For each selected item, five dose-response models (linear, Hill, exponential, 
Gauss-probit and log-Gauss-probit, see Larras et al. ???? for their definition) 
are fitted by non linear regression. 
The best one is chosen as the one giving the lowest AIC value. 
Items with the best AIC value not lower than the AIC value of the null model
(constant model) minus 2 are eliminated. 
Items with the best fit showing a global significant quadratic trend of the 
residuals as a function of the dose (in rank-scale) are also eliminated (the 
best fit is considered as not reliable in such cases). 
Each retained item is classified in a twelve class typology depending of the
chosen model and of its parameter values :

- "H.inc" for increasing Hill curves
- "H.dec" for decreasing Hill curves
- "L.inc" for increasing linear curves
- "L.dec" for decreasing linear curves
- "E.inc.convex" for increasing convex exponential curves
- "E.dec.concave" for decreasing concave exponential curves
- "E.inc.concave" for increasing concave exponential curves
- "E.dec.convex" for decreasing convex exponential curves
- "GP.U" for U-shape Gauss-probit curves
- "GP.bell" for bell-shape Gauss-probit curves
- "lGP.U" for U-shape log-Gauss-probit curves
- "lGP.bell" for bell-shape log-Gauss-probit curves

Larras, ..., ES&T ?