Coding Example

Loading...

En provenance du cours de Johns Hopkins University

Regression Models

2123 notes

Johns Hopkins University

Regression Models

2123 notes

Cours 7 sur 10 dans Specialization Data Science

Linear models, as their name implies, relates an outcome to a set of predictors of interest using linear assumptions. Regression models, a subset of linear models, are the most important statistical analysis tool in a data scientist’s toolkit. This course covers regression analysis, least squares and inference using regression models. Special cases of the regression model, ANOVA and ANCOVA will be covered as well. Analysis of residuals and variability will be investigated. The course will cover modern thinking on model selection and novel uses of regression models including scatterplot smoothing.

À partir de la leçon

Week 2: Linear Regression & Multivariable Regression

This week, we will work through the remainder of linear regression and then turn to the first part of multivariable regression.

Inference in Regression5:13

Coding Example6:57

Rencontrer les enseignants

Brian Caffo, PhD
Professor, Biostatistics
Bloomberg School of Public Health
Roger D. Peng, PhD
Associate Professor, Biostatistics
Bloomberg School of Public Health
Jeff Leek, PhD
Associate Professor, Biostatistics
Bloomberg School of Public Health

Okay.

Let's go through and just demonstrate that the formulas I'm giving are exactly

the formulas that r is using when it performs its calculations.

So, I'm going to use the data diamond diamond dataset and

that's in the UsingR library, so I'm going to load those.

Define y, x and n like I always do with this dataset, so

I don't have to type so much.

My beta one is the correlation between the y and

the x times the standard deviation of y divided by the standard deviation of x.

And my beta nojught is the mean of the y's minus beta1 times the mean of the x.

My residuals, oop, my residuals are my response y minus the predicted values,

beta nought plus beta1 x, so here I've just carried the subtraction through.

There's my residuals.

My estimate of the standard deviate of the,

standard deviation around the regression line,

the variability around the regression is the average of the residuals,

except remember, we're dividing by n minus 2 instead of n and to get it to be

a standard deviation rather than a variance, I'm going to square root it.

There's my sigma.

That's my estimate of sigma from the previous several pages.

My sums of my squares of x's is just the numerator of the variance calculation,

but let's just do it directly.

It's x minus its mean squared and sum of the, adding those up.

So let me get that and

then my standard error from my beta nought from my intercept.

If I just simply plug into the formula, there, I've written it out.

And my standard error for my beta1, which is I think the more useful and,

and mentally important formula to commit to memory is the sigma,

the variance, the standard deviation around the regression

line divided by the sum of the squares of the x's.

The sum of the square deviations of the x's around their mean.

Okay.

There we go.

There's the standard error for beta1, I'm going to create the two t statistics.

Which if you're testing a hypothesis that beta1 is zero or

beta nought is zero, then that's just going to be the estimate.

Here's my estimate divided by its standard error.

And so we don't have to subtract off the true value,

because the true value is assumed to be zero under this hypothesis.

And so for my beta1 hypothesis, it's going to be beta1 divided

by its standard error and then I'm going to calculate my two p values.

Again, if you've taken the inference class,

then how you go from a t statistic to a p value should not be a great leap.

But in this case, it's going to be twice my t probability where

in both these cases, the estimates are larger than zero so the,

so I'm going to look at the probability of being this statistic or

larger and I'm going to double those two t probabilities [SOUND] and

them I'm going to create my coefficient table created manually by

me without having done any lm or any built in higher level R function.

And I'm going to give it, its rownames and its column names, so

that it looks like R's output.

And then I'm, now I'm going to show you the easy way to do it.

Okay.

So here let me print it out.

There it is printed out and then now, I'm just going to do fit, lm,

my response y is a linear relationship with my predictor x and

then I'm going to get the summary of the coefficients.

You'll see everything is exactly the same.

So, if you didn't follow all this, it's, it's not too big of a deal.

I just wanted to show you for

once that the formulas that I'm giving you are exactly the right formulas and

we verified them computationally by checking them out on a dataset.

Maybe check them out on another dataset just to verify for yourself.

Let's go through getting a confidence interval.

So my fit, the variable fit was the output of lm.

Let's see what happens if I type summary fit.

You get the full output from lm.

So it gives you facts about the residuals, it give you the coefficients,

the estimated coefficients.

The Standard Errors, the t values and the associated p values.

These are all for test of whether or not the intercept is 0 or the slope is 0.

Now of course 0 intercept may not be of interest, but

almost always a test of a 0 slope is of interest.

But it also gives you the residual standard deviation,

the degrees of freedom The r-squared, the adjusted r-squared,

which we'll talk about later on in the class.

Let's go ahead and generate our confidence intervals for

our intercept and for our slope.

So now I've assigned that summary.

The summary from the lm statement to an object in R, but

I just wanted the table part, so I'm grabbing the, just the coefficient.

So, it's right here.

See, I'm just grabbing the coefficient table.

So let me just print out that variable, just to show you what it looks like.

It's just that table part, just the estimates, the standard errors,

the t values, and the p values.

Okay.

So, our confidence interval is just going to be estimate.

So here's for the intercept, estimate plus or minus

the 97.5 t quantile degrees of freedom, just to make

sure I get the degrees of freedom right, so I don't even have to think at all,

I'm grabbing fit$ degrees of freedom and then times the standard error.

There, I'm grabbing the standard error.

And then I'm doing it for the slope, which is in the second row of the first column.

So the slope estimate plus or

minus the releventy quantile times its standard error,

which is in the two two cell of this table here.

And then the, slope is going to be interpreted in change in Singatore,

Singapore dollar price per one mass increase in carats,

in carat, the unit for the x, the predictor variable.

But I, I might want it for a 0.1 increase, because as we mentioned earlier or

looking at the data, a one carat increase was kind of a big increase.

So why don't I divide my whole intervil, interval by ten.

So let's run both of them.

And particularly, focus in on this later one,

the confidence interval for the slope.

So, it's saying, with 95% confidence,

we estimate that a 0.1 carat increase in diamond size is going

to result in a 356 to 389 increase in price in, in Singapore dollars.

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.

© 2018 Coursera Inc. Tous droits réservés.

Télécharger dans l'App Store

Disponible sur Google Play