Reporting Analytical Results

Every certificate of analysis issued by a pharmaceutical quality-control lab reports an assay result as mean ± uncertainty with a stated confidence level, because regulatory submissions (FDA, EMA) are audited specifically for whether the reported digits are consistent with the validated method's precision — a mismatch (too many digits, or an uncertainty rounded to three figures) can trigger a data-integrity finding. Clinical chemistry reference intervals (e.g., serum creatinine, glucose) are published with explicit uncertainty budgets so that a physician can tell whether two sequential patient results differ meaningfully or are within measurement noise. Environmental monitoring under EPA methods reports contaminant concentrations (e.g., lead in drinking water, $\mu$g/L) together with method detection limits and uncertainty, since a result reported without correct rounding can trigger unwarranted regulatory action or, conversely, mask a real exceedance. Forensic toxicology and blood-alcohol reporting are legally required to state measurement uncertainty explicitly (some jurisdictions mandate a stated confidence interval on a defendant's blood alcohol content), making the rounding convention here a matter of legal, not just scientific, consequence. National metrology institutes (NIST, BIPM) issue certified reference materials with certificate values reported using exactly the funnel described above — uncertainty rounded to one or two significant figures, the certified value rounded to match — and this practice, formalized in the international GUM standard, is what every analytical laboratory's quality system is ultimately built on. The same discipline anticipates the more formal machinery of Confidence Intervals, where the width of the reported interval is tied explicitly to a stated probability level rather than a bare "±."

Ask the tutor: How do I decide, for a specific dataset, whether to report the standard error of the mean or a full 95% confidence interval?

Every analytical result is a random variable with a location (the best estimate, usually a mean) and a spread (the uncertainty). The physical act of reporting is choosing how finely to slice that spread. Slice too coarsely and you throw away real precision the experiment achieved; slice too finely and you claim precision the experiment never had. The convention that fixes this is beautifully simple: figure out how precisely you know the uncertainty itself, and let that number set the resolution for everything else. Consider the iron-ore titration. Four replicate %Fe values scatter around a mean because of small, uncontrollable variations — a half-drop of titrant here, a slightly-off endpoint color there. The standard deviation $s$ quantifies that scatter. But $s$, computed from only 4 points, is itself an estimate with huge relative uncertainty — the relative standard deviation of $s$ for $n=4$ replicates is roughly $1/\sqrt{2(n-1)} \approx 41\%$. It would be absurd to write $s = 0.03464$ and treat every digit as meaningful when the number itself is uncertain by nearly half its own value. So $s$ gets rounded hard, to one significant figure, and only then is the mean rounded to match it. The second figure below shows the anatomy of a properly reported result: three components (value, uncertainty, units/confidence) that must always travel together, and a "rounding funnel" that forces the uncertainty to be decided before the mean is finalized. The third figure contrasts three reports of the same underlying data — one under-reported, one over-reported, one correct — visualizing why the middle one is the only defensible choice.

Mean and sample standard deviation of replicates.

where: $x_i$ are the $n$ replicate measurements, $\bar x$ is the sample mean (best estimate of the true value $\mu$), $s$ is the sample standard deviation (a measure of scatter between individual replicates), and $n - 1$ is the degrees of freedom lost to estimating $\bar x$ from the same data. Reading: $s$ describes how far a single new measurement is likely to fall from $\bar x$; it does not describe how well you know $\bar x$ itself — that is the job of the standard error.

Standard error of the mean.

where: $s_{\bar x}$ is the standard error (uncertainty in the mean as an estimate of the true value), shrinking as $\sqrt n$ grows — this is the number that normally accompanies a reported mean of replicate measurements. Reading: doubling the number of replicates does not halve the uncertainty; it shrinks it by a factor $1/\sqrt2 \approx 0.71$. Diminishing returns are built into the square root.

The rounding rule (the deliverable of this whole topic).

where: $\sigma$ stands generically for whichever uncertainty is being reported ($s$, $s_{\bar x}$, or a propagated uncertainty $u_y$), and "decimal place" means literal place value (tenths, hundredths, …), not a count of significant figures in $\bar x$. Reading: the uncertainty decides the number of decimal places in the whole report; the mean is a slave to that decision, never rounded independently.

Propagated uncertainty feeding the same rule. For $y = f(a,b,\dots)$, per Propagation of Uncertainty,

where: $u_a, u_b,\dots$ are the uncertainties in each measured input, and $u_y$ is the resulting uncertainty in $y$, which then goes through the exact same rounding funnel above before $y$ itself is reported.

Example 1 (foundational) — replicate titrations, finishing the calculation above with an alternate confidence-level report. Data: 15.67, 15.71, 15.63, 15.69 %Fe ($n=4$). We already found $\bar x = 15.675$, $s = 0.03416$, $s_{\bar x} = 0.01708$. Suppose the report must instead state a 95% confidence interval, using Student's $t$ for $n-1 = 3$ degrees of freedom, $t_{0.05,3} = 3.182$ (standard $t$-table value):

Round the confidence half-width to one significant figure: $0.0544 \to 0.05$ (tenths-of-hundredths place, i.e. hundredths place after rounding). Round the mean to match:

Note the wider interval than the plain standard-error report ($\pm0.02$) — the $t$-multiplier inflates the interval to account for the small sample size, foreshadowing the formal treatment of Confidence Intervals. Both "$15.68\pm0.02$" and "$15.68\pm0.05$ (95% CI)" are correct; they answer different questions ("how well do I know the mean" vs. "where would 95% of such intervals capture the true mean"), and a report must say which one it is giving.

