In the era of big data, many sequencing-based molecular datasets are compositional, meaning they are expressed as percentages. While microbiome data is the most well-known example, single-cell subtype abundance data is also compositional in nature. These datasets are often sparse, containing numerous zero values due to the large number of features and limited sequencing depth. Compositional analysis typically assumes that only a small proportion of taxa are differentially abundant, while the ratios of relative abundances among the remaining taxa remain stable. Most existing methods rely on log-transformed data; however, log-transformation becomes problematic when zero counts are pervasive, often resulting in poor control of the false discovery rate (FDR). To address these challenges, we propose Logistic Compositional Analysis (LOCOM) — a robust logistic regression-based approach for compositional data analysis that eliminates the need for pseudocounts. LOCOM leverages permutation-based inference to account for overdispersion and small sample sizes. Additionally, it employs an asymptotic approach to enhance computational efficiency for large-sample datasets. To mitigate batch effects — commonly arising from systematic differences in sequencing depth in large-sample studies — LOCOM appropriately weights samples. Our simulations demonstrate that LOCOM consistently maintains FDR control while achieving significantly improved sensitivity compared to existing methods.