Hospital costs associated with psychiatric comorbidities: a retrospective study

The study site was a German tertiary-care university hospital that provides services for 70,000 inpatient hospital episodes and 800,000 outpatients per year. The dataset included all 338,162 inpatient episodes consecutively discharged between 2011 and 2014 at the study site. We excluded 47,335 episodes because they were younger than 18 years old and 8744 episodes because they had a psychiatric comorbidity possibly caused by the hospital stay. We excluded 37,504 episodes, because they had a main diagnosis in which no episode allowed statistical comparison in case-control analyses. Moreover, 84,025 episodes were treated in the psychiatric department of the university hospital and we excluded them because they were not reimbursed with DRGs. Furthermore, we excluded 1852 episodes because they had missing or implausible data (0.5%).

Implementation of the study required (1) data describing the characteristics of hospital episodes and (2) episode-specific resource use data. We obtained patient characteristics from the data-warehouse of the hospital. These data included all main and secondary diagnoses, age, sex, length of stay, intensive care hours and discharge status. We defined the presence of psychiatric comorbidities as the presence of at least one code from Chapter V (F0-F9) of the International Statistical Classification Of Diseases And Related Health Problems, 10th revision, German Modification (ICD-10-GM) as secondary diagnoses.

We obtained detailed resource use data from the department of management and used them to calculate episode-specific costs according to the general guidelines for unit costing for the German Institute for Hospital Reimbursement [42]. We valued all services a patient received during a stay by their full operating costs to the hospital, i.e. including variable and fixed cost shares. For instance, this included a daily documentation of the intensity of nursing care for each patient and a detailed documentation of all expensive pharmaceuticals and medical devices, which was used to calculate differences in per diem cost. We excluded capital costs, because they are not reimbursed by DRG in the German health care system. We adjusted all costs and reimbursements to € of 2014. Therefore, we used the annually reported average increases in health care costs in German as published by the German Federal Statistical Office [43].

The analyses of costs associated with psychiatric comorbidities were subject to several methodological challenges. First, hospital costs were highly clustered within diagnostic groups. Therefore, we controlled for main-diagnoses, i.e. the retrospectively coded reason for the admission, to prevent comparing patients from entirely different diagnostic and therapeutic regimes.

Second, there were no time-stamps attached to the comorbidity data. This complicated the analyses of costs associated with psychiatric comorbidities, as some conditions might have resulted from resource-intensive diagnostic and therapeutic measures during the hospital episode, thereby blurring the attributability of costs. Therefore, we excluded episodes with psychiatric comorbidities that might have occurred during hospital stays (Table 1).

Third, increased length of stay itself might have been associated with increased probability of identification and documentation of preexisting psychiatric comorbidities. This sort of time-related bias may inflate estimates if patients with a longer hospital stay and therefore higher costs are then more likely to have a documented psychiatric comorbidity. As time-stamps for secondary diagnosis were unavailable, directly accounting for the time-dependency was not feasible [44]. However, to get a lower bound of our incremental cost estimates in case time-dependency was an issue, we additionally ran our regression adjusting for length of stay in a sensitivity analysis.

We defined cases as episodes that had a non-psychiatric main diagnosis and at least one documented psychiatric secondary diagnosis other than those excluded. Eligible controls were patients that had no documented psychiatric secondary diagnosis. We applied stratification by 4-digit ICD10 main-diagnosis code using fixed effects to ensure the comparability of different diagnostic and therapeutic regimes. We excluded cases with a main diagnosis from a group in which no controls were available. Vice versa, we excluded potential controls with a main diagnosis from a group without cases. We additionally controlled for age, sex, intensive care hours and other somatic comorbidities using the Charlson Comorbidity Index (CCI) [45].

We analyzed two primary outcomes: (1) costs per episode and (2) reimbursement per episode. Furthermore, we used costs and reimbursement adjusted for length of stay in a sensitivity analysis. As hospital cost and reimbursement were non-normally distributed and right-skewed we used a generalized-linear model [46] with a log link and a Gamma distribution, which was determined using a modified park test [47].

