HTA - lecture 7 - modelling and uncertainty Flashcards
types of uncertainty
- Variability (natural spreading)
o If we look at survival, not everyone has the same survival
o If we look at age, not everyone has the same age
o If we observe differ people, we get different outcomes - Heterogeneity (variability that is understood)
o We understand that age in different countries is different - Uncertainty (incomplete information)
sensitivity analysis to address uncertainty, goal
find out how sensitive results are to changes in parameters.
* If outcome is sensitive to changes in specific parameter, this may guide in areas for further research
types of sensitivity analysis
Deterministic sensitivity analysis
* Univariate sensitivity analysis
o vary only 1 or 2 parameters at the time
* Multivariate sensitivity analysis
o vary 2, more or all parameters at the time
o Worst case’ & ‘best case’ scenario
o You change 1 parameter and you recalculate the outcome
Probabilistic sensitivity analysis
probabilistic sensitivity analysis
Recommended in many guidelines (UK, Canada, NL etc.)
Most informative method, since it presents extreme outcomes, but also likelihood of outcomes
Here: vary all parameters at the same time
Steps
* Define probability distribution for each variable
o For each parameter we need to create a distribution on which it can occur
* Draw random number from each distribution and calculate ICER
* Repeat many times (1000-5000)
Input parameters are uncertain distributions
normal distribution
Normal distribution
- 2 parameters; mean and SD
always a candidate (Central Limit Theorem (CLT): sampling distribution of the mean will always be normal, whatever the underlying distribution of the data, when the sample size is large enough
alternative for normal distribution if:
- CLT (normal distribution) does not hold
- Logical constraints on the parameter
o Costs >0, can never be negative
o Probabilities 0-1, can never be smaller then 0 and bigger then 1
distribution for probability parameters
- Probabilities are constrained on the interval 0-1
- Probabilities must sum to 1, because we cannot lose patients
- Probabilities often estimated from proportions
o Data are binomially distributed - Beta distribution is good option
fitting the beta distribution
- Beta distribution has two parameters:
o Beta(alpha, beta) - Alpha = no. of events
- Beta = no. of non-events
distribution for cost parameters
- Costs are constrained to be zero or positive
- If normal distribution unreasonable, use no negative values
o Gamma
o Lognormal
SE
o New treatment is showing effects with less costs
o Generating more health with less costs
NW
o Generating less health with more money
NE
o When it is below the threshold value, we accept
SW
o Negative health effects and negative costs
o Accepted if ICER is higher than threshold showing money that you saved for losing health effects. This should be as high as possible. Is we save 1 million and lose 1 QALY. We gain money for losing health. We save a lot and just lose a small amount of health. If we wanted to pay 80.000 per QALY, we want at least 80.000 for a loss of health.
constructing acccurve
- If all simulations on NE and SE: curve increasing
- If all simulations on SW and SE: curve decreasing
net monetary benefits
- Instead of cost-effectiveness ratio, the cost-benefit may be calculated, using the threshold ICER (l) as valuation of effects
- INMB =
- Remember lecture 1
Different threshold values can be used to evaluate outcomes
* Is ratio below or above the willingness to pay for a QALY V threshold
* Is ratio below or above our current healthcare production? K threshold
If net monetary benefits are low, it is difficult for treatments to be accepted
