{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Conditional Average Treatment Effects (CATE) with DoWhy and EconML\n",
    "\n",
    "This is an experimental feature where we use [EconML](https://github.com/microsoft/econml) methods from DoWhy. Using EconML allows CATE estimation using different methods. \n",
    "\n",
    "All four steps of causal inference in DoWhy remain the same: model, identify, estimate, and refute. The key difference is that we now call econml methods in the estimation step. There is also a simpler example using linear regression to understand the intuition behind CATE estimators. \n",
    "\n",
    "All datasets are generated using linear structural equations.\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pandas as pd\n",
    "import logging\n",
    "\n",
    "import dowhy\n",
    "from dowhy import CausalModel\n",
    "import dowhy.datasets\n",
    "\n",
    "import econml\n",
    "import warnings\n",
    "warnings.filterwarnings('ignore')\n",
    "\n",
    "BETA = 10"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "data = dowhy.datasets.linear_dataset(BETA, num_common_causes=4, num_samples=10000,\n",
    "                                    num_instruments=2, num_effect_modifiers=2,\n",
    "                                     num_treatments=1,\n",
    "                                    treatment_is_binary=False,\n",
    "                                    num_discrete_common_causes=2,\n",
    "                                    num_discrete_effect_modifiers=0,\n",
    "                                    one_hot_encode=False)\n",
    "df=data['df']\n",
    "print(df.head())\n",
    "print(\"True causal estimate is\", data[\"ate\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = CausalModel(data=data[\"df\"], \n",
    "                    treatment=data[\"treatment_name\"], outcome=data[\"outcome_name\"], \n",
    "                    graph=data[\"gml_graph\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.view_model()\n",
    "from IPython.display import Image, display\n",
    "display(Image(filename=\"causal_model.png\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "identified_estimand= model.identify_effect(proceed_when_unidentifiable=True)\n",
    "print(identified_estimand)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Linear Model \n",
    "First, let us build some intuition using a linear model for estimating CATE. The effect modifiers (that lead to a heterogeneous treatment effect) can be modeled as interaction terms with the treatment. Thus, their value modulates the effect of treatment. \n",
    "\n",
    "Below the estimated effect of changing treatment from 0 to 1. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "linear_estimate = model.estimate_effect(identified_estimand, \n",
    "                                        method_name=\"backdoor.linear_regression\",\n",
    "                                       control_value=0,\n",
    "                                       treatment_value=1)\n",
    "print(linear_estimate) "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## EconML methods\n",
    "We now move to the more advanced methods from the EconML package for estimating CATE.\n",
    "\n",
    "First, let us look at the double machine learning estimator. Method_name corresponds to the fully qualified name of the class that we want to use. For double ML, it is \"econml.dml.DML\". \n",
    "\n",
    "Target units defines the units over which the causal estimate is to be computed. This can be a lambda function filter on the original dataframe, a new Pandas dataframe, or a string corresponding to the three main kinds of target units (\"ate\", \"att\" and \"atc\"). Below we show an example of a lambda function. \n",
    "\n",
    "Method_params are passed directly to EconML. For details on allowed parameters, refer to the EconML documentation. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.preprocessing import PolynomialFeatures\n",
    "from sklearn.linear_model import LassoCV\n",
    "from sklearn.ensemble import GradientBoostingRegressor\n",
    "dml_estimate = model.estimate_effect(identified_estimand, method_name=\"backdoor.econml.dml.DML\",\n",
    "                                     control_value = 0,\n",
    "                                     treatment_value = 1,\n",
    "                                 target_units = lambda df: df[\"X0\"]>1,  # condition used for CATE\n",
    "                                 confidence_intervals=False,\n",
    "                                method_params={\"init_params\":{'model_y':GradientBoostingRegressor(),\n",
    "                                                              'model_t': GradientBoostingRegressor(),\n",
    "                                                              \"model_final\":LassoCV(fit_intercept=False), \n",
    "                                                              'featurizer':PolynomialFeatures(degree=1, include_bias=False)},\n",
    "                                               \"fit_params\":{}})\n",
    "print(dml_estimate)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(\"True causal estimate is\", data[\"ate\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dml_estimate = model.estimate_effect(identified_estimand, method_name=\"backdoor.econml.dml.DML\",\n",
    "                                     control_value = 0,\n",
    "                                     treatment_value = 1,\n",
    "                                 target_units = 1,  # condition used for CATE\n",
    "                                 confidence_intervals=False,\n",
    "                                method_params={\"init_params\":{'model_y':GradientBoostingRegressor(),\n",
    "                                                              'model_t': GradientBoostingRegressor(),\n",
