The standard approach to flood frequency analysis — fit a log-Pearson III, read off the quantiles, report confidence intervals — is familiar to every practitioner. It works. But it carries assumptions that Bayesian methods handle more honestly.
This post is a practitioner-level introduction: what Bayesian FFA offers, when it matters, and what the workflow looks like.
The Classical FFA Problem
Given a record of annual maximum flows, we want to estimate the magnitude of the 1% AEP (100-year) flood. The classical (frequentist) approach:
- Fit a parametric distribution (LP3 in Australia per ARR)
- Estimate parameters using method of moments or MLE
- Derive confidence intervals from the parameter covariance matrix
The fundamental limitation is data scarcity. A 30-year record is considered good for Australian catchments. But we’re extrapolating to the 0.1% or even 0.01% AEP — events with return periods 10–100x longer than our record. The uncertainty in those estimates is enormous, and classical confidence intervals tend to understate it.
What Bayesian FFA Adds
Bayesian methods treat the distribution parameters as random variables with probability distributions rather than fixed unknowns. This has two practical consequences:
1. Prior information can be formally incorporated. Regional flood frequency data, paleoflood evidence, or expert judgment about likely parameter ranges can be encoded as prior distributions. The posterior combines this with your gauged record.
2. Uncertainty propagates correctly. Because you’re sampling from a joint posterior distribution over all parameters, the uncertainty in your quantile estimates naturally reflects both record length and extrapolation distance.
The result is quantile estimates with credible intervals — not confidence intervals. The distinction matters: a 90% credible interval means “there is a 90% probability the true value lies here, given the data and priors,” which is actually what practitioners and planners want to know.
A Simple Example in R
library(nsRFA)
library(rstan)
# Fit LP3 via Bayesian MCMC (using Stan or JAGS)
# This sketch uses the log-space GEV for illustration
annual_max <- c(240, 380, 195, 440, 310, ...) # your data
# Stan model would encode:
# mu ~ Normal(prior_mean, prior_sd)
# sigma ~ Half-Normal(0, prior_scale)
# xi ~ Normal(0, 0.2) # shape — informative prior from regional data
# y ~ GEV(mu, sigma, xi)
# Posterior samples give you the full distribution over quantiles
I’ve written more detail on this in my Bayesian Coin Toss post which walks through the mechanics of Bayesian updating — a useful foundation before tackling flood frequency.
When Is It Worth It?
Bayesian FFA is most valuable when:
- Record length is short (< 20 years) and regional prior information is available
- Extrapolation is extreme (estimating 0.01% AEP events for dam safety)
- Uncertainty communication matters (clients or regulators need to understand risk, not just a point estimate)
- Multiple data sources exist — historic floods, botanical evidence, paleoflood records that can’t be incorporated in classical methods
For routine flood impact assessments with adequate records, the added complexity is usually not warranted.
The Regulatory Landscape
ARR 2019 acknowledges Bayesian methods but stops short of prescribing them. In Queensland, PFRA guidelines are increasingly receptive to uncertainty-aware approaches, particularly for infrastructure design. The practical constraint is that regulators need to review and accept the method — which means documentation and plain-language explanation are as important as the statistics themselves.
More detail on my flood frequency analysis work is in the Data Science section. For questions on specific applications, connect via LinkedIn.