I am reading the 800-90B document. In particular, I'm looking at Chapter 5, the chapter on checking that samples conform to IID. There are 11 tests such as Excursion Test Statistic and Number of Directional Runs. All these tests can be performed on binary data as well as non-binary data.
In the case of some tests, in the case of binary data, the test is performed through conversion in one of two ways.
Conversion I partitions the sequences into eight-bit non-overlapping blocks, and counts the number of ones in each block. Zeroes are appended when the last block has less than eight bits. For example, let the 20-bit input be (1,0,0,0,1,1,1,0,1,1,0,1,1,0,1,1,0,0,1,1). The first and the second eight-bit blocks include four and six ones, respectively. The last block, which is not complete, includes two ones. The output sequence is (4, 6, 2).
Conversion II partitions the sequences into eight-bit non-overlapping blocks, and calculates the integer value of each block. For example, let the input message be (1,0,0,0,1,1,1,0,1,1,0,1,1,0,1,1,0,0,1,1). The integer values of the first two blocks are 142, and 219. Zeroes are appended when the last block has less than eight bits. Then, the last block becomes (0,0,1,1,0,0,0,0) with an integer value of 48. The output sequence is (142, 219, 48).
From my point of view, conversion 1 follows a normal distribution, and conversion 2 seems to follow a uniform distribution.
But I don't understand why the conversion method is different depending on the test. For example, average collision test and maximum collision test use conversion 2, and tests such as number of directional runs test and length of directional runs tests use conversion 1 (even excursion tests do not require conversion).
In summary, I would like to know why NIST proposed two conversion methods and why the conversion method is different for each test.