    "                                                              \"model_final\":LassoCV(fit_intercept=False), \n",
    "                                                              'featurizer':PolynomialFeatures(degree=1, include_bias=True)},\n",
    "                                               \"fit_params\":{}})\n",
    "print(dml_estimate)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### CATE Object and Confidence Intervals\n",
    "EconML provides its own methods to compute confidence intervals. Using BootstrapInference in the example below. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.preprocessing import PolynomialFeatures\n",
    "from sklearn.linear_model import LassoCV\n",
    "from sklearn.ensemble import GradientBoostingRegressor\n",
    "from econml.inference import BootstrapInference\n",
    "dml_estimate = model.estimate_effect(identified_estimand, \n",
    "                                     method_name=\"backdoor.econml.dml.DML\",\n",
    "                                     target_units = \"ate\",\n",
    "                                     confidence_intervals=True,\n",
    "                                     method_params={\"init_params\":{'model_y':GradientBoostingRegressor(),\n",
    "                                                              'model_t': GradientBoostingRegressor(),\n",
    "                                                              \"model_final\": LassoCV(fit_intercept=False), \n",
    "                                                              'featurizer':PolynomialFeatures(degree=1, include_bias=True)},\n",
    "                                               \"fit_params\":{\n",
    "                                                               'inference': BootstrapInference(n_bootstrap_samples=100, n_jobs=-1),\n",
    "                                                            }\n",
    "                                              })\n",
    "print(dml_estimate)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Can provide a new inputs as target units and estimate CATE on them."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "test_cols= data['effect_modifier_names'] # only need effect modifiers' values\n",
    "test_arr = [np.random.uniform(0,1, 10) for _ in range(len(test_cols))] # all variables are sampled uniformly, sample of 10\n",
    "test_df = pd.DataFrame(np.array(test_arr).transpose(), columns=test_cols)\n",
    "dml_estimate = model.estimate_effect(identified_estimand, \n",
    "                                     method_name=\"backdoor.econml.dml.DML\",\n",
    "                                     target_units = test_df,\n",
    "                                     confidence_intervals=False,\n",
    "                                     method_params={\"init_params\":{'model_y':GradientBoostingRegressor(),\n",
    "                                                              'model_t': GradientBoostingRegressor(),\n",
    "                                                              \"model_final\":LassoCV(), \n",
    "                                                              'featurizer':PolynomialFeatures(degree=1, include_bias=True)},\n",
    "                                               \"fit_params\":{}\n",
    "                                              })\n",
    "print(dml_estimate.cate_estimates)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Can also retrieve the raw EconML estimator object for any further operations"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(dml_estimate._estimator_object)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Works with any EconML method\n",
    "In addition to double machine learning, below we example analyses using orthogonal forests, DRLearner (bug to fix), and neural network-based instrumental variables. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Binary treatment, Binary outcome"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_binary = dowhy.datasets.linear_dataset(BETA, num_common_causes=4, num_samples=10000,\n",
    "                                    num_instruments=1, num_effect_modifiers=2,\n",
    "                                    treatment_is_binary=True, outcome_is_binary=True)\n",
    "# convert boolean values to {0,1} numeric\n",
    "data_binary['df'].v0 = data_binary['df'].v0.astype(int)\n",
    "data_binary['df'].y = data_binary['df'].y.astype(int)\n",
    "print(data_binary['df'])\n",
    "\n",
    "model_binary = CausalModel(data=data_binary[\"df\"], \n",
    "                    treatment=data_binary[\"treatment_name\"], outcome=data_binary[\"outcome_name\"], \n",
    "                    graph=data_binary[\"gml_graph\"])\n",
    "identified_estimand_binary = model_binary.identify_effect(proceed_when_unidentifiable=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Using DRLearner estimator"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.linear_model import LogisticRegressionCV\n",
    "#todo needs binary y\n",
    "drlearner_estimate = model_binary.estimate_effect(identified_estimand_binary, \n",
    "                                method_name=\"backdoor.econml.dr.LinearDRLearner\",\n",
    "                                confidence_intervals=False,\n",
    "                                method_params={\"init_params\":{\n",
    "                                                    'model_propensity': LogisticRegressionCV(cv=3, solver='lbfgs', multi_class='auto')\n",
    "                                                    },\n",
    "                                               \"fit_params\":{}\n",
    "                                              })\n",
    "print(drlearner_estimate)\n",
    "print(\"True causal estimate is\", data_binary[\"ate\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Instrumental Variable Method"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "dmliv_estimate = model.estimate_effect(identified_estimand, \n",
