An economic evaluation, such as the cost-effectiveness analysis, aims to estimate the effect of an intervention to the life of a cohort of patients, in terms of quality of life, years of VEN. The random clinical trials, on the other hand, son an important source of data, but, in some occasions, insufficient son, for example, has limited follow -up, or because the primary variable is a variable surrog.
To help decision -making in health, health economists usually resort to the construction of analytical models, since they allow, for example, to extrapolate the results of a clinical trial to a longer temporal horizon, or estimate the years of life, or to perform uncertainty analysis. Modellos Popular Analytics in Economic Evaluation are Markov, Survival Party, and Individual Survival models. These feed with input data (Tickets) of the event rates (for example, progression of the disease, mortality, etc.) along the temporal horizon of the model.
Although the models have these advantages over clinical trials, long -term extrapolia has some challenges. A recommended method to perform this East, with the Delsayo Clinico data, «a parametric survival model», which consists of the esteem through a rurresion.
However, for this you have to choose the best formula (Weibull, log-logistic, log-normal, gamma, etc.) for representative the distribution of Dangers (Event risks). Several guides have been published, of which NICE is very respected and, in addition, didactics (TSD 14). It recommends a series of actions to guide and validate the construction of a survival model, which are based on the following two questions:
- What is the best formula to predict the distribution of event risks over time, both during the clinical trial (internal validity) and later (external validity)?
- Is the event rate in the intervention arm a multiple constant of the event rate in the control arm (proportional risks)? Or have behaviors Different?
It is said that Niels Bohr, Nobel Prize in Physics and father of the atomic model, said: «Predict is very difficult, especially if it’s the future!» (Figure 1). This appointment serves as the importation of the importance of testing a prediction model outside the masestra. It is often easy to find a model that adjusts well to the observed data, perhaps too well! But another thing is a model that correctly identifies the characteristics of the past data that will repeat in the future. Therefore, the analyst should debate both internal validity (the goodness of the selected distribution with the observed data) and external validity (the plausibility of the short -term predictions and long term).
Figure 1. Particle collision in the large Hadron collider
The TSD 14 guide recommends some statistical tools to evaluate internal validity, for example, AKike’s information criterion (AIC). On the other hand, to evaluate external validity, it is recommended to visualize the predictions of the model, in the face of the observed data, and elicitate the opinion of independent experts.
However, the construction of survival models, the visualization of its graphic results, and the execution of static tests, requires skill with codes of R, STA or other statistical software. RS a powerful program, but has an inclined learning curve, and should be updated to the versions of the new ones. In addition, in many occasions, analysts do not have individual data, but secondary data published in a magazine, and in the format of a Kaplan-Meier curve.
To facilitate the survival analysis work, Daniel Pérez Troncoso and myself, both members of the Eecona Interest Group, have developed free and free codes. Concretement, it is an applied GITHUB. Although it is made in R, the user does not have to have R installed, and neither did he know how to use R.
Once the hard drive folder is installed and decompressed, the user clicks on the tab< run.bat >> And the app opens in the browser through a visual interface built in< Shiny >>. The user provides the input data, which are the digitized coordinates of a Kaplan-Meier curve (the website can be consulted Weblootdigitalizer To learn how to dignity a curve) and data on the number of people at risk (normal disonvigable also in clinical publications). The download folder< pronóstico >> Contains an Excel template and examples of entry data in the required digital format.
From there,< pronóstico >> You can rebuilding individual data (applying the algorithm of Guyot), Modeling the parameters and risks associated with 11 alternative distributions (exponential, weibull, log-logític, log-normal, etc.), graphic representative the predictions of the predictions of the Dangers (Figure 2) and the survival (Figure 3) of the parametric model, and execute statstic tests of the adjustment goodness (AIC).
Figure 2. Example of the observed distribution of risks (Dangers) In an arm of a clinical trial and predictions of alternate parametric models
Figure 3. Example of the observed survival distribution (Kaplan-Meier) in an arm of a clinical trial and predictions of alternative parametric models
Also,< pronóstico >> You can lower the results of a statistical test to verify whether the risks in the intervention group son proportional to those of control. If the proportional child risks, the logarithm curves of the cumulate risks were parallel in the two groups (Figure 4) (the Nice Guide is suggested TSD 14 To obtain more detail to respect).
Figure 4. Example of cumulative risk logarithm
The utility of< pronóstico >>
It should be noted that the app< pronóstico >> Without pretending the best risk function for the user, but providing the tools to select it by yourself analysts in charge of performing Markov models (example water), Survival model departure (example water) Or other types of economic evaluation.
We would like to have comment From the community of analysts about their experiences with the applying, so we invite you to comment on this publication with their presses, suggestions and doubts.
