Change control is essential for good server management, and most administrators do not apply random fixes to servers for experimentation purposes. However, when troubleshooting certain classes of problems, updating the involved binaries makes a lot of sense. (We will ignore issues of memory leaks or heap corruption; in those cases, it is unlikely that an argument can be made against updating if any such fixes have been made in more recent builds.)
Take for instance, deadlocks. Deadlocks could be caused by anything in the application chain: the application, the top level API (CDO), the glue layer (MAPI32), or the providers (EMSMDB/EMSABP in this instance). Updating MAPI32 (the dll for the glue layer), CDO.dll (the .dll for the top level API), the providers (EMSMDB or EMSABP), or the application would be appropriate.
Next, consider disappearing profiles, which often manifest as failure to open information stores. Referencing the data above, you should see that profile services are implemented in MAPI32. Therefore, MAPI32 would be a possibility for updating, as would be the application, because the application might be deleting its own profile.
What about profiles not being deleted? Profile deletion might be the work of the program (if it calls MAPI directly), the top level API (CDO), or, on rare occasions, the provider (MAPI). In this case, the application, CDO.dll, and MAPI32.dll are all good candidates for updates.
Next, consider problems with context specific data such as meeting requests, free/busy, or appointments. Updating MAPI32, the glue layer that provides interfaces to the providers, but does not store or interpret data is not likely to resolve the issue. EMSMDB, the agnostic storage provider, is also not a likely candidate. What about CDO? CDO implements the sets of properties that define specific classes of messages such as meeting requests. It has to agree with Outlook about how these properties work, and the behavior can be unpredictable. In this case, updating CDO.dll makes sense. Finally, there's the application. A number of applications have attempted to operate directly against the calendar data, often with disastrous results. Always consider the application.