Model Sensitivity Checks

A sensitivity check asks a simple question: if one input were different, how much would the result move. It is the most useful check a non-specialist can ask for, because it separates robust conclusions from fragile ones. A conclusion is robust if it holds across the plausible range of uncertain inputs; it is fragile if it depends on one input taking exactly one value. Retrofit decisions should prefer robust conclusions.

Consider a home where wall construction is uncertain. If the recommended measure looks best whether the wall is solid or cavity-filled, the recommendation survives the uncertainty. If it looks best only under one wall assumption and worst under the other, the wall question must be resolved first. Sensitivity checking does not remove uncertainty; it shows where uncertainty matters and where it does not.

Simple sensitivity checks are enough for most retrofit decisions. Change one input at a time across its plausible range, for example wall U-value, window type, heating system efficiency or ventilation rate, and record the result each time. Presenting those results as a small table or a range is far more honest than a single figure with no indication of spread. A result quoted as a range, with the inputs that drive the range named, is easier to act on than a precise figure built on guesses.

More formal approaches exist, including varying several inputs together or running many combinations, and they suit complex buildings or high-value decisions. For a typical home, one-at-a-time checks on the three or four most uncertain inputs usually reveal the decision-relevant uncertainty. The goal is not statistical sophistication; it is knowing whether the decision would change if the guesses were wrong.

Ask the modeller for a sensitivity table for the inputs you flagged as uncertain, and check the ranges used. A range should reflect plausible reality, not an extreme designed to make a point: a wall U-value range should span the constructions actually possible for that home, and a heating efficiency range should span the condition actually possible for that system. If a range looks too narrow, the check understates uncertainty; if it looks absurdly wide, it overstates it.

Then test the decision, not just the number. For each plausible input combination, ask whether the recommended measure still ranks first, still looks worthwhile, or still suits the building. A measure that ranks first in most combinations but not all is a reasonable choice with a noted risk; a measure that ranks first in only one combination is a gamble on that combination being right.

Sensitivity practice is similar across EU and UK systems, because it is a property of modelling rather than of one method. SAP-based advice in the UK and national-method advice in EU member states both benefit from showing ranges where inputs are uncertain, even though the formal requirements for such checks differ. Where a certificate is concerned, the official result remains the single standardised figure; sensitivity information sits alongside it as advice, and should not be confused with the certificate itself.

For complex or high-value projects, a professional may also test interactions, where two uncertain inputs change together. That is worth requesting when measures interact strongly, for example insulation plus a heating system change, because the combined effect can differ from either effect alone. For straightforward choices, one-at-a-time checks remain sufficient.

A common mistake is to treat a sensitivity check as a way to pick the most flattering input. Running checks and then reporting only the best case is not sensitivity analysis; it is cherry-picking. The honest output shows the full range and names the input values that produce each end of it. Another mistake is to run sensitivity checks on minor inputs while leaving a major uncertain input untested, which produces a reassuringly narrow range that hides the real risk.

A further mistake is to demand certainty that modelling cannot provide. A sensitivity check that shows a wide range is not a failed model; it is a successful diagnosis of what needs to be measured or verified. The correct response to a wide range is investigation, not a more detailed model built on the same unknowns.

Keep the sensitivity table with the model report, with the plausible range for each input and its evidence noted. When a later decision revisits the same home, that table shows immediately which uncertainties were already tested and which remain open.

Questions to ask include: which input does this result depend on most, over what plausible range did you test it, and would my decision change anywhere in that range. If the answer to the last question is yes, the next step is evidence, not commitment.