Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

statsmodels: rename contrasts if they cannot be uniquely identified downstream #861

Merged
merged 12 commits into from
Jun 2, 2022
Merged

statsmodels: rename contrasts if they cannot be uniquely identified downstream #861

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
92a23a0
Make model and group_by required
adelavega May 20, 2022
f1f3616
Rename contrasts if needed
adelavega May 21, 2022
ebffe99
Clean up dummy contrasts
adelavega May 21, 2022
2fae6ed
Pass unique variable name in Node to _build_contrasts, and simply ren…
adelavega May 21, 2022
441b072
only if contrast in incoming columns
adelavega May 21, 2022
4ddd074
And not or
adelavega May 21, 2022
3b9f6c7
Use in_contrasts which includes dummy contrasts
adelavega May 21, 2022
ac6e67a
Fix dummy contrasts
adelavega May 23, 2022
bd40eac
Fix tests not expecting intercept
adelavega May 23, 2022
a7de00e
Simplify logic
adelavega May 23, 2022
31dd461
Add test for contrast preprending w/ previous contrast name
adelavega May 23, 2022
50429e1
Modify example model
adelavega Jun 1, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
120 changes: 64 additions & 56 deletions bids/modeling/statsmodels.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
from collections import namedtuple, OrderedDict, Counter, defaultdict
import itertools
from functools import reduce
from multiprocessing.sharedctypes import Value
import re
import fnmatch

Expand Down Expand Up @@ -263,11 +264,11 @@ class BIDSStatsModelsNode:
overridden if one is passed when run() is called on a node.
"""

def __init__(self, level, name, transformations=None, model=None,
contrasts=None, dummy_contrasts=False, group_by=None):
def __init__(self, level, name, model, group_by, transformations=None,
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I made model and group_by mandatory, because they are mandatory in the model spec

contrasts=None, dummy_contrasts=False):
self.level = level.lower()
self.name = name
self.model = model or {}
self.model = model
if transformations is None:
transformations = {"transformer": "pybids-transforms-v1",
"instructions": []}
Expand All @@ -279,13 +280,7 @@ def __init__(self, level, name, transformations=None, model=None,
self.children = []
self.parents = []
if group_by is None:
group_by = []
# Loop over contrasts after first level
if self.level != "run":
group_by.append("contrast")
# Loop over node level of this node
if self.level != "dataset":
group_by.append(self.level)
raise ValueError(f"group_by is not defined for Node: {name}")
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We should not be handling if group_by is None, because it should always be defined

self.group_by = group_by

# Check for intercept only run level model and throw an error
Expand Down Expand Up @@ -592,25 +587,18 @@ def __init__(self, node, entities={}, collections=None, inputs=None,

var_names = list(self.node.model['x'])

# Handle the special 1 construct. If it's present, we add a
# column of 1's to the design matrix. But behavior varies:
# * If there's only a single contrast across all of the inputs,
# the intercept column is given the same name as the input contrast.
# It may already exist, in which case we do nothing.
# * Otherwise, we name the column 'intercept'.
int_name = None
# Handle the special 1 construct.
# Add column of 1's to the design matrix called "intercept"
if 1 in var_names:
if ('contrast' not in df.columns or df['contrast'].nunique() > 1):
int_name = 'intercept'
else:
int_name = df['contrast'].unique()[0]

var_names.remove(1)
var_names = ['intercept' if i == 1 else i for i in var_names]
if 'intercept' not in df.columns:
df.insert(0, 'intercept', 1)

if int_name not in df.columns:
df.insert(0, int_name, 1)
else:
var_names.append(int_name)
# If a single incoming contrast
if ('contrast' in df.columns and df['contrast'].nunique() == 1):
unique_in_contrast = df['contrast'].unique()[0]
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this variable is only set if a Node iteration is grouped by contrast.

we later use this to rename contrasts with concatenation

else:
unique_in_contrast = None

var_names = expand_wildcards(var_names, df.columns)

