🧬 biology

Hydroclimate associations do not ensure temporal or spatial transferability of positive Anopheles catch intensity in southwestern Colombia

Despite identifying statistically significant hydroclimate associations with Anopheles mosquito catch intensity in southwestern Colombia, the study demonstrates that these relationships lack robust temporal and spatial transferability, highlighting the need for multiplicity-aware inference and blocked validation before operational deployment of climate-informed vector models.

Original authors: Johan Ceballos, Andrea Camila Silva Cordero

Published 2026-10-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Johan Ceballos, Andrea Camila Silva Cordero

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Mosquitoes are more than just a summer nuisance; in many parts of the world, they are the primary carriers of diseases like malaria. To stop these diseases, public health officials need to know when and where mosquito populations will surge. For decades, scientists have looked to the weather for answers, knowing that temperature and rain act as powerful levers for mosquito life. Warmth can speed up their growth, while rain creates the pools of water where they lay their eggs. It seems logical that if we can track the weather, we can predict the mosquitoes. However, nature is rarely simple. A relationship that holds true in one village might vanish in the next, or work in January but fail in July. The challenge lies in knowing whether a pattern spotted in historical data is a reliable rule or just a temporary coincidence.

In the humid lowlands of southwestern Colombia, a team of researchers set out to test this very idea. They wanted to see if they could use daily weather data to predict the intensity of mosquito catches in rural communities. They focused on two specific types of mosquitoes: Anopheles albimanus, which is common along the coast and tolerates many environments, and Anopheles neivai, which is closely tied to forest plants and mangroves. The team gathered a massive collection of records from 25 different villages, covering over 229,000 mosquitoes caught between 2021 and 2023. They linked every catch to the specific temperature, humidity, and rainfall recorded on those days. Their goal was to build a model that could not only explain past catches but also predict future ones in different places and different years.

The researchers approached this with a strict set of rules to ensure their findings were real and not just lucky guesses. They split their data into three parts: they used the records from 2021 to build their models, the records from 2022 to check if the models worked, and the records from the first three months of 2023 to see if the models could predict the future. They also tested whether the models could work in villages they had never seen before. This rigorous testing is crucial because a model that fits past data perfectly often fails when faced with new situations. They also had to be careful with their statistics, checking that the patterns they found were strong enough to stand up to the fact that they were testing many different weather factors at once.

When they looked at the results, the story was mixed and revealed a surprising complexity. For the forest-dwelling Anopheles neivai, the weather data showed a statistically detectable signal: higher temperatures and rainfall in the two weeks leading up to a catch were associated with more mosquitoes. This makes biological sense, as rain fills the bromeliad plants where these mosquitoes breed. However, for the coastal Anopheles albimanus, the picture was less certain. While the data suggested that cooler temperatures might be linked to more mosquitoes, this pattern was not strong enough to be considered a solid rule after accounting for all the other factors they tested.

The most critical finding, however, came when they tried to use these weather-based models to predict the future. Even for the species where the weather links seemed statistically detectable, the models struggled when applied to new time periods or new locations. The study concluded that detectable hydroclimate associations did not ensure temporal or spatial transferability. When the team tried to predict the mosquito counts for early 2023 using models trained on 2021 data, the predictions were often wildly off. For the coastal mosquito, the model predicted a catch rate that was more than six times higher than what was actually observed. Similarly, when they tried to predict catches in villages they had not included in their training, the models could only rank the villages in the correct order about one-third of the time. Furthermore, the association between rainfall and An. neivai was not robust under all conditions; when the analysis was restricted to shorter time intervals, the statistical confidence in this link weakened and crossed zero.

This outcome suggests that while weather is undeniably connected to mosquito activity, it is not a standalone crystal ball. The researchers found that the relationship between rain, heat, and mosquitoes is deeply local and changes over time. Factors that the weather data could not capture—such as changes in how the mosquitoes were being caught, local interventions to control them, or subtle differences in the landscape—played a huge role in the actual numbers. The study concludes that simply finding a statistical link between the weather and mosquitoes is not enough to build a reliable surveillance tool. Before such models can be used to guide public health decisions, they must prove they can work consistently across different years and different places, surviving the test of time and geography. Until then, the weather remains a piece of the puzzle, but not the whole picture.

