It was argued that real word data are more informative and useful for decision makers than results from controlled trial settings when evaluating medical technologies. This is particularly relevant in situations in which clinical practice may significantly differ from clinical guidelines owing to the presence of barriers to the full implementation of technologies in clinical practice . This effect has been observed in the introduction of ICDs for the prevention of sudden cardiac death. To our knowledge, this is the first empirical study that has attempted to determine the impact of ICDs using an administrative data set from a real world setting in Europe. Administrative databases are recognized as valid sources of data for identifying the outcomes of rare events and for assessing the economic impact of various interventions [11, 18]. We believe that this analysis contributes to the existing literature in two ways. First, it provides additional evidence regarding the impact of ICDs and estimates the magnitude of this impact across four different dimensions. Second, and even more importantly, it represents a novel study design in which an exact matching method was used.
In this study, ICD treatment was associated with significantly lower mortality, slightly higher re-hospitalization rate and significantly higher regional expenditure. In our sample, mortality at 1 year was reduced by between 9% and 10%, with a relative risk reduction of 0.59. The hazard ratios for the two subsamples were 0.80 and 0.85, respectively. These findings are in line with those reported from RCTs and collected in recent reviews, though direct comparisons should be made with caution, considering the different study designs . A more significant difference is observed in comparison with the results obtained from a meta-analysis of other observational studies in which ICDs reduced all-cause mortality by 46% (CI, 32% to 57%) . This greater difference can be partially explained by the high heterogeneity of the studies selected for this review. Indeed, among 11 observational studies with a contemporaneous control group, only two can be compared with our present study in terms of sample size and the method used to determine the benefits of ICDs [19, 20]. In line with our findings, the relative risk reduction in all-cause mortality in these two studies was 0.67 (CI 0.63–072 ) and 0.71 (CI 0.51–0.97), respectively.
Regarding re-hospitalization rate, the difference between the two groups was positive across all samples. This result could have been expected, because ICD patients are more likely to return hospital for monitoring of their device, though such visits do not necessarily become hospitalization events.
The study of Groeneveld et al.  analyzed Medicare patients given an ICD prophylactically and found increases in both survival and expenditure compared with propensity score-matched elderly patients who did not receive an ICD. This analysis differs from ours in terms of the method, the patients’ age (mean 76 years) and the context. The purchase price of an ICD is higher in the United States than in Europe, and specifically in Italy [20, 21], and therefore American data cannot be directly related to the setting of our health care system. However, it is noteworthy that these two studies led to similar conclusions with regard to ICD costs and benefits. More recently, another report derived from Medicare data showed how geographical areas in which the prophylactic use of ICDs increased over time showed greater improvements in survival, stressing the need for programs designed to increase the evidence-based use of ICDs .
In our study, regional expenditure was significantly higher for ICD patients both at the index hospitalization and during follow up. Both differences were driven mainly by the fixed difference between the two predominant DRG tariffs in the two groups. These estimates provide an insight into the economic impact of ICDs. In fact, given that cost assessment represents essential part for defining value of DRG tariff, the latter is frequently used as the “proxy” of hospital costs of specific patient group. However, it has been argued that costing is not the only ingredient in determining DRG tariffs and that DRG tariffs may not fully reflect the actual resource consumption associated with patient’s management . To estimate actual cost per patient, we would need further data on the resources used and their unit costs, which were not available in the present study.
The finding of similar LOSs and costs during follow-up in patients who did and did not receive an ICD is of interest, because in ICD patients prevention of sudden cardiac death by termination of ventricular tachyarrhythmia has been found to be associated with a subsequent increased risk of heart failure, with a potential risk of increased hospitalization for heart failure . In RCTs, the proportion of non-sudden deaths (including heart failure) showed a relative increase in patients given a prophylactic ICD compared with controls, though the absolute number of heart failure deaths was not increased .
A criticism of our study is related to the interpretation of the results in terms of the causal relationship between ICDs and mortality. Clinicians treat patients with ICDs according to unobservable factors, some of which might be correlated with pre-treatment mortality. Thus, because we do not know everything about clinicians’ decisions, our treatment and control groups might differ in some unobservable characteristic. This possibility cannot be excluded and is an important issue when interpreting the evidence. However, two features of our study must be highlighted in this regard.
First, though selection based on unobservables is an intrinsic problem of non-experimental evidence, the relevance of our results does not depend on the correct identification of a causality relationship (which can be achieved only through proper clinical trial designs). From a health policy perspective, our results should be interpreted as important evidence that the real world use of a new medical technology, the ICD, is in line with the clinical evidence and that the health system is not significantly distorting the application of this effective but expensive preventive technique. It is noteworthy that a systematic review and meta-analysis published by Ezekowitz et al. found that mortality over time was similar between ICD patients enrolled in RCTs and those in observational studies (both prospective and retrospective) .
Second, if the external validity of an RCT must be verified, observational data represent the best description of the physicians’ behavior and patients’ outcomes. A certain degree of bias, excluded by the design of RCTs, is thus unavoidable in this context. Alternatively, some might prefer to use a more structured approach in which the treatment is assumed to be assigned according to pre-specified criteria, generally coming from a theoretical model (e.g. see Heckman ). However, these methodologies were primarily developed for cases in which the analyst has substantial information about each individual (e.g. obtained from a survey) and can use extra matching variables to predict selection for treatment. Here, we know only what we could observe from administrative datasets and used all the available information for matching. Hence, we see no further advantage in introducing heavy discretionary assumptions about the selection process in this context.