Source code for pymc_marketing.mmm.causal

#   Copyright 2022 - 2025 The PyMC Labs Developers
#
#   Licensed under the Apache License, Version 2.0 (the "License");
#   you may not use this file except in compliance with the License.
#   You may obtain a copy of the License at
#
#       http://www.apache.org/licenses/LICENSE-2.0
#
#   Unless required by applicable law or agreed to in writing, software
#   distributed under the License is distributed on an "AS IS" BASIS,
#   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#   See the License for the specific language governing permissions and
#   limitations under the License.
"""Causal module."""

from __future__ import annotations

import itertools as it
import re
import warnings
from collections.abc import Sequence
from typing import Annotated, Literal, NotRequired, TypedDict

try:
    import networkx as nx
except ImportError:  # Optional dependency
    nx = None  # type: ignore[assignment]

import numpy as np
import pandas as pd
import pymc as pm
import pytensor
import pytensor.tensor as pt
from pydantic import Field, InstanceOf, validate_call
from pymc_extras.prior import Prior

try:
    from dowhy import CausalModel
except ImportError:

    class LazyCausalModel:
        """Lazy import of dowhy's CausalModel."""

        def __init__(self, *args, **kwargs):
            msg = (
                "To use Causal Graph functionality, please install the optional dependencies with: "
                "pip install pymc-marketing[dag]"
            )
            raise ImportError(msg)

    CausalModel = LazyCausalModel


[docs] class TestResult(TypedDict): """Conditional independence test statistics recorded during fitting.""" bic0: float bic1: float delta_bic: float logBF10: float BF10: float independent: bool conditioning_set: list[str] forced: NotRequired[bool]
EMPTY_CONDITION_SET: frozenset[str] = frozenset()
[docs] class BuildModelFromDAG: """Build a PyMC probabilistic model directly from a Causal DAG and a tabular dataset. The class interprets a Directed Acyclic Graph (DAG) where each node is a column in the provided `df`. For every edge ``A -> B`` it creates a slope prior for the contribution of ``A`` into the mean of ``B``. Each node receives a likelihood prior. Dims and coords are used to align and index observed data via ``pm.Data`` and xarray. Parameters ---------- dag : str DAG in DOT format (e.g. ``digraph { A -> B; B -> C; }``) or as a simple comma/newline separated list of edges (e.g. ``"A->B, B->C"``). df : pandas.DataFrame DataFrame that contains a column for every node present in the DAG and all columns named by the provided ``dims``. target : str Name of the target node present in both the DAG and ``df``. This is not used to restrict modeling but is validated to exist in the DAG. dims : tuple[str, ...] Dims for the observed variables and likelihoods (e.g. ``("date", "channel")``). coords : dict Mapping from dim names to coordinate values. All coord keys must exist as columns in ``df`` and will be used to pivot the data to match dims. model_config : dict, optional Optional configuration with priors for keys ``"intercept"``, ``"slope"`` and ``"likelihood"``. Values should be ``pymc_extras.prior.Prior`` instances. Missing keys fall back to :pyattr:`default_model_config`. Examples -------- Minimal example using DOT format: .. code-block:: python import numpy as np import pandas as pd from pymc_marketing.mmm.causal import BuildModelFromDAG dates = pd.date_range("2024-01-01", periods=5, freq="D") df = pd.DataFrame( { "date": dates, "X": np.random.normal(size=5), "Y": np.random.normal(size=5), } ) dag = "digraph { X -> Y; }" dims = ("date",) coords = {"date": dates} builder = BuildModelFromDAG( dag=dag, df=df, target="Y", dims=dims, coords=coords ) model = builder.build() Edge-list format and custom likelihood prior: .. code-block:: python from pymc_extras.prior import Prior dag = "X->Y" # equivalent to the DOT example above model_config = { "likelihood": Prior( "StudentT", nu=5, sigma=Prior("HalfNormal", sigma=1), dims=("date",) ), } builder = BuildModelFromDAG( dag=dag, df=df, target="Y", dims=("date",), coords={"date": dates}, model_config=model_config, ) model = builder.build() """
