We introduce a differential abundance analysis method for the analysis of sparse high-throughput data from large-scale surveys of marker genes for microbial communities. Our approach relies on cumulative sum scaling (CSS) normalization - a count data normalization technique - and the zero-inflated Gaussian (ZIG) model as a statistical method for detecting differential abundance of taxonomic features. ZIG differential abundance detection method accounts for bias introduced by the under-sampling of microbial communities commonly found in large-scale marker gene studies.  We have implemented these methods in the publicly available metagenomeSeq bioconductor package. In addition we highlight the utility of the method in a large scale study characterizing the diarrheal microbiome in young children from developing children. Diarrhea, a major cause of mortality and morbidity in young children from developing countries, leading to as many as 15% of all deaths in children under 5 years of age. While many causes of this disease are already known, conventional diagnostic approaches fail to detect a pathogen in up to 60% of diarrheal cases. Using our novel methodology Streptococci were found in our study to be statistically associated with diarrheal disease in general and more severe forms (such as dysentery) in particular.