def _hopkins_ems(regression_model, v=False):
"""
Returns a table that can be used
v=True prints E(MS) components (False by default)
"""
assert ismodel(regression_model)
# check that no var is in model
if any([type(f) == var for f in regression_model.factors]):
return [None] * len(regression_model.effects)
# E(MS) table (after Hopkins, 1976)
E_MS_table = []
for e in regression_model.effects:
E_MS_row = []
for e2 in regression_model.effects:
if np.all([(f in e or f.random) for f in e2.factors]) and np.all(
[(f in e2 or e2.nestedin(f)) for f in e.factors]
):
E_MS_row.append(True)
else:
E_MS_row.append(False)
E_MS_table.append(E_MS_row)
E_MS = np.array(E_MS_table, dtype=bool)
if v:
print "E(MS) component table:\n", E_MS
# read MS denominator for F tests from table
MS_denominators = []
for i, f in enumerate(regression_model.effects):
e_ms_den = deepcopy(E_MS[i])
e_ms_den[i] = False
match = np.all(E_MS == e_ms_den, axis=1)
if v:
print f.name, ":", match
if match.sum() == 1:
match_i = np.where(match)[0][0]
MS_denominators.append(match_i)
elif match.sum() == 0:
MS_denominators.append(None)
else:
raise NotImplementedError("too many matches")
return MS_denominators
def correlations(Y, Xs, cat=None, levels=[.05, .01, .001], diff=None, sub=None,
pmax=None, nan=True): # , match=None):
"""
:arg var Y: first variable
:arg var X: second variable (or list of variables)
:arg cat: show correlations separately for different groups in the
data. Can be a ``factor`` (the correlation for each level is shown
separately) or an array of ``bool`` values (e.g. from a comparison like
``Stim==1``)
:arg list levels: significance levels to mark
:arg diff: (factor, cat_1, cat_2)
:arg sub: use only a subset of the data
:arg pmax: (None) don't show correlations with p>pmax
:arg nan: ``True``: display correlation which yield NAN;
``False``: hide NANs but mention occurrence in summary (not
implemented);
``None``: don't mention NANs
:rtype: Table
"""
levels = np.array(levels)
if isvar(Xs):
Xs = [Xs]
# SUB
if sub is not None:
Y = Y[sub]
Xs = [X[sub] for X in Xs]
if ismodel(cat) or isfactor(cat):
cat = cat[sub]
if diff is not None:
raise NotImplementedError
if cat is None:
table = fmtxt.Table('l' * 4)
table.cells('Variable', 'r', 'p', 'n')
else:
assert iscategorial(cat)
table = fmtxt.Table('l' * 5)
table.cells('Variable', 'Category', 'r', 'p', 'n')
table.midrule()
table.title("Correlations with %s" % (Y.name))
table._my_nan_count = 0
for X in Xs:
if cat is None:
_corr_to_table(table, Y, X, cat, levels, pmax=pmax, nan=nan)
else:
printXname = True
for cell in cat.cells:
tlen = len(table)
sub = (cat == cell)
_corr_to_table(table, Y, X, sub, levels, pmax=pmax, nan=nan,
printXname=printXname, label=cell_label(cell))
if len(table) > tlen:
printXname = False
# last row
if pmax is None:
p_text = ''
else:
p_text = 'all other p>{p}'.format(p=pmax)
if nan is False and table._my_nan_count > 0:
nan_text = '%s NANs' % table._my_nan_count
else:
nan_text = ''
if p_text or nan_text:
if p_text and nan_text:
text = ', '.join([p_text, nan_text])
else:
text = ''.join([p_text, nan_text])
table.cell("(%s)" % text)
return table
