Once again!!: How long will you trust in Chemometrics?
and their methods (PLS, PLSR, PLS2, PCA, PCR, FA, MCA, etc.), as long as all the 'Teachers' are not able to really
validate your particular analysis, every day and every hour, in a way you can believe in, and you are able to communicate
to others !!!
All Prophets with all their magic numbers (RSS, SEE, PRESS, MCC, MSEP, RMSEC, RMSEP, SED, RMSECV, EIV,
and quite a few more) have NEVER the power to HIDE the biggest, disqualifying lack in this discipline!
P. Forster Jun 2005/Oct 1992. |
| Introductions: |
Probably (I think) the most important
part about Chemometrics at all! |
You may have already realized, that I call my own a very special decided relationship to all what is, or
maybe will be called Chemometrics.
But I'm absolutely sure, if you are interested in a "Real Believable Chemometrics", in which you can not
only trust during your daily routine "Analytics/Work" (Quality Control, Kettle charge, CU, DR, etc.), you soon
will participate on this, my decided relationship, too.
|
The First golden Rule of Chemometrics:
MOST IMPORTANT FOR CELEBRATING CHEMOMETRICS IS,
— to be a very good Analytical Chemist.
WHY? : It's easy to play with a bidirectional gun, but hard to survive!
And your gun has ONE trigger for every finger of your hand - at least!
For understanding we must start to think!:
| |
A. ) What every Chemometrics MUST DO for you, to survive!
B. ) What today Chemometrics CAN DO for you, to survive!
C. ) Compare both results, and you will easily identify your gun!
D. ) What requirements your data and 'you' have to fulfil!
BUT: Don't fear, there is a Way out of the Dark.
- And if you are really will, I go this Way with you!
(Who's wind I sowed — whose storm I gather !) |
Don't fear, there is a way out of the Dark. |
|
|
Let's start directly in medias res:
Our absolute first task must be to answer the four questions above with great satisfaction, otherwise we have
already pulled all triggers simultaneously. As we form, at least together, a good modern Analytical Chemist, it should
not be very difficult to do so. If ever we may be called "very" good modern Analytical Chemists, that has to be verified
and validated later.
A. ) What every Chemometrics MUST DO for you, to survive!
As first you tell me, that you must have 'your' concentration result or 'your' quality control result from 'your'
chemometrics method! So I have to say, that's wrong, because this, 'your results' are the absolutely last important
results 'ALL' chemometrics methods have to give. But let's count it anyway, but on their correct Priority level.
Before it is - ever - allowed to take not even the first view on 'your' desired results, YOU HAVE TO TEST,
IF it is allowed to take use of 'your' desired results in any other way, as to erase them!
I know it sounds strength, but as good modern Analytical Chemist you know, it's the only true!
To accomplish this absolutely indispensable goal, Chemometrics has to do a few
things for you!
For the first moment, I will only list it, and discuss them in detail a little bit later:
- You (of course, 'your' Chemometrics) has to be evidence for, that it is allowed,
to analyse your particular sample with this particular 'method' (calibration), every day, and every hour!
This is called 'SMV': Sample Method Validation !!
— And that CANNOT be done in the "Mean Centred Space!" — It's for humans sense, too!
— And that CANNOT be done in the "Std.Dev. Scaled Space!" — It's for humans sense, too!
- As long as ever possible (and it is), Chemometrics must show you, how the 'numerical uniformity' of your
estimated results is. This we call 'SURE': Statistical Uniformity of Result
Estimation !!
- Chemometrics must give you the right, correct, reliable, handable tools to validate
your estimation results:
Have you ever thought about:
- What's the real value of the "hat matrix", which is NOT VALIDATED ?
- What's the real value of the "Leverage", based on a NOT VALIDATED "hat matrix" ?
- What's the real value of the "Old EIV", based —
on a NOT VALIDATED "Leverage", based on a NOT ... ?
- What's the real value of the "New EIV", based —
on a NOT VALIDATED "Leverage", based on a NOT ... ?
- What's the real value of the "expanded SEP", based —
on a NOT VALIDATED "Leverage", based on a NOT ... ?
- What's the real value of the "SSSEP", based —
on a NOT VALIDATED "Leverage", based on a NOT ... ?
- What is the "Uncertainty" meaning, based —
on a NOT VALIDATED "Leverage", based on a NOT ... ?
- Wh.... .....
- Wha... ..... .....
- What'. ..... ..... .....
- Apropos: Have you ever seen a VALIDATED "hat matrix" If, Please let me know!!
