Multi-instance partial-label learning (MIPL) is a specialized weakly supervised learning paradigm. The challenge arises from the fact that the supervision information for each sample (bag) is provided as a candidate label set (CLS), with only one being the true label. Bag-level CLS becomes a confounding factor in MIPL, and existing methods only utilize a joint loss of CLS to implicitly mitigate the impact of these factors. Moreover, CLS also contains weak supervision information that indicates the propensity of instances, a fact ignored by existing methods. In this paper, we propose a propensity scoring for multi-instance partial-label learning framework (PSMIPL) to address this problem. PSMIPL is the first model to introduce partial labeling information into the instance inference process in MIPL, which consists of two key components. The propensity scoring component (PSC) establishes a control group and an experiment group to estimate the contribution (i.e., propensity) of each instance to each label in CLS. This process yields a fused bag representation in which confounding factors are explicitly mitigated by the deconfounding method in casual inference, denoted as do(x). Building upon this, the classification module leverages propensity scores to guide model training and produce the final predicted CLS. Experiment results validate that PSMIPL exhibits excellent performance across multiple MIPL datasets. The source code is available at https://github.com/lfslcq/PSIMIPL .