v0.3.0

• add feature for creating topographic grids
• default to Poisson boundary in LRC
• now requires Python>=3.8, pandas>=2.0

v0.2.7

• discounting according to Walter1975 for log-ratio
• make Poisson boundary the default for LRC
• major performance improvement for LRC with Poisson boundary
• further performance improvement: only calculate scores once for each frequency signature

v0.2.6

• force np.vectorize to return float, otherwise conservative log ratio might be rounded to integer

v0.2.5

• repair measuring performance
• do not calculate binomial likelihood by default

v0.2.4

make calculation of conservative log ratio with Poisson boundary robust against observations with O11=O21=0: return 0.

v0.2.3

• setup.py: installation under Windows should work properly now
• deprecated calculate_measures()
• correction in liddell()
• allow integers to be passed to observed_frequencies(); extend functionality
• don't use methods to be tested in conftest.py

v0.2.2

• new AM: conservative log ratio with correct CI boundary from Poisson distribution (Evert 2022)
• include wheel as build dependency in pyproject.toml
• use score() rather than calculate_measures() in tests; extend propagation of score() parameters
• include pytest.ini
• simplify setup.py

v0.2.1

update requirements

• maximum of version numbers s.t. python3.6
• specify wheel as requirement (for building)

two new measures:

• minimum sensitivity
• Liddell

v0.2.0

• new possible input: "keyword-friendly" corpus frequencies notation (f1, N1, f2, N2)
• new score wrapper also allows constant integer counts (N1, N2 for keyword notation; f1, N for frequency signatures) to be given as parameters
• keyword arguments are now passed from calculate_measures() (and score()) to underlying measures

v0.1.7

compatibility:

• require scipy instead of python3.8

measures:

• local MI
• simple LL
• extend parameters of conservative log ratio (Sidak correction)

sort & categorize measures:

• asymptotic hypothesis tests
• point estimates of association strength
• (likelihood measures)
• information theory
• conservative estimates

universal discounting for zeros:

• O11 and O12 are set to a small value where they're 0 (0.001 by default; except for Hardie's dubious discounting of 0.5)
• this makes additional definition of phi-function obsolete
• NB discounting does not have any effect for some measures (log-likelihood, local MI)