Chapter 1: Introduction to Casual Inference Flashcards
What is the difference between correlation and causation?
Correlation occurs when two quantities or random variables move together, where as causation is when a change in one variable changes the other.
What is causal inference?
Causal inference is the science of inferring causation from association and understanding when and why they differ
Why do we do causal inference?
We do causal inference because we want to know cause-and-effect relationships so that we can intervene on the cause to bring upon a desired effect. E.g. if we know certain marketing spend increases sales profits, managers can leverage that to increase profits.
What is the fundamental problem of causal inference?
The fundamental problem of causal inference is that you can never observe the same unit with and without treatment.
What is potential outcome?
A potential outcome with reference to an outcome i, is what i’s outcome with be if treated with treatment t.
What are assumptions in casual inference?
Assumptions are statements you make when expressing a belief about how the data was generated. The catch is that they usually can’t be verified with the data; that’s why you need to assume them.
What is the consistency assumption in casual inference?
The consistency assumption in casual inference is that the potential outcome is consistent with the treatment, i.e. there are no hidden multiple versions then the ones you specific.
For example, in marketing, if you are modelling the effect an advert has on the likelihood of a customer buying a product and you model ‘shown advert’ as a binary treatment i.e. shown advert yes, shown advert no, the consisitency assumption can be violated if there are multiple versions of the advert which might have an effect on the customers likelihood of buying.
What is the stable unit of treatment value assumption?
The stable unit of treatment value assumption is that there the treatment on one unit has no effect on another unit, the treatment has no spillover or network effect.
For example, when measuring the effectivness of vaccines, a vaccinated person can reduce the risk of those unvaccinated around them, therefore the SUTVA is violated. Violating the SUTVA can make treatment effects appear smaller than they are as the difference between treated and untreated is smaller due to the spillover of treatment.
How can violations of the constiency assumption and the single unit of treatment value assumption be mitigated?
Constiency: Include all versions of the treatment in ones analysis.
SUTVA: Expanding the definition of the treatment effect to include the effect that comes with other other units and by using more flexible models.
What is the average treatment effect (ATE)?
The average treatment effect represents the impact the treatment T would have on average. Some units will be more impacted by it, some less, and you can never know the individual impact on a unit.
What is average treatment effect on the treated?
Average treatment effect on the treated (ATT) is the treatment effect observed on the population that recieved the treatment.
For example, if you did an offline marketing campaign in a city and you want to know how many extra customers this campaign brought you in that city, this would be the ATT: the effect of marketing on the city where the campaign was implemented.
What the conditional treatment effect?
Conditional treatment effect is the effect observed in a specific group.
For example, you might want to know the effect of an email on customers that are older than 45 years and on those that are younger than that. Conditional average treatment effect is invaluable for personalization, since it allows you to know which type of unit responds better to an intervention.
Causal quanity example:
You can never know the effect of cutting prices have on sales on individual businesses as you cann’t observe both potential outcomes. Given this limit, what is possible to estimate that is useful in this instance?
ATE: The average impact of price cuts on amount sold.
ATT, how the business engaged in price cuts increased sales .
CATE which is the impact of having sales during the week Christmas.
What does bias refer to?
TECHNICAL DEFINITION
You can say that an estimator is biased if it differs from the parameter it is trying to estimate.
Bias = E[β^-β], where
β^ is the estimate.
β the thing it is trying to estimate—the estimand.
For example, an estimator for the average treatment effect is biased if it’s systematically under- or overestimating the true ATE.
What does identification refer to in causal inference?
Figuring out how to recover causal quantities from observed data.
