Quantitation,

regression Analysis

13.11.2024 1
 Absolute and relative methods
• In the context of quantitative analysis
(A somewhat simplified point of view)
• Absolute methods: calibration with the analyte unnecessary.
• Relative methods: calibration with the analyte is necessary.
– A reference material is required

13.11.2024 2
 Absolute methods
• Titrimetry
– Acid-base titration: The content of any acid can be determined using a
solution of the same base.
• A sample of the acid is not needed for the analysis.
• No need to prepare standard solutions of that particular acid.

• Gravimetry
– e.g. Gravimetric determination of moisture content (by drying).
• No need to prepare standard solutions with different concentrations.

13.11.2024 3
 Most methods are not absolute!

Most methods require calibration with the

analyte.

13.11.2024 4
 Calibration and addition
• Two major calibration methods:

• The calibration graph method.

– Several solutions with varying concentrations of the analyte are prepared
and used to plot a calibration graph, from which the content of the analyte
in the sample is inferred.
• Standard addition method.
– Fixed amounts of the analyte are added to the sample solution followed by
measurement of the resultant analytical signal.

13.11.2024 5
 Internal standards
• Both, calibration graph method and standard addition method can be
applied either
– In absolute terms.
• Without internal standard
• On the axes of the graph there are concentration and analytical signal.

– With internal standard.

• On the axes of the graph there are ratios of analytical signals vs concentrations.

• We will start with absolute calibration.

13.11.2024 6
 Linear regression
• Both calibration graph and standard addition methods are in most
cases based on linear regression.
• It is a statistical method that allows to fit experimental data points to
a straight line.
• The best fit is found by minimizing the sum of squared residuals.
– Linear regression is one of the so-called least squares methods.

13.11.2024 7
 Linear regression: model
• The linear regression model:
Ai = Ci · b1 + b0

• The analyte concentration CX in the sample solution is found from

the sample solution analytical signal AX:

AX − b0
CX =
b1

13.11.2024 8
 Calibration graph (absolute) obtained by
200
linear regression
180
Slope b1
160 R is the correlation coefficient
Analytical signal, A

140

120 A = 9.4286 C + 11.714

2
100 R = 0.974

A80
x
Residuals
60

40

20 Intercept b0
0
Cx
0 5 10 15 20
13.11.2024 Analyte concentration, C 9
 Slope and Intercept
• Analytical expressions for Slope and Intercept:
A· Ci2 − C · Ci Ai
b0 = i i In matrix notation:
C b0 
2
2
− n·C
b =   = ( X ' X ) X 'Y
i −1
i
 b1 
 Ai Ci − n·A·C https://online.stat.psu.edu/stat462/
b1 = i node/132/

C
2
i
2
− n·C
i

• The regression line always passes the point defined by the averages
of A and C:
A = C  b1 + b0
13.11.2024 10
 Linear regression: assumptions
• Assumption 1:
• linear relationship exists between the analytical signal and analyte
concentration.
• Many analytical methods are highly linear:
– UV-Vis photometry, LC-UV, GC-FID, GC-MS, ICP-MS, potentiometry
(with respect to logC), XRF, molecular fluorescence.

• Some have limited linearity:

– AAS (especially GF-AAS), flame photometry

13.11.2024 11
 Linear regression: assumptions
• Assumption 2:
• uniform scatter of data points throughout the used concentration
range (homoscedasticity)
– In most methods the scatter is larger at higher concentrations.
• At moderate concentrations scatter is often almost proportional to concentration

• In analytical chemistry this usually does not invoke big problems.

– Usually more uncertainty comes from systematic effects than from scatter
of data points.
• If needed, weighted regression can be used.

13.11.2024 12
 Linear regression: assumptions
• Assumption 3:
• The scatter of data points originates from inaccuracy of Ai, while Ci
are very accurate.
– The minimization is carried out in the A axis direction.
– In most cases this holds in analytical chemistry.
– In most cases, if one switches the axes, the regression lines will be very
similar.

13.11.2024 13
 The calibration graph method
• Basic assumption: The analyte in the standard solution influences the
measuring instrument exactly as the analyte in the sample solution.
– This assumption is often reasonably valid.
– Matrix effects can sometimes make it invalid.

• If matrix effects are present, then matrix-matched calibration is often

recommended.
– Calibration solutions are prepared in blank matrix.
– This works fine if the sample matrix is reasonably well defined.
• e.g. XRF analysis in steel industry.

13.11.2024 14
 Calibration curve method: advantages
• The calibration curve need not be linear
– Although it is good if it is
• The calibration curve need not pass 0,0
– Although it is good if it passes
• Usually yields a lower uncertainty than standard addition
– Calibration curve method is interpolative.
– Standard addition method is extrapolative.
• Less work than standard addition.
– Especially if many samples are analyzed.

