Estimation of uncertainty and predictive error for multimetric indices. Improving the ecological status of fish communities in inland waters. - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2009

Estimation of uncertainty and predictive error for multimetric indices. Improving the ecological status of fish communities in inland waters.

Didier Pont
  • Fonction : Auteur
  • PersonId : 1093083
Maxime Logez

Résumé

To estimate the predictive error associated with individual and global fish bio-indicator scores, we propose a hybrid approach based on three elements: i) theoretical knowledge on generalized linear model (GLM), ii) simulation procedure and iii) principle of the error propagation. For one single metric, the model provides expected values and standard errors. By extending the classical regression propriety, a random sampling procedure based on normal, expected values and standard errors produces an empirical distribution of the expected values in the link space. After the inverse link transformation, the computation of the standardized distance between the quantile values (e.g. 0.1 and 0.9) and the observed ones provides a good approximation of the predictive intervals. This approach is commonly uses in epidemiological studies to estimate error interval of odds-ratio. Concerning multimetric indices, the computation of theoretical variance and the estimation of predictive error is more complex because it involves the addition of non-independent variables (metrics). To reduce these difficulties, we generalize the previous results obtained for one single metric and we adapt a simulation strategy to estimate an empirical distribution of aggregated scores. For each metric, we randomly generate empirical values from normal distribution based on expected values and expected variances and we calculate the new scores. After the aggregation of the simulated values, the consideration of quantile values (e.g. 0.1 and 0.9) easily completes the construction of the predictive interval. For cyprinid river type index, the size of the 80% tolerance interval is on average close to 0.30 units and it corresponds more or less to one class.
Fichier non déposé

Dates et versions

hal-02592968 , version 1 (15-05-2020)

Identifiants

Citer

Pierre Bady, Didier Pont, Maxime Logez. Estimation of uncertainty and predictive error for multimetric indices. Improving the ecological status of fish communities in inland waters.. International Symposium and EFI+ Workshop,, Mar 2009, Hull, United Kingdom. ⟨hal-02592968⟩
4 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More