Skip to contents

Note

Une version française de cet article est disponible : Principes méthodologiques de l’analyse d’incertitude et de sensibilité.

Context and scope

Cost-benefit analysis based on avoided damages (CBA) is widely used to appraise flood management projects. In France, it is the recommended approach for appraising flood management projects involving structural measures whose investment exceeds 2.5 million euros, where a global sensitivity analysis is a mandatory step.

Estimating the main CBA indicator — Net Present Value (NPV) — requires chaining several modeling stages:

  • hydrological modeling to characterize flood scenarios;
  • hydraulic modeling to spatialize flood scenarios;
  • digital terrain modeling to represent topography and protection structures;
  • GIS processing to locate exposed stakes;
  • classification of stakes according to their nature and vulnerability;
  • vulnerability modeling via damage functions.

Each stage relies on assumptions and input data that introduce uncertainty into the resulting NPV. Performing a full uncertainty analysis across the entire hydrological-hydraulic-damage chain is computationally prohibitive for routine operational appraisals. However, as formalized in Grelot (2017) (GrelotF2017b), a practical first assessment can be carried out by building a simplified meta-model around the intermediate outputs of the nominal CBA.

From floodam.cba.sa to floodam.cba

Originally implemented in the CBA-DE-AS scripts and later in the standalone package floodam.cba.sa, this meta-model is now integrated directly into floodam.cba (from version 1.3.0.0). It reuses the nominal intermediate quantities already computed in a Case object to perform:

  1. Verification of nominal CBA indicators (npv and benefit);
  2. Monte Carlo uncertainty analysis (UA) by sampling uncertain input parameters and evaluating the empirical distribution of npv and benefit;
  3. Global sensitivity analysis (SA) by estimating first-order (SiS_i) and total (TiT_i) Sobol indices using the sensitivity package (by default sensitivity::sobolmartinez).

Because floodam.cba does not rerun hydraulic simulations or spatial damage intersections (which are handled by upstream tools such as floodam.spatial), uncertainty on spatial hazard and vulnerability is represented through aggregated multipliers at the flood-scenario and stake-type levels.

Meta-model formulation

Net Present Value (npv)

In floodam.cba, the Net Present Value of a project is evaluated over a time horizon HH from normalized annual schedules of benefit activation and cost disbursements:

NPV=Benefit⋅∑t=0Hwtbendt−∑i∈ICi∑t=0Hwi,tcostdt\text{NPV} = \text{Benefit} \cdot \sum_{t=0}^{H} w_t^{\text{ben}} \, d_t - \sum_{i \in I} C_i \sum_{t=0}^{H} w_{i,t}^{\text{cost}} \, d_t

where:

  • Benefit\text{Benefit} is the expected annual avoided damage (EAAD, or benefit) once the project is fully operational;
  • wtben∈[0,1]w_t^{\text{ben}} \in [0, 1] is the normalized benefit activation ratio in year tt (derived from the project’s state calendar);
  • CiC_i is the total nominal or sampled cost for cost item i∈Ii \in I (investment or annual maintenance for each project action);
  • wi,tcostw_{i,t}^{\text{cost}} is the annual disbursement weight for cost item ii in year tt (normalized so that ∑twi,tcost=1\sum_t w_{i,t}^{\text{cost}} = 1);
  • dt=∏u=0t(1+ru)−1d_t = \prod_{u=0}^{t} (1 + r_u)^{-1} is the discount factor at year tt for discount rate schedule rur_u (computed by discount_weight()).

Note

In the current implementation (1.3.0.0), project costs CiC_i are evaluated assuming zero costs in the counterfactual reference situation (reference). Differential cost accounting between project and reference is planned for a future release.

Expected annual avoided damage (benefit)

Expected annual avoided damage is approximated by trapezoidal integration of avoided damages along exceedance frequency fs=1/Tsf_s = 1 / T_s:

Benefit=∑s∈Sps(𝑻)∑j∈Jij,s⋅αj⋅hs\text{Benefit} = \sum_{s \in S} p_s(\mathbf{T}) \sum_{j \in J} i_{j,s} \cdot \alpha_j \cdot h_s

where:

  • SS is the set of flood scenarios, ordered by increasing return period TsT_s;
  • ps(𝑻)p_s(\mathbf{T}) is the trapezoidal integration weight of scenario ss computed from the vector of return periods 𝑻\mathbf{T} via calculate_weight();
  • JJ is the set of stake categories;
  • ij,si_{j,s} is the nominal avoided damage for stake jj and flood scenario ss between the project’s initial and final states (which may be positive or negative);
  • αj\alpha_j is a multiplicative uncertainty factor on damage estimation for stake jj (nominal value αj=1\alpha_j = 1);
  • hsh_s is a modeling proxy representing uncertainty on the project’s hydraulic efficiency for flood scenario ss (nominal value hs=1h_s = 1).

Note

If Benefit<0\text{Benefit} < 0 for a specific stake or scenario, the project induces a net disbenefit rather than an avoided damage. This has no effect on NPV\text{NPV}, though it can affect benefit-cost ratio interpretations.

Flood scenario weights

Assuming N≥2N \ge 2 flood scenarios sorted in increasing order of return period T1<T2<…<TNT_1 < T_2 < \dots < T_N (with exceedance frequencies fs=1/Tsf_s = 1 / T_s), the trapezoidal weights ps(𝑻)p_s(\mathbf{T}) are given by:

  • p1=f1−f22p_1 = \frac{f_1 - f_2}{2}
  • ps=fs−1−fs+12p_s = \frac{f_{s-1} - f_{s+1}}{2} for 1<s<N1 < s < N
  • pN=fN−1−fN2p_N = \frac{f_{N-1} - f_N}{2}

Because return periods TsT_s are themselves subject to uncertainty, the weights ps(𝑻)p_s(\mathbf{T}) are recomputed dynamically for each Monte Carlo draw. To preserve the physical ordering T1<T2<…<TNT_1 < T_2 < \dots < T_N, prepare_flood_range() automatically truncates overlapping return-period bounds at the midpoint between adjacent nominal return periods.

Uncertainty modeling and Sobol indices

Uncertainties are assigned to four parameter families:

  1. Project costs (CiC_i, investment and maintenance per action);
  2. Return periods (TsT_s, per flood scenario);
  3. Hydraulic efficiency (hsh_s, per flood scenario);
  4. Damage estimations (αj\alpha_j, per stake category).

Each uncertain parameter is modeled as an independent uniform random variable between a lower bound min and an upper bound max (specified either as relative variations around the nominal value or as absolute bounds).

Global sensitivity analysis decomposes the output variance Var(Y)\text{Var}(Y) (for Y=NPVY = \text{NPV} or Y=BenefitY = \text{Benefit}) into contributions from individual parameters and their interactions:

  • First-order Sobol index Sk=Var(𝔼[Y∣Xk])Var(Y)S_k = \frac{\text{Var}(\mathbb{E}[Y \mid X_k])}{\text{Var}(Y)}: fraction of output variance explained by parameter XkX_k alone;
  • Total Sobol index Tk=1−Var(𝔼[Y∣X−k])Var(Y)T_k = 1 - \frac{\text{Var}(\mathbb{E}[Y \mid X_{-k}])}{\text{Var}(Y)}: fraction of output variance explained by XkX_k including all its interactions with other parameters.

Reference

  • Grelot, F. (2017). Méthode simplifiée d’analyse d’incertitude et de sensibilité globale pour les analyses coût-bénéfice des projets de gestion des inondations (GrelotF2017b).