Compared to RNA-sequencing transcript differential analysis, gene-level differential expression analysis is more robust and experimentally actionable. However, the use of gene counts for statistical analysis can mask transcript-level dynamics. We demonstrate that ‘analysis first, aggregation second,’ where the p values derived from transcript analysis are aggregated to obtain gene-level results, increase sensitivity and accuracy. The method we propose can also be applied to transcript compatibility counts obtained from pseudoalignment of reads, which circumvents the need for quantification and is fast, accurate, and model-free. The method generalizes to various levels of biology and we showcase an application to gene ontologies.