Evaluating organizational change in health care: the patient-centered hospital model

MDC codes are internationally recognized thanks to their adoption in the United States medical care reimbursement system. They are formed mapping all the DRG codes into 25 mutually exclusive diagnosis areas.

We estimate log-linear models of the outcome means considering that the outcomes that we use are strictly non-negative (e.g. means of count variables or rates), not over-dispersed and do not raise zero inflation concerns ([47], p. 645)

The coefficient of interest, γ, refers to a dummy variable, PC_h,t, that is equal to one for those hospitals that adopted a PC model in the years immediately after their organizational change and zero otherwise. This is clearly equivalent to including a standard interaction term between the treatment variable and a post-reform dummy. Also notice that there is no need to include a treatment dummy, as we have the full set of hospital fixed effects, or a post-reform dummy variable, as we have the full set of year fixed effects

Data are provided by the Health Care Department of the Lombardy Region and are processed in collaboration with CRISP - the Inter-university Research Centre on Public Services at the University of Milan-Bicocca (Italy). Individual HDC records are not publicly available under the Italian privacy law. The Health Care Department of the Lombardy Region must be contacted to discuss the provision of the data

The diagnosis-related group (DRG) code is a standard classification ([48]) adopted in the Lombardy Region of Italy since 1995. The DRG classifies hospital discharge charts depending on patients’ diagnoses, procedures, complications, co-morbidity and demographic factors (such as age and gender)

In fact, HDC data trace the department that is in charge of each patient and record the total number of departmental transfers of each HDC, but not whether a transfer is in fact a bed change within the same hospital or, more simply, a change of the administratively responsible department.

An important efficiency measure that we do not observe is the cost of single HDCs as we have no information on the composition and cost of the physical and human resources used. In fact, we are provided with the cost of reimbursement by the Lombardy Health Care System to hospitals for each HDC, but this variable is unsuitable for use as a cost measure as it is affected by DRG up-coding practices, discretionality of the regional policy makers in deciding the price of the duration and the DRG of each HDC, allowing for strategic behaviour of hospital managers. For an extensive analysis of the reimbursement mechanism adopted in the Lombardy Health Care System, see [49]

The attractiveness of the Lombardy Health Care System is indeed relevant, with a proportion of hospitalized patients from other regions close to 10% ([49]) of the yearly provision. The main reason for dropping the HDCs of patients with residence outside Lombardy is because they might be occasional users of the Lombardy Health Care System and we lack relevant information about them regarding their possible re-hospitalization and death. For instance, as we know the date of death of Lombardy residents only, including non-Lombardy patients would bias the average mortality rate of patients downward by an unpredictable amount. We also dropped one-day-long and subacute HDCs due to comparability issues.

A similar approach was used by [50]

Some robustness checks assessing the relevance of this selection rule are provided in Tables 7 and 8.

We developed this test (results in Table 4) for all the models that we estimated in Table 5 (columns 1 to 4), starting from the basic equation (Eq. 1) to the saturated equation (Eq. 2), as follows. First, we computed each outcome variable of interest after partialling out the contribution of all the independent variables except for PC_h,t. Hence, we regressed each of them on a fourth-degree polynomial time trend, allowing all the coefficients to differ between the PC and the traditionally organized hospitals (unrestricted model), and we regressed the same dependent variable on a fourth-degree polynomial time trend in which only the intercept is allowed to differ between the two groups considered. Finally, we computed the statistic \(\left (\left (\left (R^{2}_{UR}-R^{2}_{R}\right)/r\right)/\left (1-R^{2}_{UR}\right)\right)\), which is distributed as an F-distribution with (r,n−k) degrees of freedom and in which \(R^{2}_{R}\) and \(R^{2}_{UR}\) are respectively the R² of the restricted and unrestricted models, r is the number of restrictions imposed and n−k is the number of degrees of freedom of the unrestricted model.