PCF Sensitivity Analysis and Hotspot Identification
What to do once the product carbon footprint exists: find what drives it (contribution analysis and hotspots), test how robust it is to your choices (sensitivity analysis), and quantify how confident you can be in the result (uncertainty analysis). Covers the dominance principle, methodological-choice sensitivity, the pedigree matrix and Monte Carlo propagation, the refinement loop that ties them together, and the explicit ISO 14044 / ISO 14067 / GHG Protocol requirements behind each.
A product carbon footprint is not finished when the total is computed — that is when the most valuable work begins. Three distinct questions follow every PCF: which parts of the lifecycle dominate the number, how much would the number move if a key choice were different, and how wide is the confidence interval around it. These are contribution, sensitivity, and uncertainty analysis respectively, and conflating them — as most reporting does — is the single most common analytical weakness a verifier finds. This page keeps them separate and shows how to use all three to turn a static number into a decision-grade result.
Contribution analysis ranks where emissions come from; sensitivity analysis tests how far the result moves when an input or a methodological choice changes; uncertainty analysis propagates the spread to express the footprint as a range rather than a point.
Three Distinct Questions: Contribution, Sensitivity, Uncertainty
The three techniques are routinely used interchangeably in practice, and that imprecision is itself a defect. They answer different questions, require different methods, and satisfy different standards clauses. Holding them apart is the foundation of defensible post-calculation analysis.
| Technique | Question it answers | Method | Output | Primary use |
|---|---|---|---|---|
| Contribution analysis | Where do the emissions come from? | Decompose the total by stage, process, or substance and rank | A ranked breakdown; the hotspots | Targeting reductions; scoping data effort |
| Sensitivity analysis | How much does the result move if I change a choice? | Vary one input or methodological choice; observe the swing | A swing magnitude or sensitivity ratio per variable | Finding which assumptions matter; defending choices |
| Uncertainty analysis | How confident can I be in the number? | Propagate the spread in all inputs (e.g. Monte Carlo) | A confidence range (e.g. P5–P95) around the total | Reporting honestly; supporting comparative claims |
The relationship between them is sequential and reinforcing: contribution tells you where to look, sensitivity tells you which assumptions at those hotspots matter, and uncertainty tells you how much the unavoidable spread at those points widens the final answer. A PCF that reports only a point total, with none of the three, is a number without a confidence statement — and increasingly one that fails scrutiny under GHG Protocol Product assurance and EU green-claims review.
Provider directory
Need help keeping this up all year?
See who does this work. Every listing names the standards it works to, and paid placements are labelled. Including Anthesis Group and KERAMIDA.
Browse 6 carbon accounting & inventory providers →Do this work? A listing is US$390 a year. Get listed →
Contribution Analysis: Finding the Hotspots
Contribution analysis decomposes the total footprint into its constituent parts and ranks them. It is the cheapest and highest-leverage of the three techniques, because it tells you immediately where reduction effort and data-quality effort should go. The governing observation is the dominance principle.
The dominance principle
In almost every product system, a small number of inputs or stages account for the large majority of the footprint — the carbon analogue of the Pareto distribution. Identifying that dominant minority is the entire purpose of contribution analysis. Once the top contributors are known, every later technique (sensitivity, uncertainty, data refinement) focuses on them, because effort spent on a 1%-of-footprint input rarely changes a decision.
Contribution by lifecycle stage
The first decomposition is by lifecycle stage: raw materials, inbound transport, manufacturing, distribution, use, and end-of-life. Stage-level contribution reveals whether a product is upstream-dominated (most consumer goods, where raw materials lead) or use-phase-dominated (energy-using products, where the use phase swamps everything else). The dominant stage determines reduction strategy entirely — a use-phase-dominated product cannot be meaningfully decarbonised by changing its packaging.
Contribution by process and substance
Beneath the stage view sit two finer decompositions. Process contribution attributes the footprint to specific unit processes — a particular ingredient, a specific energy carrier, one transport leg — which is what reduction projects actually target. Substance contribution splits the total by greenhouse gas (CO₂, CH₄, N₂O), which matters because a CH₄- or N₂O-dominated footprint behaves differently under a change of GWP basis than a CO₂-dominated one — a fact the GWP-basis sensitivity (below) depends on.
