Are word-level affect lexicons useful in detecting emotions at sentence level? Some prior research finds no gain over and above what is obtained with ngram features---arguably the most widely used features in text classification. Here, we experiment with two very different emotion lexicons and show that even in supervised settings, an affect lexicon can provide significant gains. We further show that while ngram features tend to be accurate, they are often unsuitable for use in new domains. On the other hand, affect lexicon features tend to generalize and produce better results than ngrams when applied to a new domain.