    "                                        method_name=\"iv.econml.iv.dml.DMLIV\",\n",
    "                                        target_units = lambda df: df[\"X0\"]>-1, \n",
    "                                        confidence_intervals=False,\n",
    "                                method_params={\"init_params\":{\n",
    "                                                              'discrete_treatment':False,\n",
    "                                                              'discrete_instrument':False\n",
    "                                                             },\n",
    "                                               \"fit_params\":{}})\n",
    "print(dmliv_estimate)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Metalearners"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_experiment = dowhy.datasets.linear_dataset(BETA, num_common_causes=5, num_samples=10000,\n",
    "                                    num_instruments=2, num_effect_modifiers=5,\n",
    "                                    treatment_is_binary=True, outcome_is_binary=False)\n",
    "# convert boolean values to {0,1} numeric\n",
    "data_experiment['df'].v0 = data_experiment['df'].v0.astype(int)\n",
    "print(data_experiment['df'])\n",
    "model_experiment = CausalModel(data=data_experiment[\"df\"], \n",
    "                    treatment=data_experiment[\"treatment_name\"], outcome=data_experiment[\"outcome_name\"], \n",
    "                    graph=data_experiment[\"gml_graph\"])\n",
    "identified_estimand_experiment = model_experiment.identify_effect(proceed_when_unidentifiable=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.ensemble import RandomForestRegressor\n",
    "metalearner_estimate = model_experiment.estimate_effect(identified_estimand_experiment, \n",
    "                                method_name=\"backdoor.econml.metalearners.TLearner\",\n",
    "                                confidence_intervals=False,\n",
    "                                method_params={\"init_params\":{\n",
    "                                                    'models': RandomForestRegressor()\n",
    "                                                    },\n",
    "                                               \"fit_params\":{}\n",
    "                                              })\n",
    "print(metalearner_estimate)\n",
    "print(\"True causal estimate is\", data_experiment[\"ate\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Avoiding retraining the estimator "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Once an estimator is fitted, it can be reused to estimate effect on different data points. In this case, you can pass `fit_estimator=False` to `estimate_effect`. This works for any EconML estimator. We show an example for the T-learner below."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# For metalearners, need to provide all the features (except treatmeant and outcome)\n",
    "metalearner_estimate = model_experiment.estimate_effect(identified_estimand_experiment, \n",
    "                                method_name=\"backdoor.econml.metalearners.TLearner\",\n",
    "                                confidence_intervals=False,\n",
    "                                fit_estimator=False,\n",
    "                                target_units=data_experiment[\"df\"].drop([\"v0\",\"y\", \"Z0\", \"Z1\"], axis=1)[9995:],                        \n",
    "                                method_params={})\n",
    "print(metalearner_estimate)\n",
    "print(\"True causal estimate is\", data_experiment[\"ate\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Refuting the estimate"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Adding a random common cause variable"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "res_random=model.refute_estimate(identified_estimand, dml_estimate, method_name=\"random_common_cause\")\n",
    "print(res_random)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Adding an unobserved common cause variable"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "res_unobserved=model.refute_estimate(identified_estimand, dml_estimate, method_name=\"add_unobserved_common_cause\",\n",
    "                                     confounders_effect_on_treatment=\"linear\", confounders_effect_on_outcome=\"linear\",\n",
    "                                    effect_strength_on_treatment=0.01, effect_strength_on_outcome=0.02)\n",
    "print(res_unobserved)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Replacing treatment with a random (placebo) variable"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "res_placebo=model.refute_estimate(identified_estimand, dml_estimate,\n",
    "        method_name=\"placebo_treatment_refuter\", placebo_type=\"permute\",\n",
    "        num_simulations=10 # at least 100 is good, setting to 10 for speed \n",
    "        ) \n",
    "print(res_placebo)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Removing a random subset of the data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "res_subset=model.refute_estimate(identified_estimand, dml_estimate,\n",
    "        method_name=\"data_subset_refuter\", subset_fraction=0.8,\n",
    "        num_simulations=10)\n",
    "print(res_subset)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "More refutation methods to come, especially specific to the CATE estimators."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.10"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": false,
   "sideBar": true,
   "skip_h1_title": true,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": false
  },
  "vscode": {
   "interpreter": {
    "hash": "dcb481ad5d98e2afacd650b2c07afac80a299b7b701b553e333fc82865502500"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}