Setting a materiality threshold
Contribution analysis requires a stated materiality (cut-off) threshold, fixed in the goal and scope before results are seen, so that excluding small flows is a documented rule rather than a post-hoc convenience. A common pattern excludes individual flows below 1% of the total provided the sum of excluded flows stays under a stated ceiling. The threshold must be disclosed: a PCF that quietly drops “small” inputs without a stated rule has made an undocumented scope decision, which is a verification finding.
Contribution analysis is usually framed as a reduction-targeting tool, and it is — but its more rigorous use is to direct data-collection effort. A hotspot computed from a low-quality generic factor is a number you do not yet trust at the one place you most need to. The correct reading of a contribution analysis is therefore two-dimensional: a stage is a priority if it is large (drives the total) and currently uncertain (rests on weak data). That two-dimensional reading is formalised in the effort-prioritisation quadrant below, and it is what separates a screening hotspot list from a decision-grade one.
Sensitivity Analysis: Testing What Matters
Sensitivity analysis asks how far the result moves when a single input value or a methodological choice is changed, holding everything else fixed. It separates the assumptions that matter — those that swing the answer — from those that do not, and it is the technique that makes a PCF’s conclusions defensible against “but what if you had assumed X?”
One-at-a-time (OAT) parameter sensitivity
The workhorse method varies each significant parameter in turn — by a fixed percentage, or across its plausible range — and records the resulting change in the total. Plotting the swings ranks parameters by influence (the familiar “tornado” ordering). OAT is simple, transparent, and the right starting point; its limitation is that it cannot capture interactions between parameters, which is where uncertainty analysis (below) takes over.
The sensitivity ratio
To compare parameters of different units and scales, normalise each swing into a dimensionless sensitivity ratio (an elasticity): the fractional change in the result divided by the fractional change in the input.
A parameter with a sensitivity ratio near zero can be left on a generic value with no loss of decision quality; a parameter with a high ratio must be measured well, because the result tracks it closely. The sensitivity ratio is what converts the contribution ranking into a data-collection priority list.
Sensitivity to methodological choices — the under-tested kind
Most practitioners run parameter sensitivity and stop. The more revealing test is sensitivity to the methodological choices the analyst made — the choices that the earlier pages in this cluster cover:
- Allocation method. Re-run the footprint under mass versus economic allocation and record the swing. For a multi-output process this is frequently the single largest sensitivity, dwarfing any parameter. The decision is detailed on the PCF allocation methodology page.
- System boundary. Re-run cradle-to-gate versus cradle-to-grave. The swing is the entire downstream lifecycle — usually decisive. Boundary mechanics are on the cradle-to-gate methodology page.
- GWP basis. Re-run on AR5 versus AR6 GWP-100 values. The swing depends on the gas mix from the substance contribution: a CH₄- or N₂O-heavy footprint moves materially, a CO₂-only one barely at all. The current corporate-reporting anchor is AR6 — biogenic methane at
27on the AR6 GWP-100 basis — and the full dataset is on the IPCC AR6 GWP values data page. - Framework choice. Re-run under the recycling and data rules of each framework the PCF must satisfy — see the PCF frameworks comparison page.
Methodological-choice sensitivity is what tells you whether your headline number is robust or an artefact of a defensible-but-arbitrary decision. A footprint that doubles when allocation flips from mass to economic is not wrong — but it cannot be reported as a single confident figure without disclosing that sensitivity.
If the conclusion of a comparative PCF (“Product A beats Product B”) holds under mass allocation but reverses under economic allocation, the comparison is not robust and must not be published as settled. ISO 14044 is explicit that comparative assertions disclosed to the public require the sensitivity of the conclusion to methodological choices to be tested and reported. Running only parameter sensitivity, and skipping the choice sensitivity, is the most common way an apparently rigorous comparison turns out to be indefensible.
