In the last video, we looked at how to run a logistic regression and

get odds ratio estimates in the R package.

So, now let's do the same thing in the Stata, so

I am going to use the same regression as in the R example.

So, I'm regressing school wide, which is an indicatorr variable for

whether a school wide growth target in API score was met.

And the covariants I'm dealing with are enrollment,

college grad, the percentage of parents with college degrees and

whether a school is a year-round school or not.

Now, here's this Stata code that's appropriate for doing this.

So, the first thing we do is we load up the data set with use apistrat.dta, clear.

And this is a state of format version of the API strap data set,

that I wrote out using the right.DTA

function in the foreign package in R.

So, I've created exactly the same data, but

in this neat format that Stata understands.

Then, we used svyset to specify the design,

cds is a counter variable.

We've got no PSUs, but we need an IED variable.

The design wait is called pw, the strata are school types, s type,

and I've got to find that population correction factor in this design.

Now, the next thing I want to do is recode school wide,

which i need to do in Stata,

in order to get it to understand that I'm dealing with a 01 variable.

So, I generate a new variable called sw01 and that's based on school wide.

Now, notice when the DTA form of the data set got written out,

it changed the dot that was in the R variable name to underscore.

So, it took care of that for me automatically because Stata doesn't like

dots in the middle of its variable names.

And then, I recode sw01, my new variable,

by saying, if it was 1, set is equal to 0,

if it was 2 then set it equal to 1.

So, that way I'm sure of modelling the probability

that school wide target has been met.

You know, if I didn't get that quite right,

I'll model the indicator for whether it has not been met.

Which, numerically, what it will cause is the coefficients will

be negative in a way that I don't expect.

It'll seem like the directions is the wrong way we influence, but

it's just because I don't have the package interpreting what true and false is.

All right, so here are the coefficients,

and if you look back at the R output, these are the same.

The standard errors are very slightly different, so

internally there's a slightly different formula being used to get

these standard errors, but it won't have any substitute effect.

Now, notice on year round, that's down here,

the confidence level along the parameter

goes from -0.046 up to 3.11,

which is similar to what we got in R.

That's saying there's not much evidence that year-round

schools are different from non-year-round schools, at least,

our sample doesn't appear to be precise enough to pin that down.

Now, we can get confidence intervals in Stata.

So, you'll notice on the last page, let's just check that again.

What I did was call the model

with svy: logit.

And what that does is result in getting me parameter

estimates and confidence intervals on those.

Now, in order to get Stata to give me confidence intervals on odds ratios,

what I need to do is use a slightly different call.

So instead of logit, I do SVY Logistic.

I fit the same model, it just gives me slightly different output.

That's the convention and

status, so it's just a matter of learning how to talk to the package.

So, here are the odds ratios, and

those are essentially the same as what we saw in R.

And notice, for year-round,

the odds ratio is 3.76 to two decimal places.

Implying that it looks like the year-round schools are likely

to meet the targets, more so than non-year-round schools.

But when we do the confidence interval,

we also get about the same thing as we saw in R,

0.63 up to 2.56 or thereabouts.

Even though the value of the odds ratio itself right here,

is suggestive that year-round schools may be more likely to hit the target.

Our sample is not precise enough,

our data are not strong enough to give us real evidence that that's the case.

But on the whole, we get the same numerical and

qualitative information by running the model in Stata, as we do in R.

And that, of course, is what we would expect to happen.

Logistic Regression in Stata

Combining and Analyzing Complex Data

Conheça os instrutores