The cost of resistance: incremental cost of methicillin-resistant Staphylococcus aureus (MRSA) in German hospitals

Our data set consists of hospital stays, not patients. In the G-DRG system, there are elaborate rules that result in two cases being merged into one if they occur within a short timeframe and are either likely to be caused by the same underlying condition or if the second is likely to be caused by complications incurred during the first stay. Cases that were merged according to these rules were excluded from the analysis to avoid contamination with this problem and because no 100% clear-cut length of stay can be determined. Thus, the remaining cases can be regarded as independent for the purpose of our analyses.

This figure varied between 4.1 and 92.3% across the surveyed hospitals.

The maximum proportion of U80.0!-codes without a positive laboratory result was 0.3%.

This figure is slightly lower than the 0.56% that can be computed from data supplied by the National Reference Centre for the Surveillance of Nosocomial Infections based on reports from 65 hospitals [17].

The percentage of bias reduction should be as equal as possible over the covariates, as otherwise bias may be increased for some function of these covariates, even if univariate bias has been reduced for each of them. This property is known as “equal percentage bias reducing” or EPBR [22].

Here, we are at odds with Ho et al. [14], who claim that the choice of the matching algorithm is merely a mechanical decision of picking the one with the lowest bias (i.e. the best balance). This would require knowledge of which variables best represent the universe of the relevant pretreatment covariates—known to be measured or not. Also, there is more than one criterion for assessing the quality of the matching, and these do not always point in the same direction. The outcome, it is claimed, has no part to play in the decision as this could lead to “stacking the deck”. While this danger is real, in our view the decision for one matching algorithm will always involve some judgment—even with regard to the plausibility of the estimated effects. Here, we present the results of different datasets to make the influence of our decision on the reported results as transparent as possible.

We have chosen to carry out this subgroup analysis because MV is generously reimbursed in the G-DRG system. As reimbursement is based on average cost across the German hospital system, we decided that this subgroup merits closer analysis. MV is potentially associated with higher costs in a number of ways:

(1) MV increases length of stay.

(2) MV is an indicator of a more severe course of disease and is thus potentially associated with increased LOS and higher treatment costs.

(3) MV is a potential cause of nosocomial infection and its associated costs.

(4) According to G-DRG cost accounting regulations, some cost (nurses and medical technology) are apportioned to patients with mechanical ventilation using a much higher weighting factor for hours (e.g. 1.71) as compared to treatment (e.g. 1) or monitoring hours (e.g. 0.57) [6].

Unfortunately, in our retrospective study, we can neither determine the relative importance of these influences nor the direction of causation.

Subgroups were analysed only in matching 1 for reasons of sample size. Also, we deem these results more accurate and thus present results from the second matching only to demonstrate the robustness of our overall results.