Residuals distributional results

Loading...

Del curso dictado por Johns Hopkins University

Advanced Linear Models for Data Science 2: Statistical Linear Models

25 calificaciones

Johns Hopkins University

Advanced Linear Models for Data Science 2: Statistical Linear Models

25 calificaciones

Welcome to the Advanced Linear Models for Data Science Class 2: Statistical Linear Models. This class is an introduction to least squares from a linear algebraic and mathematical perspective. Before beginning the class make sure that you have the following: - A basic understanding of linear algebra and multivariate calculus. - A basic understanding of statistics and regression models. - At least a little familiarity with proof based mathematics. - Basic knowledge of the R programming language. After taking this course, students will have a firm foundation in a linear algebraic treatment of regression modeling. This will greatly augment applied data scientists' general understanding of regression models.

De la lección

Residuals

In this module we will revisit residuals and consider their distributional results. We also consider the so-called PRESS residuals and show how they can be calculated without re-fitting the model.

Residuals distributional results5:14

Code demonstration3:15

Leave one out residuals8:37

Press residuals14:36

Conoce a los instructores

Brian Caffo, PhD
Professor, Biostatistics
Bloomberg School of Public Health

So recall that our residuals were defined as e equal to y minus y hat

which we could write as I minus the hat matrix times y.

And again, now, we're in the context where we're going to assume y is normally

distributed with mean x beta in variance sigma squared times I.

So the residuals then are singularly normal,

distributed because this is not a full rank transformation.

So the residuals are singularly normal distributed mean.

Expected as the residuals is I minus H of x,

times X beta which is zero.

Because for the reasons that we've discussed at length.

The variance of the residuals is then I minus

H of x sigma squared I minus H of x and

then which is just I minus H of x sigma squared since

the matrix I minus H of x is [INAUDIBLE].

So, the residuals have this issue or not issue, but

they have this property that they inherit the units of y.

So, of course, that's a good thing.

They're a linear transformation y.

But that often means that if we do things like a plot of our residuals,

which we often plot.

Say for example, our residuals versus our fit of values, that's a terrible Y.

The scale, the vertical scale,

when we look at that, isn't comparable across experiments.

And so, if you wanted to standardize your residuals,

that's some of the things that we're going to talk about.

In this lecture, so what you could do is simply take

the diagonal of the variance of E,

that matrix to the one half power.

So, what I mean by that, not writing this out perfectly, but

the square root of the diagonal of the variance at.

Let me write it out here.

To the minus one-half power.

So I just mean the diagonal matrix

that has the rate, the inverse of standard deviations.

If we were to multiply that times our e, then we would have now,

random variables that have unit variance.

So of course, we'd have to estimate to sigma squared.

So that is a nice way to do things.

And of course, it satisfies our criteria for creating a unit free quantity.

However, these quantities.

These so called standardized residuals have the problem that you

can't easily establish boundaries for calling standardized residuals outliers.

Just because the distribution of this quantity

isn't a terribly easy distribution to work with.

It's not a t distribution because the numerator and

the denominator are not independent in the same way that our typical statistic,

t statistics that we derive for, say, the coefficients of ours So we're going to

talk a little bit more about how to get more T like residual statistics here

when we talk more about different ways in which we can standardize our residuals.

But this is by no means a terrible thing to do to take our residual, say e i, and

then divide it by the, let's say square root 1 minus h i i

where HII and then times S.

Where hii is the eyes diagonal element of the matrix

that's often quite useful thing to do.

Another little comment that we might need later on.

Is because the variance of e is

equal to i minus h of x times sigma square.

We know that that has to be greater or equal to zero.

So, we know that any element, I should say any element of this,

the diagonal of this matrix has to be greater than or equal to 0.

So 1- hii, this quantity in the denominator of our statistic over there,

has to be greater or equal to 0 so that hii has to be less than or equal to 1.

So our diagonals of our hat matrix always have to be bounded from above by 1.

This is a fact that we're going to use later on

when we use our hat diagonals as a diagnostic measure.

Okay, so let's just go through a coding exercise

where we create standardized residuals and look at them.

Explora nuestro catálogo

Inscríbete de manera gratuita y obtén recomendaciones personalizadas, actualizaciones y ofertas.

Coursera brinda acceso universal a la mejor educación del mundo, al asociarse con las mejores universidades y organizaciones, para ofrecer cursos en línea.

© 2018 Coursera Inc. Todos los derechos reservados.

Descargar en la App Store

Consíguelo en Google Play