5.2 Analytical Data Evaluation: Errors, Uncertainty, Statistics, and Calibration

Analytical Data Evaluation: Errors, Uncertainty, Statistics, and Calibration

Analytical chemistry provides qualitative and quantitative chemical information. However, no quantitative measurement is physically meaningful without an assessment of its experimental error and uncertainty. For the PRC Chemist Licensure Examination, mastery of analytical data evaluation requires differentiating error classifications, applying hypothesis testing, propagating measurement uncertainties, and selecting appropriate calibration models.

---

1. Classification of Analytical Errors: Determinate vs. Indeterminate

Every analytical measurement contains errors arising from instruments, chemical reagents, method limitations, or operational execution.

Subtypes of Determinate Errors

1. Instrumental Errors: Caused by imperfections in measuring devices, uncalibrated glassware (volumetric flasks, pipets), electrical drift in spectrophotometers, or worn mechanical components.
2. Methodic Errors: The most severe and difficult errors to detect. They stem from non-ideal physical or chemical behavior of analytical systems, including incomplete precipitation, coprecipitation of impurities, reagent decomposition, side reactions, or non-stoichiometric reactions.
3. Personal (Operative) Errors: Result from personal judgment limitations of the analyst, such as consistently reading a buret meniscus from an angle (parallax error), overshooting visual titration endpoints, or improper mechanical sample transfer.

Constant vs. Proportional Determinate Errors

• Constant Errors: The absolute magnitude of the error remains independent of the sample mass or analyte concentration analyzed. As sample size increases, the relative error decreases inversely ($E_r \propto 1/w$). A classic example is a constant excess of titrant ($0.05\text{ mL}$) required to trigger an indicator color change.
• Proportional Errors: The absolute magnitude of the error scales directly with the sample size analyzed, keeping the relative error invariant. A common example is the presence of an interfering contaminant in a standardized titrant or reagent whose consumption is proportional to sample volume.

Diagnostic Test: To diagnose whether a systematic error is constant or proportional, run replicate determinations across varying sample masses ($0.1\text{ g}, 0.2\text{ g}, 0.5\text{ g}, 1.0\text{ g}$). If the calculated percent analyte varies with sample weight, the error is constant; if the percent analyte remains identical but systematically incorrect, the error is proportional.

Detection and Elimination of Determinate Errors

• Standard Reference Materials (SRMs / CRMs): Certified reference matrices analyzed alongside unknowns to detect recovery biases.
• Independent Analytical Methods: Validating a titration result using an orthogonal technique, such as inductively coupled plasma (ICP) spectrometry or gravimetry.
• Blank Determinations:
  • Reagent Blank: Contains solvent and all reagents except the analyte, taken through every preparatory step. Subtracting the reagent blank corrects for background impurities.
  • Matrix Blank: Simulates the entire matrix of the sample without the analyte, accounting for matrix absorption or concomitant ion interferences.

---

2. Precision, Accuracy, and Fundamental Statistics

• Accuracy: The degree of agreement between an experimental result (or the mean of replicate measurements) and the true or accepted reference value ($x_t$ or $\mu$).
• Precision: The degree of mutual agreement or reproducibility among independent measurements obtained under specified identical conditions.

Mathematical Metrics of Central Tendency and Dispersion

For a finite sample dataset containing $n$ replicate observations ($x_1, x_2, \dots, x_n$):

• Sample Mean ($\bar{x}$): $\bar{x} = \frac{\sum_{i=1}^n x_i}{n}$
• Median: The central value of an array of data ranked in ascending order. For an odd $n$, it is the middle observation; for an even $n$, it is the arithmetic average of the two central observations. The median provides a robust measure of central tendency resistant to extreme outliers.
• Sample Standard Deviation ($s$): $s = \sqrt{\frac{\sum_{i=1}^n (x_i - \bar{x})^2}{n - 1}}$ where $(n - 1)$ represents the degrees of freedom ($\nu$). The divisor $(n - 1)$ provides an unbiased estimator of the population standard deviation (Bessel's correction).
• Sample Variance ($s^2$): The square of the standard deviation, representing the additive property of dispersion.
• Relative Standard Deviation (%RSD / Coefficient of Variation, $CV$): $\%RSD = CV = \frac{s}{\bar{x}} \times 100\%$ For parts-per-thousand precision reporting: $\text{RSD}_{ppt} = \frac{s}{\bar{x}} \times 1000\text{ ppt}$