Uncertainty Analysis: Quantifying Confidence
Uncertainty analysis propagates the spread in the inputs through to the result, expressing the footprint as a range with a stated confidence rather than a single deceptive point. Where sensitivity asks “what if this one thing changed?”, uncertainty asks “given that everything is imperfectly known at once, how wide is the believable band?”
Three sources of uncertainty
Parameter uncertainty
The inputs themselves are imperfectly known — an emission factor, an activity quantity, an efficiency. This is the most tractable source and the one Monte Carlo handles directly, given a distribution for each input.
Scenario uncertainty
The defensible choices — allocation method, boundary, end-of-life assumption, electricity mix — that could each reasonably be set differently. This is the source that sensitivity analysis explores and that no single distribution captures.
Model uncertainty
The structural simplifications in the model itself — linearity assumptions, the choice of characterisation factors, omitted feedbacks. The hardest to quantify and usually addressed qualitatively, by documenting the model’s known limitations.
The pedigree matrix and data-quality scoring
Parameter uncertainty is rarely known from measurement, so LCA uses a structured proxy: the pedigree matrix, which scores each data point on several quality indicators and converts those scores into an uncertainty distribution. The indicators are the backbone of data-quality assessment in the ecoinvent database and underpin the data-quality requirements of ISO 14067.
| Pedigree indicator | High quality (low added uncertainty) | Low quality (high added uncertainty) |
|---|---|---|
| Reliability | Verified, measured data | Non-verified estimate or expert judgement |
| Completeness | Representative sample over an adequate period | Small or unrepresentative sample |
| Temporal correlation | Data from the reporting year | Data many years out of date |
| Geographical correlation | Data from the exact production region | Data from a different, dissimilar region |
| Technological correlation | Data from the exact process and technology | Data from a different process or scale |
Each indicator is scored on a band scale, and the combined score maps to a multiplicative uncertainty factor applied around the input’s value. A data point that is verified, current, local, and technology-matched carries a narrow distribution; a generic, dated, foreign proxy carries a wide one. This is how qualitative data-quality judgements become quantitative inputs to the propagation step.
Monte Carlo propagation
With a distribution attached to each input, Monte Carlo simulation samples all of them simultaneously across many thousands of iterations, recomputing the footprint each time, and builds the distribution of the result. The output is not a point but a histogram, from which a confidence range is read.
Reporting the result as a range
The honest output of a PCF is a central estimate with a confidence interval and a stated basis — for example, “X kg CO₂e (P5–P95: lower–upper), parameter uncertainty only, pedigree-matrix-derived.” Reporting only the point estimate hides the very information a decision-maker needs: whether two products genuinely differ or merely differ within noise. A difference smaller than the overlap of the two confidence intervals is not a real difference, however precise the point estimates look.
The most consequential use of uncertainty analysis is adjudicating comparisons. If Product A’s footprint is 10 with a P5–P95 of 8–12, and Product B’s is 11 with 9–13, the two are statistically indistinguishable — the intervals overlap heavily, and any claim that A is “lower-carbon than” B is unsupportable. This is precisely the situation a green-claims regulator probes. Run the uncertainty analysis before drafting any comparative claim, not after, so the claim is calibrated to what the data can actually support rather than to the point estimates’ false precision.
How the Three Fit Together: The Refinement Loop
Contribution, sensitivity, and uncertainty are not a one-pass checklist — they form an iterative loop that progressively tightens the footprint where tightening matters. Each pass spends data-collection effort exactly where the previous pass showed it would change the answer.
Contribution
Decompose the total; rank stages and processes; identify hotspots.
Sensitivity
At the hotspots, test which parameters and choices actually swing the result.
Uncertainty
Score data quality on the sensitive hotspots; propagate to a confidence range.
Refine & re-run
Replace the weakest high-influence data with primary data; loop back to Step 1.
The loop terminates when the confidence range is narrow enough to support the decision the PCF exists to inform — not when the data is “perfect,” which it never is. A screening footprint might run the loop once; a footprint underpinning a public comparative claim might run it several times, each pass targeting the next most influential, least certain input. This is the practical meaning of the ISO 14044 instruction that data quality be improved iteratively where it affects the conclusions.
