Nature of cost variability

Patient-level cost drivers may be classified as either predictable or non-predictable, based on whether relevant information is available prior to the hospital admission. We propose that predictable cost drivers contain information independent of the care provided in hospital and thus explain warranted variability of patient-level cost of care in hospital. In this scenario, the hospital has no way of easily influencing these factors, with the associated cost variability being ‘pre-determined’ and therefore warranted from the hospital’s perspective. For example, a patient’s prior health status (i.e., co-morbidities), socioeconomic status as well as demographic factors may justifiably lead to varying consumption of health resources while in hospital. As a further example, the widely used DRG (Diagnosis-Related Group) system, which classifies hospital admissions into similar groups, is a variable that links the clinical characteristics of a patient with the expected resource requirements; it can also be viewed as a cost driver of warranted cost variability. Each DRG (e.g., “circulatory disorders without acute myocardial infarction, with invasive cardiac procedure(s), without catastrophic or severe complications”) describes the patient’s condition as well as its severity [1], features that are largely known prior to hospital admission, despite the DRG classification being typically “coded” following completion of the actual episode of care.

Cost drivers that are not known or predictable prior to admission do not relate to the underlying unique patient health predicaments. Such factors may however relate to aspects of patient care provided during hospital admission and may thus contribute to unwarranted (i.e., potentially preventable) variability of cost of care. Large cost variability has been observed between hospitals [2, 3] and it is the unwarranted portion of this variability which may result in no clinical benefit to patients [4]. In fact, large amounts of cost variability with no correlation to the quality of care have been observed previously [5].

The concept of unwarranted nature of hospital cost variability, or equivalently preventability thereof is not yet consistently defined. Useful parallels can be drawn with preventability of patient harm in health care, because harm is known to be strongly correlated with increased cost [6, 7]. Like cost preventability, preventability of patient harm is also inconsistently defined [8]. This suggests a conservative inclination to declaring a particular cost driver as predictable a-priori and thus leading to warranted (or non-preventable) cost variability. Such an approach avoids under-estimating the potential residual unwarranted cost variability, as adhered to in this study.

Analysis of cost variability

Cost variability in health care—both warranted and not—has been typically modeled with patient-related predictors [9–12] using regression-based risk adjustment methods. There are however two recognised shortcomings of such an approach, which may be why there is no single model yet which is considered as optimal [13, 14]. Firstly, these methods tend to omit some cost drivers that are non-predictable from information available prior to commencement of care. This narrow focus may stem from the imprecise nature of the link between these cost drivers and unwarranted cost variability. Examples of such often ignored cost drivers are: the choice of the admitting and/or discharging unit, the admitting and/or discharging doctor, as well as the timing of the initial clinical encounter as a surrogate for inconsistent (or otherwise) matching of clinical resources to demand, e.g., in after-hours periods. These additional predictors, potentially related to unwarranted cost variability, ought to be included in any model designed to analyse patient-level costs. Secondly, fundamental technical shortcomings of the currently prevailing methods relate to “possible omitted variable biases associated with correlations between unobservable individual specific effects and included covariates” [10]. This is particularly pertinent for patient data which may have many strongly correlated variables that drive costs (e.g., age and co-morbidity).

The first issue is remedied explicitly by the design of this study (see Table 1). A large number of non-predictable cost predictors (not known prior to hospital admission) were included in the model. Clinical judgment was exercised when assigning these variables into either the a-priori predictable or the a-priori non-predictable category, with a general caution to not declare any cost predictor as predictable without sufficient confidence. For example, variables such as day of the week, month, as well as the hour of the day were classified as non-predictable as patients do not plan on presenting at any particular time.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. All potential cost predictors analysed. (See S1 Text).

