Regression for prediction and regression for inference
ComingThe same fitted line does two different jobs, and the job decides what you check.
About 5 minutes when recorded.
Short reviews of the big ideas behind data work in economics, one idea per video, 3 to 6 minutes each. Each one stands alone: watch it when the idea comes up in a lab, a project or an interview, in any order. They are grouped by the stage of the data science process where the idea first matters.
What do you want to know, for whom, and is it a prediction or an effect?
The same fitted line does two different jobs, and the job decides what you check.
About 5 minutes when recorded.
Why a coin flip turns a difference in means into an estimate of a causal effect.
About 5 minutes when recorded.
How many observations you need before you look, and what an underpowered study can and cannot find.
About 5 minutes when recorded.
Where the rows came from, who is in them, and who is missing.
Why a huge sample can still give the wrong answer when the people in it chose to be there.
About 4 minutes when recorded.
Types, dates, missing values, joins, and the test set kept out of sight.
Why a model must be scored on data it never saw, and the quiet ways test information leaks into training.
About 5 minutes when recorded.
Look before you model, and know what a picture can and cannot tell you.
How to draw the story behind the data and read off which variables to control for, and which not to.
About 6 minutes when recorded.
The estimator or the learner, and the assumption each one rests on.
How two groups and two periods estimate an effect when nothing was randomized, and the parallel-trends bet it rests on.
About 6 minutes when recorded.
How a variable that moves the treatment and nothing else recovers a causal effect, and why a weak one is worse than none.
About 6 minutes when recorded.
Why a cutoff rule is a natural experiment for the people just either side of it.
About 5 minutes when recorded.
How shrinking coefficients buys a little bias for a lot less variance, and why you standardize first.
About 5 minutes when recorded.
How to model a yes-or-no outcome as a probability and read a coefficient on the odds scale.
About 5 minutes when recorded.
How wrong could you be, and on data the model never saw?
What a p-value measures, what it does not, and why 0.049 and 0.051 are the same news.
About 4 minutes when recorded.
How shuffling the labels gives you a p-value with no distribution assumed.
About 4 minutes when recorded.
Why a model that fits the data you have can predict the data you do not have worse, and where the sweet spot sits.
About 5 minutes when recorded.
Why accuracy misleads when one class is rare, and how to choose a cutoff for the mistake you care about.
About 5 minutes when recorded.
One number with its interval, one chart that does not mislead.
How to put an interval around an estimate when there is no formula for its standard error.
About 5 minutes when recorded.