Skip to content
CGIAR Climate Action CGIAR Climate Data Hub — Wikis

Bias correction — what it does and why it matters

What this page covers

What bias correction means, the three methods you’ll encounter in African adaptation work, why methodologists are sometimes nervous about it, and what to know when citing a bias-corrected product.

What is bias?

To be written. A model’s raw output systematically differs from observations — too cold, too wet, wrong seasonality. Bias is the systematic part of the model-vs-observation gap, separate from natural variability.

The two-step transformation

To be written. (1) Map raw model historical to observed historical, learning the bias function. (2) Apply the same function to model future output. Result: future projection adjusted for the model’s historical bias.

Methods you’ll encounter

To be written.

  • Delta change — simplest. Add the model’s projected anomaly to observed historical climate. Conservative; preserves observed climatology.
  • Quantile mapping — map each quantile of the model distribution to the corresponding quantile of the observation distribution. The “BCSD” / standard method (Wood 2004; Cannon 2015).
  • ISIMIP3BASD (Lange 2019) — trend-preserving variant of quantile mapping. Preserves the model’s projected change while correcting historical bias. Used by ISIMIP3b and CHELSA-CMIP6.

Why methodologists are sometimes nervous

To be written. Sources: Ehret et al. 2012; Maraun 2016. Key points:

  • Bias correction assumes the bias function is stationary — i.e. that the model’s behaviour relative to observations stays the same in a warmer world. Hard to verify.
  • Bias correction can break physical consistency between variables (a precipitation correction may not match the temperature correction; the corrected fields can be physically incoherent).
  • Over-correction risks: if the observed reference dataset has its own biases, you’re correcting toward the wrong target.

What the Atlas uses

To be written. NEX-GDDP-CMIP6 uses BCSD against the GMFD reference. ISIMIP3b / CHELSA-CMIP6 use ISIMIP3BASD against W5E5. Cross-link to Dataset landscape.

Practical guidance

To be written. When you cite a bias-corrected dataset, the bias-correction method matters less than understanding it exists. The honest framing in a proposal: “We use NEX-GDDP-CMIP6, statistically downscaled and bias-corrected against the GMFD reference.”

Further reading