13.11.2024 15
 The calibration curve method: drawback

• Not applicable when a matrix effect influences the slope of the

calibration curve and this cannot be taken into account by matrix-
matched calibration.

13.11.2024 16
 Calibration curve (absolute) obtained by
200
linear regression
180
Slope b1
160 R is the correlation coefficient
Analytical signal, A

140

120 A = 9.4286 C + 11.714

2
100 R = 0.974
S = 9.9
A80
x
Residuals
60

40

20 Intercept b0
0
Cx
0 5 10 15 20
13.11.2024 Analyte concentration, C 17
 Finding analyte concentration in sample
solution
• The analyte concentration CX in the sample solution is found from
the sample solution analytical signal AX:

AX − b0
CX =
b1
• Important:
– At least one calibration point has to be located below and at least one
above the value of AX

13.11.2024 18
 Scatter
• Scatter of points can be numerically represented by R2 or S of the
regression
• The scatter in the example is exaggerated to be able to demonstrate residuals

• R2 (correlation coefficient) is a unitless quantity

– Different curves can be compared
– With the majority of analysis methods the R2 value should be at least 0.99
• S (residual standard deviation) has the unit of the ordinate axis
– S is the standard deviation of the points around the calibration line
– It is not possible to set any "generally reasonable" S because its value
depends on the ordinate units

http://www.cma4ch.org/chemo/ftp/7004x0993.pdf
13.11.2024 19
 Linearity
• Generally scatter by itself does not show linearity
– A graph with highly scattered points can be linear
• The graph that we just looked at is a good example
– A graph with low scatter can be nonlinear

• Determination of linearity needs analysis of residuals

• There are algorithms for this but for us at the moment visual
assessment is sufficient

13.11.2024 20
 Residuals: Case 1
15

10

Residuals
5
Residuals

0
0 5 10 15 20

-5 Residuals

Residuals are randomly

-10 distributed ➔ linear!

-15
Analyte concentration, C
13.11.2024 21
 Calibration graph: case 2
200

180
y = 9.7212x + 10.849
2
Analytical signal, A 160 R = 0.991

140

120

100

80

60

40

20

0
-2 3 8 13 18 23
13.11.2024 Analyte concentration, C 22
 Residuals: case 2
8

2
Residuals

0
0 5 10 15 20
-2

-4
The actual calibration
-6 function is clearly
nonlinear!
-8

-10
13.11.2024 Analyte concentration, C 23
 Calibration graph: case 2
• At medium and high concentrations the linear approximation works
well

• At low concntrations:
• If the value of the analytical signal is Ax = 9
• We get Cx = -0.19

What is wrong?!?

13.11.2024 24
 60
A
Calibration graph: case 2
The general regression line taking into account all points is fully

40 inadequate at low concentrations!

20 The actual calibration

function at low
Ax concentrations!

0 Cx Cx

13.11.2024
-2 0 3 25
 Case 2: Possible solutions
200

180
y = 9.7212x + 10.849
2
160 R = 0.991
Analytical signal, A
140

120
1. Different regression lines for
100 different parts of the graph
80 2. Interpolation (if the data points
are reasonably accurate)
60
3. Nonlinear regression
40 4. If very high accuracy is not
required then at medium or high
20
concentrations nonlinearity can
0 be ignored
-2 3 8 13 18 23
Analyte concentration, C
13.11.2024 26
 Standard addition method
• Main requirements:
• The calibration curve must be linear
• The calibration curve must pass 0,0
– There requirements are strict – if these are not met significant additional
uncertainty will be introduced
– There is a wide range of methods for which these assumptions are not
applicable

• Advantages:
– Makes it possible to work in a situation where matrix effects yield curves
with different slopes for sample and standard solutions.

13.11.2024 27
 Standard addition
• Drawbacks:
– Not applicable when the calibration graph is nonlinear or the intercept is
not zero.
– Not applicable when some matrix effect influences the intercept of the
calibration plot.
– More work-intensive than the calibration graph approach.
– Due to extrapolative nature uncertainty is higher than for calibration graph
method.

13.11.2024 28
 Absolute calibration and the internal standard method

• We have been discussing the application of absolute calibration.

• Often, relative calibrations are favored, ie relative to an internal
standard.
– Abscissa – analyte to IS concentration ratio.
– Ordinate – analyte to IS signal ratio.
• Makes it possible to account for
– Inaccuracy in sample introduction.
– Systematic effects in sample preparation.

13.11.2024 29