Probabilistic Sensitivity Analysis
Steps for Probabilistic Sensitivity Analysis:
Step 1: Assign Probability Distributions:
Assign a probability distribution to each parameter in your model.
Spreadsheets will be provided to assist in this task.
Step 2: Draw Random Numbers:
Use a computer to draw a random number from each assigned distribution.
Step 3: Model Rerunning:
Input the random numbers into your model.
Rerun the model to obtain the output.
Step 4: ICER Calculation:
Calculate the Incremental Cost-Effectiveness Ratio (ICER) for the event before the intervention.
The ICER can be calculated for any outcome, but the primary focus is on the ICER.
Step 5: Plotting:
Save the calculated ICER.
Plot the ICER values on a graph.
Dependent variable (e.g., cost) on the Y-axis, independent variable (e.g., effect) on the X-axis.
Repeat the procedure multiple times.
Step 6: Cost-Effectiveness Plane:
Plot all ICERs on the cost-effectiveness plane (as discussed previously).
Step 7: Statistical Analysis:
Calculate statistics such as probabilities or frequency distributions.
Determine the percentage of times the ICER falls below or above an acceptance threshold.
Assess the frequency of occurrences where the ICER plots in an undesirable quadrant.
Cloud Diagram Interpretation:
The result is a cloud diagram, which visually represents the uncertainty in the model.
A larger cloud indicates greater uncertainty.
A smaller, denser cloud indicates greater reliability or certainty in the outputs.
Probability Distributions:
Normal Distribution:
While commonly used in statistics, it is often unsuitable for cost-effectiveness analysis.
Costs cannot be normally distributed because they must be greater than zero and tend to be right-skewed.
Prices for inputs may sometimes be normally distributed, but negative prices are unlikely.
Utilities and probabilities (ranging from 0% to 100%) cannot be normally distributed either.
Common Distributions:
Gamma distribution
Beta distribution
Log-normal distribution
Poisson distributions
Matching Distribution to Parameter:
Crucially, the selected distribution must align with the characteristics of the parameter being modeled.
The use of inappropriate distributions will be challenged (e.g., in a thesis).
Normal distributions may not be suitable for integers like the number of tests or doses.
Graphical Examples of Distributions:
Normal Distribution:
Bell-shaped curve with clear peak and tails.
Gamma Distribution:
Defined by shape and scale parameters.
Tends to lean to the left, with higher frequency outcomes toward the left.
Beta Distribution:
Resembles a normal distribution but leans more toward the left.
Log-Normal Distribution:
Defined by mean and standard deviation.
Applying Distributions:
Gamma Distribution:
Parameters: shape and scale.
Scale: ranges from greater than zero to infinity.
Use case: simulating costs because costs are positive and tend to skew.
Beta Distribution:
Parameters: aimed at shape.
Support values: 0 to 1.
Use case: simulating measures of utility and probabilities.
Bounded between zero and one.
Log-Normal Distribution:
Exponent of a normal distribution used for very large numbers.
Log transformation reduces the scale of numbers and brings widely distributed values closer together.
Domain (or Support):
The range of values for which the distribution is defined.
Support = domain = range of values.
Simulation Implementation:
Software:
Statistical packages like SAS or Stata.
Excel (with additional steps).
Excel:
Requires calculating a random value from each distribution in multiple steps.
For gamma and beta distributions, using a statistical package is recommended or seeking assistance with the two-step process in Excel.
Random Number Generators:
Most software includes a random number generator.
Inverse Transform Method:
Used to simulate values from continuous distributions.