[*]
|
- Chemometrics must give you the correct tools to have all time the best
possible "Overview" about your data and estimation in every state!
- Chemometrics must give you reliable, more reliable procedures to handle extraordinary results!
- Chemometrics must give you when ever possible (it is) a better/clearer and more descriptive estimation of factors
(Principal Components)!
- Chemometrics must give you at least for spectra the possibility (it is) to rotate the factors back into real world
spectra!
- 'SMV', 'SURE', , , and so on, must be easily 'interpretable'/'controllable'
for everybody in a unitary manner !!
- At last, when all previous statements are passed with flying colours, we come to 'your'
longing asked results for concentration or quality parameters or so.
This List is preferred to be expanded!!
B. ) What today Chemometrics CAN DO for you, to survive!
I give this List B.) into your custody to be completed in the same manner as I did it with List A.).
Don't be shy. Even when it may be hard to do, - try to be as precise as possible.
C. ) Compare both results, and you will easily identify your gun!
If you finish List B.) with success, what I hope, then you have already identified your gun/your triggers too!
Yes, — And that's good this way!!
D. ) What requirements your data and 'you' have to fulfil!
There is an older german wisdom/proverb:
"Wer misst, misst Mist.
Was weiter nicht schlimm ist,
solange er sich bewusst ist,
was für ein Mist er misst."
Addendum:
"Wer viel misst, mistet viel."
|
Wer -> Who, misst -> measures, Mist -> crap/rubbish,
Was -> What, weiter -> in additon, nicht -> not,
schlimm -> worse/bad, ist -> is, solange -> as long as,
er -> he, bewusst ist -> is aware,
für -> for, ein Mist -> a kind of crap/rubbish,
viel -> a lot, mistet viel -> produce a lot of 'Mist'.
|
What have we to learn from this proverb?: (It's also true for other MVDA!)
- What ever we measure, it will contain some errors:
- We must know as much as possible about our measurement errors.
- We must reduce our errors as much as ever possible.
- Think!: All error in measurements will recover in the estimation results, only in higher order.
- Try never to get a better precision as the errors you have measured.
|
- We must know as much as ever possible about our measurements:
- To ensure a correct and reliable validation of the estimation.
- To optimise our measurements in case it will be required.
- To select the most adequate measurements for estimation.
|
- We must know as much as ever possible about our possible measurements:
- To build the best, most adequate observation model.
- Think!: What is not measured, cannot be estimated.
- But!: A lot of measured garbage results in a lot of estimated garbage.
- It's easy to shrink the underlying data model at the right moment.
|
- Take very much care to your data and their use:
- Do the right thing at the right moment.
- Do for the human, what for the human is, it's a lot.
- Do for the computer, what for the computer is.
- Take special care to what for the human is!
|
Let's summarise our "new" knowledge in a couple of Thesis, valid not only for spectra:
- Thesis: Reliable estimation (results) requires an absolute reliable validation.
- Thesis: Reliable estimation (results) based only on exact "modeling".
- Thesis: Only a reliable Selection also results in a reliable estimation (results).
- Thesis: What is not measured, cannot be estimated. But you can 'all' estimate - or:
- Thesis: Measured garbage results in a lot of estimated garbage.
- Thesis: Only reliable measurements can result in a reliable estimation.
- Thesis: There are no "outliers", except you have an absolute valid evidence for.
- Thesis: Ask never for a better precision as the errors you have measured.
- Thesis: Do for human what for the human is, do for computer what for it is. But, do it!
- Thesis: The real world is a little bit more decorated for 'good modern spectroscopists'.
- Thesis: A Data Analysis needs to pas an "Adequateness Test" before its use"!
- Thesis: A Calibration needs to pas a "Validation for its future use" Test on Samples!
- Thesis: For unexceptional all future Samples, you need an adequate SMV!
- Thesis: Never trust results, they look like the once you prefer! - Validat twice!
SMV: "Sample Method Validation"
|
Don't try to use these Theses as a Checklist for MVA-Programs/-Methods:
It's on your own Risk!
|
Will we bring these Theses to real Live?:
(But be aware, that not many chemometricans, statisticians will prefer the content of
this chapter. The same is for sure also true for some organisations.
Don't be surprised, when quite several muskets will not only be loaded!) |
An other note, too:
Our discussion will not strictly follow the presented list, because some Theses are more or less related. On the other
hand it's also not the best idea to use for explanation a defined example, as our discussion will influence our selection
of the required tools and our methodic steps, too.
How many examples I will add to our conversation, I have not decided yet.
|
|
|
Home