Population vs. Sample Statistics

When the number of observations approaches infinity ($N \rightarrow \infty$), the population mean is denoted as $\mu$ and the population standard deviation is: $\sigma = \sqrt{\frac{\sum_{i=1}^N (x_i - \mu)^2}{N}}$ In a standardized normal (Gaussian) distribution, the probability of an observation falling within specified standard deviation intervals is:

• $\mu \pm 1\sigma$: contains $68.27\%$ of all observations.
• $\mu \pm 2\sigma$: contains $95.45\%$ of all observations.
• $\mu \pm 3\sigma$: contains $99.73\%$ of all observations.

Confidence Intervals ($CI$)

The confidence interval defines the range within which the true population mean $\mu$ is expected to lie with a specified probability ($1 - \alpha$):

• When $\sigma$ is known (large sample limit, $N > 30$): $CI = \bar{x} \pm \frac{z \sigma}{\sqrt{n}}$ where $z = 1.960$ for $95\%$ confidence, and $z = 2.576$ for $99\%$ confidence.
• When $\sigma$ is unknown (small experimental datasets, $n < 30$): $CI = \bar{x} \pm \frac{t s}{\sqrt{n}}$ where $t$ is Student's $t$-statistic evaluated at $\nu = n - 1$ degrees of freedom at the chosen confidence level (e.g., $95\%$).

Worked Example: Statistical Evaluation of Titrimetric Data

An analyst performs 5 replicate titrations of a standardized alkali solution, obtaining volumes: $25.12, 25.18, 25.15, 25.22$, and $25.13\text{ mL}$.

1. Mean ($\bar{x}$): $\bar{x} = \frac{25.12 + 25.18 + 25.15 + 25.22 + 25.13}{5} = \frac{125.80}{5} = 25.160\text{ mL}$
2. Deviations $(x_i - \bar{x})$ and Squared Deviations:
   • $(25.12 - 25.160)^2 = (-0.040)^2 = 0.001600$
   • $(25.18 - 25.160)^2 = (+0.020)^2 = 0.000400$
   • $(25.15 - 25.160)^2 = (-0.010)^2 = 0.000100$
   • $(25.22 - 25.160)^2 = (+0.060)^2 = 0.003600$
   • $(25.13 - 25.160)^2 = (-0.030)^2 = 0.000900$
   • $\sum (x_i - \bar{x})^2 = 0.006600\text{ mL}^2$
3. Sample Standard Deviation ($s$): $s = \sqrt{\frac{0.006600}{5 - 1}} = \sqrt{\frac{0.006600}{4}} = \sqrt{0.001650} = 0.04062\text{ mL} \approx 0.041\text{ mL}$
4. Relative Standard Deviation (%RSD): $\%RSD = \frac{0.04062}{25.160} \times 100\% = 0.161\%$
5. $95\%$ Confidence Interval (using $t_{0.05, 4} = 2.776$): $CI = 25.160 \pm \frac{2.776 \times 0.04062}{\sqrt{5}} = 25.160 \pm \frac{0.11276}{2.2361} = 25.160 \pm 0.050\text{ mL}$ The true mean lies in the range $[25.110\text{ mL}, 25.210\text{ mL}]$ at the $95\%$ confidence level.

---

3. Significance Hypothesis Testing and Outlier Rejection

Significance testing assesses whether an experimental discrepancy between two numbers reflects genuine physical divergence (determinate bias) or expected random fluctuation (indeterminate error).

Case 1: Comparing an Experimental Mean to a Known True Value (Bias Test)

To evaluate whether a new method possesses systematic bias relative to a certified reference material value ($\mu$): $t_{calc} = \frac{|\bar{x} - \mu|\sqrt{n}}{s}$

• If $t_{calc} > t_{crit}$ (from $t$-table at $\nu = n - 1$), reject the null hypothesis ($H_0: \bar{x} = \mu$). A statistically significant determinate error is present.
• If $t_{calc} \le t_{crit}$, retain $H_0$. No significant bias is detected at the chosen confidence level.

