causalis.dgp.multicausaldata.functional

Notes

The first treatment column is treated as the control arm by convention. When return_causal_data=True, the returned :class:MultiCausalData object uses that first arm as control_treatment.

Examples

from causalis.dgp.multicausaldata.functional import generate_multitreatment data = generate_multitreatment( … n=1000, … n_treatments=3, … theta=[0.0, 0.5, 1.0], … return_causal_data=True, … ) data.treatment_names [‘d_0’, ‘d_1’, ‘d_2’]

High-level generators for multi-arm treatment datasets.

The helpers in this module build datasets where treatment assignment is a full one-hot encoding over multiple arms, with arm-specific structural effects and optional oracle columns.

Module Contents

Functions

generate_multitreatment

Generate a multi-treatment dataset using MultiCausalDatasetGenerator.

API

causalis.dgp.multicausaldata.functional.generate_multitreatment(n: int = 10000, n_treatments: int = 3, outcome_type: str = 'continuous', sigma_y: float = 1.0, alpha_y: float = 0.0, gamma_shape: float = 2.0, tau: Optional[Any] = None, target_d_rate: Optional[Union[List[float], Any]] = None, confounder_specs: Optional[List[Dict[str, Any]]] = None, beta_y: Optional[Any] = None, g_y: Optional[Callable[[numpy.ndarray], numpy.ndarray]] = None, beta_d: Optional[Any] = None, g_d: Optional[Any] = None, theta: Optional[Any] = None, random_state: Optional[int] = 42, k: int = 0, x_sampler: Optional[Any] = None, use_copula: bool = False, copula_corr: Optional[Any] = None, include_oracle: bool = True, return_causal_data: bool = False, d_names: Optional[List[str]] = None) → Union[pandas.DataFrame, causalis.data_contracts.multicausaldata.MultiCausalData]

Generate a multi-treatment dataset using MultiCausalDatasetGenerator.

Parameters

n : int, default=10_000 Number of samples. n_treatments : int, default=3 Number of treatment classes (including control). outcome_type : {“continuous”, “binary”, “poisson”, “gamma”}, default=”continuous” Outcome family. sigma_y : float, default=1.0 Noise level for continuous outcomes. alpha_y : float, default=0.0 Outcome intercept on the structural link scale. gamma_shape : float, default=2.0 Shape parameter when outcome_type="gamma". tau : callable or list of callables, optional Heterogeneous treatment effects per class. When provided with theta, the structural effect is additive on link scale (theta + tau(X)). target_d_rate : array-like, optional Target marginal class probabilities (length K). confounder_specs : list of dict, optional Schema for confounder distributions. beta_y : array-like, optional Linear coefficients for outcome model. g_y : callable, optional Nonlinear baseline outcome function added on the link scale. beta_d : array-like, optional Linear coefficients for treatment model. g_d : callable or list of callables, optional Nonlinear treatment score function(s) added on the softmax-link scale. theta : float or array-like, optional Constant treatment effects per class (adds to tau when both are provided). random_state : int, optional Random seed. k : int, default=0 Number of confounders if confounder_specs is None. x_sampler : callable, optional Custom sampler for confounders. use_copula : bool, default=False Whether to use Gaussian copula sampling for confounders. copula_corr : array-like, optional Correlation matrix used when use_copula=True. include_oracle : bool, default=True Whether to include oracle columns. return_causal_data : bool, default=False Whether to return a MultiCausalData object. d_names : list of str, optional Names of treatment columns.

Returns

pd.DataFrame or MultiCausalData

Notes

Treatment columns form a mutually exclusive one-hot representation:

• exactly one treatment column equals 1 for each row;

• d_names[0] is the control arm;

• arm-specific structural effects are defined on the outcome link scale and are mapped back to the natural outcome scale in oracle outputs.

Examples

from causalis.dgp.multicausaldata.functional import generate_multitreatment df = generate_multitreatment( … n=500, … n_treatments=4, … outcome_type=”continuous”, … theta=[0.0, 0.2, 0.5, 0.8], … ) treatment_cols = [“d_0”, “d_1”, “d_2”, “d_3”] int(df[treatment_cols].sum(axis=1).eq(1).all()) 1