Heterogeneous UAV swarms demonstrate exceptional practical value and scalability in complex scenarios. However, challenges such as the difficulty in modeling heterogeneous UAV swarm tasks and the unclear intrinsic collaboration mechanisms have hindered the optimization of swarm configurations. To address these issues, we propose TB-DML4HS, a task-based, node-level modeling and causal effect analysis method using Double Machine Learning (DML) for heterogeneous UAV swarms. Specifically, TB-DML4HS first decomposes task objectives and events for heterogeneous swarm tasks and constructs a node-level network-structured model. It then employs the DML method to estimate causal effects among various events during task execution, analyzes the contribution rates of heterogeneous nodes within the swarm, and provides optimization suggestions for node configurations to ultimately enhance swarm task performance. Experimental results demonstrate that TB-DML4HS effectively estimates the overall performance with an expected error of less than 5%, and the proposed optimizations significantly improve the specific performance of UAV swarms in current tasks.