Other distributions than Weibull

1. Other distributions than Weibull

The Weibull model assumes that the outcome has a certain distribution. In this video, we want to look into this assumption and the alternatives.

2. survreg() options

What does it matter, which distribution to choose? Well, the result that you get will differ depending on the distribution you assume for your outcome. Here is one example using the GBSG2 data set. Depending on whether we assume an exponential, log-normal, or Weibull distribution, the survival curve will look different. The hormonal Therapy is more beneficial in all cases, but the curves do look different.

3. survreg() options

To compute the different models in R you can just specify the dist argument. Weibull is the default. To compute an exponential model, set dist to "exponential". To compute a lognormal model, set dist to "lognormal". Plotting can be done the same way as in the Weibull model. More distributions are available. Just take a look at the survreg help page via ?survreg. Now the difficult question is: which distribution should I use? Sometimes it is quite clear that for a certain time-to-event outcome a certain distribution should be used. You can, for example, use the same as what other people have used in their projects. Also, I would usually use the Weibull model instead of the exponential model, because the Exponential model is a special case of the Weibull model and the Weibull model is more flexible. In simple models such as in the example here it is also an option to fit Kaplan-Meier curves and compare the survival curves with the model fit to see if it is reasonable. A general note on modeling choices: Many choices we make in statistical modeling are philosophical. They mirror what we think the world is like. This is not only the case in survival analysis but in any statistical modeling. Remember, for example, that in linear models we assume a normal distribution. A lot of people use the linear model without thinking about this fact.

4. Let's try working with different models

Let's try working with different models

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.