[docs] @validate_call def __init__( self, *, dag: str = Field(..., description="DAG in DOT string format or A->B list"), df: InstanceOf[pd.DataFrame] = Field( ..., description="DataFrame containing all DAG node columns" ), target: str = Field(..., description="Target node name present in DAG and df"), dims: tuple[str, ...] = Field( ..., description="Dims for observed/likelihood variables" ), coords: dict = Field( ..., description=( "Required coords mapping for dims and priors. All coord keys must exist as columns in df." ), ), model_config: dict | None = Field( None, description=( "Optional model config with Priors for 'intercept', 'slope' and " "'likelihood'. Keys not supplied fall back to defaults." ), ), ) -> None: self.dag = dag self.df = df self.target = target self.dims = dims self.coords = coords # Parse graph and validate target self.graph = self._parse_dag(self.dag) self.nodes = list(nx.topological_sort(self.graph)) if self.target not in self.nodes: raise ValueError(f"Target '{self.target}' not in DAG nodes: {self.nodes}") # Merge provided model_config with defaults provided = model_config self.model_config = self.default_model_config if provided is not None: self.model_config.update(provided) # Validate required priors are present and of correct type self._validate_model_config_priors() # Validate coords are present and consistent with dims, priors, and df self._validate_coords_required_are_consistent() # Validate prior dims consistency early (does not require building the model) self._warning_if_slope_dims_dont_match_likelihood_dims() self._validate_intercept_dims_match_slope_dims()
@property def default_model_config(self) -> dict[str, Prior]: """Default priors for intercepts, slopes and likelihood using ``pymc_extras.Prior``. Returns ------- dict Dictionary with keys ``"intercept"``, ``"slope"`` and ``"likelihood"`` mapping to ``Prior`` instances with dims derived from :pyattr:`dims`. """ slope_dims = tuple(dim for dim in (self.dims or ()) if dim != "date") return { "intercept": Prior("Normal", mu=0, sigma=2, dims=slope_dims), "slope": Prior("Normal", mu=0, sigma=2, dims=slope_dims), "likelihood": Prior( "Normal", sigma=Prior("HalfNormal", sigma=2), dims=self.dims, ), } @staticmethod def _parse_dag(dag_str: str) -> nx.DiGraph: """Parse DOT digraph or edge-list string into a directed acyclic graph.""" if nx is None: raise ImportError( "To use Causal Graph functionality, please install the optional dependencies with: " "pip install pymc-marketing[dag]" ) # Primary format: DOT digraph s = dag_str.strip() g = nx.DiGraph() if s.lower().startswith("digraph"): # Extract content within the first top-level {...} brace_start = s.find("{") brace_end = s.rfind("}") if brace_start == -1 or brace_end == -1 or brace_end <= brace_start: raise ValueError("Malformed DOT digraph: missing braces") body = s[brace_start + 1 : brace_end] # Remove comments (// ... or # ... at line end) lines = [] for raw_line in body.splitlines(): line = re.split(r"//|#", raw_line, maxsplit=1)[0].strip() if line: lines.append(line) body = "\n".join(lines) # Find edges "A -> B" possibly ending with ';' for m in re.finditer( r"\b([A-Za-z0-9_]+)\s*->\s*([A-Za-z0-9_]+)\s*;?", body ): a, b = m.group(1), m.group(2) g.add_edge(a, b) # Find standalone node declarations (lines with single identifier, optional ';') for raw_line in body.splitlines(): line = raw_line.strip().rstrip(";") if not line or "->" in line or "[" in line or "]" in line: continue mnode = re.match(r"^([A-Za-z0-9_]+)$", line) if mnode: g.add_node(mnode.group(1)) else: # Fallback: simple comma/newline-separated "A->B" tokens edges: list[tuple[str, str]] = [] for token in re.split(r"[,\n]+", s): token = token.strip().rstrip(";") if not token: continue medge = re.match(r"^([A-Za-z0-9_]+)\s*->\s*([A-Za-z0-9_]+)$", token) if not medge: raise ValueError(f"Invalid edge token: '{token}'") a, b = medge.group(1), medge.group(2) edges.append((a, b)) g.add_edges_from(edges) if not nx.is_directed_acyclic_graph(g): raise ValueError("Provided graph is not a DAG.") return g def _warning_if_slope_dims_dont_match_likelihood_dims(self) -> None: """Warn if slope prior dims differ from likelihood dims without the 'date' dim.""" slope_prior = self.model_config["slope"] likelihood_prior = self.model_config["likelihood"] like_dims = getattr(likelihood_prior, "dims", None) if isinstance(like_dims, str): like_dims = (like_dims,) elif isinstance(like_dims, list): like_dims = tuple(like_dims) # Guard against None dims (treat as empty) if like_dims is None: expected_slope_dims = () else: expected_slope_dims = tuple(dim for dim in like_dims if dim != "date") slope_dims = getattr(slope_prior, "dims", None) if slope_dims is None or not isinstance(slope_dims, tuple): slope_dims = () elif isinstance(slope_dims, str): slope_dims = (slope_dims,) elif isinstance(slope_dims, list): slope_dims = tuple(slope_dims) if slope_dims != expected_slope_dims: warnings.warn( ( "Slope prior dims " f"{slope_dims if slope_dims else '()'} do not match expected dims " f"{expected_slope_dims} (likelihood dims without 'date')." ), stacklevel=2, ) def _validate_intercept_dims_match_slope_dims(self) -> None: """Ensure intercept prior dims match slope prior dims exactly.""" def _to_tuple(maybe_dims): if maybe_dims is None: return tuple() if isinstance(maybe_dims, str): return (maybe_dims,) if isinstance(maybe_dims, list | tuple): return