Phonologists all over the world are familiar with a set of Japanese consonant alternations called rendaku 'sequential voicing'. In many cases, a morpheme that begins with a voiceless obstruent word-initially has an allomorph beginning with a voiced obstruent that appears (at least sometimes) when that morpheme is a non-initial element in a compound. For example, /tako/ 'callus' appears with initial /d/ in /peN+dako/ 'writing callus', and the appearance of /dako/ is an instance of rendaku. Although rendaku is fundamentally irregular, more than a century of intensive research, beginning with Lyman (1894), has shown that several factors influence how likely rendaku is to occur. By the end of the twentieth century, it seemed unlikely that there was much left to discover, but Eric Rosen, in his University of British Columbia doctoral dissertation (Rosen 2001) and an article based on it (Rosen 2003), made a strikingly original contribution to rendaku research. Rosen's most important claim is that in non-coordinate, two-element compounds in which both elements are native Japanese nouns and at least one of the two is three moras or longer, rendaku is predictable. To state the claim more explicitly, in a compound A+B that meets these criteria, as long as B begins with a voiceless obstruent as a word on its own and is not immune to rendaku, A+B will have rendaku. There are exceptions to Rosen's Rule, but it is a very strong tendency, even if elements that are not native Japanese nouns are taken into consideration. This paper has three goals. First, it offers a brief introduction to Rosen's work. Next, it looks at compounds containing a wider range of element types than those that Rosen considered and shows that many such compounds conform to his predictions. Finally, the paper examines the theoretical explanation that Rosen proposes. This explanation is dubious, since it rests on the idea that voicing is marked in the environment V-([j])V or N-([j])V.