GC650 Lecture 5 - CEA and CUA with Sensitivity Analyses

Willingness to Pay

Arbitrary threshold of whether something is a good value.

When we say an intervention is ‘cost-effective’ we mean that the cost per QALY is below some threshold of Willingness to Pay.

Cost Effectiveness Plane

Y-axis: Cost

X-axis: Effectiveness

Dominated

If the new treatment is less effective and more expensive than the old, it is dominated by the old treatment.

Cost Effectiveness Acceptability Curve

A graphical tool used in health economics to represent the probability that an intervention is cost-effective compared to alternatives across various willingness-to-pay (WTP) thresholds

Sensitivity Analysis

Used to help us understand how much the data quality issues may impact the outcomes of the model

Allow you to test “What-If” situations when you know a set of model inputs may not be entirely correct or if you suspect something may change in the near future

One-way

Vary one uncertainty input at a time.

Tornado Diagram

A specialized horizontal bar chart used in sensitivity analysis to visually compare the impact of various input variables on a specific outcome

Two-way (Multi-way)

Vary two (or multiple) inputs at the same time

The graphic axes are the minimum to maximum values of each parameter

The “inner” part of the graphic is shaded where one strategy is more cost effective than the other

Probabilistic Sensitivity Analyses

Rather than use the MEAN of each input, we use the distribution

CEA and CUA Relationship

CUA is a subset of CEA

Result of of CUA is QALY

Uncertainty

Uncertainty reflects variability in model inputs (costs, outcomes, probabilities)

Characterizing uncertainty enhances reliability and transparency.

Types of Uncertainty (2)

1. Structural - model design choices

2. Parameter - input variability

Methods to Address Uncertainty (3)

1. One-way Sensitivity Analysis

2. Two-Way Sensitivity Analysis

3. Probabilistic Sensitivity Analysis (PSA) - Monte Carlo simulations

What Can Cause Error in Mean Measurements (3)

1. Endpoints can have a lot of variability (large SD)

2. Quality issues in the input studies

3. Lack of data on some data points

Quality Issues in Models (4)

1. The data best suited for the model often do not exist

2. Prior studies are of lower quality than desirable

3. Prior studies may not be generalizable

4. Some data points may not be available at all

Ways the Data Available May Not Fit the Model (3)

1. Prior studies conducted on a different population

2. Prior studies conducted with a different comparator

3. Prior studies conducted under a different treatment paradigm

Single Deterministic One-Way Analysis

Shows the impact of changing one model input (or parameter) on a key outcome (cost or effectiveness) for each treatment strategy

Multiple Deterministic using Tornado Diagram One-Way Analysis

Shows the impact of changing a series of inputs, one at a time, on the ICER and plots them in order of biggest impact to smallest impact

Types of One-Way Analyses

1. Single Deterministic

2. Multiple Deterministic using Tornado Diagram

Natural Variation in Patient Experience

All data are derived from a pool of observations that represent natural variations in experience, when we summarize data – especially with models – we often focus on or use the mean or median of the data and may not perfectly reflect the real world experience

Probabilistic Sensitivity Analyses Take into Account (3)

1. That the mean does not represent everyone, there is variation among individual patients

2. That the mean has error in its measurement

3. That the incremental cost per outcome will also vary by individual and with measurement error

Steps of a Probabilistic Sensitivity Analysis (3)

1. Distribution is specified by the user based on SE and expected standard distribution (gamma, weibull, etc…)

2. The model randomly selects a point from the distribution of each input parameter

3. We run the model roughly 1,000 times

Negatively Skewed, Normal Distribution, Positively Skewed Distribution Appearances

Monte Carlo Simulations

We run the model roughly 1,000-10,000 times, ach time randomly selecting a point off the distribution of every input parameter (this is the monte carlo part)

Then we plot the points on the CE Plane

Discounting

Current and future dollars are not valued the same 🡪 TIME PREFERENCE FOR MONEY AND TREATMENT BENEFITS

Future dollars must be discounted to reflect present value when program extends beyond one year

What Can be Discounted (4)

1. Costs

2. Utilities

3. QALYs

4. Benefits in dollars (CBA)

Discounting Calculation

3% per year is the recommended discount rate

Should you discount benefits?

Yes - because there is a preference for having additional life years or utility gains NOW versus later

No - Some argue against if there is no trade-off between having additional life years/utility gains now versus later

Final recommendation - Future benefits SHOULD be discounted, at the same rate as costs, if they are in units of QALYS or utilities

C/U Ratio if Benefits are Not Discounted

If you only discount costs, then C/U Ratio will always improve over time.

Inflation

Future costs, both direct and indirect, will increase each year.

We generally inflate costs but not benefits/outcomes.

Discounting With Inflation Steps (2 and a Substep)

1. Adjust historical costs for inflation first, to bring them to present value

2. Then apply an inflation-adjusted discount rate for each year of analysis in the future.

   1. If using a 3% discount rate, and inflation rate for hc goods and services is 2%, then inflation adjusted discount rate is 1%