📊 statistics

Regression approaches for modelling genotype-environment interaction and making predictions into unseen environments

This paper reviews and unifies various linear mixed model approaches for modeling genotype-environment interaction and predicting performance in unseen environments, demonstrating how methods ranging from Finlay-Wilkinson regression to kernel-based techniques fit within a common framework for assessing prediction uncertainty, illustrated with a long-term rice trial dataset from Bangladesh.

Original authors: Maksym Hrachov, Hans-Peter Piepho, Niaz Md. Farhat Rahman, Waqas Ahmed Malik

Published 2026-03-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Maksym Hrachov, Hans-Peter Piepho, Niaz Md. Farhat Rahman, Waqas Ahmed Malik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a farmer trying to decide which variety of rice to plant next year. You have data from past years showing how different rice varieties performed in different fields. But here's the catch: next year's weather is a mystery, and the specific field you want to plant in might be a brand-new location you've never tested before.

This paper is like a guidebook for statisticians and plant breeders trying to solve that exact problem. They are asking: "How can we use what we know about the past (weather, soil, location) to predict how a specific rice seed will perform in a future, unseen world, and how sure can we be about that prediction?"

Here is the breakdown of their journey, using some everyday analogies.

1. The Problem: The "One-Size-Fits-All" Trap

Traditionally, breeders looked at how a rice variety did on average. But rice is like a person: some people thrive in the rain, others in the sun. A variety that is great in a wet year might fail in a dry one. This is called Genotype-by-Environment Interaction.

The authors wanted to move beyond simple averages. They wanted to build a "smart map" that says: "If it rains 20mm more than usual, this specific rice variety will grow 5% taller."

2. The Toolkit: Different Ways to Draw the Map

The paper reviews several mathematical "recipes" (models) to draw this map. Think of these as different ways to navigate a city:

The Basic Map (Baseline): Just looks at the average performance. It's like saying, "This car gets 30 miles per gallon," without knowing if you're driving in the city or on a mountain.
The Weather-Dependent Map (Factorial Regression): This model says, "The car's mileage changes based on the road type." It connects the rice's performance directly to specific weather data (like temperature or rain).
The "Synthetic" Map (Reduced Rank Regression / FW-US): Sometimes, you have too many weather variables (rain, wind, humidity, soil temp, etc.), and it gets messy. This method creates a "super-weather" variable. Imagine combining "rain," "humidity," and "cloud cover" into a single score called "Moisture Index." It simplifies the map without losing the important details.
The "Similarity" Map (Kernel Approach): Instead of measuring the weather directly, this looks at how similar the new field is to old fields. It's like saying, "This new field feels just like Field A and Field B, so the rice will probably do well there."

The Big Discovery: The authors found that these seemingly different maps are actually just different angles of the same mountain. They are all mathematically linked!

3. The Twist: Predicting the Unknown (The "Crystal Ball" Problem)

This is the most creative part of the paper. Most studies test their models by looking at past data they already have. But in real life, you don't have the weather data for next year yet.

The authors defined four scenarios for prediction, like different levels of a video game:

The Long-Term Average: Predicting how the rice does in a "typical" year at a known farm. (Easy mode).
The New Year at a Known Farm: Predicting a specific future year, but we don't know the exact weather yet. We only know the average weather for that farm. (Medium mode).
The New Farm: Predicting how it will do at a brand-new farm, but using the long-term average weather for that farm. (Hard mode).
The New Farm, New Year: The ultimate challenge. A new location, in a future year, with unknown weather. (Expert mode).

The Analogy: Imagine you are betting on a horse race.

Scenario 1: You know the horse and the track.
Scenario 4: You are betting on a horse running on a track you've never seen, on a day next year, and you don't know if it will rain or shine.

4. The Innovation: Measuring the "Wobble" (Uncertainty)

The authors didn't just want to make a prediction; they wanted to know how shaky that prediction is.

If I tell you, "This rice will yield 5 tons," that's a number. But if I say, "It will likely be between 4 and 6 tons, but there's a 20% chance it could be 3 tons because next year's rain is a mystery," that is uncertainty.

They developed a new mathematical way to calculate this "wobble."

Old way: "We guessed the weather, so our guess is perfect." (Naive).
New way: "We guessed the weather, but since the weather is random, our guess has a built-in error margin. Here is exactly how big that error margin is."

