Hi everyone
What is the importance of each parameter: R, R2, RMSE, p, F, T-test Q2 LOO, Q2 boot, Q2 EXT ..... of the QSAR model by the MLR statistical method ? and among these parameters which are the parameters which are satisfactory to say that this model is well correlated ???
Hi,
By doing Multiple Linear Regression you want to see how predictors (x variables) can explain the response variable (y). The linearity of model determines by R-square or coefficient of multiple determination, which is a between 0-1. The R-square is determined by subtracting the standard error of the equation. If you choose extra predictor variables the Adjusted-R-square will be lower, even if the predictors are desirable. F determines if the predictors are significantly associated with the model. The null hypothesis is that none of the predictors are significant. By F < 0.05 we can reject null hypothesis. Leave-one-Out (LOO) or Leave-N-Out (LNO) cross validations are used to estimate how valid is our model. However, if the sample size is small, the LOO or LNO cross-validation are no reliable. By LOO cross-validation, we discard one of the observation and try to resample again and see if the model stands still. The root mean square error of cross-validation is some other validation and obtain by residuals differences. Bootstrapping is something like LOO and LNO, but a sample may be discard only once or several time or never. When the model is built based on training set, one may want to externally validated by applying it on test set. Here, Q2 is step forward. Q2 is interpreted like R2 but the value of Q2 in robust linear models is low.
Hi
Late responce butt hope will be still usefull. Some example of LOO validation (and apropriate references) you can find in my paper:
Article Models describing metronidazole pharmacokinetics in relation...
squared cross-validated correlation coefficient (Q2)
- measures of the internal model performance
differences between Q2 and R2
- model predictive ability
difference between asymptotic squared cross-validated correlation coefficient (Q2asym) and Q2
- difference of ability between fitting and predictive ability
total sum of squares (SS).
- value closer to 0 indicate that the model has a smaller random error, and fit will be moreoptimal for prediction.
predicted residual sums of squares (PRESS)
- lowest value of PRESS indicating not over-fitted model
Validation acceptance criteria
Q2 ≥ 0.65
Q2ext ≥ 0.65
R2 ≥ 0.85
Q2-R2 < 0.3
Q2asym-Q2 > 0
Hope it will be helpful,
Best regards
Tomasz