An uncertainty budget is not a table that simply makes the error margin large

When people are asked to attach uncertainty to a carbon-reduction result, they often take the accuracy from a sensor specification—for example, ±2%—and apply it unchanged to the final reduction amount. But a reduction amount is not a single sensor reading. It is the result of subtracting project-period emissions from baseline-period emissions and, in some cases, accounting for leakage, project emissions, changes in herd size, and changes in production. Even with a highly accurate concentration analyzer, the reduction amount can be far off if ventilation is estimated incorrectly or the baseline does not represent the comparison period.

The International Vocabulary of Metrology (VIM) defines an uncertainty budget as a statement of measurement uncertainty, its components, and their calculation and combination. It includes the measurement model, estimates and uncertainties for each input quantity, covariances, probability distributions, degrees of freedom, evaluation types, and, when needed, coverage factors. An uncertainty budget is therefore neither a vague “safety margin” nor a disclaimer that weakens the result. It is a technical document showing what information the final number depends on and which inputs must be improved to achieve a real improvement in the result.

Another misconception is to use uncertainty and error as if they meant the same thing. An incorrectly entered herd size, confusion between ppm and %, mismatched clocks, and duplicate application of a calibration factor are errors that must be found and corrected. By contrast, uncertainty expresses dispersion that cannot be completely eliminated with the information currently available, such as the residual limits of sensor calibration after correction, natural variation among animals, and how well a sample represents the population. Merely widening an uncertainty interval without correcting known bias is not good design.

The first step is to define the measurand in one sentence before writing the equation

An uncertainty budget must begin by defining what is being measured: the measurand. “Methane reduction” alone is insufficient. It should specify the location, herd, period, boundary, unit, and comparison rule—for example: “Monthly reduction in methane mass (kg CH₄) emitted by 120 finishing cattle in Barn A from June 1 through 30, 2026, compared with a baseline normalized to the same unit of production.” This sentence must also connect to whether only enteric fermentation is included, whether methane from manure storage is included, and how the external background concentration is subtracted.

Next, express the final result as a function of the input quantities. A simplified barn emission rate can be written as follows.

E = Q × (x_out − x_in) × (M_CH4 / V_m) × t

Here, Q is the volumetric flow rate on a dry basis; x_out and x_in are the methane mole fractions at the outlet and inlet; M_CH4 is the molecular weight of methane; V_m is the molar volume at the relevant temperature and pressure; and t is the integration time. A ppm concentration is converted to a dimensionless mole fraction by multiplying it by 10⁻⁶, and the dry or wet basis of the concentration and flow rate must match. The reduction amount R is calculated from the difference between baseline emissions E_base and project emissions E_project, with adjustments for additional emissions or leakage within the boundary. A naturally ventilated barn has no single outlet and may experience flow reversal, so rather than applying this equation unchanged, document the full model for the adopted method, such as a tracer-gas method, CO₂ balance, mass balance, or dispersion model.

Writing the model first reveals omissions. It shows whether the budget includes only the concentration sensor but leaves out ventilation, whether the herd-size adjustments for the baseline and project periods use a separate model, and whether predicted values used to fill missing intervals are being treated like observations. The measurement-model version and its period of application must also be retained, because changing the equation changes both the result and its uncertainty even when the same raw data are used.

Find uncertainty sources across the entire measurement chain

In practice, dividing the sources into six groups helps prevent omissions.

  1. Instrumentation and calibration: certified values of standard gases, zero and span, linearity, resolution, repeatability, long-term drift, cross-sensitivity, temperature, humidity and pressure effects, and sample-line leakage and adsorption.

  2. Spatial sampling: whether the sensor represents the air actually leaving the barn, whether it is biased toward local concentrations near feeders, manure openings, or fans, and whether inlet and outlet points have been classified correctly.

  3. Temporal sampling: whether rumination, feeding, manure agitation, and ventilation-control cycles were captured adequately, and how unmeasured nights, weekends, and seasons were handled.

  4. Activity data: errors in the actual daily herd size, body weight, intake, feed composition, milk yield or weight gain, manure-management period, and equipment operating time.

  5. Models and conversions: uncertainty from estimating ventilation rates, subtracting background concentration, dry and wet bases, conversion to standard conditions, time integration, missing-data substitution, and extrapolation.

  6. Baseline and comparability: differences in weather, herd composition, productivity, and non-feed management changes between the baseline and project periods.

Components obtained from the statistics of repeated measurements are generally close to Type A evaluations, while components derived from calibration certificates, manufacturer information, previous research, and expert judgment are generally close to Type B evaluations. This does not mean that Type A is “good uncertainty” and Type B is a “poor estimate.” Both must be converted to the common form of standard uncertainty before combination, and the key is to document data sources and assumptions transparently.

Do not simply add all the percentages

Sensitivity coefficients are needed to express each input’s uncertainty in the unit of the final result. A sensitivity coefficient indicates how much the result changes when an input changes slightly. If the inputs are independent and the model is approximately linear, square each contribution obtained by multiplying its standard uncertainty by its sensitivity coefficient, add the squares, and take the square root. A typical uncertainty budget presented by NIST likewise manages the components, sensitivity coefficients, standard deviations, and degrees of freedom in one table.