Expand All @@ -626,7 +614,7 @@ def __init__(self, node, entities={}, collections=None, inputs=None,

# Create ModelSpec and build contrasts
self.model_spec = create_model_spec(self.data, node.model, self.metadata)
self.contrasts = self._build_contrasts(int_name)
self.contrasts = self._build_contrasts(unique_in_contrast)

def _collections_to_dfs(self, collections):
"""Merges collections and converts them to a pandas DataFrame."""
Expand Down Expand Up @@ -690,17 +678,55 @@ def _inputs_to_df(self, inputs):
input_df.loc[input_df.index[i], con.name] = 1
return input_df

def _build_contrasts(self, int_name):
"""Contrast list of ContrastInfo objects based on current state."""
contrasts = {}
def _build_contrasts(self, unique_in_contrast=None):
"""Contrast list of ContrastInfo objects based on current state.

Parameters
----------
unique_in_contrast : string
Name of unique incoming contrast inputs (i.e. if there is only 1)
"""
in_contrasts = self.node.contrasts.copy()
col_names = set(self.X.columns)
for con in self.node.contrasts:
name = con["name"]

# Create dummy contrasts as regular contrasts
dummies = self.node.dummy_contrasts
if dummies:
if 'conditionlist' in dummies:
conditions = set(dummies['condition_list'])
else:
conditions = col_names

for col_name in conditions:
in_contrasts.insert(0,
{
'name': col_name,
'condition_list': [col_name],
'weights': [1],
'test': dummies.get('test')
}
)

# Process all contrasts, starting with dummy contrasts
# Dummy contrasts are replaced if a contrast is defined with same name
contrasts = {}
for con in in_contrasts:
condition_list = list(con["condition_list"])
if 1 in condition_list and int_name is not None:
condition_list[condition_list.index(1)] = int_name
if name == 1 and int_name is not None:
name = int_name

# Rename special 1 construct
condition_list = ['intercept' if i == 1 else i for i in condition_list]
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this is going to fail:

{
  ...
  "DummyContrasts": {"Test": "t"}
}

The reason is that instead of changing the 1 in the design matrix to match the column name, it's now going to be "intercept", and when you expand the dummy contrasts to proper contrasts, then the "name" will be "intercept" so we will never pick the contrast name back up. There's also nothing stopping somebody from explicitly creating a contrast named "intercept", so detecting this case will now be harder.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ah, you're right. I am passing the name of the unique variable to this function, so I can rename the "intercept" contrast the name of the unique contrast grouping variable in this special case. i think i meant to do that but forgot.

however, about catching an explicitly named "intercept" contrast, i have two suggestions:

  1. the user can deal w/ the consequences
  2. throw an error if any explicitly named contrast conflict w/ dummycontrasts.

the spec currently says that explicilt contrasts over-ride dummycontrasts, so 2) would be most consistent


name = con["name"]

# Rename contrast name
if name == 1:
name = unique_in_contrast or 'intercept'
else:
# If Node has single contrast input, as is grouped by contrast
# Rename contrast to append incoming contrast name
if unique_in_contrast:
name = f"{unique_in_contrast}_{name}"

missing_vars = set(condition_list) - col_names
if missing_vars:
if self.invalid_contrasts == 'error':
Expand All @@ -711,31 +737,13 @@ def _build_contrasts(self, int_name):
elif self.invalid_contrasts == 'drop':
continue
weights = np.atleast_2d(con['weights'])

# Add contrast name to entities; can be used in grouping downstream
entities = {**self.entities, 'contrast': name}
ci = ContrastInfo(name, condition_list,
con['weights'], con.get("test"), entities)
contrasts[name] = ci

dummies = self.node.dummy_contrasts
if dummies:
conditions = col_names
if 'conditions' in dummies:
conds = set(dummies['conditions'])
if 1 in conds and int_name is not None:
conds.discard(1)
conds.add(int_name)
conditions &= conds
conditions -= set(c.name for c in contrasts.values())

for col_name in conditions:
if col_name in contrasts:
continue
entities = {**self.entities, 'contrast': col_name}
ci = ContrastInfo(col_name, [col_name], [1], dummies.get("test"),
entities)
contrasts[col_name] = ci

return list(contrasts.values())

