First a correction to the previous post. Replication does not require a table to have an identity column. For transactional replication, all that is required is a declared primary key. It is up to you to deal with how new primary key values are created. One option is to identity but you can choose any method you like. In fact, the primary key does not even need to be numeric.

If you are using an identity column for the primary key, you should also declare it as "NOT FOR REPLICATION". This tells replication that when it inserts a record into the subscriber, do not generate a brand new identity value (similar to using SET INDENTITY_INSERT ON/OFF).

It sounds like there is update activity on the subscriber (I am assuming that you are using transactional replication with updateable subscribers). You will need to manage identity ranges if this is happening (well, actually only if you are inserting on the subscriber). The replication sproc sp_Addarticle allows for this with parameters @auto_identity_range , @pub_identity_range , @identity_range and @threshold. If you are using the new publication wizard, these are available from the article properties in the section "Identity Range Management".

If you are using Merge Replication, I probably would not bother with identity columns at all since Merge Replication requires you to have a row guid column. If you were to have an identity column as well, you would effectively have two completely separate unique columns on the table which is a bit of a waste.