Example: Nonlinear Standard Addition Curves
This is the standard addition data, collected using atomic absorption (a technique notorious for nonlinear
calibration curves).
x 0.00 0.10 0.20 0.30 0.40 ( )
T
volume (mL) of added standard
y 0.238 0.445 0.622 0.776 0.895 ( )
T
measured absorbance peak areas
V
a
25.00 C
std
10.0 n 5
vol (mL) of aliquots conc (ppm) of standard spikes number of data pts
We must correct for the effects of dilution (due to the std spike) on the signal:
i 0 n 1 .. y
corr
i
y
i
V
a
x
i
V
a
.
y
corr
T
0.2380 0.4468 0.6270 0.7853 0.9093 =
(MathCAD needs subscript)
corrected for dilution due to spikes
Now let's look at the standard addition plot, along with the best-fit line of the data.
b
0
intercept x y
corr
, b
0
0.2650 = b
1
slope x y
corr
, b
1
1.6812 =
Let's calculate the std deviation of the residuals for model comparison purposes
y
hat
b
1
x
.
b
0
res y
corr
y
hat
n 2 s
res
1
i
res
i
2
.
s
res
0.0296 =
degrees of freedom
residuals
Standard Additions Plot
y = 1.6812x + 0.265
R
2
= 0.9908
0
0.2
0.4
0.6
0.8
1
0 0.1 0.2 0.3 0.4
Spike Volume
S
i
g
n
a
l
-0.04
-0.02
0
0.02
0.04
0 0.1 0.2 0.3 0.4
The standard additions plot is
shown here with a best-fit line.
The inset plot shows the
residuals. There definitely
seems to be some structure to
the plot (ie, the residuals are
nonrandom), suggesting that a
linear regression model is not
appropriate.
So let's try a second-order
polynomial fit of the std add'n
data.
1
Let's do a second-order fit of the data, then calculate the std dev of the residuals
f x ( ) 1 x x
2
T
b linfit x y
corr
, f , b
T
0.2377 2.2280 1.3672 =
(defines regression
model in MathCAD)
determine LS estimates
(2nd-order fit)
LS estimates of polynomial coefficients
y
hat
i
b
2
x
i
2
.
b
1
x
i
.
b
0
res y
corr
y
hat
n 3 s
res
1
i
res
i
2
.
s
res
2.0690 10
3
=
degrees of freedom
This is better than
the best-fit line.
Here is the data again, along with the LS quadratic fit.
Standard Additions Plot
y = -1.3672x
2
+ 2.228x + 0.2377
R
2
= 1
0
0.2
0.4
0.6
0.8
1
0 0.1 0.2 0.3 0.4
Spike Volume
S
i
g
n
a
l
-0.004
-0.002
0
0.002
0.004
0 0.1 0.2 0.3 0.4
This time, the fit looks better. The residuals are more randomly distributed about the fit, and the s
res
statistic also
favors the quadratic model. The problem we run into now is this: the standard additions method for analyte
quantitation assumes a linear relationship between signal and analyte concentration. To be more precise, the
method assumes that the calibration function is linear when extrapolated to the baseline. The following figures
illustrate this point.
2
volume added standard
linear
extrapolation
assumed
linear
extrapolation
assumed
volume added standard
linearity verified for this range not linear!
V
/
V
/
As mentioned, in order for the standard addition equations to work, extrapolation to the baseline must be linear.
This assumption is more justifiable if the calibration function is linear for the standard addition data (eg, the plot on
the left). If the standard addition data are nonlinear (above right) linear extrapolation might still be appropriate;
however, the evidence to support this assumption is much weaker.
The calibration function for atomic absorption is usually linear for small absorbances, showing curvature toward the
baseline as the concentration increases - exactly the behavior demonstrated by the data in this problem. Thus, we
may claim with some legitimacy that, even though our standard addition data is slightly nonlinear, the extrapolation
to the baseline is linear, since this encompasses smaller absorbance values. In other words, the response was still
(at least approximately) linear for the analyte present in the sample.
With this assumption, we can now extrapolate to the baseline.
V_prime
b
0
b
1
V_prime 0.1067 = volume (in mL) of added standard that contains the same amount
of analyte as was originally present in the aliquot (before spiking)
C
a
C
std
V_prime
V
a
.
C
a
0.0427 = analyte concentration (in ppm) in original aliquot
Since 1ppm is assumed equivalent to 1 g/mL, we can now determine the "extractable" concentration of Cd in the
original soil sample
mass
Cd
C
a
100
.
mass
Cd
4.2674 = mass (in g) of Cd extracted from the soil sample
C
soil
mass
Cd
1.2901
C
soil
3.3078 = concentration (in ppm) of extractable cadmium in the original soil sample
This is a point estimate. Let's calculate the standard error in this estimate.
x
bar
mean x ( ) x
bar
0.2000 = S
xx
4 Var x ( )
.
S
xx
0.1000 =
s
V
s
res
b
1
1
1
5
V_prime x
bar
2
S
xx
.
s
V
1.3586 10
3
= std error in V_prime
3
s
Ca
s
V
C
std
V
a
.
s
Ca
5.4344 10
4
= s
Csoil
s
Ca
100
1.2901
.
s
Csoil
0.0421 =
std error in Ca std error (in ppm) of
point estimate
t qt .975 2 , ( ) t 4.3027 = t s
Csoil
.
0.1812 =
t-crit with 2 df width of 95% confidence interval
The concentration of extractable cadmium in the soil sample is 3.31 0.18 ppm (95% CL)
4