Correlation in QRA

Correlation in a quantitative risk analysis model captures the tendency for certain risks or uncertainties to move together. If ground conditions are worse than expected in one area of the site, they are likely to be worse than expected across the whole site — the risks are positively correlated. If the concrete programme slips, the structural steel programme that follows it is likely to slip too — the uncertainties are correlated through the schedule logic. In a Monte Carlo simulation, ignoring these correlations means that each risk is sampled independently: in some iterations, ground conditions are bad but the concrete programme is fine and the structural steel is also fine. That combination is physically implausible, and modelling it as if it were possible compresses the spread of the simulation output below what reality would produce.

The practical consequence is that a model with no correlation will produce a narrower S-curve — apparently lower uncertainty — than the same model with realistic correlation applied. The P80 from an uncorrelated model may be significantly lower than the true P80. This is not a small technical adjustment: on complex programmes where many risks share common drivers (weather, labour market, regulatory environment, a single key supplier), the difference between an uncorrelated and a properly correlated model can be 10–20% of the total contingency requirement.

Setting correlation coefficients is an area where expert judgement is required. Most QRA tools allow analysts to specify correlation coefficients between pairs of risk items or between activity duration ranges — values between −1 (perfectly negatively correlated) and +1 (perfectly positively correlated), with 0 meaning independent. For most project risks, positive correlations of 0.5–0.8 are appropriate for risks sharing a common driver. A simple practical approach is to group risks by their root cause — weather risks, labour risks, design risks, regulatory risks — and apply uniform moderate positive correlation within each group, reflecting the fact that a bad year for weather will affect all weather-sensitive activities simultaneously.