Case 2: Comparing Two Experimental Means (Independent Datasets)

To test whether two independent analytical procedures yield equivalent mean results ($\bar{x}_1$ vs $\bar{x}_2$): First, verify that the two sample variances do not differ significantly via the $F$-test. If pooled precision is valid, calculate the pooled standard deviation ($s_{pooled}$): $s_{pooled} = \sqrt{\frac{(n_1 - 1)s_1^2 + (n_2 - 1)s_2^2}{n_1 + n_2 - 2}}$ Calculate $t_{calc}$: $t_{calc} = \frac{|\bar{x}_1 - \bar{x}_2|}{s_{pooled}}\sqrt{\frac{n_1 n_2}{n_1 + n_2}}$ Compare $t_{calc}$ against $t_{crit}$ with $\nu = n_1 + n_2 - 2$ degrees of freedom.

Case 3: Paired Student's $t$-Test

When two analytical methods are compared by analyzing multiple distinct sample matrices (e.g., comparing a new HPLC method with a standard GC method across 8 different blood plasma samples), individual sample differences must be isolated:

1. Compute the difference for each sample: $d_i = x_{1, i} - x_{2, i}$.
2. Compute the mean difference: $\bar{d} = \frac{\sum d_i}{n}$.
3. Compute the standard deviation of the differences ($s_d$): $s_d = \sqrt{\frac{\sum (d_i - \bar{d})^2}{n - 1}}$
4. Compute $t_{calc}$: $t_{calc} = \frac{|\bar{d}|\sqrt{n}}{s_d}$ Compare $t_{calc}$ against $t_{crit}$ with $\nu = n - 1$ degrees of freedom.

The $F$-Test for Comparing Precisions

The $F$-test evaluates whether two analytical methods exhibit statistically identical precision by comparing their variances ($s^2$): $F_{calc} = \frac{s_1^2}{s_2^2} \quad \text{where } s_1^2 \ge s_2^2$ By convention, the larger variance is always placed in the numerator, ensuring $F_{calc} \ge 1.0$.

• If $F_{calc} > F_{crit}$ (evaluated at $\nu_1 = n_1 - 1$ and $\nu_2 = n_2 - 1$), the hypothesis of equal precision ($H_0: \sigma_1^2 = \sigma_2^2$) is rejected. The two methods have significantly different precision.

Outlier Rejection Criteria

An outlier is an extreme data point that appears anomalous relative to the rest of the dataset.

1. Dixon's $Q$-Test (for small datasets, $3 \le n \le 10$)

1. Arrange the data in ascending numerical order: $x_1 \le x_2 \le \dots \le x_n$.
2. Identify the suspect value ($x_1$ or $x_n$).
3. Calculate $Q_{calc}$: $Q_{calc} = \frac{|x_{suspect} - x_{nearest}|}{x_{max} - x_{min}} = \frac{\text{Gap}}{\text{Range}}$
4. Compare with $Q_{crit}$ ($Q$-table at $n$ and $90\%$ or $95\%$ confidence). If $Q_{calc} > Q_{crit}$, reject the outlier; otherwise, retain it.

2. Grubbs' Test (Standardized ISO/ASTM Outlier Test)

Grubbs' test is preferred over the $Q$-test by international standards bodies because it accounts for overall distribution variance: $G_{calc} = \frac{|x_{suspect} - \bar{x}|}{s}$ where $\bar{x}$ and $s$ are calculated including the suspect point.

• If $G_{calc} > G_{crit}$, the point is rejected at the chosen significance level.

---

4. Propagation of Measurement Uncertainty

When experimental parameters with known random uncertainties are combined algebraically, their uncertainties propagate into the final calculated result ($y$).

1. Addition and Subtraction

For functions of the form $y = a + b - c$: $s_y = \sqrt{s_a^2 + s_b^2 + s_c^2}$ Rule: For addition and subtraction, absolute uncertainties combine in quadrature.

2. Multiplication and Division

For functions of the form $y = \frac{a \cdot b}{c}$: $\frac{s_y}{y} = \sqrt{\left(\frac{s_a}{a}\right)^2 + \left(\frac{s_b}{b}\right)^2 + \left(\frac{s_c}{c}\right)^2}$ Rule: For multiplication and division, relative uncertainties combine in quadrature.