The independence assumption, however, often fails for livestock data. If the same sensor measures both the baseline and project periods, calibration bias affects both values in common. If the CO₂ generation rate and feed intake used to calculate the emission rate are derived from the same activity data, they may be correlated. Multiple sensors may also share a single temperature and humidity correction value or the same standard gas. Ignoring covariance can overestimate or underestimate uncertainty. Both adding everything indiscriminately in the name of being “conservative” and assuming that a common instrument will make biases cancel require evidence.

For nonlinear models, asymmetric distributions, data near a detection limit, or ratios with a small denominator, first-order error propagation may be unstable. A Monte Carlo simulation that defines the probability distribution and correlation structure of each input can be useful in these cases. Increasing the number of simulations, however, does not correct a wrong distributional assumption or poor representativeness. Model validation and evidence for the inputs come first.

A final report cannot be interpreted if it states only ±. It must say whether the value is a combined standard uncertainty or an expanded uncertainty obtained by multiplying by a coverage factor, and what the coverage probability is. NIST describes expanded uncertainty as U = k × u_c and notes that, under a normal distribution and sufficient conditions, k=2 can correspond to an interval of approximately 95%. But it does not automatically mean 95% for every dataset. The distribution, effective degrees of freedom, and adopted method must be reported together.

Example of an uncertainty budget prepared at a farm

Suppose monthly emissions before and after the introduction of low-methane feed are compared. The concentration analyzer’s calibration uncertainty is 2%, but natural-ventilation estimation is assessed at 18%, spatial sampling of the concentration difference between openings at 10%, missing-data substitution at 6%, and daily and weekly baseline variation at 12%. These numbers must not simply be added and reported as 48%. First determine which result each item affects, whether the items are independent, and whether common bias cancels in the before-and-after comparison. In addition, the example values can be compared and combined only after conversion to standard relative uncertainties for the same output quantity; the actual assessment varies with the distributions, coverage factors, and correlations.

If sensitivity analysis shows that ventilation accounts for more than half of the total variance, replacing the concentration sensor with a more expensive one is not the priority. It may be more effective to conduct more tracer-gas tests, add airflow measurements at openings, and stratify periods of ventilation operation. Conversely, if baseline variation dominates, the first priorities are to extend the measurement period, use a concurrent control group, and normalize for herd size and intake. The value of an uncertainty budget is not limited to producing a final interval; it also determines the order of future measurement investments.

If the reduction amount is 100 kg CH₄ and the expanded uncertainty is 30 kg, do not feature only “100 kg reduction.” Report it as 100 kg CH₄, expanded uncertainty 30 kg, with the applicable period, boundary, coverage probability, and method stated. When comparing it with an eligibility criterion or a credit-issuance threshold, separately apply the conservativeness rules of that program. Calculating statistical uncertainty does not automatically determine a programmatic discount rate or verification decision.

Management criteria for an operational uncertainty budget

An uncertainty budget must be connected to the data pipeline rather than remain a one-off table at the end of a report. When a sensor calibration certificate is updated, the applicable Type B component must also change. If the missing-data rate exceeds a predefined limit, apply the established substitution model or exclude the period and record the reason for the change. Reassess baseline suitability when herd size or the feed formulation changes.

Version control requires, at minimum, the raw-data ID, sensor and standard-gas IDs, calibration date, calculation-code version, sources and versions of emission and correction factors, missing-data rules, and the history of responsible personnel and approvals. Keeping raw data unchanged and separating corrected values from calculated outputs allows recalculation when the methodology changes. Assigning an owner and reassessment cycle to each uncertainty item prevents the state in which “the table exists, but no one updates it.”

The IPCC national greenhouse-gas inventory guidelines treat uncertainty analysis not as a procedure for proving that a single number is perfect, but as a process for prioritizing improvements to the major sources of uncertainty and increasing inventory quality over time. Although a farm-scale project and a national inventory do not have the same purpose, the principles of identifying every material source, maintaining a consistent time series, and linking the work to QA/QC remain valid.

Implementation checklist

  • Has the measurand been defined in one sentence down to the gas, boundary, herd, period, unit, and reference conditions?

  • Has the final reduction amount been expressed as a function of every input and correction, with the model version fixed?

  • Does the budget include spatial and temporal sampling, ventilation, activity data, missing data, and baseline uncertainty—not just sensors?

  • Were known errors and correctable biases corrected first?

  • Are the Type A or Type B basis, distribution, standard uncertainty, and degrees of freedom documented for each component?

  • Were correlations and covariances arising from common sensors, standard gases, and activity data examined?

  • Were sensitivity coefficients used to calculate each component’s contribution to the final result?

  • Was a propagation method appropriate for nonlinearity, asymmetry, and detection-limit problems selected?

  • Were combined standard uncertainty and expanded uncertainty distinguished, and the coverage factor and coverage probability reported?

  • Were the largest contributing items connected to the next measurement-improvement plan and budget?

Conclusion: a good uncertainty budget does not hide a number’s weaknesses

Uncertainty in carbon-reduction measurement is not a sign of failure. It is a way to quantify real-world variation and the limits of available information. What matters is not attaching an arbitrarily wide interval, but fixing the measurand and model and tracing the route by which each uncertainty source enters the result.

Managing only sensor accuracy can produce a precise-looking but incorrect reduction amount. By contrast, a budget that includes spatial representativeness, ventilation, activity data, the baseline, and correlations shows what claims can be made and what must be measured further. Ultimately, an uncertainty budget is not an appendix that weakens carbon-performance results; it is a quality blueprint for turning field records into verifiable claims.

Sources