https://doi.org/10.1371/journal.pone.0204300.t001

While the amount of information is greatly enriched by this inclusion of additional variables, the need to handle correlations between them is potentiated. This is because the possibility of co-linearities increases with the inclusion of additional predictors. This study uses the technique of sparse group LASSO (Least Absolute Shrinkage and Selection Operator) to overcome this challenge. A lasso model is one that fits a linear regression via a penalised maximum likelihood [15]. It solves an l₁ optimisation problem to select out unimportant coefficients but to keep the necessary ones in the model. This results in an interpretable model that includes only the pertinent, or principal cost drivers. Moreover, an l₁ method seems appropriate, given its solution exhibits robustness-like properties [16] born out of the fact that it effectively restricts the number of coefficients in the solution, therefore diminishing the effect of outliers. Such resultant relative insensitivity to outliers is beneficial in the case of the skewed hospital cost data [17]. Naturally, there is a balance between including too few and too many predictors. Leaving an insufficient number of predictors in the model may lead to inadequate explanation of the observed cost variability, in turn resulting in a loss of accuracy. On the other hand, too many predictors will also increase error through inclusion of parameters with little additional information. It is therefore convenient that the number of residual predictors left in the lasso model is adjustable and can be set to minimise the resultant error.

The sparse group lasso optimisation is a variation of the more generic lasso technique that accounts for the fact that some variables need to be grouped [18, 19]. This is particularly important when dealing with sparse categorical variables, such as the admitting doctor. The ability to group such variables is important to ensure practical relevance of outcomes, so that each categorical variable can be either discounted or selected as an entire group. The sparse group lasso includes an additional penalty factor in the optimisation in addition to the lasso penalty that encourages such grouping [20, 21]. It is a flexible algorithm that does so with a varying force, as determined by the balance between the two penalty factors (see Eq 1). Grouping more decisively results in selecting out more variables [18] and thus might be expected to decrease the error, but only to a point beyond which the error may increase again due to insufficient number of factors left in the model. Again, the optimal balance is able to be computed based on overall error minimisation considerations.

Regression Penalty in Sparse Group Lasso: (1)

λ  tuning parameter

α  grouping parameter

m  number of coefficient groups

n  total number of coefficients

β^(j)  the j-th coefficient group

β_i  the i-th coefficient

Data pre-processing

Given the large number of data dimensions in our research question, we explored using Principal Component Analysis (PCA) as an initial dimension reducing step. An excessive number of dimensions would make the computing requirements impractical. The use of this technique to pre-process large data is well recognised [22] and it has been specifically used in grouping of binary variables [23]. Hospital-based data is frequently large and can exhibit multi-dimensionality. This leads to high potential error in relation to both the way in which it is measured and recorded. Therefore the PCA is a good choice to pre-process hospital-generated data, as it extracts the most pertinent aspects of the information, leaving out the residual noise. The output of the PCA then represents suitable input into lasso-based analysis (See S2 Text).

The sparse group lasso approach provides values of regression coefficients of the predictors left in the model. Comparison of their absolute values provides estimation of the predictors’ relative importance in terms of explaining patient-level cost variability. In the case of grouped variables, the PCA conveniently converts any group of related variables (e.g., corresponding to each DRG, or each day of the week) into a group with an orthogonal set of elements, i.e., the principal components, which are then analysed by the sparse group lasso method. The orthogonality justifies summation of the absolute values of regression coefficients of all members of each group to accurately estimate the group’s total effect on cost variability via computation of an aggregate coefficient for each group.

Hospital data

Research was undertaken in a public hospital, which is ultimately under the jurisdiction of the Queensland Department (ie Ministry) of Health. The study has been approved by a Human Research and Ethics Committee (HREC), which is the relevant Institutional Review Board. The particular committee is the Royal Brisbane and Womens’ Hospital’s HREC. The Australian approval number is HREC/16/QRBW/61. The Human Research and Ethics Committees in Australia grant approvals in accordance with the Australian National Statement on Ethical Conduct in Human Research (2007), which in turn fulfills the obligations under the Declaration of Helsinki. As the study is only a statistical analysis of large data, with negligible risk of patients being identified, the above-mentioned ethics committee has granted a waiver of patient consent. This waver has been subsequently endorsed and approved by the Queensland Department of Health. The current dates of the ethical approval are from 09 March 2016–09 March 2019.