tuple(maybe_dims) return tuple() slope_dims = _to_tuple(getattr(self.model_config["slope"], "dims", None)) intercept_dims = _to_tuple( getattr(self.model_config["intercept"], "dims", None) ) if slope_dims != intercept_dims: raise ValueError( "model_config['intercept'].dims must match model_config['slope'].dims. " f"Got intercept dims {intercept_dims or '()'} and slope dims {slope_dims or '()'}." ) def _validate_model_config_priors(self) -> None: """Ensure required model_config entries are Prior instances. Enforces that keys 'slope' and 'likelihood' exist and are Prior objects, so downstream code can safely index and call Prior helper methods. """ required_keys = ("intercept", "slope", "likelihood") for key in required_keys: if key not in self.model_config: raise ValueError(f"model_config must include '{key}' as a Prior.") for key in required_keys: if not isinstance(self.model_config[key], Prior): raise TypeError( f"model_config['{key}'] must be a Prior, got " f"{type(self.model_config[key]).__name__}." ) def _validate_coords_required_are_consistent(self) -> None: """Validate mutual consistency among dims, coords, priors, and data columns.""" if self.coords is None: raise ValueError("'coords' is required and cannot be None.") # 1) All coords keys must correspond to columns in the dataset for key in self.coords.keys(): if key not in self.df.columns: raise KeyError( f"Coordinate key '{key}' not found in DataFrame columns. Present columns: {list(self.df.columns)}" ) # 2) Ensure dims are present in coords for d in self.dims: if d not in self.coords: raise ValueError(f"Missing coordinate values for dim '{d}' in coords.") # 3) Ensure Prior.dims exist in coords (for all top-level priors we manage) def _to_tuple(maybe_dims): if isinstance(maybe_dims, str): return (maybe_dims,) if isinstance(maybe_dims, list | tuple): return tuple(maybe_dims) else: return tuple() for prior_name, prior in self.model_config.items(): if not isinstance(prior, Prior): continue for d in _to_tuple(getattr(prior, "dims", None)): if d not in self.coords: raise ValueError( f"Dim '{d}' declared in Prior '{prior_name}' must be present in coords." ) # 4) Enforce that likelihood dims match class dims exactly likelihood_prior = self.model_config["likelihood"] likelihood_dims = _to_tuple(getattr(likelihood_prior, "dims", None)) if likelihood_dims and tuple(self.dims) != likelihood_dims: raise ValueError( "Likelihood Prior dims " f"{likelihood_dims} must match class dims {tuple(self.dims)}. " "When supplying a custom model_config, ensure likelihood.dims equals the 'dims' argument." ) def _parents(self, node: str) -> list[str]: """Return the list of parent node names for the given DAG node.""" return list(self.graph.predecessors(node))
[docs] def build(self) -> pm.Model: """Construct and return the PyMC model implied by the DAG and data. The method creates a ``pm.Data`` container for every node to align the observed data with the declared ``dims``. For each edge ``A -> B``, a slope prior is instantiated from ``model_config['slope']`` and used in the mean of node ``B``'s likelihood, which is instantiated from ``model_config['likelihood']``. Returns ------- pymc.Model A fully specified model with slopes and likelihoods for all nodes. Examples -------- Build a model and sample from it: .. code-block:: python builder = BuildModelFromDAG( dag="A->B", df=df, target="B", dims=("date",), coords={"date": dates} ) model = builder.build() with model: idata = pm.sample(100, tune=100, chains=2, cores=2) Multi-dimensional dims (e.g. date and country): .. code-block:: python dims = ("date", "country") coords = {"date": dates, "country": ["Venezuela", "Colombia"]} builder = BuildModelFromDAG( dag="A->B, B->Y", df=df, target="Y", dims=dims, coords=coords ) model = builder.build() """ dims = self.dims coords = self.coords with pm.Model(coords=coords) as model: data_containers: dict[str, pm.Data] = {} for node in self.nodes: if node not in self.df.columns: raise KeyError(f"Column '{node}' not found in df.") # Ensure observed data has shape consistent with declared dims by pivoting via xarray indexed = self.df.set_index(list(dims)) xarr = indexed.to_xarray()[node] values = xarr.values data_containers[node] = pm.Data(f"_{node}", values, dims=dims) # For each node add slope priors per parent and likelihood with sigma prior slope_rvs: dict[tuple[str, str], pt.TensorVariable] = {} # Create priors in a stable deterministic order for node in self.nodes: parents = self._parents(node) # Slopes for each parent -> node mu_expr = 0 for parent in parents: slope_name = f"{parent.lower()}:{node.lower()}" slope_rv = self.model_config["slope"].create_variable(slope_name) slope_rvs[(parent, node)] = slope_rv mu_expr += slope_rv * data_containers[parent] intercept_rv = self.model_config["intercept"].create_variable( f"{node.lower()}_intercept" ) self.model_config["likelihood"].create_likelihood_variable( name=node, mu=mu_expr + intercept_rv, observed=data_containers[node], ) self.model = model return self.model