They realized that if you don't account for the fact that future weather is unknown, you are overconfident. Their new method adds a "safety buffer" to the prediction, telling breeders, "Don't bet the farm on this one; the uncertainty is too high."

5. The Real-World Test: Rice in Bangladesh

They tested all these ideas on real rice data from Bangladesh (a country where rice is life).

The Result: Using weather data (environmental covariates) generally helped predict better than just looking at averages.
The Surprise: The "Synthetic" maps (simplifying the weather data) worked almost as well as the complex ones, but were easier to use.
The Lesson: While the models improved predictions, the improvement wasn't huge. Why? Because the weather data they used was a bit "blurry" (low resolution). It's like trying to navigate a city with a map that only shows major highways but misses the side streets. To get truly perfect predictions, we need sharper, more detailed weather data.

Summary: What Should You Take Away?

Context is King: You can't predict how a plant will do without knowing the environment (weather, soil).
Simplicity Wins: You don't always need a super-complex model; sometimes a simplified "super-variable" works just as well.
Confidence Matters: The most important contribution of this paper is a new way to say, "Here is our prediction, and here is exactly how much we should doubt it."
Data Quality: Better weather data leads to better predictions. If your map is blurry, your navigation will be shaky.

In short, the authors built a better compass for plant breeders, one that not only points the way but also tells you how much the wind might blow you off course.

1. Problem Statement

In plant breeding and variety testing, predicting the performance of genotypes in Target Populations of Environments (TPE)—specifically new locations and future years—is a critical challenge. Traditional Multi-Environment Trial (MET) models often treat environments as random effects without explicitly modeling the underlying environmental drivers.

The Gap: While using environmental covariates (EC) to explain Genotype $\times$ Environment (G $\times$ E) interaction is common, most existing validation methods (e.g., leave-one-environment-out cross-validation) assume the EC values for the "unseen" environment are known. In reality, for future years or new farms, exact EC values are unknown; only long-term means or distributions are available.
The Need: There is a lack of a unified framework that:
1. Links disparate regression methods (Factorial Regression, Kernel approaches, Reduced Rank Regression) under a common model-based prediction framework.
2. Provides rigorous methods to estimate prediction uncertainty when EC values must be estimated or are unknown (i.e., accounting for the variance of the covariates themselves).

2. Methodology

A. Theoretical Framework

The authors propose a unified Linear Mixed Model (LMM) framework where environments are treated as random. The core model extends the Factorial Regression (FR) approach:
$y_{ij} = \eta_{ij} + u_j + e_{ij}$
Where $\eta_{ij}$ is the expected response for genotype $i$ in environment $j$ , modeled as a regression on ECs ( $x_{jk}$ ).

Genotype Modeling: Genotypes are modeled as random factors with genotype-specific intercepts ( $\alpha_i$ ) and slopes ( $\gamma_{ik}$ ).
Variance-Covariance Structures: The paper compares four specific implementations based on how the variance-covariance matrix ( $\Sigma$ $Σ$ ) of the random slopes is structured:
1. Random Factorial Regression (RFR): Unstructured $\Sigma$ (most flexible, many parameters).
2. Environmental Kernel Approach: Assumes a diagonal $\Sigma$ (ridge regression on ECs), leading to a kernel matrix $K_E = XX^T$ .
3. Reduced Rank Regression (RRR): Approximates $\Sigma$ using a low-rank factor-analytic structure ( $\Sigma = \Lambda\Lambda^T$ ). This creates Synthetic Covariates (SC).
4. Extended Finlay-Wilkinson (FW): A two-stage approach where synthetic covariates are derived first, then used in a regression with fixed genotype effects.

B. Prediction Scenarios and Uncertainty Quantification

A key contribution is the definition of four distinct prediction scenarios based on the availability of EC data:

Case 1: Long-term mean in the TPE (Known mean, no EC variance).
Case 2: New year at the TPE mean (EC variance due to year effects).
Case 3: Long-term mean at a new location (Known location mean, no EC variance).
Case 4: New year at a new location (Unknown future EC values; requires EC variance components).

New Approach for Prediction Variance:
The authors derive a novel estimator for the total prediction variance ( $\vartheta_i$ ), which accounts for:

Estimation Error: Uncertainty in estimating the regression coefficients ( $\hat{\gamma}$ ) and the covariate means ( $\hat{\xi}$ ).
Covariate Uncertainty: The variance of the EC itself ( $\sigma_x^2$ ) when predicting into a future year/location.
Residual Deviation: The variance of deviations from the regression line (G $\times$ L, G $\times$ Y, G $\times$ L $\times$ Y).

