Rx-Risk Index mapping

The Rx-Risk Index consists of 46 comorbidity categories. For each Rx-risk category, medicines indicative of that condition were mapped (see table 1). This mapping was performed at the ATC classification level and performed by consensus between two pharmacists. If an individual had ≥1 dispensing for a medicine in a given category, then they were considered to have been treated (using medicines) for that comorbidity. Medications in both the Department of Veterans’ Affairs (DVA) and Pharmaceutical Benefits Scheme (PBS) datasets are coded using WHO ATC Classification System14 and PBS item codes.15

Data sources

The primary data source was the Australian Government DVA’s administrative claims database. This database contains details of all prescription medicines, medical and allied health services, and hospitalisations subsidised by DVA. The current treatment population consists of 223 181 members of the Australian veteran community, which includes veterans, war widows and widowers. The median age of the DVA treatment population is 75 years and 62% are men. DVA also maintains a client file containing gender, date of birth and date of death information.

External validation of the Rx-Risk Index was conducted using the PBS 10% sample of the Australian population. This dataset contains information on the dispensing of prescription medicines, and includes basic demographic information regarding gender, year of birth and year of death. It is maintained by the Australian Government Department of Human Services. The current treatment population consists of 1 346 340 members of the Australian community. The median age of the PBS treatment population is 43 years and 48% are men.

Patient and public involvement

Patients and public were not involved in this research.

Study population

The DVA study population included individuals with at least one healthcare encounter in the 6-month period from 1 July 2013 to 31 December 2013. A healthcare encounter included any of the following: a medication dispensing, a doctor’s visit or hospitalisation. Analysis was restricted to veterans who were DVA Gold Card holders prior to 1 July 2013 (ensuring they were eligible for all DVA subsidised services; thus, the dataset captured all their health claims), and were between the ages of 65 and 100 years at 1 January 2015 (n=135 406). Inclusion criteria for the PBS cohort were people with a healthcare encounter (defined as at least one medicine dispensing) in the 6-month period from 1 July 2013 to 31 December 2013, who were aged between 65 and 100 years at 1 January 2015 (n=303 135).

The primary outcome for this study was death recorded in 2015; hence, patients were only included if they were alive as at 1 January 2015. Rx-Risk Scores were calculated separately in each dataset using prescription claims for supplied medicines over a 1-year baseline period between 1 January 2014 and 31 December 2014.

Calculating Rx-Risk Scores and prescription counts

The full Rx-Risk Index has 46 comorbidity categories; however, tuberculosis and hepatitis B and C were removed from the Rx-Risk Index for this study as the number of individuals with these conditions in the DVA cohort was less than 10 so weights could not be generated. This resulted in 43 categories in the validation study. The following forms of the Rx-Risk Score were generated. The unweighted Rx-Risk Score was calculated as the count of the number of different comorbidity categories for which an individual was treated with a possible score ranging from 0 to 43. The weighted Rx-Risk Score was calculated by adding the 43 indicator variables to a logistic regression model with mortality as the outcome including age and gender as covariates. From this model, each comorbidity category was assigned a weight according to the statistical significance and magnitude of the OR generated from the logistic regression model (table 2).7 The weighted Rx-Risk Score for an individual was then the sum of the weighted indicator variables. As an example, the unweighted Rx-Risk Score for a patient who has two comorbidities ‘pain’ and ‘renal disease’ is 2, while their weighted Rx-Risk Score is 9 that is the sum of the weight for pain (3) and renal disease (6).

Three crude prescription counts were also calculated: (1) total number of prescriptions dispensed in the baseline period (1 January 2014–31 December 2014), (2) total number of unique medicines dispensed in the baseline period based on ATC codes, and (3) total number of unique medicines dispensed during the baseline period based on PBS item codes.15 The distinction between ATC and PBS codes was made because different strengths or formulations of the same medicine have the same ATC code but different PBS codes; for example, if a patient is dispensed two different strengths of the same medicine, this will be counted once in the total number of ATC medicines dispensed but twice in the total number of PBS medicines dispensed.

Statistical analysis

Primary analysis was performed in the DVA database. A baseline logistic regression model was calculated for mortality using age and gender as predictors. The comorbidity scores, individual comorbidities and crude prescription counts were added to the baseline model separately. Age, comorbidity scores and crude prescription measures were included in the models as continuous variables assuming linear associations. Models using the individual comorbidities were developed with an indicator variable included for the presence (1) or absence (0) of each individual Rx-isk category. The overall goodness of fit of each model was compared with the baseline model using the Akaike information criterion (AIC).16 The model with the lowest AIC value is considered the best fit. The difference between the AIC values of two models must be greater than 10 for one model to be considered superior to the other. Model discrimination was compared based on the c-statistic and the relative integrated discrimination improvement (IDI).17 The value of the c-statistic can range from 0 to 1, with 1 indicating perfect prediction and 0.5 indicating chance predictions. A c-statistic between 0.8 and 0.9 is generally considered as excellent, and between 0.7 and 0.8 acceptable.18 Using 1000 bootstrap samples, 95% CIs were calculated for the c-statistics.9

Internal validation of weighted scores

We used 10-fold cross validation to internally validate the binary logistic regression model used to calculate the Rx-risk category weights. This subsets the DVA cohort, using random sampling without replacement, into 10 equal folds. Each fold is a 10% subset of the DVA cohort. The training set was ninefold and was used to calculate Rx-risk category weights. The calculated weights were then applied to the testing set (ie, fold left out of training set). A binary logistic regression model of 1-year mortality including age, gender and weighted Rx-risk was built separately for the training set and testing set and the c-statistics recorded. This process was repeated 10 times until each fold was used as the testing set once. This resulted in 20 c-statistics being calculated, 10 each on the training and testing set. The average c-statistic was recorded for each set.

Sensitivity analyses

Two sensitivity analyses were carried out. In the first sensitivity analysis, we used the lower confidence limit of the OR to determine the Rx-risk category weights rather than the estimated OR itself. We did this as some comorbidities are uncommon in this population resulting in large CIs. In the second sensitivity analysis, we calculated the Rx-risk category weights (based on the OR) generated from 5000 bootstrap samples. For each Rx-risk category, the weight was calculated as the median of the 5000 weights for that category generated in each bootstrap sample. We did this sensitivity analysis to test the stability of the weights.

External validation

To determine the external validity of the weights, the calculated Rx-risk category weights derived from the DVA dataset were applied to the PBS cohort. The AIC model fit and c-statistics were calculated to determine the validity of the model.