[docs] def model_graph(self): """Return a Graphviz visualization of the built PyMC model. Returns ------- graphviz.Source Graphviz object representing the model graph. Examples -------- .. code-block:: python model = builder.build() g = builder.model_graph() g """ if not hasattr(self, "model"): raise RuntimeError("Call build() first.") return pm.model_to_graphviz(self.model)
[docs] def dag_graph(self): """Return a copy of the parsed DAG as a NetworkX directed graph. Returns ------- networkx.DiGraph A directed acyclic graph with the same nodes and edges as the input DAG. Examples -------- .. code-block:: python g = builder.dag_graph() list(g.edges()) """ if nx is None: raise ImportError( "To use Causal Graph functionality, please install the optional dependencies with: " "pip install pymc-marketing[dag]" ) g = nx.DiGraph() g.add_nodes_from(self.graph.nodes) g.add_edges_from(self.graph.edges) return g
[docs] class TBFPC: r""" Target-first Bayes Factor PC (TBF-PC) causal discovery algorithm. This algorithm is a target-oriented variant of the Peter–Clark (PC) algorithm, using Bayes factors (via ΔBIC approximation) as the conditional independence test. For each conditional independence test of the form .. math:: H_0 : Y \perp X \mid S \quad \text{vs.} \quad H_1 : Y \not\!\perp X \mid S we compare two linear models: .. math:: M_0 : Y \sim S \\ M_1 : Y \sim S + X where :math:`S` is a conditioning set of variables. The Bayesian Information Criterion (BIC) is defined as .. math:: \mathrm{BIC}(M) = n \log\!\left(\frac{\mathrm{RSS}}{n}\right) + k \log(n), with residual sum of squares :math:`\mathrm{RSS}`, sample size :math:`n`, and number of parameters :math:`k`. The Bayes factor is approximated by .. math:: \log \mathrm{BF}_{10} \approx -\tfrac{1}{2} \left[ \mathrm{BIC}(M_1) - \mathrm{BIC}(M_0) \right]. Independence is declared if :math:`\mathrm{BF}_{10} < \tau`, where :math:`\tau` is set via the ``bf_thresh`` parameter. Target Edge Rules ----------------- Different rules govern how driver → target edges are retained: - ``"any"``: keep :math:`X \to Y` unless **any** conditioning set renders :math:`X \perp Y \mid S`. - ``"conservative"``: keep :math:`X \to Y` if **at least one** conditioning set shows dependence. - ``"fullS"``: test only with the **full set** of other drivers as :math:`S`. Examples -------- **1. Basic usage with full conditioning set** .. code-block:: python import numpy as np, pandas as pd rng = np.random.default_rng(7) n = 2000 C = rng.gamma(2,1,n) A = 0.7*C + rng.gamma(2,1,n) D = 0.5*C + rng.gamma(2,1,n) B = 0.8*A + rng.gamma(2,1,n) Y = 0.9*B + 0.6*D + 0.7*C + rng.gamma(2,1,n) df = pd.DataFrame({"A":A,"B":B,"C":C,"D":D,"Y":Y}) df = (df - df.mean())/df.std() # recommended scaling model = TBFPC(target="Y", target_edge_rule="fullS") model.fit(df, drivers=["A","B","C","D"]) print(model.get_directed_edges()) print(model.get_undirected_edges()) print(model.to_digraph()) **2. Using forbidden edges** You can specify edges that must *not* be tested or included (prior knowledge about the domain). .. code-block:: python model = TBFPC( target="Y", target_edge_rule="any", forbidden_edges=[("A","C")] # forbid A--C ) model.fit(df, drivers=["A","B","C","D"]) print(model.to_digraph()) **3. Conservative rule** Keeps driver → target edges if **any conditioning set** shows dependence. .. code-block:: python model = TBFPC(target="Y", target_edge_rule="conservative") model.fit(df, drivers=["A","B","C","D"]) print(model.to_digraph()) References ---------- - Spirtes, Glymour, Scheines (2000). *Causation, Prediction, and Search*. MIT Press. [PC algorithm] - Spirtes & Glymour (1991). "An Algorithm for Fast Recovery of Sparse Causal Graphs." - Kass, R. & Raftery, A. (1995). "Bayes Factors." """
[docs] @validate_call(config=dict(arbitrary_types_allowed=True)) def __init__( self, target: Annotated[ str, Field( min_length=1, description="Name of the outcome variable to orient the search.", ), ], *, target_edge_rule: Literal["any", "conservative", "fullS"] = "any", bf_thresh: Annotated[float, Field(gt=0.0)] = 1.0, forbidden_edges: Sequence[tuple[str, str]] | None = None, required_edges: Sequence[tuple[str, str]] | None = None, ): """Create a new TBFPC causal discovery model. Parameters ---------- target Variable name for the model outcome; must be present in the data used during fitting. target_edge_rule Rule that controls which driver → target edges are retained. Options are ``"any"``, ``"conservative"``, and ``"fullS"``. bf_thresh Positive Bayes factor threshold applied during conditional independence tests. forbidden_edges Optional sequence of