This paper presents several statistical methods for topic identification (TID): topic unigrams, cache model, TFIDF classifier, topic perplexity, and weighted model. Our work aims to improve these methods by confronting them to very different data, measuring their potential complementarity and their TID performance with simple combinations. Statistical topic identification methods depend not only on a corpus, but also on its type. This study allows to advance the cache model which achieves a TID performance of 82 %. This performance has been increased to 82.3 % with our best linear combination.