@property
Expand Down
20 changes: 16 additions & 4 deletions bids/modeling/tests/test_statsmodels.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,10 +113,22 @@ def test_entire_graph_smoketest(graph):
cis = list(chain(*[op.contrasts for op in outputs]))
assert len(cis) == 18
outputs = graph["participant"].run(cis, group_by=['subject', 'contrast'])
# 2 subjects x 3 contrasts
# 2 subjects x 3 contrasts)
assert len(outputs) == 6
# * 2 participant level contrasts = 12
cis = list(chain(*[op.contrasts for op in outputs]))
assert len(cis) == 6
assert len(cis) == 12

# Test output names for single subject
out_contrasts = [
c.entities['contrast'] for c in cis if c.entities['subject'] == '01'
]

expected_outs = [
'gain', 'gain_neg', 'RT', 'RT_neg', 'RT:gain', 'RT:gain_neg'
Copy link
Collaborator Author

@adelavega adelavega May 23, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These are the expected outputs at the participant level w/ the concatenation I implemented.

For this model: bids/tests/data/ds005/models/ds-005_type-test_model.json

cc: @jmumford

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should work and seems intuitive. This is a nice model spec to learn from. One thing, I wouldn't scale RT (no mean centering and no sd scaling) as it adds between-subject RT differences at the group level (I have been very slowly writing a paper about this).

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks! I'm happy to make that change.

I'm modifying this model as an example on the next chapter of the walkthrough.

First chapter is now live (https://bids-standard.github.io/stats-models/)-- feedback welcome

]

assert set(out_contrasts) == set(expected_outs)

# Construct new ContrastInfo objects with name updated to reflect last
# contrast. This would normally be done by the handling tool (e.g., fitlins)
Expand All @@ -140,7 +152,7 @@ def test_entire_graph_smoketest(graph):
assert model_spec.X.shape == (2, 2)
assert model_spec.Z is None
assert len(model_spec.terms) == 2
assert not set(model_spec.terms.keys()) - {"RT", "gain", "RT:gain", "sex"}
assert not set(model_spec.terms.keys()) - {"intercept", "sex"}
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

X now just had intercept not contrast name in DM


# BY-GROUP NODE
outputs = graph["by-group"].run(inputs, group_by=['contrast'])
Expand All @@ -153,7 +165,7 @@ def test_entire_graph_smoketest(graph):
assert model_spec.__class__.__name__ == "GLMMSpec"
assert model_spec.X.shape == (2, 1)
assert model_spec.Z is None
assert not set(model_spec.terms.keys()) - {"RT", "gain", "RT:gain"}
assert not set(model_spec.terms.keys()) - {"intercept"}


def test_expand_wildcards():
Expand Down
15 changes: 9 additions & 6 deletions bids/tests/data/ds005/models/ds-005_type-test_model.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,10 +25,6 @@
"Name": "Rename",
"Input": "trial_type.parametric gain",
"Output": "gain"
},
{
"Name": "Scale",
"Input": "RT"
}
],
"DummyContrasts": {
Expand All @@ -53,14 +49,21 @@
"ConditionList": [1],
"Weights": [1],
"Test": "FEMA"
},
{
"Name": "neg",
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Modified test model to add a user defined contrast (not using 1)

"ConditionList": [1],
"Weights": [-1],
"Test": "FEMA"
}
]
},
{
"Name": "by-group",
"Level": "Dataset",
"GroupBy": [
"sex"
"sex",
"contrast"
],
"Model": {
"X": [
Expand All @@ -75,7 +78,7 @@
{
"Name": "group-diff",
"Level": "Dataset",
"GroupBy": [],
"GroupBy": ["contrast"],
"Model": {
"X": [
1,
Expand Down