Example 2 (applied) — density of a solution from mass and volume, full propagation then correct rounding. A 25.00 mL volumetric flask (calibrated tolerance $\pm 0.03$ mL, treated as the uncertainty $u_V$) is weighed empty (mass $m_1 = 24.1163$ g) and full of an unknown salt solution (mass $m_2 = 55.2481$ g), each weighing done on a balance with uncertainty $u_m = \pm 0.0002$ g (repeatability, from balance calibration). Find the density and report it correctly. Mass of solution:

Propagate the mass uncertainty (subtraction — absolute uncertainties add in quadrature):

Density:

Propagate through division (relative uncertainties add in quadrature):

Compute each term: $0.000283/31.1318 = 9.09\times10^{-6}$; squared $= 8.26\times10^{-11}$ (negligible). $0.03/25.00 = 1.200\times10^{-3}$; squared $=1.440\times10^{-6}$.

So $u_\rho = (1.2000\times10^{-3})(1.245272) = 1.494\times10^{-3}\ \text{g/mL} = 0.001494\ \text{g/mL}$. Round the uncertainty to one significant figure: $0.001494 \to 0.001$ g/mL (thousandths place). Round the density to match:

Notice the volumetric-flask tolerance completely dominates the balance's contribution (the mass term was 4–5 orders of magnitude smaller inside the square root) — a common and important pattern: propagated uncertainty is almost always set by the worst contributing measurement, not by averaging all of them. Reporting all six digits of $1.245272$ would falsely claim the balance's precision transferred all the way through, when in fact the flask tolerance erased it after the third decimal.

Example 3 (exam/competition level) — calibration curve, dilution, and a systematic bias correction, reported end to end. A spectrophotometric calibration curve for Cu²⁺ gives, from least-squares regression on five standards, slope $m = 0.0512 \pm 0.0009$ L/mg (uncertainty = standard error of slope) and intercept $b = 0.0031 \pm 0.0015$ absorbance units (both from standard regression-statistics formulas, an application of Propagation of Uncertainty to least-squares fitting). An unknown, diluted 1:10 (10.00 mL to 100.00 mL volumetric, both glassware with standard tolerance $u = \pm 0.02$ mL relative uncertainty converted below), is measured at $A = 0.417 \pm 0.003$ (instrument reading noise, one standard deviation from triplicate reads).

Step 1 — invert the calibration to get diluted concentration.

Step 2 — propagate uncertainty through the inversion. Treat $c_{\text{dil}} = f(A,b,m)$; partial derivatives:

So $c_{\text{dil}} = 8.084 \pm 0.157$ mg/L before dilution correction — note the slope uncertainty term (0.0202) dominates completely, again illustrating that the calibration's own precision, not the reading noise, is the bottleneck.

Step 3 — undo the 1:10 dilution. Dilution factor $DF = 100.00/10.00 = 10.00$, with glassware tolerances $u_{V_1} = 0.02$ mL on the 10.00 mL pipet and $u_{V_2}=0.08$ mL on the 100.00 mL flask (typical Class A tolerances). Relative uncertainty in $DF$:

Original concentration:

Relative uncertainty in $c_{\text{dil}}$ is $0.1565/8.084 = 1.936\times10^{-2}$. Combine in quadrature with the dilution factor's relative uncertainty:

Step 4 — apply a known systematic bias correction. Suppose spike-recovery experiments show the method reads systematically 2.0% low (a matrix effect), independently determined with its own uncertainty of $\pm0.5\%$ (relative, one standard deviation). Correct the mean upward and add the correction's uncertainty in quadrature:

Relative uncertainty from the correction factor is $0.5\%/0.980 \approx 5.10\times10^{-3}$ (propagating the correction the same way as any multiplicative factor); combine with the existing relative uncertainty $1.948\times10^{-2}$:

Step 5 — round correctly and report. Round $u_{\text{final}}=1.661$ to one significant figure: $\to 2$ mg/L (ones place). Round $c_{\text{corrected}}=82.49$ to the same place (ones):

Every intermediate quantity in this chain carried 4–5 digits precisely so that rounding error would not compound through five propagation steps; only the final answer is truncated. This is the standard practice: round only at the end, never mid-calculation.

The modern convention for reporting a measured value with its uncertainty descends from the broader development of statistical inference for small samples, most directly through the work of William Sealy Gosset, who published under the pseudonym "Student." Gosset worked as a chemist and statistician for the Guinness brewery in Dublin, where he needed a way to judge whether small batches of barley or yeast differed meaningfully from each other — exactly the problem of a chemist running four replicate titrations and asking whether the spread reflects real variability or is just noise from a small sample. His 1908 paper "The Probable Error of a Mean" derived the distribution now called the Student's $t$-distribution, which corrects for the extra uncertainty introduced by estimating both the mean and the standard deviation from the same small dataset — precisely the correction used in Example 1's 95% confidence interval above. Gosset's practical, industrial motivation (Guinness forbade its scientists from publishing under their own names, fearing trade secrets would leak) mirrors the practical motivation behind every rule in this article: analytical chemists do not have infinite replicates, infinitely precise balances, or infinitely well-characterized calibration curves, and the rounding conventions exist to keep a report from claiming more certainty than a finite, real experiment actually delivered.