I don't remember whether there is an official NIST publication on the topic, but there are definitely advantages to having a small set of possible key sizes. Contrasting RSA and DSA is instructive in this respect. A 170-bit $q$ wouldn't be less secure than a 160-bit $q$ if implemented correctly, but offering the choice is less secure.
The lack of guidance on choosing an RSA key size has led to some users picking insecurely low sizes, or occasionally, conversely, needlessly high (and thus slow) sizes. The original DSA standard (FIPS 186-1) allowed a range of $p$ sizes from 512 to 1024; the current version imposes the higher end of the range and does not allow going further. NIST now publishes a list of correspondence between key size and security strength in SP 800-57.
When an algorithm may be instantiated with different key sizes, interoperability requires that all cooperating implementations support compatible key sizes. For example, with a signature algorithm, all verifies must support key sizes that are generated by any signer. This means additional implementation complexity (never a good thing for security, and a pain for implementers anyway) imposed on everyone. This also means that communication protocols may need an extra step to negociate key sizes, which is both extra overhead and a risk of security downgrade if one party is malicious (or due to a man-in-the-middle, but that would mean the protocol is flawed).
Allowing variable key sizes means additional implementation complexity, and more complexity means more risk of bugs. FIPS 186-1 already restricted $p$ sizes to a multiple of 32, the word size on many machines, to avoid bugs that were not all that uncommon with RSA implementations in computing the size of intermediate values, the memory requirements, etc. For example, there have been (and probably still are) RSA implementations that don't produce the expected modulus size when that expected size is odd, or that calculate the byte size from the word size by dividing by 4 and rounding down instead of up; DSA is not prone to these particular bugs.
Given the risk of bugs, every size must be tested, or at least a sufficient sample. Testing every RSA size, for example, is impractical. Testing schemes generally require testing an algorithm at each size they will be used for. This is again extra work which is not required if there is a single standard size.