Technical Summary: Hydroclimate Associations and Transferability of Anopheles Catch Intensity in Southwestern Colombia

Problem Statement
Biometeorological associations are frequently utilized as potential surveillance signals for vector abundance. However, statistical detection of an association within a specific dataset does not guarantee that the relationship will transfer across different years (temporal transferability) or geographic sites (spatial transferability). In the Colombian Pacific lowlands, characterized by high rainfall, humidity, and ecological heterogeneity, Anopheles albimanus and Anopheles neivai are key malaria vectors with distinct ecological niches. While previous studies have noted links between climate and vector abundance, simple extrapolation is hazardous because associations may be context-dependent. Furthermore, public biodiversity archives often contain only "positive" records (presence-only data), which, if treated as if zero counts were observed, violate sampling assumptions and alter the estimand. This study addresses the gap between detecting hydroclimate associations and establishing their operational utility for prediction outside the training environment.

Methodology
The authors conducted a secondary, observational analysis of 1,502 openly published positive-catch records (representing 229,394 mosquitoes) collected between January 2021 and March 2023 across 25 rural localities in Guapi, Timbiquí, and Santa Bárbara. The data were linked to daily meteorological variables (temperature, humidity, precipitation) from NASA POWER, mapped to three 0.5° grid cells.

Statistical Approach: The study employed species-specific zero-truncated negative binomial (ZTNB) models. This likelihood was chosen because the dataset lacks zero-catch events; the model conditions on a positive catch to estimate variation in positive-catch intensity rather than occurrence probability or total abundance.
Predictors: Models included standardized event temperature, relative humidity, mean precipitation, and 14-day antecedent rainfall, alongside seasonal harmonics and administrative indicators.
Validation Strategy: To rigorously test transferability, the authors avoided random cross-validation in favor of:
1. Frozen Temporal Evaluation: Training on 2021 data, validating on 2022, and testing on January–March 2023.
2. Spatial Transfer: A "leave-one-sampling-unit-out" approach where models were refitted excluding one parent event (sampling unit) to predict the omitted unit.
Inference: The study applied Benjamini–Hochberg false-discovery-rate (FDR) correction across eight species-by-climate coefficients to account for multiplicity. Cluster-robust sandwich covariance matrices were used to handle within-unit dependence.

Key Results

Coefficient Estimation:
- Anopheles neivai: Two associations remained robust after FDR correction: a positive association with event temperature (standardized $\beta = 0.453$ , $q = 0.047$ ) and 14-day antecedent rainfall ( $\beta = 0.132$ , $q = 0.044$ ).
- Anopheles albimanus: A negative association with event temperature was nominally supported ( $\beta = -1.482$ ) but did not survive FDR correction ( $q = 0.079$ ).
Predictive Transfer:
- Temporal: The inclusion of hydroclimate variables did not consistently improve predictive performance over a seasonal-administrative baseline. For An. albimanus, the full model significantly overpredicted the aggregate catch in the 2023 test period (by approximately 6.2-fold), and error metrics (RMSE, MAE) were worse than the baseline. For An. neivai, the full model showed mixed results, improving some metrics but failing to consistently outperform the baseline in rank correlation or log scores.
- Spatial: Spatial rank correlations were modest for both species (An. albimanus: 0.342; An. neivai: 0.358), indicating limited ability to generalize the fitted relationships to new sampling units.
Sensitivity: Sensitivity analyses revealed that coefficient stability was sensitive to event duration and data cleaning (e.g., removing coordinate-imputed rows), with some confidence intervals crossing zero under stricter constraints.

Significance and Claims
The paper's primary contribution is the empirical demonstration that detectable hydroclimate associations do not ensure temporal or spatial transferability of positive catch intensity. The authors argue that successful internal model fit and random cross-validation are insufficient for establishing geographic or temporal transferability.

Key claims include:

Operational Caution: A statistically significant climate coefficient should not be treated as an operational surveillance signal without structured out-of-sample evaluation. The study found that while hydroclimate terms improved in-sample fit (lower AIC), they did not yield consistent predictive superiority in forward temporal or spatial tests.
Data Limitations: The reliance on positive-only data prevents estimation of zero-catch probability, total population abundance, or intervention efficacy. The study explicitly states it cannot support claims about malaria transmission risk or causal climate effects.
Methodological Necessity: The authors advocate for "multiplicity-aware inference" and "blocked validation" (temporal and spatial) as prerequisites before the operational use of climate-informed vector models. They suggest that future models must be evaluated on complete sampling events (including zeros), tested against non-climatic baselines, and validated across full annual cycles to be credible.

In conclusion, while species-specific hydroclimate patterns exist (particularly for An. neivai), their strength is highly dependent on the inferential and validation lens. The study concludes that an association only becomes operationally credible after surviving rigorous multiplicity control, baseline comparison, and transferability testing.

Technical Summary: Hydroclimate Associations and Transferability of Anopheles Catch Intensity in Southwestern Colombia

More like this