instrumental_variables

notebooks/instrumental_variables.ipynb

@@ -17,33 +17,24 @@
    "cell_type": "markdown",
    "id": "7963ad8f-4844-4beb-b8fd-48b73a2d0505",
    "metadata": {},
-   "source": [
-    "In many set"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "66dcf13e-d058-4793-bba3-f085b9d0deda",
-   "metadata": {},
-   "outputs": [],
    "source": []
   },
   {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "832aa948-5775-4731-93d1-c86f2028777a",
+   "cell_type": "markdown",
+   "id": "ae9e0548-39f6-41ff-a682-0fca168b4fae",
    "metadata": {},
-   "outputs": [],
-   "source": []
+   "source": [
+    "## Evaluating Subscriptions"
+   ]
   },
   {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "15613d4c-484a-4188-a1e9-93e01d4c597e",
+   "cell_type": "markdown",
+   "id": "cc8daed1-fd1b-4a40-85d0-d4b92beb6c05",
    "metadata": {},
-   "outputs": [],
-   "source": []
+   "source": [
+    "In the rest of the article, we are going to use a toy example. Suppose you were an online company and you have a weekly **newsletter**. You would like to understand whether the newsletter is worth the effort and it is actually successfull in increasing sales. Unfortunately, you cannot run a standard A/B test since you cannot force customers to subscribe to the newsletter. Does it mean you cannot evaluate the newsletter? Not exactly. \n",
+    "\n"
+   ]
   },
   {
    "cell_type": "code",
@@ -832,15 +823,17 @@
     }
    },
    "source": [
-    "## Instrumental Variables and Machine Learning"
+    "## The Limits of IV"
    ]
   },
   {
    "cell_type": "markdown",
    "id": "739c0518-8d23-4e84-b101-986e1ee72827",
    "metadata": {},
    "source": [
-    "Before concluding, I would like to point towards an interesting feature of instrumental variable"
+    "The main limit of instrumental variables in an expeerimental setting, as analyzed in this article, is that they estimate a very \"special\" causal effect. As we have seen in the previous section, we have to re-scale the total effect by the number of customers decided to `subscribe` because of the `newsletter`. This means that we can estimate the effect only for those customer that we *complied* with our interventions. This set of customers is usually called **compliers** and the corresponding causal effect is someimes called the **Local Average Treatment Effect (LATE)** or **Compliers Average Causal Effect (CACE)**.\n",
+    "\n",
+    "We are unfortunately unable to say anything for what concerns customers that subscribed to the `newsletter` even without the `notification`, called the **always-takers**, and those that we couldn't convince with the `notification`, the **never-takers**."
    ]
   },
   {