and the probability of getting an infection is a highly statistically significant determinant of those costs. In particular, the model predicts that for every 10% decrease in the ex-ante probability of acquiring a LRTI, expected costs fall by 693 (about US$985). An overall risk reduction of 10 points would eliminate risk for most (95%) patients6 . Among the few higher-risk patients if the risk of LRTI were reduced from 40% to 20%, the expected total savings in hospital costs would be 1,386 ($1,970 ) per patient. If we compare these figures with the those of Plowman et al. (1999), the closest existing traditional estimates from the same hospital, we find that the ex-post estimates of the cost of an LRTI in that study are 3,050 (about US $4,500) for admissions to general medicine services and 1,765 (about US$2,650) for admissions to care of the elderly services. Thus our estimates of the potential savings from infection risk reduction are substantially larger than has been previously estimated in the literature. We think our estimates are superior since we control for endogeneity between length of stay and risk of infection, but never such a large difference with the established literature deserves further examination. In particular, the conventional wisdom has been that the endogeneity between STAY and LRTI should bias the traditional estimates upwards, not downwards. This is not necessarily the case, however. Unlike measurement error, endogeneity can include positive or negative correlation with disturbances or reverse causality. Correcting endogeneity can thus increase or decrease the effect, and predicting the sign of the impact is quite tricky without full understanding of the underlying data generating mechanisms.
6

Clearly, for most patients the cost of bringing risk to 0 will be much less.

To illustrate this point more concretely we undertake a simple simulation exercise in which we generate a simple data set which superfically has many of the same properties as our data set. We generate this data following the conventional wisdom about the underlying endogeneity between LRT1 and STAY; in other words we allow the likelihood of getting an LRTI to increase with length of stay and the length of stay to increase whenever a patient acquires an infection. In particular: Let: 1 , 2 , 3 , 4 ,~Uniform(0,1) and Then, NGi = 1 if 2 <.02; 0 otherwise OXi = 1 if

4 ~ Normal(12,4)

3 <.19; 0 otherwise BASEi = 0.03* 1 + 0.1 if OXi =1 BASEi = 0.03* 1 + 0.08 if NGi =1 PROBIi = BASEi + .05 if (30 * 1 ) > 22 PROBIi =BASEi + .1 if (30 * 1 ) > 27 LRTIi = 1 if 4 < PROBIi STAYi = 30 * 1 + 4 if LRTI = 1 STAYi = 30 * 1 if LRTI = 0
We then apply a highly simplified version of our estimation strategy to data generated in this fashion. First we estimate a nave or simple regression with endogeneity bias as: (3) STAYi = + 1 LRTIi In particular in the first stage we model the the probability of getting an infection as: (4) PROBIi = f ( NG, OX , u ) Then use this estimate in our second stage model of STAY: (5) STAYi = + 2 PROBIi Table 8 compars our estimates from regressions (3) and (5) using both our simulated (artificial) data and our actual data. Table 8: Comparison of Estimates from Simulated and Real Data

1
Simulated Data Actual Data (OLS) 18.30 (15.24) 26.30 (17.42)

2
(IV) 15.69 (2.07) 44.08 (5.35)

*t-statistics in parentheses

For the simulated data we get exactly the results that conventional wisdom would have speculated: the estimates from the IV estimation are lower than those from the nave, biased OLS regression. Thus when our assumptions about the underlying data generating process are correct, the IV estimation works as expected. However we get a very different picture when we use our actual data for the same simplified specifications. As can be read from Table 8, the IV estimate is quite a bit larger than the nave OLS estimator. Thus it is the properties in the underlying data that are driving our results rather than the estimation strategy. The primary conclusion that can be drawn at this point, then, is that the relationship between length of stay and risk of infection is likely much more complex than the simple feedback mechanism conventional wisdom may have assumed. Furthermore, given the underlying complexity of the endogeneity it is impossible, then, to guess in advance whether simple OLS estimates will produce estimates biased upwards or downwards. If our identifying assumptions are correct (namely that NG and OX are correlated with risk of infection but causally unrelated to length of stay) then our IV estimates should be much less biased and provide a more reliable estimate of the costs of infection than those in the existing literature. This in turn suggests that the costs of HAI may be quite a bit higher than has previously been reported.

