Causality, the sphere centered on understanding the relationships between trigger and impact, seeks to reply essential questions resembling ‘Why?’ and ‘What if?’. Understanding the idea of causality is essential from preventing local weather change, to our quest for happiness, together with strategic choices making.

Examples of main questions requiring causal inference:

• What influence would possibly banning gasoline automobiles have on air pollution?
• What are the causes behind the unfold of sure well being points?
• Might lowering display time result in elevated happiness?
• What’s the Return On Funding of our advert marketing campaign?

In what follows I’ll primarily refer to 2 free e-books accessible with Python code and knowledge to play with. The primary e-book affords fast overviews, whereas the second permits for a extra in-depth exploration of the content material.

1. Causal Inference for the Courageous and True by Matheus Facure

2. Causal Inference: The Mixtape by Scott Cuningham

1.1 The elemental drawback of causal inference

Let’s dive into essentially the most elementary idea obligatory to know causal inference via a state of affairs we would all be conversant in.

Think about that you’ve got been working in your laptop all day lengthy, a deadline is approaching, and also you begin to really feel a headache approaching. You continue to have a couple of hours of labor forward, so that you resolve to take a tablet. After some time, your headache is gone.

However then, you begin questioning: Was it actually the tablet that made the distinction? Or was it since you drank tea or took a break? The fascinating however finally additionally irritating half is that it’s unattainable to reply this query as all these results are confounded.

The one method to know for sure if it was the tablet that cured your headache can be to have two parallel worlds.

In one of many two worlds you are taking the tablet, and within the different, you don’t, otherwise you take a placebo ideally. You’ll be able to solely show the tablet’s causal impact for those who really feel higher on this planet the place you took the tablet, because the tablet is the one distinction between the 2 worlds.

Sadly, we do not need entry to parallel worlds to experiment with and assess causality. Therefore, many elements happen concurrently and are confounded (e.g., taking a tablet for a headache, ingesting tea, and taking a break; rising advert spending throughout peak gross sales seasons; assigning extra cops to areas with larger crime charges, and many others.).

To rapidly grasp this elementary idea in additional depth with out requiring any extra technical data, you may dive into the next article on In the direction of Information Science:

📚 Useful resource:

The Science and Artwork of Causality (half 1)

1.2 Just a little little bit of formalization: Potential Outcomes

Now that you simply perceive the essential thought, it’s time to go additional and theoretically formalize these ideas. The commonest strategy is the potential outcomes framework, which permits for the clear articulation of mannequin assumptions. These are important for specifying the issues and figuring out the options.

The central notation used on this mannequin are:

1. Yᵢ(0) represents the potential end result of particular person i with out the therapy.
2. Yᵢ(1) represents the potential end result of particular person i with the therapy.

Word that varied notations are used. The reference to the therapy (1 or 0) could seem in parentheses (as used above), in superscript, or subscript. The letter “Y” refers back to the end result of curiosity, resembling a binary variable that takes the worth one if a headache is current and nil in any other case. The subscript “i” refers back to the noticed entity (e.g., an individual, a lab rat, a metropolis, and many others.). Lastly, the time period ‘therapy’ refers back to the ‘trigger’ you have an interest in (e.g., a tablet, an commercial, a coverage, and many others.).

Utilizing this notation, we will check with the elemental drawback of causal inference by stating that it’s unattainable to watch each Yᵢ(0) and Yᵢ(1) concurrently. In different phrases, you by no means observe the result for a similar particular person with and with out the therapy on the similar time.

Whereas we can’t determine the person impact Yᵢ(1)-Yᵢ(0), we will measure the Common Therapy Impact (ATE): E[Yᵢ(1)-Yᵢ(0)]. Nonetheless, this Common Therapy Impact is biased when you’ve got systematic variations between the 2 teams aside from the therapy.

To transcend this brief introduction you may check with the 2 following chapters:

📚Sources:

1.3 Visible illustration of causal hyperlinks: Directed (Acyclic) Graphs

Visible representations are highly effective instruments for lowering psychological load, clarifying assumptions, and facilitating communication. In causal inference, we use directed graphs. Because the title suggests, these graphs depict varied parts (e.g., headache, tablet, tea) as nodes, related by unidirectional arrows that illustrate the path of causal relationships. (Word: I intentionally omit mentioning the widespread assumption of ‘Acyclicity’ related to these graphs, because it goes past the scope of this overview however is mentioned within the second reference accessible on the finish of this subsection.)

Causal inference primarly differs from predictive inference because of the assumed underlying causal relationships. These relationships are explicitly represented utilizing this particular sort of graph known as Directed (Acyclic) Graphs. This software collectively with the potential end result framework is on the core of causal inference and can enable considering clearly in regards to the potential issues and, consequently, options for assessing causality

📚Sources: