Weak instruments

Loading...

From the course by University of Pennsylvania

A Crash Course in Causality: Inferring Causal Effects from Observational Data

92 ratings

University of Pennsylvania

A Crash Course in Causality: Inferring Causal Effects from Observational Data

92 ratings

We have all heard the phrase “correlation does not equal causation.” What, then, does equal causation? This course aims to answer that question and more! Over a period of 5 weeks, you will learn how causal effects are defined, what assumptions about your data and models are necessary, and how to implement and interpret some popular statistical methods. Learners will have the opportunity to apply these methods to example data in R (free statistical software environment). At the end of the course, learners should be able to: 1. Define causal effects using potential outcomes 2. Describe the difference between association and causation 3. Express assumptions with causal graphs 4. Implement several types of causal inference methods (e.g. matching, instrumental variables, inverse probability of treatment weighting) 5. Identify which causal assumptions are necessary for each type of statistical method So join us.... and discover for yourself why modern statistical methods for estimating causal effects are indispensable in so many fields of study!

From the lesson

Instrumental Variables Methods

This module focuses on causal effect estimation using instrumental variables in both randomized trials with non-compliance and in observational studies. The ideas are illustrated with an instrumental variables analysis in R.

Introduction to instrumental variables11:19

Randomized trials with noncompliance11:58

Compliance classes16:30

Assumptions13:29

Causal effect identification and estimation16:49

IVs in observational studies17:06

Two stage least squares15:57

Weak instruments5:25

IV analysis in R16:04

Meet the Instructors

Jason A. Roy, Ph.D.
Professor of Biostatistics
Department of Biostatistics and Epidemiology

This video is on weak instruments.

By the end of the video you should be able to understand how to measure the strength of

an instrumental variable and

also you should be a lot understand why weak instruments cause problems.

So the strength of an instrumental variable

is basically how well it predicts treatment received.

So remember, the instrumental variable we can think of as encouragement,

some encouragement to receive treatment.

And so if it's a strong instrument,

then this encouragement works well and most people

who are encouraged will actually take the treatment.

So a strong instrument is highly predictive of

treatment which means encouragement greatly increases the probability of treatment.

A weak instrument is,

on the other hand, is weakly predictive of treatment.

So an encouragement might increase the probability of

treatment a little bit but not very much.

So we can actually quantify this so we can measure how strong the instrumental variable

is and a way to do that is essentially to estimate the proportion of compliers.

So this is something that we can identify from observed data,

so we can simply remember that the proportion of compliers is just

simply the probability of receiving treatment given that you were encouraged,

so that's this part here.

Expect the value of A|Z=1 so that is a probability that

A=1|D=1 which is the probability of receiving treatment

given that you were encouraged minus the probability of receiving

treatment given that you weren't encouraged so

that difference is the proportion of compliers.

So these are just, you can estimate those from the sample means in your data.

So the proportion of compliers is estimated as the observed proportion of

treated subjects for Z=1 minus the observed proportions of treated subjects for Z=0.

So it's a contrast of two proportions,

and then we'll get an actual number,

so we'll be able to estimate what proportion of

this population have compliers and that will tell us then how strong our instrument is.

So if you, if this is,

if this number is close to one it's

a strong instrument and if it's close to zero it's a weak instrument.

So if you have a weak instrument this can cause problems.

So as an example, I'm imagining that only one percent of the population are compliers.

So this, getting encouraged only bumps up

your probability at least of actually

receiving treatment by just a very small percentage.

So now remember that in instrumental variable analysis we're

interested in estimating local treatment effects,

so local causal effects or complier average causal effects.

So we're sort of focused on the subset of

compliers but if the proportion of compliers and the population is very small,

then essentially our effective sample size is also very small.

So if our original sample size,

the sample size of our dataset is n,

then one percent of n would be

how many people really are contributing information to the causal effect estimate.

So you could imagine that if you essentially have a small sample size,

a small number of compliers,

then we'll end up with noisy estimates of causal effects.

So the causal effect estimates will have a very large variance associated with it.

So we have these unstable estimates of

causal facts and also weak instruments can also lead to bias.

But, I think probably the larger issue is that these estimates will be very unstable,

so if you, your confidence interval would be very wide if you have a weak instrument.

And, you know, so then,

if you did have a weak instrument then an IV analysis might not be the best option.

So a lot of times the first thing that people do when they have

a proposed instrument is to see is it really predicting treatment.

So that's, a lot of times,

a first step you might take is you have an idea and I think this variable

might be a good instrument and then you

want to check is it really predictive of treatment.

And if it's not, you might want to, you know,

consider trying to find a different instrument or doing some other kind of analysis.

There's also an area of active research is that,

are methods to strengthen instrumental variables,

so you might have a instrumental variable that's

weak overall but there might be ways to strengthen it.

So one example is this methodology called near/far

matching and the idea is that you're matching on covariates,

so those kinds of things that you think of as confounder.

So you're making, by matching you're making them people as similar as can be on

these on these covariates but you're also

matching in such a way that the instruments as different as possible.

So a matched pair would include people who,

all their covariates are similar except they differ by a lot on this instrument.

And so that's actually similar to what you would see in

randomized trials where you should, you know,

people should, sort of, have similar covariate distributions

but their level of encouragement would differ greatly.

So there's this near/far matching approach and there's also

other newer statistical methods that are focused on ways to strengthen the instrument.

So those are a couple of options.

There's one, if you have a weak instrument,

you might want to just consider a different kind of analysis.

Alternatively you might think about ways to strengthen the instrument.

Explore our Catalog

Join for free and get personalized recommendations, updates and offers.

Coursera provides universal access to the world’s best education, partnering with top universities and organizations to offer courses online.

© 2018 Coursera Inc. All rights reserved.

Download on the App Store

Get it on Google Play