References Cavalcante, M. D., Braga, O. B., Teofilo, C. H., Oliveira, E. N. and Alves, A. S. O. (1991) Cost improvements through the establishment of prudent infection control practices in a Brazilian general hospital, 1986-1989. Infection Control Hospital Epidemiology, 12, 649-53. Celis, R., Torres, A., Gatel, J. M., Almela, M., Rodriguez-Roisin, R. and Agusti- Vida, A. (1988) Nosocomial pneumonia. A multivariate analysis of risk and prognosis. Chest, 93, 318-24. Coello, R., Glenister, H., Fereres, J., Bartlett, C., Leigh, D., Sedgwick, J. and Cooke, E. (1993) The Cost of Infection in Surgical patients: A Case Control Study. J Hosp Infect, 25, 239-250. Cohen, D., R. (1984) Economic issues in infection control. J Hosp Infect, 5 (Supplement A), 17-25. Davies, T. and Cottingham, J. (1979) The cost of Hospital Infection in Orthopaedic Patients. Journal of Infection, 1, 330-338. Drummond, M. and Davies, L. F. (1991) Evaluation of the costs and benefits of reducing hospital infection. J Hosp Infect, 18 (Supplement 18), 85-93. Emmanual, E. J. (1996) Cost savings at the end of life: what do the data show? JAMA, 275, 1907-14. Freeman, J., Rosner, B. A. and McGowan, J. E. (1979) Adverse Effects of Nosocomial Infection. Journal of Infectious Diseases, 140,5, 732-740. Girard, R., Fabry, J., Meynet, R., Lambert, D. C. and Sepetjan, M. (1983) Costs of Nosocomial infection in a neonatal unit. J Hosp Infect, 4, 361-366. Glynn, A. W., V. Wilson, J. Charlett, A. Cookson, B. Taylor, L. Cole, N. (1997) Hospital Acquired Infection: Surveillance Policies and Practice, Public Health Laboratory Service, London. Gould, D. (1994) Nurses' hand decontamination practice: results of a local study. J Hosp Infect, 28, 15-30. Graham, M. (1992) Frequency and duration of hand washing in an intensive care unit. American Journal of Infection Control, 18, 77-81. Graves, N. (2001) Estimating the Economic Effects of Hospital Acquired Infection, University of London, PhD Thesis. Green, J. and Wenzel, R., P. (1977) Postoperative wound infection: A controlled study of the increased duration of hospital stay and direct costs of hospitalization. Annals of Surgery, 185,3, 264-268.

Green, William H. (1997) Econometric Analysis Third edition. Prentice-Hall International, New Jersey Haley, R. (1985) In Hospital Infections (Eds, JV., B. and Brachman, P.) Little , Brown & Co., Boston. Haley, R., Hooton, T., Culver, D., Stanley, R. and Emori, T. (1981a) Nosocomial infections in US hospitals 1975-76. Estimated frequency by selected characeristics of patients. Am J Med, 70, 947-959. Haley, R., W. (1986) Managing Hospital Infection Control for Cost-effectiveness: A Strategy for Reducing Infectious Complications, AHA, USA. Haley, R., W. (1991) Measuring the Costs of Nosocomial Infections: Methods for Estimating Economic Burden on the Hospital. American Journal of Medicine, 91 (Supplement 3B), 32s-38s. Haley, R., W. (1992) In Hospital Infections(Eds, Brachman, P., S. and Bennett, J., V.) Little, Brown, Boston, pp. 507-532. Haley, R., W., Schaberg, D., Crossley, K., Von Allmen, S. and McGowan, J. (1981b) Extra Charges and Prolongation of Stay Attributable to Nosocomial Infections: A Prospective Inter- hospital Comparison. Am Journal Med, 70, 51-58. Hojer, H. (1978) The effect on total anti- microbial consumption and hospitalization time after prophylactic treatment with doxycycline in colorectal surgery. Acta Chirurgica Scandinavica, 144, 175-9. Hughes, J. M. S. O. (1988) Study on the efficacy of nosocomial infection control (SENIC Project): results and implications for the future. Chemotheraphy, 34, 553-61. Hyryla, M., L,J. and Sintonen, H. (1994) The Use of Health Services in the management of Wound Infection. J Hosp Infect, 26, 1-14. Johnson, H. T. and Kaplan, R., S. (1987) In Relevance lost: The rise and fall of management accountingHarvard Business School Press, Boston, pp. 227-253. Johnson, J. R., Roberts, P. L. and Olsen, R. J. (1990) Prevention of catheter associated urinary tract infections with a silver coated oxide coated catheter: clinical and microbiolic correlates. Journal of Infectious Disease, 162, 1145-50. Kaiser, A. B., Roach, A. C. and Mulherin, L. L. (1983) The cost-effectiveness of antimicrobial prophylaxis in clean vascular surgery. Journal of Infectious Disease, 147, 1103. Kaplan, R. S. and Atkinson, A. A. (1989) Advanced Management Accounting., Prentice Hall, New Jersey. Kappstein, I., Schulgen, G., Beyer, U., Geiger, K., Schumacher, M. and Daschner, F. D. (1992) Prolongation of stay and extra cost due to ventilator-associated