The data were all hospital admissions to a mid-sized hospital during the 2014/15 financial year. This totaled 53,224 admissions. Some admissions pertain to the same person, with the number of unique patients being 31,449. The hospital serves a geographically contained population and lacks any notable sub-specialisation skews or unusual service gaps which may occur in geographical areas with less defined servicing boundaries (e.g., in a large city with several hospitals in close proximity). This should justify extrapolation and wider applicability of the results.

The potential cost predictors analysed are listed in Table 1. The cost drivers were classified as predictable a-priori and thus representing potentially warranted cost variability, as well as non-predictable, as also depicted in Table 1. This process of classification followed previously used logic [24] and it also incorporated senior medical opinion, in line with the classification principles described above in the Introduction. Predictable drivers were considered as associated with warranted cost variability. Examples of such cost drivers included age and sex. It is specifically worth highlighting that DRG was considered as representing a predictable cost driver a-priori and was classified accordingly. On the other hand, cost drivers such as day of the week or time of the patient’s admission were not reasonably predictable a-priori, and were thus classified as non-predictable, with any related cost variability being potentially unwarranted. We should not expect that a patient admitted on a Wednesday should cost any more or less than a patient with similar characteristics admitted on a Sunday.

Initial exclusion of factors

Several potential cost drivers available from the hospital information system were then deliberately excluded from further analysis as they were felt to merely represent likely consequences of other variables; see the causal diagram in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Causal diagram.

https://doi.org/10.1371/journal.pone.0204300.g001

The potential driver of days spent in the ICU (Intensive Care Unit) was left out, as it depended strongly on other variables left in the model, in particular the type and severity of the patient condition (already described by DRG), as well clinical decision making (already encapsulated in both the variables of admitting unit and admitting doctor). In fact, the decision to admit to intensive care essentially reflects clinical interpretation of the nature of the patient condition, in the context of limited intensive care resources [25]. The limitation on resources, particularly the immediate availability of intensive care beds is, of course, likely to be more prominent at certain times than at others. While the model specifically includes variables of day of the week, month as well as hour of the day, it is however acknowledged that at least some of this variability is likely random and would not be included therein. Because the very aim at this stage is to identify the proportion of cost variability that is unwarranted and potentially rectifiable, stochastic variability is not modeled in the purposely deterministic model.

There are strong parallels between the decision to transfer a patient to a different facility and to accept a patient to intensive care, as discussed above. Transfer out can occur either because the clinical capacity of the hospital under analysis is thought to be exceeded, because the patient condition necessitates admission under a medical team that is not available locally, or perhaps because the patient condition is considered stable enough for transfer to a lower acuity facility. All these are dependent on clinical decision making within the context of the patient condition. Hence, the potential cost driver of “transfer out” was left out, mirroring exclusion of the variable of intensive care admission.

The flag for selected clinical incidents was left out because clinical incidents were related to other variables left in the model. Clearly, rates of clinical incidents would be necessarily associated with the patient condition (as described by the DRG). For instance, patients with conditions that require multiple medications would be more likely to experience medication-related adverse than those who may take fewer medications. Also, those with medications that are known to decrease blood pressure or cause confusion would be more likely to experience in-hospital falls. Furthermore, medical conditions that limit mobility may render a patient more likely to develop pressure injuries.

There is also likely to be a random component that contributes to when clinical incidents take place. As mentioned above, as the aim at this stage was to identify the proportion of deterministic and thus potentially rectifiable cost variability that is unwarranted, these variables were left out of the model.

Clinical incidents are known to be associated with problems in clinical assessment, communication and decision making, the nature and staff profile of the relevant clinical unit as well as the physical attributes of the surroundings [26]. These are likely to be associated with random events, especially given the emerging theories of accident causation in complex sociotechnical systems [27]. However, any additional deterministic effect would likely be encapsulated by variables of admitting ward, admitting unit, admitting doctor, discharge ward, discharge unit and discharge doctor, which were all left in the model.

Parallels exist between the CHADx flag (which is a flag for hospital-acquired complications) and clinical incidents. Information encapsulated in the CHADx variable is likely to lead to varying DRG assignment and was left out of the model on the basis of the DRG staying in. This is supported by the fact that the CHADx uses routinely abstracted hospital diagnosis and condition-onset information about in-hospital complications [28].