Worked Example: Hotspotting a Coffee Product
A complete pass through all three techniques for one product. All figures are fixed — an illustrative audit record at the snapshot date of this methodology revision, with simple inputs chosen to make the method visible rather than to publish a real coffee footprint.
A cradle-to-gate footprint of 5.00 kg CO₂e per kg of roasted coffee decomposes across five contributors. The analysis runs contribution first, then one-at-a-time sensitivity on the top contributor, then a pedigree-informed uncertainty range. (Figures illustrative; a live PCF would draw factors from the agricultural and energy factor libraries.)
| Step 1 — Contribution | kg CO₂e/kg | Share of total | Cumulative |
|---|---|---|---|
| Green coffee cultivation (incl. N₂O) | 3.05 | 61% | 61% |
| Roasting energy | 1.00 | 20% | 81% |
| Inbound shipping | 0.55 | 11% | 92% |
| Packaging | 0.30 | 6% | 98% |
| Processing water & waste | 0.10 | 2% | 100% |
Contribution verdict: cultivation and roasting together explain 81% of the footprint — the two hotspots. Packaging and water/waste together are 8% and can be left on generic data without affecting any decision. The dominance principle holds cleanly.
| Step 2 — Sensitivity on the cultivation hotspot | Change applied | Effect on 5.00 total | Sensitivity ratio |
|---|---|---|---|
| Fertiliser N₂O factor | +20% | 5.00 → 5.37 (+7.4%) | 0.37 |
| Roasting grid factor | +20% | 5.00 → 5.20 (+4.0%) | 0.20 |
| Allocation: economic vs mass (cherry→green) | method flip | 5.00 → 5.65 (+13.0%) | n/a (choice) |
Sensitivity verdict: the largest single mover is not a parameter at all — it is the allocation choice at the cherry-to-green-bean processing step (+13% from a method flip), which exceeds the +7.4% from a 20% error in the N₂O factor. This is the methodological-choice sensitivity that parameter-only analysis would have missed entirely, and it tells CoffeeCo to document and defend its allocation basis before worrying about fertiliser-factor precision.
Assigning pedigree-derived distributions to the inputs — a wide band on the generic cultivation N₂O factor (dated, regional proxy), a narrow band on the metered roasting energy (current, site-specific) — and propagating over 10,000 iterations yields an illustrative result of 5.00 kg CO₂e/kg with a P5–P95 range of 4.10–6.15. The range is driven almost entirely by the cultivation hotspot’s weak data, confirming where primary-data collection should go next.
Loop verdict: all three techniques point to the same action — collect primary data on green-coffee cultivation and pin down the processing allocation basis. That single refinement would shrink the range and harden the number more than any other available effort. A second pass would re-run contribution on the improved inventory and check whether roasting has become the new leading uncertainty.
Data-Quality-Driven Refinement: Where to Spend Effort
The practical payoff of running all three techniques is a defensible answer to the question every PCF programme faces: with limited budget, which inputs do we measure properly? The answer is a two-dimensional reading — influence (from contribution and sensitivity) against data quality (from the pedigree assessment).
Priority 1 — collect primary data now
These inputs drive the total and rest on weak data, so they dominate both the result and its uncertainty. Every unit of data-collection budget spent here buys the most confidence. The CoffeeCo cultivation hotspot sits here.
Maintain — already solid
Important to the total but already well-measured. No further data effort needed; just keep the data current at each review. CoffeeCo’s metered roasting energy sits here.
Watch — accept the proxy
Weak data, but it barely moves the total, so improving it changes nothing. Leave on a generic factor and document the cut-off. Spending budget here is the classic misallocation.
Ignore — over-invested already
Well-measured but immaterial. If anything, effort here was over-spent; do not add more. A generic factor would have sufficed.
This quadrant is the single most useful artefact a sensitivity-and-hotspot analysis produces, because it converts three abstract techniques into one concrete budget decision. It also pre-empts the most common verifier challenge — “why did you use a generic factor here?” — by showing the answer is deliberate: the input is in the low-influence band, where a proxy is the correct, documented choice.
