Subscribe Bookmark RSS Feed

f-test

How do I do an f-test in jmp?
8 REPLIES
There's an "F Distribution" function, along with an assortment of other functions for calculating quantiles from statistics and vice versa, all of which are described in the online help under "Probability Functions". If you run the following lines of script you'll get a couple of the familiar results from performing tests with the F, t and Normal distributions falling out, showing how several of them work:



The 1 and 10000 are my numerator and denominator degrees of freedom for the F-test; an F-test on (1 and N) df is equivalent to a t-test on the square root of the F statistic on N df (but note that the F-test is one-tailed, whereas the t-test is two-tailed - and only one tail is reported in the result), and if N is large enough, the t-test approximates to a Normal test.
F-tests appear in several places throughout JMP. Do you mean an F-test for an ANOVA? Or an F-test for comparing two variances? What you are looking for is probably already part of JMP, and you don't have to do the calculations by hand.
David and Jonathan,

Thanks for the responses.
I want to compare two variances. I ran the following script:

F Distribution(1.14, 581, 568)

The value of "1.14" being the ratio of the squares of the two variances. It returned a p-value of 0.941. From this I conclude that the variances are not statistically different at the 95% confidence level.

Is this right? I'm a beginner with regards to statistics (and jmp), so forgive me if this seems like a very basic question.
Trex,

You started along the right path, but there are a few things that are incorrect in your last post.

1. You said 1.14 is the ratio of the square of two variances. I think I know what you meant to say, but want to make sure. The statistic to use is the ratio of the square of the std dev's. In other words, the ratio of the variances, not the ratio of the square of the variances.

2. Your script was almost correct. If the ratio of variances is 1.14, and the degrees of freedom are 581 and 568 (the df are n1-1 and n2-1), then the p-value for a two-sided test is found by
2*(1-F Distribution(1.14, 581, 568)) = 0.1168

Alternatively, you can compare 1.14 to an F quantile. If you use an alpha level of 0.05, then the quantile is found by
F Quantile(.975,581,568) = 1.18. The 0.975 = 1-(0.05/2).

Since the p-value>.05 and 1.14 < 1.18, we can conclude equal variances at the 0.05 level.


If you have the raw data, you can feed it into Fit Y by X and do the F-test.
Trex,

Could you tell us what the data actually is, please? You've got two variances, and even though either could theoretically be larger than the other one, it might not make sense to test both alternatives. For example, if you had the results of 10 types of chemical reaction, each of them replicated two or three times, and you wanted to know whether there was a difference between the types of reaction, you would only test to see whether the between-reaction-type variance was larger than the within-reaction-type variance (because a result in the other direction, though curious, would necessarily have to be a fluke). That would effectively be a one-way analysis of variance you'd be doing there, and as has already been noted above, in such a case you wouldn't need to do a separate F test anyway, because the result of the test would be supplied by JMP's Fit Model procedure.

Also, is there anything that ties pairs, or groups, of observations together here? The note above about feeding it through the Fit Y by X procedure would only apply if there were - and even then, that would tell you something about the relationship between the two sets of observations, not whether one of them was more variable than the other. So really we need to know what the structure of the data actually is before attempting to describe how to apply an F test to it - because we could be missing something important about what type of test is most appropriate.
Clarification about my note to put the data into Fit Y by X:
Use Fit Y by X, Oneway. You can use the Unequal Variances command to test for equal variances. If there are two groups of data, then that produces the F test I described above. I was not referring to the ANOVA F test for comparing means.
From jg: "1. You said 1.14 is the ratio of the square of two variances. I think I know what you meant to say, but want to make sure. The statistic to use is the ratio of the square of the std dev's. In other words, the ratio of the variances, not the ratio of the square of the variances."

Yes, that's what I meant. Thanks for the correction and for the correct script.

Here's what I'm trying to do:

I built a model using the fit model platform using data from one process run from one process tool. I built this model from half, selected randomly, of my data set. As a way of validating the model, I used the resulting prediction formula to predict the other half of the data set. I obtained two sets of residuals from the two subsets of data. I want to test if the two distributions of residuals are equivalent by doing an f-test (for equal variances) and a t-test (for equal means).

I performed the Fit Y by X, Oneway, Unequal Variances command as suggested by jg. For the 2-sided F-test, it gave a p-value of 0.1139, which is different from the p-value of 0.1168 given in jg's post. Should they be the same? Did I do something wrong?

Let me know if this is not enough background info to give the gist of what I'm trying to do.
The p-value is slightly different (0.1139 vs 0.1168) due to rounding. The jsl script use an F statistic=1.14. The Fit Y by X platform uses an F statistic with more significant digits. So, the p-value is slightly different, but it's ok. You didn't do anything wrong.

Not only do you want the residuals to have similar means, but you want the means to be 0, for a good model. Also, you don't want there to be any trends over time, etc. You can use the residual plots to assess these things.

I don't know anything about the model your building, or your industry or application, but some other things you might need to consider. You said you built the model from data on one run from one tool. Would a different process run produce a wildly different model? Would a different tool produce a different model? If you're going to use this model to predict future runs across diffferent tools, you may want the model to be based on data that includes run-to-run variability and tool-to-tool variability, and any other source of variability that may be important.

I'm glad you understand and are using the concept of testing a model on independent data.