2.2. Statistical modelling
We used multi-variate regression models to assess the association of outsourcing with MRSA incidence rates, as follows:
MRSAit = α + βOutsourcei + γTrustit + μi + nt + εit (1)
Here i is Trust and t is year. MRSA is the MRSA incidence rate per 100,000 hospital beds; Outsource is a dummy for whether the Trust outsourced cleaning services or retained them in-house; Trust is a series of variables controlling for Trust differences, including the number of beds in the Trusts and the average length of stay in the Trust; μ adjusts for four regional dummies (North, South, East, and West), and n is a set of year dummies to control for geo-spatial correlation, such as periods of MRSA outbreaks. ε is the error term.
To further adjust for potential confounding and facilitate comparability across Trusts, in a subsequent step we matched hospitals within geographic regions on dimensions of size (measured by number of hospital beds), complexity (measured as numbers of specialist and multiservice sites hospital within each Trust i) and case mix using propensity score matching (Rosenbaum and Rubin, 1983). Importantly, we match the two dimensions separately with respect to complexity, to take account of the possibility that differences in the number of specialist and multiservice sites might confound the results. Our ability to adjust for patient case mix is constrained by the absence of any severity measure based on diagnostic codes or something similar that predicts hospital acquired infection (as opposed to, for example and with caveats, the well-established case mix predictors of mortality). Propensity Score matching reduces potential confounding by comparing hospitals operating in similar regions, with matching size and complexity, but differing their management’s choice of cleaning operation. It is used in policy evaluation because it reduces confounding compared with simple OLS models (Imbens, 2004). At this stage the 126 Trusts that had data on both MRSA rates in at least one year and sufficient information on complexity to enable matching were analysed. As a further robustness check we also implement coarsened exact matching (Iacus et al., 2011), which further address potential sources of residual confounding. The comparative advantage of coarsened exact matching vis-a-vis propensity score matching is that it ensures multivariate balancing between treated and control group.
All data and models were estimated using Stata version 13. All t-tests were two-tailed assuming unequal variances. Standard errors were bootstrapped and clustered by Trust to account for non-independence of sampling (Abadie and Imbens, 2009).
3.1. Unadjusted comparison of outsource and in-house cleaning provision
Fig. 1 compares the pattern of MRSA incidence per 100,000 hospital bed-days in outsourced and in-house hospitals in 2010. The mean MRSA incidence in outsourced hospitals is 2.28 per 100,000 bed-days, almost 50% greater than the observed mean of 1.46 per 100,000 bed days in those that retained in-house cleaning (Stone et al.). Indeed, as shown in Fig. 3 in the web appendix, the entire MRSA risk distribution is greater in outsourced hospitals, which reflect the high levels of MRSA risk.
MRSA Incidence Rate by type of cleaning service in 2010.
Next, we evaluated patient perceptions of cleanliness of bedrooms and bathrooms (web appendix Fig. 4a and b). Fewer patients in Trusts with outsourced services (57.6%) compared to in-house services (59.7%) described the cleanliness of the bedrooms as ‘excellent’ (t-test: 2.55, p = 0.01). We also observe a similar pattern for bathroom cleanliness (67.0% for outsourced hospitals compared with 68.5% for in-house hospitals; t-test = 2.04, p = 0.04).
In web appendix Fig. 5 we present the distribution of the percentage of staff who report access to hand-washing material across Trusts. 63.0% of staff who work in Trusts with outsourced cleaning services report that hand-washing materials are always available compared with 68.0% in Trusts with in-house cleaning (t-test: 3.47 p=
3.2. Adjusted association of outsourcing with MRSA incidence rates
Table 1 shows the results of our statistical models, which can be interpreted as the average variation in MRSA incidence rate between Trusts which outsourced their cleaning services and those which retained their cleaning services in house. (In web appendix table 4, we also present the results using log-outcomes). Using simple OLS models we estimate that Trusts which outsourced their cleaning services tend to report on average 0.42 more cases of MRSA bacteraemia per 100,000 bed-days (95% CI: 0.24 to 0.61, p-value = 0.001). To translate this number into the original framework, we estimate the level of MRSA infection in two scenarios when cleaning services for the Trust i are outsourced vis-à-vis when they are provided in house. Accordingly, while outsourced Trusts will report an average rate of MRSA bacteraemia of to 1.44 cases per 100,000 bed days, their counterpart with in-house cleaning will report an average MRSA bacteraemia rate of 1.02.