3. Powers and Exponentials

• For power functions $y = x^n$: $\frac{s_y}{y} = |n| \left(\frac{s_x}{x}\right)$
• For common logarithms $y = \log_{10}(x)$: $s_y = 0.4343 \left(\frac{s_x}{x}\right)$
• For natural logarithms $y = \ln(x)$: $s_y = \frac{s_x}{x}$
• For antilogarithms $y = 10^x$: $\frac{s_y}{y} = 2.3026 s_x$

Worked Uncertainty Propagation Calculation

Calculate the molarity ($M$) of a sodium chloride solution prepared by dissolving $m = 2.9220 \pm 0.0004\text{ g}$ of pure $NaCl$ ($FW = 58.4430 \pm 0.0005\text{ g/mol}$) into a volumetric flask with calibrated volume $V = 500.00 \pm 0.05\text{ mL}$:

1. Calculated Value: $M = \frac{m / FW}{V} = \frac{2.9220 / 58.4430}{0.50000} = 0.0999949\text{ M}$
2. Relative Standard Deviations:
   • $\frac{s_m}{m} = \frac{0.0004}{2.9220} = 1.369 \times 10^{-4}$
   • $\frac{s_{FW}}{FW} = \frac{0.0005}{58.4430} = 8.555 \times 10^{-6}$
   • $\frac{s_V}{V} = \frac{0.05}{500.00} = 1.000 \times 10^{-4}$
3. Combined Relative Uncertainty: $\frac{s_M}{M} = \sqrt{(1.369 \times 10^{-4})^2 + (8.555 \times 10^{-6})^2 + (1.000 \times 10^{-4})^2} = \sqrt{1.874 \times 10^{-8} + 7.32 \times 10^{-11} + 1.000 \times 10^{-8}} = 1.697 \times 10^{-4}$
4. Absolute Uncertainty ($s_M$): $s_M = M \times (1.697 \times 10^{-4}) = 0.0999949 \times 1.697 \times 10^{-4} = 0.0000170\text{ M}$
5. Reported Concentration: $0.09999 \pm 0.00002\text{ M}$.

---

5. Quantitative Calibration Strategies and Figures of Merit

Quantitative analytical instrumentation measures an analytical signal ($S$) that must be correlated with analyte concentration ($C$).

1. External Standard Calibration and Linear Regression

A series of standard solutions containing known analyte concentrations is prepared in pure solvent. Linear least-squares regression fits the equation $y = mx + b$:

• Slope ($m$ - Calibration Sensitivity): $m = \frac{\sum (x_i - \bar{x})(y_i - \bar{y})}{\sum (x_i - \bar{x})^2}$
• $y$-Intercept ($b$): $b = \bar{y} - m\bar{x}$
• Coefficient of Determination ($R^2$): Measures the proportion of variance in signal explained by analyte concentration ($R^2 > 0.995$ is typically required in regulatory testing).
• Limitation: External calibration assumes that the sample matrix produces no signal enhancement or suppression (zero matrix interference).

2. Standard Addition Method (Spike Recovery)

Used when the chemical matrix of the sample is complex and produces matrix effects (such as viscosity, ionic strength, or concomitants) that alter analyte sensitivity.

• Aliquots of the sample solution are spiked with increasing, precisely known quantities of analyte standard.
• In a multiple-addition plot where signal ($S$) is plotted against added concentration ($C_{added}$): $S = m C_{added} + S_0$ Setting $S = 0$ gives the negative $x$-intercept: $C_{unknown} = \frac{S_0}{m} = -x_{\text{intercept}}$

3. Internal Standard Method

Used when experimental variables cause instrument drift, volatilization, or injection volume fluctuations (e.g., in Gas Chromatography or ICP-MS):

• A fixed, unvarying concentration of an internal standard (a distinct compound chemically and physically similar to the analyte, absent in the original sample) is added to all blanks, standards, and unknowns.
• Calibration plots the ratio of analyte signal to internal standard signal ($\frac{S_{analyte}}{S_{internal\_std}}$) against analyte concentration.
• Any variation in sample injection volume or nebulizer efficiency affects both analyte and internal standard identically, preserving the signal ratio.

Analytical Figures of Merit