Causal Effect Exploration of Landslide Susceptibility in Chongqing Using a T-Learner

In the rugged, mountainous landscapes of southwestern China, the earth sometimes shifts with sudden, destructive force. Landslides are a persistent threat to the lives, homes, and infrastructure of the people living in these steep valleys. For decades, scientists and engineers have relied on computer models to predict where these disasters might strike next. These models are excellent at spotting patterns; they can look at a map of a hillside and say, "This spot looks dangerous," by analyzing factors like how steep the slope is, how much rain falls there, or what kind of rock lies beneath the soil. However, there is a crucial difference between spotting a pattern and understanding cause. Knowing that a landslide often happens where the ground is steep does not automatically tell us what would happen if we could somehow change the steepness, or if we could alter the amount of water in the soil. To truly manage risk, we need to know not just what is associated with a landslide, but what would actually cause the risk to go up or down if we intervened.

This is the challenge that a team of researchers from Chongqing Polytechnic University of Electronic Technology, Chongqing University of Education, and Peking University set out to address. They focused their study on Chongqing, a city famous for its dramatic topography and rapid urban growth, where the intersection of nature and human activity creates a perfect laboratory for testing new ways to understand geological hazards. Instead of simply asking which factors predict a landslide, they asked a different, more difficult question: if we could change a specific environmental factor, how would the risk of a landslide change? To answer this, they moved beyond standard prediction tools and applied a method known as a T-learner. This approach allows researchers to simulate a "what if" scenario, estimating the difference in risk between a high level of a factor, such as heavy rainfall, and a low level, while carefully accounting for all the other conditions present in the landscape.

The researchers began with a massive dataset containing over 30,000 records, half of which were locations where landslides had occurred and half where they had not. They gathered 17 different pieces of information for each location, ranging from the height of the land and the angle of the slope to the amount of water evaporating from the soil, the distance to rivers, and the type of rock underneath. Using a powerful computer learning algorithm, they first trained a model to recognize the patterns that distinguish a landslide site from a safe one. This model was highly accurate, correctly identifying landslide locations in its test data with a high degree of success. However, the team knew that high accuracy alone does not reveal cause. A model might learn that landslides happen near roads simply because roads are built on steep slopes, not because the roads themselves trigger the slide. To untangle this, they used the T-learner to estimate the average effect of changing each factor.

The results offered a nuanced view of what drives landslide risk in this region. The analysis suggested that the three factors with the largest estimated counterfactual risk differences were evapotranspiration, precipitation, and elevation. Evapotranspiration is a measure of how much water is moving from the soil and plants into the air; in this study, areas with higher levels of this process were associated with a lower predicted risk of landslides. Similarly, higher elevation and higher total precipitation were also linked to lower risk in their specific model. The researchers cautioned that this finding regarding precipitation should not be directly translated into an intervention recommendation of increasing precipitation. They explained that this result likely reflects long-term conditions rather than immediate triggers; perhaps areas with high annual rainfall have vegetation that is better adapted to hold the soil, or the data captures a different aspect of the water cycle than the sudden storms that cause immediate failure. In contrast, the intensity of precipitation—the speed at which rain falls—showed a positive effect, meaning that faster, more intense rain was associated with higher risk, which aligns with common understanding of how landslides are triggered.

The study also highlighted a critical distinction between what a computer model finds important and what actually causes a change in risk. The researchers compared their new "what if" estimates with a standard method called SHAP, which simply ranks factors based on how much they help the model make predictions. In the standard ranking, evapotranspiration was the most important factor, followed by precipitation and elevation. While the order was similar, the meaning was different. The standard method said these factors were the best at guessing where a landslide would happen. The new method suggested that if we could change these factors, the risk would shift significantly. For instance, the study found that the estimated risk difference for evapotranspiration was larger than that for slope, a factor that is traditionally considered the most critical in landslide science. However, the researchers explicitly noted that this pattern does not imply that these variables have been proven to have stronger direct causal effects than slope. The treatment definitions, spatial covariation, and unmeasured confounding may all influence the relative ranking, meaning these results should be viewed as hypotheses for further verification rather than definitive proof of causal strength.

Despite these intriguing findings, the researchers were careful to define the limits of their work. They emphasized that their study was a snapshot in time, using data collected all at once rather than tracking changes over years. Because of this, they could not prove that one factor definitely causes another, nor could they rule out the possibility that some hidden factor, like the type of soil or past human engineering, influenced both the environmental conditions and the landslide risk. They also noted that their results are specific to the mountainous, humid environment of Chongqing and might not apply to dry deserts or flat plains. The study serves as a powerful exploratory framework, showing that we can move beyond simple prediction to ask deeper questions about cause and effect. It suggests that future efforts to prevent landslides should look closely at the water cycle and vegetation, but it stops short of telling engineers exactly what to build or where to dig. Instead, it provides a new set of hypotheses to test, urging scientists to design future studies that can track changes over time and confirm whether these estimated effects hold true in the real world.

The ultimate value of this research lies in its ability to separate the signal of prediction from the signal of cause. By applying these advanced statistical tools to a real-world problem, the team demonstrated that the factors we use to predict a disaster are not always the same factors we should target to prevent it. They showed that while a computer can learn to spot a dangerous hillside with great precision, understanding why that hillside is dangerous requires a different kind of thinking. The study concludes that while we cannot yet claim to have solved the mystery of landslide causation in Chongqing, we have taken a significant step toward understanding the complex relationships between rain, soil, and slope. This new perspective offers a roadmap for future research, one that prioritizes verifying these causal links with better data and more rigorous testing, ultimately leading to safer communities in these challenging landscapes.