What the Standards Mandate
Sensitivity and uncertainty analysis are not optional polish — they are requirements wherever they affect the study’s conclusions. The table below maps each technique to where the governing standards require it.
| Standard | Contribution / dominance | Sensitivity | Uncertainty |
|---|---|---|---|
| ISO 14040/14044 | Dominance analysis part of interpretation | Required in interpretation; mandatory for public comparative assertions | Required where it affects conclusions; described qualitatively or quantitatively |
| ISO 14067 | Inherits ISO 14044; data-quality assessment mandated at hotspots | Inherits ISO 14044; carbon-specific choices (biogenic, GWP) tested | Inherits ISO 14044; data-quality scoring expected |
| GHG Protocol Product | Recommended; informs the inventory boundary | Recommended for significant assumptions | Dedicated uncertainty guidance; qualitative + quantitative assessment |
| EU PEF v3.1 | Required; most-relevant processes/flows identified per PEFCR | Prescribed for the most relevant contributors | Data-quality rating mandatory per category rule |
The common thread is conditionality: each technique is required where it matters, which is precisely why contribution analysis comes first — it is the evidence that establishes which parts of the study trigger the sensitivity and uncertainty requirements. Skipping contribution analysis therefore also undermines the defence that sensitivity was unnecessary elsewhere.
Common Errors in Sensitivity and Hotspot Work
Each error below weakens or invalidates the post-calculation analysis, and each appears regularly in PCF verification findings.
| Error | What happens | How to avoid |
|---|---|---|
| Conflating sensitivity and uncertainty | Reporting an OAT swing as if it were a confidence interval, or vice versa — two different claims presented as one. | Keep them separate: sensitivity = “if this changed”; uncertainty = “how wide the believable band is”. |
| Parameter-only sensitivity | Testing input values but never the allocation, boundary, or GWP-basis choices — missing the largest swings. | Always include methodological-choice sensitivity, especially for comparative claims. |
| Hotspotting on weak data | Declaring a hotspot from a low-quality generic factor, then trusting the ranking it produces. | Read hotspots two-dimensionally (influence × data quality); refine the weak high-influence inputs first. |
| Undocumented cut-off | Dropping “small” flows without a stated materiality threshold set in goal & scope. | Fix and disclose the cut-off rule before seeing results. |
| Point-estimate comparison | Claiming Product A < Product B from point totals whose uncertainty intervals overlap. | Compare intervals, not points; no claim if the intervals overlap materially. |
| Monte Carlo on independent inputs only | Ignoring correlations between inputs, understating or overstating the true spread. | Model known correlations; note residual independence assumptions. |
| Stopping the loop too early | Accepting the first range without refining the high-influence, low-quality inputs it exposed. | Iterate the refinement loop until the range supports the decision at hand. |
| Reporting precision the data can’t support | Quoting a footprint to several significant figures when the P5–P95 range spans ±25%. | Round the central estimate to the precision the uncertainty range justifies. |
How This Flows into the Calculator
A PCF calculator can automate the mechanics of all three techniques once the inputs and distributions are supplied — but the judgements that make the analysis meaningful remain the analyst’s. The split is the same one that governs allocation and framework choice across this cluster.
What a calculator handles
Decomposing the total into stage and process contributions and ranking them; running one-at-a-time sensitivity sweeps and computing sensitivity ratios; applying pedigree-derived distributions and running the Monte Carlo propagation; and returning the contribution table, the sensitivity ranking, and the confidence range together.
What you must decide first
The materiality cut-off; which methodological choices to test (allocation, boundary, GWP basis, framework); the pedigree scores for each input; which correlations to model; and — the decision the techniques exist to inform — how narrow the range must be before the result is fit for its purpose. None are computable.
The decisions these techniques test are set on the companion pages: the allocation basis on the PCF allocation methodology page, the framework and recycling rules on the PCF frameworks comparison page, and the boundary on the cradle-to-gate methodology page. Sensitivity and hotspot analysis is what proves whether those upstream decisions actually moved the CO₂e result — and which of them deserve the most scrutiny.
