| Author | Message | ||
     Pierre Dardenne (dardenne) Senior Member Username: dardenne Post Number: 58 Registered: 3-2002 |
Hi, one more idea. RPD is used for calibration development. In calibration monitoring, only RMSEP is valid. Low errors whatever the range is. Or do use SEP/SD_calibration_set and not SEP/SD_val. Pierre | ||
| Fernando Morgado (fmorgado) Intermediate Member Username: fmorgado Post Number: 20 Registered: 12-2005 |
Dear All: Thanks for your opinion about RPD. I read several publications and realy RPD seems only usefull for comparate models maked with diferent preprocesing for the same database. Is logig SD is a constant here. The formula Ya look interesting , I will meake some calculations with that. Thanks again | ||
| Daniel Alomar (dalomar) Junior Member Username: dalomar Post Number: 7 Registered: 2-2009 |
Hi Fernando, RPD is dependent of SD and a problem may arise when a calibration set is "forced" to include a wider set of values than you can expect in normal operation. This is even more so when you use RER (range/secv) which is improved if a few extreme, "far" values are included. In this way a better equation can be obtained, but, will it represent most of your future predictions? Other term I find useful is CV (secv/average). In your example, assuming an average of, say 55%, your CV should be near 0.022, which looks better to me than your RPD, although that again is dependent on how much error can be accepted in your application. Greetings! Daniel | ||
| Bob Jordan (jordan) New member Username: jordan Post Number: 4 Registered: 3-2003 |
Fernando, You need to note that RPD is a simple function of R^2, or in fact adjusted R^2 so is no more dependent on the SD than is R^2adj. In fact 1/RPD=sqrt(1-R^2adj) It came about as a rough measure of how many 'categories' you might sort a population into. So if you were sorting fruit for total sugars and the NIR calibration R^2adj was 75% then the RPD would be 2. Moving up to 89% in R^2adj space gives you an RPD of 3. It is as simple as that. Personally I prefer to use the term that is the inverse of RPD and we call this Ya (the reverse Cyrillic R character) such that Ya^2=1-R^2adj. If you use the correct character you will understand the choice. This makes Ya=SE/SD. Ya behaves more 'linearly' compared to RPD which has a parabolic form. So an RPD of 1.76 corresponds to an R^2adj of 67%. This value comes from a relationship that is definitely highly significant but is at the low end of usefulness - BUT usefulness will depend on the application. As they say - Your mileage may vary Bob J. | ||
| Fernando Morgado (fmorgado) Intermediate Member Username: fmorgado Post Number: 19 Registered: 12-2005 |
Dear Sirs : RPD is the standard desviation of the calibration samples divided for SECV. In some place I read RPD values more bigger than 3 are nice models. But, RPD calculate always have a strong dependency to the standard desviation. For example at this moment I have a model with 46 samples ( range 49% to 61%) and SECV calculate using the correct factor numbers is 1.238 with high R ( 0.944). The standard desviation of the chemical data used for created the model is 2.19. If I calculate RPD the result is 1.76 Can I say this Model is not nice for to have a RPD below 3? ( I am not confused with H value, dont worry) I think the RPD only is valid if the database have some standard desviation or some number of samples. I hope you understand my qustion. Nice weekend to all. Fernando |