Understanding Simulations in TreeAge Pro
TreeAge Pro supports four kinds of Monte Carlo simulation. It can be a little confusing understanding which to run and when. This article describes each form of simulation and when you would use it.
- Samples/Probabilistic Sensitivity Analysis (PSA)
- Samples & Trials/PSA with Microsimulation
- EVPPI Sampling
Microsimulation runs individual patients through your model. Each iteration in the output represents the final values for a single patient after the person finished its path through the model. The aggregated mean values from this simulation represent a single calculation of the model to determine the average values (i.e., cost and QALYs) per patient for each strategy. Those mean values can then be compared with other strategies via decision analysis (i.e., CEA).
Only models that use individual patient data need to be analyzed via Microsimulation. This would be true if you added either of these capabilities into your model.
- Heterogeneity through trial-level distributions or through bootstrapping.
- Event tracking through trackers.
If you don't use either of these techniques, then standard cohort analysis expected values are fine and Microsimulation is unnecessary.
Samples/Probabilistic Sensitivity Analysis (PSA)
PSA recalculates the model many times under different data scenarios to measure the impact of combined parameter uncertainty. Each iteration reflects a full recalculation of the model with different parameters drawn from EV-level distributions.
The mean values from the analysis do not provide much value; however, review of individual iterations/recalculations can help measure the robustness of your base case analysis. When more individual calculations choose the same optimal strategy as your base case analysis, the more confident you can be in that conclusion. Typically, secondary reports like Acceptability Curve and ICE Scatterplot provide PSA outputs typically included in publications.
This analysis should not be run on models that require Microsimulation (individual patients).
Samples & Trials/PSA with Microsimulation
This analysis is run for the same purpose as Sampling alone, and generates data in the same format. However, if your model requires Microsimulation, you should run this analysis instead.
The analysis starts with PSA by drawing parameters from EV-level distributions; however, once the parameters are inserted into the model, individual patients are run through the model using the Microsimulation analysis described earlier. The results of the Microsimulation trial set are then aggregated into mean values for each strategy and payoff set. This process is then repeated many times for different sets of sampled parameters.
In essence, this is a combination of the prior two types of analyses, which allows you to run PSA on a model that requires patient data.
EVPPI is a special form of PSA, which isolates the uncertainty related to 1 (or a couple) of distributions from the uncertainty generated by the other EV-level distributions. To achieve this goal, the PSA analysis is broken into two calculation loops - VOI and PSA.
In the VOI loop, the isolated distribution is sampled. In the PSA loop all the other parameter distributions are sampled, then the model is calculated. Model calculations can take the form of cohort analysis or Microsimulation depending on whether your model requires patient data. The PSA loop then continues with another set of those samples and another recalculation of the model. Once many PSA-level calculations are complete, those values are aggregated into a mean. Then the analysis returns to the VOI loop to sample the isolated parameter again, followed by another PSA loop with sampling and model recalculations.
The resulting dataset is in the same format as the PSA output described above, except that the only uncertainty reflected in the results is that caused by the isolated distribution.