This page closes the PCF methodology cluster: allocation sets the contested decision, framework choice sets the rules, and sensitivity-and-hotspot analysis proves how robust the resulting number is.
PCF methodology series. Core methods: cradle-to-gate, cradle-to-grave, allocation, frameworks comparison, sensitivity & hotspots. Sector methods: food, apparel & textile, electronics, packaging, beverages.
Frequently Asked Questions
They answer different questions and are frequently confused. Sensitivity analysis changes one input or choice at a time and records how far the result moves — it identifies which assumptions matter. Uncertainty analysis propagates the spread in all inputs simultaneously to produce a confidence range around the total — it tells you how wide the believable band is. Sensitivity is “what if this one thing were different?”; uncertainty is “given that everything is imperfectly known at once, how confident can I be?” A complete PCF analysis uses both: sensitivity to find the influential inputs, uncertainty to quantify the resulting spread.
As many as it takes to explain the large majority of the footprint — typically a handful. The dominance principle means a small number of stages or inputs usually account for around 80% of the total, and those are the hotspots. There is no fixed count; the right test is cumulative coverage. Keep adding contributors to the hotspot list until they collectively explain enough of the footprint to support the decisions the PCF informs — often the top three to five. Inputs below the materiality cut-off are documented and set aside, not investigated further.
Not always — the depth of uncertainty analysis scales with the stakes of the decision. A screening footprint used for internal hotspotting can document uncertainty qualitatively (which inputs are weak, how weak) without a full simulation. A footprint underpinning a public comparative claim, or one facing assurance, generally needs quantitative propagation — usually Monte Carlo on pedigree-derived distributions — because the claim depends on whether two products genuinely differ or merely differ within noise. ISO 14044 requires uncertainty to be addressed where it affects the conclusions; how rigorously is proportionate to how much the conclusions depend on it.
Because allocation is a methodological choice that can move the result more than any single parameter, and testing that movement is exactly what methodological-choice sensitivity is for. In the coffee worked example on this page, flipping the processing allocation from mass to economic moved the footprint by 13% — more than a 20% error in the dominant fertiliser factor. Parameter-only sensitivity would have missed this entirely. The lesson is to treat allocation, boundary, and GWP basis as first-class sensitivity objects alongside the input parameters. The allocation decision itself is covered in full on the dedicated PCF allocation methodology page.
The pedigree matrix scores each data point on several quality indicators — reliability, completeness, and temporal, geographical, and technological correlation — on a band scale. Those scores combine into a multiplicative uncertainty factor that sets the width of the distribution placed around the input’s value. A data point that is verified, current, local, and technology-matched gets a narrow distribution; a generic, dated, foreign proxy gets a wide one. This converts a qualitative judgement about data quality into the quantitative input that Monte Carlo propagation needs. It is the mechanism underpinning data-quality scoring in the ecoinvent database and the data-quality requirements of ISO 14067.
On the inputs that are both high-influence and currently low-quality — the priority quadrant. Reading contribution and sensitivity together gives you influence; reading the pedigree assessment gives you data quality. The intersection — important inputs resting on weak data — is where each unit of budget buys the most confidence, because those inputs dominate both the result and its uncertainty. Inputs that are low-influence should stay on generic factors regardless of their data quality, because improving them changes no decision. This two-dimensional reading is the single most useful output of the whole analysis, and it pre-empts the verifier’s “why a generic factor here?” by making the choice demonstrably deliberate.
Yes, and recognising this is the most important discipline in comparative PCF work. If Product A is 10 with a P5–P95 of 8–12 and Product B is 11 with 9–13, their point estimates differ but their confidence intervals overlap heavily — the difference is within noise, and a claim that A is lower-carbon than B is unsupportable. Point estimates carry a false precision that uncertainty analysis strips away. Always compare the ranges, not the points, and run the uncertainty analysis before drafting any comparative claim so the claim is calibrated to what the data can actually support. This is exactly what a green-claims regulator examines.