Matching directly on confounders

Loading...

En provenance du cours de University of Pennsylvania

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

132 notes

University of Pennsylvania

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

132 notes

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!

À partir de la leçon

Matching and Propensity Scores

An overview of matching methods for estimating causal effects is presented, including matching directly on confounders and matching on the propensity score. The ideas are illustrated with data analysis examples in R.

Observational studies15:48

Overview of matching12:35

Matching directly on confounders13:21

Greedy (nearest-neighbor) matching17:12

Optimal matching10:40

Assessing balance11:17

Analyzing data after matching20:20

Sensitivity analysis10:28

Data example in R16:53

Rencontrer les enseignants

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

Hi, this video is on matching directly on confounders.

We're going to begin by understanding why a distance matrix is needed,

and understand options for distance matrices.

And then we'll move on to discussions on how distance matrices could be use for

matching.

Here we'll begin to think about how to match.

In order to find close matches, we'll need some metric of closeness.

So we'll at two of them.

Mahalanobis distance, which I'll just call M distance for

short, and robust M distance.

And these are two distance measures that will look at how

similar two sets of covariates are from each other.

As a reminder, we won't be able to match exactly,

because we'll typically have a lot of variables that we need to control for.

And it's very likely that no treated person will have exactly the same set of

variables or values of their variables are as the control subject.

So we'll look at these two different distance measures,

which will look at how similar.

It'll measure the degree to which these sets of variables are similar to

each other.

So now we look at M distance.

And we'll begin with some notations.

So first, X subscript j is just a vector of covariates for subject j.

So j is a person in your data sets.

Subject j is a person in your data set.

So you could think of it as one row in your data.

And X is some collection covariates that you decided ahead of time you needed

to control for.

So these are the set of variables that are sufficient to control for confounding.

So just picture X as values of those variables.

So it could be age, sex, blood pressure, and so on.

So it's a set of values of those variables.

So then, the distance then between subject i and

subject j on the covariates we'll denote with D.

So D(Xi, Xj) just means that's a function that's going to,

as input, you submit the values of the covariates for subject i and subject j.

And as output, it'll give you a score, a single value, which is a distance.

And so we can look at this formula, we'll try to break it down in some detail.

So, first, the inner part here,

this is just the difference between the covariates, so Xi and Xj.

So, remember it's a vector of covariates.

So, for each of the covariates, let's say for example blood pressure,

we would take the difference in blood pressure between subject i and subject j.

If the next variable is age, we would take the difference in age between subject i

and subject j, and so on.

We would run through the whole list.

So then we'd have a vector of differences.

The T here, that's just a transpose.

So that's just a transpose of a vector.

And then here we have the difference again.

So in matrix and vector world, if you have a vector

transpose times another vector like this,

what you're essentially doing is summing these squared differences.

So you could roughly think of this as squaring the differences and

then adding them up.

But there's one additional complication is this part here of this S inverse.

So S, that's just the covariance of X.

So remember X is a bunch of random variables.

So X is a random vector, and those will have some covariants.

So of course,

each of the Xs has a variance but they also might covariate together.

So S is just your standard kind of covariance matrix, and we're inverting it.

So that's what S inverse is.

So you could roughly think of that as just a scaling kind of thing,

where we're scaling by the unit of measure.

And the reason we want to do that is because we have different kinds of

variables here.

So one unit in difference in age will mean something a lot

different than a one unit difference in diabetes status,

where diabetes is a 01 variable.

So for example, if you compare somebody age 67 or

somebody age 68, their age differs by one year.

If you compare somebody who has diabetes versus who doesn't,

that variable differs by one.

But we don't necessarily think those are the same thing.

And differing by one unit of age probably, if it's years, it's probably

not very significant, whereas differing on diabetes status probably is.

This S inverse is going to scale things.

You could roughly think of this as if you had one variable,

what we would do is we would take the difference, let's say, it's age.

We would take the difference in age, square it, divide by the variance, and

then take the square root.

So what we're really doing then is just scaling.

We're scaling each variable, essentially sort of by its variance.

And that's basically,

then, putting all of these different variables on the same kind of scale.

So that hopefully, once we rescale, a difference in one on one

variable means something similar to a difference in one on another variable.

So that's the main idea here.

So that's what the formula is.

You could roughly think of it as a sum of squared

differences that are scaled, and then we take the square root.

But the details maybe aren't quite as important as just the big picture is

if these collection of variables differ by a lot, this distance value will be bigger.

If they don't differ by very much, this distance measure will be small.

So we'll get this value of D for any pair of subjects.

In particular, we're going to get it for

a given treated subject, compared with all the different controls.

So we'll have a list of distances and

then we can use that to help us find good matches.

So we can walk through an example here where imagine that we just have 3

covariates.

So we have age, COPD,

which is just equal to 1=yes, 0=no.

So that's chronic obstructive pulmonary disorder, COPD, and

then female which equals one if yes, zero equals no.

