Add FAQ entry for DML vs R-Learner (#44)

docs/faq.rst

@@ -0,0 +1,34 @@
+FAQ
+========
+
+* **What's the difference between Double Machine Learning, Debiased Machine Learning,
+  Orthogonal Machine Learning, Doubly-Robust Machine Learning and the R-Learner?**
+    The Double Machine Learning blueprint relies on estimating two nuisance models in its
+    first stage: a propensity model as well as an outcome model, both depending only on X.
+    These models are then used to calculate the treatment and outcome residuals and finally
+    the outcome residuals are regressed against the treatment residuals. See
+    `Chernozhukov et al. (2016) <https://arxiv.org/abs/1608.00060>`_ for more details.
+    Some implementations of Double Machine Learning are the
+    `DoubleML Library <https://docs.doubleml.org/stable/index.html>`_ or the
+    `DML module of EconML <https://econml.azurewebsites.net/_autosummary/econml.dml.DML.html>`_.
+
+    Debiased Machine Learning and Orthogonal Machine Learning usually refer to the same
+    algorithms as Double Machine Learning as their goal is to *debias* the used nuisance
+    models with *orthogonalization*.
+
+    The R-Learner is a generalization of the Double Machine Learning framework where instead
+    of regressing the outcome residuals against the treatment residuals, usually with
+    Linear Regression in DML, these are used
+    to build a loss function which can be used with any Machine Learning which supports
+    weighted loss functions. See `Nie and Wager (2017) <https://arxiv.org/abs/1712.04912>`_
+    for more details. The main drawback of this method against Double Machine Learning is
+    the fact that in its standard form it can only be used with binary or single-dimensional
+    continuous treatments, some adaptations can be done to work with categorical treatments
+    but model choice for the final model becomes highly restricted. On the other hand,
+    Double Machine Learning can be used with categorical or continuous treatment.
+
+    On the other hand, Doubly-Robust Machine Learning differs from all the previous methods
+    by the fact that the outcome nuisance models depends not only on the variables X,
+    but also on the treatments T. Then these models are used to build pseudo outcomes which
+    their expected value is the true CATE and a final model is used to learn the CATE
+    from them. See `Kennedy (2020) <https://arxiv.org/abs/2004.14497>`_ for more details.

docs/index.rst

@@ -9,6 +9,7 @@ Welcome to metalearners's documentation!
    Background <background.rst>
    Glossary <glossary.rst>
    Examples <examples/examples.rst>
+   FAQ <faq.rst>
    API Reference <api/modules>
    Change Log <changelog.rst>