The formula for the variance of a product of random variables is used to correct for bias in the naive plug-in estimators:
$est. var(\hat{\gamma}\hat{\xi}) = \hat{\gamma}^2 var(\hat{\xi}) + \hat{\xi}^2 var(\hat{\gamma}) - var(\hat{\gamma})var(\hat{\xi})$

C. Data and Implementation

Dataset: Long-term rice variety trials from the Bangladesh Rice Research Institute (BRRI), covering 2001–2022. Two datasets: Irrigated Winter Rice and Rainfed Summer Rice.
Covariates: 8 weather variables (temperature, precipitation, radiation) aggregated by growing season from the AgERA5 database.
Software: Models fitted using ASReml-R 4.2.
Validation:
- LOEO (Leave-One-Environment-Out): Uses observed EC values (optimistic).
- LYLO (Leave-One-Year-and-Location-Out): Uses long-term mean EC values for the left-out location/year (realistic/unseen scenario).

3. Key Contributions

Unification of Methods: Demonstrates that RFR, Kernel approaches, RRR, and Extended Finlay-Wilkinson regression are mathematically linked within a single LMM framework, differing primarily in their variance-covariance assumptions.
Realistic Prediction Framework: Moves beyond standard cross-validation by explicitly modeling the uncertainty of future environmental conditions (Case 4), rather than assuming known covariates.
Novel Variance Estimator: Introduces a bias-corrected method to estimate the total prediction variance, incorporating the uncertainty of the environmental covariates themselves. This allows breeders to quantify confidence intervals for recommendations in unseen environments.
Synthetic Covariates: Validates the use of Reduced Rank Regression (RRR) to derive synthetic covariates, which often provide a parsimonious and robust alternative to using raw, high-dimensional weather data.

4. Results

Model Fit:
- Models incorporating ECs generally reduced the residual G $\times$ L and G $\times$ Y variance components compared to the baseline (no ECs).
- RRR2 (Rank 2) and FW1-US (1 Synthetic Covariate) achieved the lowest Akaike Information Criterion (AIC), balancing fit and complexity.
- The "Main EC Effect" (fixed slope) did not significantly improve AIC in all cases but was crucial for unbiased predictions in unseen environments.
Predictive Performance (Cross-Validation):
- LOEO Scenario: EC models slightly outperformed the baseline, but differences were marginal.
- LYLO Scenario (Unseen Environments): The inclusion of ECs provided more consistent improvements. Notably, models with the main EC effect performed better in LYLO, confirming that modeling the mean regression is essential when future EC values are unknown.
- Winter vs. Summer Rice: Summer rice (rainfed) showed clearer benefits from EC modeling. Winter rice (irrigated) showed weaker links between weather covariates and yield, likely due to irrigation buffering environmental stress.
Variance Estimation:
- The model-based prediction variance (MVP) closely approximated the Mean Squared Prediction Error (MSPE) from cross-validation.
- The new variance estimator successfully captured the increased uncertainty associated with predicting into new years/locations (Case 4), where the variance of the covariates contributes significantly to the total error.

5. Significance and Implications

For Plant Breeders: This work provides a statistically rigorous method to quantify the risk of recommending a variety for a new location or future season. It moves beyond point estimates to provide confidence intervals that account for weather uncertainty.
Methodological Advancement: By linking disparate regression techniques, the paper clarifies the trade-offs between model complexity (RFR) and parsimony (Kernel/RRR). It suggests that Reduced Rank Regression or Extended Finlay-Wilkinson with synthetic covariates often offers the best balance of accuracy and stability.
Data Requirements: The study highlights that while ECs improve predictions, the quality of the data is paramount. The modest gains in the irrigated winter rice dataset suggest that coarse weather data may not capture micro-climatic nuances (like irrigation management) critical for certain crops.
Future Directions: The framework is extensible to include genomic data (via kinship matrices) and non-linear terms. The authors call for the availability of "variance rasters" for environmental covariates to further refine uncertainty estimates.

In summary, this paper establishes a robust, model-based foundation for enviromics, enabling breeders to make more informed, risk-aware decisions when deploying varieties into untested environments.