So imagine now we have this treated subject on the left and

we want to find a match for them in the controls.

And imagine there's six controls available just for simplicity, okay?

So this person is age 78.17.

They don't have COPD and they're female.

So as a candidate match, we could look at this first person.

So this person is fairly close in age, they're 70.25, but they have COPD,

whereas the treated person didn't.

And they're male, whereas the treated person's female.

So this is not really a great match overall.

They differ on COPD and sex.

And you'll see the distance, if you used that formula on a previous page,

we would end up with the distance of a little over 4.

Here's a different person.

This person, they actually match on COPD and

sex, but they differ on age by a lot.

So this is an 18 year old compared to a 78 year old.

So they're not a good match, and you see that reflected in the distance.

The distance there is 3.6, so that wouldn't be a good match.

Whereas this person, you'll see that they agree on COPD and sex.

And on age, they don't differ by very much, so

it's a 75 year old compared to a 78 year old.

So it's a close match, and out of this list, you would probably say, okay,

this is the best match.

The distance between their sets of covariates is the smallest, and

we would probably be pretty happy with this particular one.

So that distance, you can use a formula on the previous page,

you can calculate it pretty easily with software such as our.

So now we'll look at an alternative, which is this robust version of a distance.

And the motivation here is that it has to do with outliers.

So outliers could create a great distance between subjects,

even if their covariates are otherwise similar.

So a treated subject and

a control subject who maybe all of their covariates are very close.

But maybe this control subject has one really extreme value of one

of the covariates, unusually large value, for example.

Well, then the distance would be very large here.

Even though they'd seem like otherwise they're a good match.

So an outlier can create a great distance, maybe greater than we would want it to.

So an alternative is to use ranks.

So it's like robust statistics with classical of kind of robust statistics,

where instead of using the original values you use ranks.

You can do the same kind of thing here where just for

the purpose of matching we could replace all of our variables with their ranks.

So if we looked at age, for example, and whatever software you use,

you would ask it to rank the age variable.

So a person who was the youngest in that data set might get a rank of one.

If you had 100 people in your data set,

then the person who was the oldest would maybe get a rank of 100 and so on.

So you would replace the original values with ranks.

And this should help with the issue of outliers.

So now the difference between the largest and

second largest value would be just a difference of one in the ranks.

Whereas if you kept the original values, that could be very extreme

difference between the largest and the second largest.

So that's that the basic idea is that if you want to use this robust

distance measure, you could replace the original values with ranks.

And then apply the same kind of distance measure to it.

So that's the big picture idea, is you replace them with the ranks.

And then there's a subtle thing having to do with the diagonal of the covariates

matrix.

Remember we have that S inverse.

Well, if you're going to use a ranked centipical,

you would want the diagonal of covariates matrix to be a constant.

And that's because, now ranks should all be on the same scale now.

So, you really don't want to weight one variable more than others in that sense.

So again, there's statistical software that could just automatically do

this for you.

But that's the main idea.

And so then you could use this distance matrix on the ranks.

So those were just two distance matrices that you could use, or distance metrics.

And hopefully, it's very clear why we need distance, right?

We need distance because we need to figure out what does close mean?

When we're trying to find a match,

we want to know, how similar are these covariates?

So you could use either of the two measures that I just talked about.

There's other things you can do as well.

So one of the nice things about matching is you can do some kind of clever

things to get exact matches on some variables.

And be more willing to tolerate inexact matches on others.

So, if you want an exact match on a few important covariates,

what you could then is essentially set the distance to infinity if

they don't match on those variables.

So, you could apply the M distance kind of formula.

But then you could take an additional step which says,

we found these important variables.

They're not an exact match.

Replace the original distance with infinity,

essentially,to make it impossible for them to match.

So, sometimes there's going to be a couple of covariates that you think are the most

critical, the biggest confounders.

And you might want to do exact matching on those and then tolerate and

exact matching, but good matching on others.

So that's one thing you could do.

There's also a common approach is to use a propensity score and

do a distance match on the propensity score.

So this is a topic for another video, but

I just wanted to raise awareness of it here.

So where are we going from here?

So we figured out that we need to be like calculate a distance

between sets of covariates on different subjects.

But then once you have that, how do you actually select matches?

So we'll look at a couple of algorithms, and these are not the only ones but

these are probably the most popular.

So we'll look at greedy matching, which is also known as nearest neighbour.

And this has the advantage of being very computationally fast.

So this you can apply to large data sets without a problem typically.

Then there's optimal matching, which should be better but

it's very computationally demanding.

And so, you might have more trouble with this on large data sets.

Explorer notre catalogue

Rejoignez-nous gratuitement et obtenez des recommendations, des mises à jour et des offres personnalisées.

Coursera propose un accès universel à la meilleure formation au monde, en partenariat avec des universités et des organisations du plus haut niveau, pour proposer des cours en ligne.

© 2019 Coursera Inc. Tous droits réservés.

Télécharger dans l'App Store

Disponible sur Google Play