Matthew Dumas already answered from a historical perspective, explaining what scientists have done. Now I will answer from a linguistic one, explaining why it was done.
No person or people "decide" when to name a model after an author; What you see from today's perspective is a result of a long process of lexicalization, whereby new words (lexical items) or — more loosely — terms† are added to the language of a community: Communication is a collaborative process, and language is always changing even without intentionally thinking "I'm going to use this other person's term because others are more familiar with it" (even Queen Elizabeth's language has changed over time, and she doesn't need to impress many people).
Naming scientific methods is analogous to other forms of lexicalization, such as
- cleaning the floors with a vacuum cleaner/Hoover → vacuuming/Hoovering the floor → vacuuming/Hoovering
- communicating via electronic mail → writing an e-mail → emailing
In other words, descriptions of relatively "new" concepts are initially very analytic in structure, i.e. the meaning can be inferred directly from the structure of the language. Consider the examples below:
We interpolate the data using the frequency estimation of Good (1953).
Here, the concepts signified by frequency estimation and Good (1953) are already accepted well enough to be understandable by those in the relevant linguistic community; I imagine that this sentence would have been perfectly acceptable even in an article from 1954.
Over time, however, common concepts tend to be signified by more synthetic constructions, such as:
Smoothing was done using Good-Turing frequency estimation.
Here, Good-Turing frequency estimation signifies the same thing but is slightly less compositional: Yes, it is still denotes a "type" of frequency estimation, but in an associative manner rather than an ascriptive one (such as seen in effective frequency estimation). In other words, it is frequency estimation associated with Good and Turing. Jump forward even further in time and you have phrases like Good-Turing smoothing which are even less amenable to structural analysis.
At the end of the road, you finally have terms which are so firmly lexicalized that their "original" meaning is no longer apparent to any but the most historically-aware people. Exemplars of this in regards to scientists are e.g. watts, volts and angstroms: I knew what they were from a very young age, but only much later did I find out that there were actually people with those names, and I'm sure I'm not the only one.
Eponymous names aren't invented; They form over time as people become accustomed to the work denoted by the name. In other words, language conventions aren't decided by anyone but rather emerge from the collective behavior of a large group of people.
† What constitutes a "word" is itself debatable and is not a binary distinction of being either "100% word" or "0% word". The more "word-like" something is, however, the more stable the relationship between the signifier (the word) and the signified is, and the less compositional the meaning is: e.g. White House can be considered one word because its meaning is clearly not a "house which is white".