Neural Machine Translation (NMT) has made significant progress by taking advantage of subword segmentation to address the open-vocabulary problem. While Byte Pair Encoding (BPE) deterministically splits words into unique sequences, stochastic subword segmentation methods (e.g., subword regularization (SR) and BPE dropout) show the ability to produce multiple rational segmentation of the same word, hence increase model robustness. However, previous subword segmentation methods primarily work on the word level, leading to being restrictive and ambiguous. Those techniques learn the compositionality of the units up to word level but cannot model the lexical units beyond words, such as collocations, multiword expressions, or phrases. This paper introduces Collocation Dropout, a simple and effective segmentation method that allows NMT models to learn to translate a collocation or a phrase better. We conducted experiments on various typical NMT tasks: WMT14 English-German, ASPEC English-Japanese, IWSLT15 English-Vietnamese, English-Chinese, and IWSLT17 English-German, English-French to demonstrate the effectiveness of the proposed method. Experimental results also show that combining our method and other subword segmentation, including both deterministic and stochastic methods, further improves NMT systems.