pneumonia in an intensive care unit. European Journal of Clinical microbiology and Infectious Disease, 11(6), 504-508. Kulcher, F. and Golan, E. (1999) Assigning Values to Life: Comparing Methods for Valuing Health Risks, United States Department of Agriculture, Washington. Larson, E. (1988) A causal link between hand-washing and risk of infection? Examination of the evidence. Infection Control Hospital Epidemiology, 9, 2835. Liedberg, H. and Lundeberg, T. (1990a) Refinements in the coating of urethral catheters reduces the incidence of catheter-associated bacteriuria. An experimental and clinical study. European Urology, 17, 236-40. Liedberg, H. and Lundeberg, T. (1990b) Silver alloy coated catheters reduce catheter associated bacteriuria. British Journal of Urology, 65, 379-81. Lundeberg, T. (1986) Prevention of catheter associated urinary tract infection by use of silver impregnated catheters (letter). Lancet, 2, 1031. McGowan, J. E. (1982) Cost and benefit - a critical issue for hospital infection control. American Journal of Infection Control, 10, 100-108. McGrail, K., Green, B., Barer, M., Evans, R. G., Hertzman, C. and Normand, C. N. (2000) Age, costs of acute and long-term care and proximity to death: evidence for 1987-88 and 1994-95 in British Columbia. Age and Ageing, 29, 249-253. Mendleson, D. N. and Schwartz, W. B. (1993) Effects of ageing and population growth on health care costs. Health Affairs, 12, 119-25. National Audit Office (2000) The management and control of hospital acquired infection in acute NHS trusts in England : executive summary and recommendations., HMSO, London. Persson, U., Montgomery, F., Carlsson, A., Lindgren, B. and Ahnfelt, L. (1988) How far prophylaxis against infection in total joint replacement offset its cost? BMJ, 296, 99-102. Pittet, D., Tarara, D. and Wenzel, R., P. (1994) Nosocomial Bloodstream Infection in critically Ill Patients. Excess length of stay, extra costs and attributable mortality. Journal of the American Medical Association, 271, 1598-1601. Plowman, R. P., Graves, N., Griffin, M., Roberts, J. A., Swan, A. V., Cookson, B. C. and Taylor, L. (1999) The Socioecomic Burden of Hospital Acquired Infection, Public Health Laboratory Service, London. Plowman, R. P., Graves, N. and Roberts, J. A. (1997) Hospital Acquired Infection, Office of Health Economics, London. Poulson, K. B., Bremmelgaard, A., Sorensen, A. I., Raahave, D. and Petersen, J. V. (1994) Estimated Costs of Postoperative Wound Infections. A case-control

study of marginal hospital costs and social security costs. Epidemiology of Infection, 113, 282-295. Rebollo, M. H., Bernal, J. M., Llorca, J., Rabasa, J. M. and Revuelta, J. M. (1996) Nosocomial infections in patients having cardiovascular operations: a multivariate analysis of risk factors. Journal of Thoracic Cardiovascular Surgery, 112, 908-13. Reybrouck, K. G. (1983) The role of hands in the spread of nosocomial infections. J Hosp Infect, 4, 103-11. Riley, D. K., Classen, D. C., Stevens, L. E. and Burke, J. P. (1995) A large randomised clinical trial of a silver impregnated urinary catheter: lack of efficacy and staphylococcal superinfection. American Journal of Medicine, 98, 349-56. Rubinstein, E., Green, M., Modan, M., Amit, P., Bernstein, L. and Rubenstein, A. (1982) The Effect of Nosocomial Infections on the Length and Costs of Hospital Stay. Journal of Anti-microbial Chemotherapy, 9 (Supplement A), 93100. Schaeffer, A. J., Story, K. O. and Johnson, S. M. (1988) Effect of silver oxide chloroisocyanuric acid antimicrobial urinary drainage system on catheterassociated bacteriuria. Journal of Urology, 139, 69-73. Scheckler, W. E. (1978) Septicaemia and nosocomial infections in a community hospital. Annals of Internal medicine., 89, 754-756. Scitovsky, A. A. (1984) The high cost of dying: what do the data show? Millbank Memorial Fund Q Health Soc, 62, 591-608. Scitovsky, A. A. (1994) The high cost of dying revisited. Millbank Q, 72, 561-91. Shapiro, M., Schoenbaum, S., C., Tager, I., B., Munoz, A. and Polk, B., F. (1983) Benefit-Cost Analysis of Antimicrobial Prophylaxis in Abdominal and Vaginal Hysterectomy. Journal of the American Medical Association, 249;10, 12901294. Takeuchi, H., Hida, S., Yoshida, O. and Ueda, T. (1993) Clinical study on efficacy of a Foley catheter coated with silver protein in prevention of urinary tract infections. Hinyokika Kiyo, 39, 293-8. Thornton, G. F. and Andriole, V. T. (1970) Bacteria during in-dwelling catheter drainage II: effect of a closed drainage system. Journal of the American Medical Association, 214, 339-42. Wakefield, D. S., Pfaller, M. A., Hammons, G. T. and Massanari, R. M. (1987) Use of the Appropriateness Evaluation Protocol for Estimating the Incremental costs Associated with Nosocomial Infections. Medical care, 25,6, 481-488.