node pairs that must not be connected in the learned graph. required_edges Optional sequence of directed ``(u, v)`` pairs that must be present in the learned graph as ``u -> v``. """ warnings.warn( "TBFPC is experimental and its API may change; use with caution.", UserWarning, stacklevel=2, ) self.target = target self.target_edge_rule = target_edge_rule self.bf_thresh = float(bf_thresh) self.forbidden_edges: set[tuple[str, str]] = set(forbidden_edges or []) self.required_edges: set[tuple[str, str]] = set(required_edges or []) conflicts = [ (u, v) for (u, v) in self.required_edges if (u, v) in self.forbidden_edges or (v, u) in self.forbidden_edges ] if conflicts: conflict_str = ", ".join(f"{u}->{v}" for u, v in conflicts) raise ValueError( f"Required edges conflict with forbidden edges: {conflict_str}" ) conflicts = [ (u, v) for (u, v) in self.required_edges if (u, v) in self.forbidden_edges or (v, u) in self.forbidden_edges ] if conflicts: conflict_str = ", ".join(f"{u}->{v}" for u, v in conflicts) raise ValueError( f"Required edges conflict with forbidden edges: {conflict_str}" ) # Internal state self.sep_sets: dict[tuple[str, str], set[str]] = {} self._adj_directed: set[tuple[str, str]] = set() self._adj_undirected: set[tuple[str, str]] = set() self.nodes_: list[str] = [] self.test_results: dict[tuple[str, str, frozenset[str]], TestResult] = {} # Shared response vector for symbolic BIC computation # Initialized with placeholder; will be updated with actual data during fitting self.y_sh = pytensor.shared(np.zeros(1, dtype="float64"), name="y_sh") self._bic_fn = self._build_symbolic_bic_fn()
@staticmethod def _bitmasks(k: int): """Yield tuples of 0/1 of length k (fast product without importing itertools).""" # Equivalent to itertools.product([0,1], repeat=k) but minimal if k == 0: yield () return stack = [0] * k i = 0 while True: if i < k: stack[i] = 0 i += 1 continue yield tuple(stack) # increment like binary counter i -= 1 while i >= 0 and stack[i] == 1: i -= 1 if i < 0: break stack[i] = 1 i += 1 @staticmethod def _parse_cpdag_dot( dot: str, ) -> tuple[set[str], set[tuple[str, str]], set[tuple[str, str]]]: """Minimal DOT parser for a single CPDAG block.""" import re digraph = re.search(r"digraph\b[^{}]*\{(.*?)\}", dot, flags=re.DOTALL) if not digraph: raise ValueError("No 'digraph { ... }' block found.") body = digraph.group(1) nodes: set[str] = set() directed: set[tuple[str, str]] = set() undirected: set[tuple[str, str]] = set() node_re = re.compile(r'^\s*"([^"]+)"\s*(?:\[[^\]]*\])?\s*;\s*$') edge_re = re.compile(r'^\s*"([^"]+)"\s*->\s*"([^"]+)"\s*(\[[^\]]*\])?\s*;\s*$') def is_undirected(attrs: str | None) -> bool: if not attrs: return False low = attrs.lower() return "style=dashed" in low and "dir=none" in low for raw in body.splitlines(): line = raw.strip() if not line or line.startswith("//"): continue # node? m = node_re.match(line) if m: nodes.add(m.group(1)) continue # edge? m = edge_re.match(line) if m: u, v, attrs = m.group(1), m.group(2), m.group(3) nodes.update((u, v)) if is_undirected(attrs): undirected.add((u, v) if u <= v else (v, u)) else: directed.add((u, v)) continue # ignore other lines (global styles etc.) return nodes, directed, undirected @staticmethod def _is_acyclic( nodes: set[str], edges: list[tuple[str, str]] | set[tuple[str, str]] ) -> bool: """DFS cycle check.""" adj: dict[str, list[str]] = {u: [] for u in nodes} for u, v in edges: adj.setdefault(u, []).append(v) adj.setdefault(v, []) # ensure key exists state = {u: 0 for u in nodes} # 0=unseen, 1=visiting, 2=done def dfs(u: str) -> bool: state[u] = 1 for w in adj[u]: if state[w] == 1: return False if state[w] == 0 and not dfs(w): return False state[u] = 2 return True return all(state[u] or dfs(u) for u in nodes) def _dot_from_edges( self, nodes: set[str], edges: list[tuple[str, str]] | set[tuple[str, str]] ) -> str: """Render a fully directed graph to DOT; highlights target if present.""" lines = ["digraph G {", " node [shape=ellipse];"] for n in sorted(nodes): if hasattr(self, "target") and n == self.target: lines.append(f' "{n}" [style=filled, fillcolor="#eef5ff"];') else: lines.append(f' "{n}";') for u, v in sorted(edges): lines.append(f' "{u}" -> "{v}";') lines.append("}") return "\n".join(lines).replace("\\n'", "\\n").replace("'\\n", "\\n") # hygiene def _key(self, u: str, v: str) -> tuple[str, str]: """Return a sorted 2-tuple key for an undirected edge between ``u`` and ``v``.""" return (u, v) if u <= v else (v, u) def _set_sep(self, u: str, v: str, S: Sequence[str]) -> None: """Record the separation set ``S`` for the node pair ``(u, v)``.""" self.sep_sets[self._key(u, v)] = set(S) def _has_forbidden(self, u: str, v: str) -> bool: """Return True if edge ``u—v`` is forbidden in either direction.""" return (u, v) in self.forbidden_edges or (v, u) in self.forbidden_edges def _is_required(self, u: str, v: str) -> bool: """Return True if the directed edge ``u -> v`` is required.""" return (u, v) in self.required_edges def _add_directed(self, u: str, v: str) -> None: """Add a directed edge ``u -> v`` if not forbidden; drop undirected if present.""" if not self._has_forbidden(u, v): self._adj_undirected.discard(self._key(u, v)) self._adj_directed.add((u, v)) def _add_undirected(self, u: str, v: str) -> None: """Add an undirected edge ``u -- v`` if allowed and not already directed.""" if ( not self._has_forbidden(u, v) and (u, v) not in self._adj_directed and (v, u) not in self._adj_directed and not self._is_required(u, v) and not self._is_required(v, u) ): self._adj_undirected.add(self._key(u, v)) def _remove_all(self, u: str, v: str) -> None: """Remove any edge (directed or undirected) between ``u`` and ``v``.""" if self._is_required(u, v) or self._is_required(v, u): return self._adj_undirected.discard(self._key(u, v)) self._adj_directed.discard((u, v)) self._adj_directed.discard((v, u)) def _enforce_required_edges(self) -> None: """Force required edges to appear as directed adjacencies.""" for u, v in self.required_edges: self._adj_undirected.discard(self._key(u, v)) self._adj_directed.discard((v, u)) self._adj_directed.add((u, v)) self.test_results[(u, v, EMPTY_CONDITION_SET)] = { "bic0": float("nan"), "bic1": float("nan"), "delta_bic": float("nan"), "logBF10": float("nan"), "BF10": float("nan"), "independent": False, "conditioning_set": [], "forced": True, } def _validate_required_nodes(self, drivers: Sequence[str]) -> None: """Ensure required edges reference known nodes.""" allowed = set(drivers) | {self.target} missing: set[str] = set() for u, v in self.required_edges: if u not in allowed: missing.add(u) if v not in allowed: missing.add(v) if missing: raise ValueError( "Required edges reference unknown nodes: " + ", ".join(sorted(missing)) ) def _build_symbolic_bic_fn(self): """Build a BIC callable using a fast solver with a pseudoinverse fallback.""" X = pt.matrix("X") n = pt.iscalar("n") xtx = pt.dot(X.T, X) xty = pt.dot(X.T, self.y_sh) beta_solve = pt.linalg.solve(xtx, xty) resid_solve = self.y_sh - pt.dot(X, beta_solve) rss_solve = pt.sum(resid_solve**2) beta_pinv = pt.nlinalg.pinv(X) @ self.y_sh resid_pinv = self.y_sh - pt.dot(X, beta_pinv) rss_pinv = pt.sum(resid_pinv**2) k = X.shape[1] nf = pt.cast(n, "float64") rss_solve_safe = pt.maximum(rss_solve, np.finfo("float64").tiny) rss_pinv_safe = pt.maximum(rss_pinv, np.finfo("float64").tiny) bic_solve = nf * pt.log(rss_solve_safe / nf) + k * pt.log(nf) bic_pinv = nf * pt.log(rss_pinv_safe / nf) + k * pt.log(nf) bic_solve_fn = pytensor.function( [X, n], [bic_solve, rss_solve], on_unused_input="ignore", mode="FAST_RUN" ) bic_pinv_fn = pytensor.function( [X, n], bic_pinv, on_unused_input="ignore", mode="FAST_RUN" ) def bic_fn(X_val: np.ndarray, n_val: int) -> float: try: bic_value, rss_value = bic_solve_fn(X_val, n_val) if np.isfinite(rss_value) and rss_value > np.finfo("float64").tiny: return float(bic_value) except (np.linalg.LinAlgError, RuntimeError, ValueError): pass return float(bic_pinv_fn(X_val, n_val)) return bic_fn def _ci_independent( self, df: pd.DataFrame, x: str, y: str, cond: Sequence[str] ) -> bool: """Return True if ΔBIC indicates independence of ``x`` and ``y`` given ``cond``.""" if self._has_forbidden(x, y): return True if self._is_required(x, y) or self._is_required(y, x): self.test_results[(x, y, frozenset(cond))] = TestResult( bic0=float("nan"), bic1=float("nan"), delta_bic=float("nan"), logBF10=float("nan"), BF10=float("nan"), independent=False, conditioning_set=list(cond), forced=True, ) return False n = len(df) self.y_sh.set_value(df[y].to_numpy().astype("float64")) if len(cond) == 0: X0 = np.ones((n, 1)) else: X0 = np.column_stack([np.ones(n), df[list(cond)].to_numpy()]) X1 = np.column_stack([X0, df[x].to_numpy()]) bic0 = float(self._bic_fn(X0, n)) bic1 = float(self._bic_fn(X1, n)) delta_bic = bic1 - bic0 logBF10 = -0.5 * delta_bic BF10 = np.exp(logBF10) independence = BF10 < self.bf_thresh result: TestResult = { "bic0": bic0, "bic1": bic1, "delta_bic": delta_bic, "logBF10": logBF10, "BF10": BF10, "independent": independence, "conditioning_set": list(cond), } self.test_results[(x, y, frozenset(cond))] = result return independence def _test_target_edges(self, df: pd.DataFrame, drivers: Sequence[str]) -> None: """Phase 1: test driver→target edges according to ``target_edge_rule``.""" for xi in drivers: neighbor_sets = [d for d in drivers if d != xi] max_k = min(3, len(neighbor_sets)) all_sets = [ tuple(S) for k in range(max_k + 1) for S in it.combinations(neighbor_sets, k) ] if self.target_edge_rule == "any": keep = True for S in all_sets: if self._ci_independent(df, xi, self.target, S): self._set_sep(xi, self.target, S) keep = False break if keep: self._add_directed(xi, self.target) else: self._remove_all(xi, self.target) elif self.target_edge_rule == "conservative": indep_all = True for S in all_sets: if not self._ci_independent(df, xi, self.target, S): indep_all = False else: self._set_sep(xi, self.target, S) if indep_all: self._remove_all(xi, self.target) else: self._add_directed(xi, self.target) elif self.target_edge_rule == "fullS": S = tuple(neighbor_sets) if self._ci_independent(df, xi, self.target, S): self._set_sep(xi, self.target, S) self._remove_all(xi, self.target) else: self._add_directed(xi, self.target) def _test_driver_skeleton(self, df: pd.DataFrame, drivers: Sequence[str]) -> None: """Phase 2: build the undirected driver skeleton via pairwise CI tests.""" for xi, xj in it.combinations(drivers, 2): others = [d for d in drivers if d not in (xi, xj)] max_k = min(3, len(others)) dependent = True sep_rec = False for k in range(max_k + 1): for S in it.combinations(others, k): if self._ci_independent(df, xi, xj, S): self._set_sep(xi, xj, S) dependent = False sep_rec = True break if sep_rec: break if dependent: self._add_undirected(xi, xj) else: self._remove_all(xi, xj)
[docs] def fit(self, df: pd.DataFrame, drivers: Sequence[str]): """Fit the TBFPC procedure to the supplied dataframe. Parameters ---------- df : pandas.DataFrame Dataset containing the target column and every candidate driver. drivers : Sequence[str] Iterable of column names to treat as potential drivers of the target. Returns ------- TBFPC The fitted instance (``self``) with internal adjacency structures populated. Examples -------- .. code-block:: python model = TBFPC(target="Y", target_edge_rule="fullS") model.fit(df, drivers=["A", "B", "C"]) """ self._validate_required_nodes(drivers) self.sep_sets.clear() self._adj_directed.clear() self._adj_undirected.clear() self.test_results.clear() self._enforce_required_edges() self._test_target_edges(df, drivers) self._test_driver_skeleton(df, drivers) self._enforce_required_edges() self.nodes_ = [*list(drivers), self.target] return self
[docs] def get_directed_edges(self) -> list[tuple[str, str]]: """Return directed edges learned by the algorithm. Returns ------- list[tuple[str, str]] Sorted list of ``(u, v)`` pairs representing oriented edges. Examples -------- .. code-block:: python directed = model.get_directed_edges() """ return sorted(self._adj_directed)
[docs] def get_undirected_edges(self) -> list[tuple[str, str]]: """Return undirected edges remaining after orientation. Returns ------- list[tuple[str, str]] Sorted list of ``(u, v)`` pairs for unresolved adjacencies. Examples -------- .. code-block:: python skeleton = model.get_undirected_edges() """ return sorted(self._adj_undirected)
[docs] def get_test_results(self, x: str, y: str) -> list[TestResult]: """Return ΔBIC diagnostics for the unordered pair ``(x, y)``. Parameters ---------- x : str Name of the first variable in the pair. y : str Name of the second variable in the pair. Returns ------- list[dict[str, float]] Each dictionary contains ``bic0``, ``bic1``, ``delta_bic``, ``logBF10``, ``BF10``, and the conditioning set used during the test. Examples -------- .. code-block:: python stats = model.get_test_results("A", "Y") """ return [v for (xi, yi, _), v in self.test_results.items() if {xi, yi} == {x, y}]
[docs] def summary(self) -> str: """Render a text summary of the learned graph and test count. Returns ------- str Multiline string describing directed edges, undirected edges, and the number of conditional independence tests executed. Examples -------- .. code-block:: python print(model.summary()) """ lines = ["=== Directed edges ==="] for u, v in self.get_directed_edges(): suffix = " [required]" if self._is_required(u, v) else "" lines.append(f"{u} -> {v}{suffix}") lines.append("=== Undirected edges ===") for u, v in self.get_undirected_edges(): lines.append(f"{u} -- {v}") lines.append("=== Number of CI tests run ===") lines.append(str(len(self.test_results))) return "\n".join(lines)
[docs] def to_digraph(self) -> str: """Return the learned graph encoded in DOT format. Returns ------- str DOT string compatible with Graphviz rendering utilities. Examples -------- .. code-block:: python dot_str = model.to_digraph() """ lines = ["digraph G {", " node [shape=ellipse];"] for n in self.nodes_: if n == self.target: lines.append(f' "{n}" [style=filled, fillcolor="#eef5ff"];') else: lines.append(f' "{n}";') for u, v in self.get_directed_edges(): attrs = " [color=darkgreen, penwidth=2]" if self._is_required(u, v) else "" lines.append(f' "{u}" -> "{v}"{attrs};') for u, v in self.get_undirected_edges(): lines.append(f' "{u}" -> "{v}" [style=dashed, dir=none];') lines.append("}") return "\n".join(lines)
