Title: Nonlinear Regression Analysis with Fitter Software Application
1Non-linear Regression Analysiswith Fitter
Software Application
Alexey Pomerantsev Semenov Institute of Chemical
PhysicsRussian Chemometrics Society
2Agenda
- Introduction
- TGA Example
- NLR Basics
- Multicollinearity
- Prediction
- Testing
- Bayesian Estimation
- Conclusions
31. Introduction
4Linear and Non-linear Regressions
2
Close relatives?
52. Thermo Gravimetric Analysis Example
Lets see it!
6TGA Experiment and Data
TGA Experiment
TGA Data
7TGA Example Variables
Small sizeproblem!
8Plasticizer Evaporation Model
Diffusion isnot relevant!
9Fitter Worksheet for TGA Example
10Service Life Prediction by TGA Data
113. NLR Basics
12Data and Errors
Weight isan effectiveinstrument!
13Model f(x,a)
Presentation at worksheet
14Data Model Prepared for Fitter
Apply Fitter!
15Objective Function Q(a)
Objective function Qis a sum of squaresand may
be more
Parameter estimates
Weighted variance estimate
16Very Important Matrix A
Matrix A is the cause of troubles..
17Quality of Estimation
Matrix A is the measure of quality!
18Search by Gradient Method
Matrix A is the key to search!
194. Multicollinearity
20Multicollinearity View
Multicollinearity is degradation of matrix A
Objective function Q(a)
1
N(A)
2
4
5
6
7
21Multicollinearity Source
22Data Model Preprocessing
((a b) c) d ? a (b (c d)) as
110 20 1
23Example The Arrhenius Law
24Derivative Calculation and Precision
255. Prediction
26Reliable Prediction
Forecast shouldinclude uncertainties!
27Nonlinearity and Simulation
Non-linear models callfor special methods
ofreliable prediction!
28Prediction Example
Model ofrubber aging
Accelerated aging tests
Upper confidence limits
296. Testing
30Hypotheses Testing
Test statistics x is compared with critical
value t (a)
Test dont prove a model! It just shows that
the hypothesis is accepted or rejected!
31Lack-of-Fit and Variances Tests
These hypotheses are based on variances and they
cant be tested without replicas!
Lack-of-Fitis a wily test!
32Outlier and Series Tests
These hypotheses are based on residuals and they
can be tested without replicas
Series test isvery sensitive!
337. Bayesian Estimation
34Bayesian Estimation
How to eat awayan elephant?Slice by slice!
35Posterior and Prior Information. Type I
The same error ineach portion of data!
36Posterior and Prior Information. Type II
Different errors in each portion of data!
378. Conclusions
Mysterious Nature
LR Model
NLR Model
Thankyou!