Log in
Register|
Geokad ☆ Canada, 2008-10-03 19:19 (6472 d 19:29 ago) Posting: # 2474 Views: 4,512 |
|
|
Hello, If you please, can anyone give me some idea/criteria, on how to choose the best regression algorithm? — Regards, Geokad |
|
martin ★★ Austria, 2008-10-03 21:23 (6472 d 17:25 ago) @ Geokad Posting: # 2475 Views: 3,750 |
|
|
Dear geokad ! Statistical regression analysis is a very broad topic (non-linear fitting, multiple regression, ... ). there is no general best or general worst method. I may can give some suggestions in the case that you can specify your query in more detail as your questions sounds like “how can I earn some money?” best regards martin |
|
Helmut ★★★   Vienna, Austria, 2008-10-04 02:09 (6472 d 12:39 ago) @ Geokad Posting: # 2477 Views: 3,863 |
|
|
Dear Geokad, a few comments adding to Martin's. Since you posted in the 'Bioanalytics' category, I guess you are rather interested in calibration, which in specialized statistical textbooks (e.g. Draper & Smith) is called 'inverse regression'. Apart from statistical tools (e.g., running an F-test on the 'full model' = quadratic vs the 'reduced model' = linear) mostly the distribution of residuals (no trend, no funnel shape, etc.) and accuracy/precision of back-calculated calibrators and QC-samples are used. Cave: although often seen, the coefficient of determination R2 or the coefficient of correlation R are useless in model selection.  — Dif-tor heh smusma 🖖🏼 Довге життя Україна! ![[image]](https://static.bebac.at/pics/Blue_and_yellow_ribbon_UA.png) Helmut Schütz ![[image]](https://static.bebac.at/img/CC by.png) The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
|
Dr. Harish L. Rao ☆ India, 2008-10-04 06:28 (6472 d 08:20 ago) @ Geokad Posting: # 2478 Views: 3,726 |
|
|
Dear Geokad, In a simple approach, for each algorithm, you need to calculate the Sum of the Absolute Difference (ignoring the sign) between 100 and the % Accuracy for each Calibration Standard. The algorithm which yields the Lowest Sum is the Best Regression Algorithm. Best regards, Dr. Harish L. Rao |
|
Aceto81 ★ Belgium, 2008-10-06 11:52 (6470 d 02:56 ago) @ Dr. Harish L. Rao Posting: # 2485 Views: 3,714 |
|
|
❝ In a simple approach, for each algorithm, you need to calculate the Sum of the Absolute Difference (ignoring the sign) between 100 and the % Accuracy for each Calibration Standard. The algorithm which yields the Lowest Sum is the Best Regression Algorithm. I do not agree with this approach (at least not without a few modifications). If you got 10 points to calibrate on, and use a 9th order polynomal, you get a sum of absolute difference of 0, perfect fit, but this an overfitted model, as you also fit the error. I think the above approach is usefull, but also keep in mind that the algorith with the minimum predictor which gives you an acceptable result is prefferable. There are also AIC, BIC, ... criteria to make your selection, but I think it's usefull to get a good book on regression and model selection Ace |
|
vpardhasaradhi ☆ 2008-10-08 20:10 (6467 d 18:38 ago) @ Geokad Posting: # 2495 Views: 3,701 |
|
|
Dear Geokad, Please see if the following article helps you: Linear Regression for Calibration Lines Revisited: Weighting schemes for bioanalytical methods by A.M.Almeida, M.M.Castel-Branco and A.C.Falcao in Journal of Chromatography B. Volume 774, Issue 2, 15 July 2002, Pages 215 - 222. With best regards, V.Pardhasaradhi |
Ing. Helmut Schütz