The amount of health care received by patients, when standardised by patient related characteristics is known to be largely driven by clinical decision making [29]. The amount of total clinical care received during an episode of care is however a function of the amount of care provided per unit time, as well as the total duration of the episode of care. Because of inclusion of variables pertaining to doctor as well as the medical unit in the model, both the amount of care per unit time and the length of stay in hospital were therefore excluded from the model. The exclusion of the hospital length of stay (LOS) from the model mirrors other recent work on cost driver analysis in hospitals [24]. Furthermore, the LOS may be also dependent on a wide suite of determinants, some pertaining to patient flow dynamics, rather than to processes related to clinical care.

In summary, the white boxes represent such potential cost drivers, not likely to represent the primary drivers of patient-level costs. This exclusion specifically applied to the hospital length of stay (LOS) and its non-inclusion in the model mirrors other recent work on cost driver analysis in hospitals [24]. The grey boxes represent the variables left in the model, as they were considered to represent, at least in substantial part, primary cost drivers. This preliminary choice of variables does not negate the need to analyse and eliminate correlations between the variables that remain in the model. Accordingly, Table 1 indicates the factors included in the subsequent sparse group lasso analysis, following the causal analysis.

Subsequent analysis

The CCI (Charlson Comorbidity Index) has been shown to have a predictive association with health care costs [30]. It was obtained from the Health Roundtable, an organisation which computes the index for subscribing health providers.

The Index of Relative Socio-economic Advantage and Disadvantage (IRSAD) and the Index of Relative Socio-economic Disadvantage (IRSD) were obtained from the Australian Bureau of Statistics SEIFA 2011 [31] which were added by cross-referencing against the patient’s postal code. Postal codes were missing in just 326 of the 53,224 patients. To prevent loss of other potentially valuable information related to these patients, these were given the average values of both indices, calculated from the remaining data, rather than being excluded.

Other than the postal code data, missing data was negligible. Only one patient had missing data for private insurance (both private insurance as well as DVA or veteran insurance status), smoking, obesity or alcohol flags. All these fields were given modal values, calculated based on all remaining information, in line with the categorical nature of the data.

All categorical variables, except for the CCI, were first represented as sets of binary variables. The relationship of the CCI (ordinal integer values), IRSAD, IRSD and age were first both assumed as linear with respect to the patient hospital costs. These assumptions were felt to be justified in relation to the IRSAD and the IRSD, given the design of both these indices. Appropriateness of these assumptions in relation to the CCI and age was however analysed further with the sparse group lasso re-run, but this time including a squared term for each of these two variables. Computational efficiency necessitated that the re-run was carried out on a random 10% of the data; however it was repeated three times.

PCA was used to pre-process the data and reduce its size without losing vital information. See S3 Text for further description of use of this technique. Output from application of the PCA to categorical groups of factors was used as input into the sparse group lasso. Please see S4 Text on how the optimal lasso parameters of α and λ were derived.

Once λ (the lasso penalty) was determined, the corresponding coefficients of the selected variables were used to compare their respective influences on patient-level hospital costs variability. Absolute values of all regression coefficients were first determined, in keeping with the focus being on cost variability rather than the absolute cost values. Absolute values of all coefficients within groups were added together to provide aggregate coefficients for each multi-factor cost driver (e.g., DRG or discharge ward). Each aggregate coefficient is therefore a resultant estimation of the relative contribution of each grouped variable to patient-level cost variability.

Please see the S5 Text for the method used to compare respective effects of categorical and continuous variables. Please see the S6 Text for an explanation of how confidence intervals were constructed. S7 Text describes the computing times associated with the calculations.

Comparison with simple regression

The results of the sparse group lasso were paralleled by also analysing the data using simple linear regression, for comparison purposes. This used the command lm, available in the stats package in R. Like sparse group lasso, this analysis was also performed on the data pre-processed by the PCA, to facilitate ease of comparison. Grouping of binary parameter estimates pertaining to the same cost driver and handling of the age, IRSAD, IRSD as well as the CCI coefficients was carried out in a manner identical to the case of the sparse group lasso.