We are receiving a lot of questions about relevant analyses in the Analyse-it Method Validation edition to help in evaluating new diagnostic tests in the fight against COVID-19. Below are some quick links that will help, but contact us if you have questions - we are working as normal.
Also see our latest blog post: Sensitivity/Specificity and The Importance of Predictive Values for a COVID-19 test
R² and similar statistics measure how much variability is explained by the model.
R² is the proportion of variability in the response explained by the model. It is 1 when the model fits the data perfectly, though it can only attain this value when all sets of predictors are different. Zero indicates the model fits no better than the mean of the response. You should not use R² when the model does not include a constant term, as the interpretation is undefined.
For models with more than a single term, R² can be deceptive as it increases as you add more parameters to the model, eventually reaching saturation at 1 when the number of parameters equals the number of observations. Adjusted R² is a modification of R² that adjusts for the number of parameters in the model. It only increases when the terms added to the model improve the fit more than would be expected by chance. It is preferred when building and comparing models with a different number of
For example, if you fit a straight-line model, and then add a quadratic term to the model, the value of R² increases. If you continued to add more the polynomial terms until there are as many parameters as the number of observations, then the R² value would be 1. The adjusted R² statistic is designed to take into account the number of parameters in the model and ensures that adding the new term has some useful purpose rather than simply due to the number of parameters approaching
In cases where each set of predictor values are not unique, it may be impossible for the R² statistic to reach 1. A statistic called the maximum attainable R² indicates the maximum value that R² can achieve even if the model fitted perfectly. It is related to the pure error discussed in the lack of fit test.
The root mean square error (RMSE) of the fit, is an estimate of the standard deviation of the true unknown random error (it is the square root of the residual mean square). If the model fitted is not the correct model, the estimate is larger than the true random error, as it includes the error due to lack of fit of the model as well as the random errors.