[docs] def get_all_cdags_from_cpdag(self, dot_cpdag: str | None = None) -> list[str]: """ Enumerate all acyclic orientations (consistent extensions) of the CPDAG. Parameters ---------- dot_cpdag : str | None If provided, parse the CPDAG from this DOT string (expects undirected edges encoded as `[style=dashed, dir=none]`). If None, use the model's current CPDAG from `self.get_directed_edges()` and `self.get_undirected_edges()`. Returns ------- list[str] A list of DOT strings, each representing a fully oriented DAG (no dashed edges). """ nodes, fixed_dir, undirected = ( self._parse_cpdag_dot(dot_cpdag) if dot_cpdag is not None else ( set(self.nodes_), set(self.get_directed_edges()), set(self.get_undirected_edges()), ) ) if not undirected: # Already a DAG: validate acyclicity and return it edges = sorted(fixed_dir) if self._is_acyclic(nodes, edges): return [self._dot_from_edges(nodes, edges)] return [] cdags: list[str] = [] und = sorted({self._key(u, v) for (u, v) in undirected}) # canonical ordering for mask in self._bitmasks(len(und)): oriented = list(fixed_dir) # bit 0 -> u->v, bit 1 -> v->u oriented.extend( (u, v) if b == 0 else (v, u) for b, (u, v) in zip(mask, und, strict=False) ) if self._is_acyclic(nodes, oriented): cdags.append(self._dot_from_edges(nodes, oriented)) return cdags
[docs] class CausalGraphModel: """Represent a causal model based on a Directed Acyclic Graph (DAG). Provides methods to analyze causal relationships and determine the minimal adjustment set for backdoor adjustment between treatment and outcome variables. Parameters ---------- causal_model : CausalModel An instance of dowhy's CausalModel, representing the causal graph and its relationships. treatment : list[str] A list of treatment variable names. outcome : str The outcome variable name. References ---------- .. [1] https://github.com/microsoft/dowhy """
[docs] def __init__( self, causal_model: CausalModel, treatment: list[str] | tuple[str], outcome: str ) -> None: self.causal_model = causal_model self.treatment = treatment self.outcome = outcome
[docs] @classmethod def build_graphical_model( cls, graph: str, treatment: list[str] | tuple[str], outcome: str ) -> CausalGraphModel: """Create a CausalGraphModel from a string representation of a graph. Parameters ---------- graph : str A string representation of the graph (e.g., String in DOT format). treatment : list[str] A list of treatment variable names. outcome : str The outcome variable name. Returns ------- CausalGraphModel An instance of CausalGraphModel constructed from the given graph string. """ causal_model = CausalModel( data=pd.DataFrame(), graph=graph, treatment=treatment, outcome=outcome ) return cls(causal_model, treatment, outcome)
[docs] def get_backdoor_paths(self) -> list[list[str]]: """Find all backdoor paths between the combined treatment and outcome variables. Returns ------- list[list[str]] A list of backdoor paths, where each path is represented as a list of variable names. References ---------- .. [1] Causal Inference in Statistics: A Primer By Judea Pearl, Madelyn Glymour, Nicholas P. Jewell · 2016 """ # Use DoWhy's internal method to get backdoor paths for all treatments combined return self.causal_model._graph.get_backdoor_paths( nodes1=self.treatment, nodes2=[self.outcome] )
[docs] def get_unique_adjustment_nodes(self) -> list[str]: """Compute the minimal adjustment set required for backdoor adjustment across all treatments. Returns ------- list[str] A list of unique adjustment variables needed to block all backdoor paths. """ paths = self.get_backdoor_paths() # Flatten paths and exclude treatments and outcome from adjustment set adjustment_nodes = set( node for path in paths for node in path if node not in self.treatment and node != self.outcome ) return list(adjustment_nodes)
[docs] def compute_adjustment_sets( self, channel_columns: list[str] | tuple[str], control_columns: list[str] | None = None, ) -> list[str] | None: """Compute minimal adjustment sets and handle warnings.""" channel_columns = list(channel_columns) if control_columns is None: return control_columns self.adjustment_set = self.get_unique_adjustment_nodes() common_controls = set(control_columns).intersection(self.adjustment_set) unique_controls = set(control_columns) - set(self.adjustment_set) if unique_controls: warnings.warn( f"Columns {unique_controls} are not in the adjustment set. Controls are being modified.", stacklevel=2, ) control_columns = list(common_controls - set(channel_columns)) self.minimal_adjustment_set = control_columns + list(channel_columns) for column in self.adjustment_set: if column not in control_columns and column not in channel_columns: warnings.warn( f"""Column {column} in adjustment set not found in data. Not controlling for this may induce bias in treatment effect estimates.""", stacklevel=2, ) return control_columns