To the best of our knowledge, our work is the first to discuss both abovementioned issues of possible bias caused by time-dependency in the context of psychiatric comorbidities retrospectively obtained from routine diagnosis data. By excluding cases with secondary psychiatric diagnoses that could potentially occur as a complication rather than as an independent comorbidity, we avoided blurring the attributability of cost differences. Leaving these cases in the regression model resulted in substantially larger estimates (data not shown).

We also discussed the possibility that psychiatric comorbidity detection might be a result rather than a cause of length of stay and showed results after controlling for length of stay. These figures should be regarded as a lower bound of the real differences, since psychiatric comorbidities are likely to be associated with increased length of stay. The calculation of differences in per diem costs was possible because we obtained comparably detailed resource use data, such as daily acuity scores and consumption of expensive pharmaceutical and other materials.

Furthermore, the retrospective design of the study allowed the inclusion of about 160 thousand episodes that were treated during four consecutive years in statistical analyses. The high number of observations and the relatively long study period should improve the confidence in the robustness of the findings to cross-sectional and longitudinal variations in clinical practice. However, relatively little high quality work exists the validity of administrative diagnosis data for research purposes [48], and studies have found both poor [49] and good validity [50]. Furthermore, research has found substantial differences in proclivity of health care providers for making and documenting diagnoses in routine clinical care [51]. It was not possible to investigate the validity of the analysed diagnosis data. However, the comprehensive employment of specialised staff for documentation and verification of diagnosis coding for reimbursement purposes at the study site should strengthen the confidence in the presented results. Nonetheless, the detection of psychiatric comorbidities was probably less consistent than could have been achieved in a prospective study.

The results of our study can be compared to previous research. Creed et al. [25] prospectively studied the association of depression and anxiety with hospital costs in 263 medical inpatients. They found a statistically non-significant increase of 19% in costs per episode after controlling for gender, social class, social benefit, and severity of physical illness. This cost difference lies between the results presented in Table 3 and the lower bound presented in our sensitivity analyses. However, the study of Creed et al. [25] only investigated medical inpatients. Furthermore, their study was not exclusively dedicated to hospital costs, but included other services, such as primary care und community costs. Hence, they applied fixed unit costs to hospital services, a method which should be less able to detect cost difference.

Egede et al. [27] retrospectively studied the association between comorbid depression and hospital costs in about 21,500 individuals with diabetes. They found that the presence of comorbid disorders was associated with a statistically non-significant increase in inpatient costs per case of 31%. This figure is relatively close to the results presented in Table 3 and substantially higher than our lower bound estimation provided in the sensitivity analyses. However, Egede et al. [27] used hospital charges instead of costs, which entails methodological caveats [52]. Furthermore, Egede et al. [27] did not account for the potential time-related-bias that the likelihood of detecting depression is likely to increase with increased length of stay.

Hochlehnert et al. [29] retrospectively studied the association of psychiatric comorbidity and hospital costs and reimbursement in 940 cardiovascular inpatients. They found that psychiatric comorbidities were associated with a 49% increase in hospital costs per case after adjusting for age and sex. Furthermore, they found a reimbursement-cost difference in patients with psychiatric comorbidities that was 901 € less than the reimbursement-cost difference in controls. Both of these values were higher than the results presented in Table 3. However, Hochlehnert et al. [29] did not account for the potential time-related biases associated with (1) the potential occurrence of psychiatric secondary diagnoses as results of resource intensive treatments, such as delirium, and (2) the higher likelihood of detecting psychiatric comorbidities in patients that stay longer. These biases could have led to an overestimation of the effect on hospital costs.

This study has found that the presence of psychiatric comorbidities in somatic inpatients is associated with higher costs per episode and per day. This cost difference is not sufficiently accounted for in the German DRG-system. If differences in hospital costs between patient groups are not adequately controlled for in DRG-systems, i.e. the payment is either too high or too low for many patients, strong incentives are created to game the system [11]. Therefore, we suggest intensifying the inclusion of psychiatric comorbidities in the German DRG system. Future research should investigate the appropriate inclusion of psychiatric comorbidities in other health care systems’ payment schemes. In analysis of the cost of somatic diseases in general, psychiatric comorbidities should be controlled for as a relevant cost-increasing factor. Methodologically, we would like to stress the importance of taking into account the two sources of time-related bias discussed in this paper when conducting retrospective analyses.