Title: Identification, Measurement and Decision in Analytical Chemistry
1Identification, Measurement and Decision in
Analytical Chemistry
- Steve EllisonLGC, England
2Introduction
- What is Identification?
- Why does it matter?
- Where does measurement fit in?
- Quality in identification
- How sure are you?
- characterising uncertainty and method performance
3What is identification?
- Classification according to specific criteria
- Above or Below a limit
- Within Spec.
- Red
- Classification into ranges (lt2 2-5 5-10 gt10)
- Molecular species by NMR, IR, MS..
- Material or ingredient (Rubber, Fat)
- Origin or authenticity
4Why does it matter?
- Classification contributes to decisions
- Decisions cost money
- Incorrect batch rejection incurs reprocessing
costs - Incorrect acceptance risks litigation and loses
business - False positives may generate spurious
prosecutions - Costs are directly related to false
classification probabilities - Know probabilities - optimise cost
5Where does measurement fit in?
- Measurement contributes to most identifications
- Comparison with limits
- Consistency of values (wavelength, mass, sequence
length) - But not all
- Relative Pattern identification (?)
- Colour matching by eye
- Identity parades.
6Interpretation Against Limits
Measurement result
Limit
(a)
(b)
(c)
(d)
(e)
(f)
(g)
7Controlling Identification
- Good practice guidance
- Stated criteria
- Trained staff
- Controlled and calibrated instruments
- Traceability!
- Validated methods
- .. etc
8How sure are you?
- Does Measurement Uncertainty apply?
- If not, what does?
9Does Measurement Uncertainty Apply?
- NO
- at least, not for the classification result
10Uncertainty and classification
Uncertainty in the measurement result contributes
to uncertainty about classification
11Uncertainty and classification
- Uncertainty in the measurement result contributes
to uncertainty about classification - Uncertainties in test conditions lead to
uncertainty in classification - Uncertainties should be controlled to have little
effect on the test result
12Characterising uncertainty in identification
- False response rates
- What is a false response rate?
- How is it determined?
- Alternative expressions of method performance or
uncertainty
13False response rates
Observed
Actual
14False response rates
Observed
Actual
15False negative rates
- Fraction of observed negatives which are false
- Fraction of true positives reading negative
- Fraction of all results which are incorrectly
read as negative
16False negative rates
- Fraction of observed negatives which are false
- Fraction of true positives reading negative
- Fraction of all results which are incorrectly
read as negative
AOAC Definition(clinical)
17False negative rates
- Fraction of observed negatives which are false
- Fraction of true positives reading negative
- Fraction of all results which are incorrectly
read as negative
AOAC Definition(clinical)
The one that affects costs directly
18False response rates Example
Observed
Actual
Nortestosterone in urine screening method
validation data.Actual rates from confirmatory
test
19False negative rates Example
- Fraction of observed negatives which are false
- 2/17 11
- Fraction of true positives reading negative
- 2/16 13
- Fraction of all results which are incorrectly
read as negative - 2/32 6
20False response rates Example
Observed
Actual
EMIT test for cocaine in urine Ferrara et al,
J. Anal. Toxicol., 1994, 18, 278
21False negative rates Example
- Fraction of observed negatives which are false
- 7/429 1.6
- Fraction of true positives reading negative
- 7/126 5.6
- Fraction of all results which are incorrectly
read as negative - 7/522 1.3
22False response rates - how much data?
- Observed 7/126 (5.6)
- 95 confidence interval (binomial)
- 1.6 to 9.5
- 95 CI proportional to 1/?nobs
- needs a LOT of false responses for precise
figures - but false responses are rare for good methods.
- Most useful direct studies are worst case or
near 50 false response levels
23False responses Estimation from thresholds
24False responses From probabilities
- Spectroscopic identification study
- S.L.R. Ellison, S.L. Gregory, Anal. Chim. Acta.,
1998 370 181. - Calculated chance FT-IR match probabilities
- probabilities based on match-binning - hits
within set distance - required hypergeometric distribution (n matches
of m taken from population) - Compared with actual hits on IR database
25False responses From probabilities
- Theoretical predictions very sensitive to
probability assumptions - 10 changes in p make large differences in
predictions - Best performance within factor of 3-10
- (Improved over binomial probabilities by gt106)
- Probability information must be excellent for
good predictions
26False response rates from databases
- Most spectral databases contain 1 of each
material - most populations do not!
- Population data must account for sub-populations
- cf. DNA profiles for racially inhomogeneous
populations
27Alternative performance indicators
28Conclusions
- Classification needs control to save money
- INPUT uncertainties need control
- False response rates and derived measures are
useful performance indicators - Definitions vary and make a (big) difference
- Sufficient data are hard to get except for
carefully chosen analyte levels - Databases suspect unless built for the purpose
- Theoretical predictions usable with great care
- Unwise to expect precise numbers!
29Best practice
- Consider costs of false responses
- Control qualitative test conditions via traceable
calibration of equipment - Check most critical false response rate
- preferably both
- Use worst-case and likely interferent studies
to show limits of method performance - Use APPROPRIATE population data
- Report with caution
- particularly on probability estimates
30(No Transcript)