55529_en-US

liliconcom

Difference between JMP and Excel in dummy variable regression outputs

Hi all,

Newbie question here: I'm doing dummy variable regression using JMP for a school project. When I cross-referenced to Excel's version, I realized the coefficients are completely different even though the t-stats are the same. 

 

JMP: 

 

These two images are of JMP and as you can see as compared to the Excel's version, the figures are different even though I'm using the same attributes.

And it's not because of the interaction effect as I did with just the variables, without the interaction effect (e.g. COO x Coupons) and the results are still different.

 

Strangely, Excel's version seems more correct than the JMP's version, aligning more closely with my secondary and other primary research. 

 

Any advice?

Re: Difference between JMP and Excel in dummy variable regression outputs

The standard error is 1 by definition.

Re: Difference between JMP and Excel in dummy variable regression outputs

Hi Mark, 

 

All the questions are answered except for "how to find the standard error of beta coefficient for each independent variable" after running a regression on Fit Model in JMP. 

 

Can you point me in the right direction? Thanks! 

Re: Difference between JMP and Excel in dummy variable regression outputs

There are no attached images.

The link takes me to a page that says "Timed out."

Which question is answered? What is the answer? (Thanks!)

Re: Difference between JMP and Excel in dummy variable regression outputs

Hi Mark,

Thanks for getting back so promptly.

If you refer to the attached images, that's the analysis output after running a regression. So my question is, is there a way for me to find the "total standard error of estimate", if there's such a thing since currently all I can see about the standard error is of each coefficient estimate.

Thanks. I understand where you're coming from. 

 

Another thing is, is there a way to find the standard error of beta coefficient for each independent variable in the Fit Model's Analysis? I understand this is the method used to find beta coefficient: http://www.jmp.com/support/notes/36/009.html

 

Regards,
Fred

[Edited as I've found the answer to one of the questions] 

Re: Difference between JMP and Excel in dummy variable regression outputs

I need some clarification.

1. What do you mean by the "estimate of the overall model?" You estimate parameters. Can you show me an example where it is available? Then I might recognize it.
2. We generally do not compare parameter estimates between models. We instead use model selection criteria (e.g., adjusted R square, AICc, BIC, Mallow's Cp) to choose a model with minimum bias and variance and then infer about the parameter estimates of the selected model. We would not want to compare two estimates from different models with a hypothesis test for a difference. Estimates are expected to depend on the other terms in the model except for the rare case that all the estimates are orthogonal (uncorrelated), so different models (i.e., different sets of terms) would usually yield different estimates for any given term.

Re: Difference between JMP and Excel in dummy variable regression outputs

Hi Mark,

 

Thanks for your help! You've helped me so much on how to use JMP. 

 

I'm at the finishing line, so just two more questions: 

1) I can't seem to find the standard error of estimate for the overall model after running the regression (I can find only the standard error for each independent variable). Can you point me in the right direction for this? 

2) Let's say I've coefficient estimates generated from one regression, and those generated from another regression. Am I able to run a t-test of the coefficient estimates in JMP? If so, do you have any advice how I can approach that? 

 

Really appreciate your help on this! Thanks. 

 

Re: Difference between JMP and Excel in dummy variable regression outputs

First, as I said before, you don't need to create dummy variables in JMP. It is a statistical analysis and modeling software.

Second, create data columns as I described for the factors (not separate dummy variable columns for each factor level) so that each factor uses one data column and enter the value directly for each row.

Third, create a column (Respondent, for example) and populate it with values that identify the respondent for each row.

Now select Analyze > Fit Model. Select the response column and click Y. Select the factor columns and click Add or use Cross or a macro to add interaction effects to your model. Select the Respondent effect (not the data column), click the red triangle next to Attributes and select Random Effect. That change is all you have to do. You are simply distinguishing the type of effect of Respondent (random) from that of the other factors (fixed).

Re: Difference between JMP and Excel in dummy variable regression outputs

Hi Mark, 

 

Thanks for the quick reply. Apologise that I wasn't specific enough in my explanation. 

 

Yes you're right. A respondent contributes more than one observation, but the analysis interprets each observation as independent and as one respondent when the actual case it's not true. 

 

What I want to do is to adjust for standard errors for such clustered correlation, in JMP's Fit Model. 

 

How can do I from the console below (if it's from here)? (: 

 

 

https://www.gsb.stanford.edu/faculty-research/working-papers/when-should-you-adjust-standard-errors-clustering

 

Regards,

Fred

Re: Difference between JMP and Excel in dummy variable regression outputs

I cannot picture your study and its data well enough to answer you latest question. (Note: I didn't say 'last' question.) What do you mean by 'factor in clustered correlation?' I know that it is related to 'dummy variables are all treated as independent entries' and 'done by the same respondent.'

A row in a JMP data table represents all the values that make up one observation. That is, you know the value of every variable (independent and dependent) for a complete record. Are you saying that there is another variable, Respondent, that identifies the observation, that respondents contribute more than one observation, and you want to account for Respondent variability in the response? If so, enter this variable in your data table and add it as a term in Fit Model. Be sure to select it, click the red triangle for Attributes, and select Random Effect.

Re: Difference between JMP and Excel in dummy variable regression outputs

Hi Mark, 

 

Thanks for your help once more. I appreciate your prompt reply and really help me better understand how to use JMP for my work. 

 

I do have one last question: Is there an option for me to factor in clustered correlation. The issue is, my dummy variables are all treated as independent entries when some of them are done by the same respondent, which means I need to factor that in. 

 

I understand other statistical software allows that through a unique variable, but is there a way I can do so with the Fit Model in JMP. 

 

Thanks for your help. 

 

Regards,

Fred

Re: Difference between JMP and Excel in dummy variable regression outputs

Correct, select the data columns for the factors involved in the interaction and click Cross.

See the Macros button for many convenient shortcuts, such as Factorial to Degree. With degree set to 2, it will add all possible two-factors using the selected data columns.

Re: Difference between JMP and Excel in dummy variable regression outputs

The parameter is interpreted as the change in the response from the mean. A negative value does, in fact, indicate a drop in the response variable.

Re: Difference between JMP and Excel in dummy variable regression outputs

Adding to the newbie question (pardon me), I'm also looking at the interaction between the different variables with the country of origin. 

 

From a software's perspective, is this how we set up the model (Use the 'Cross' Button) to study interaction effect between Updates and COO - USA?

 

Thanks in advance, Mark.

 

Regards,

Fred

Re: Difference between JMP and Excel in dummy variable regression outputs

Hi Mark, 

 

Thanks for the quick reply. I think that's exceedingly helpful. 

 

Just want to take this chance to ask too: Do the negative estimates mean that the presence of those variables reduces preference (assuming preference is the LHS variable and it is coded as low = not preferred and high = preferred), relative to the baseline?

 

This will be an odd discovery as compared to Excel as those variables are expected to increase preference. 

Re: Difference between JMP and Excel in dummy variable regression outputs

Excel is not more correct. They are just different. It has to do with the parameterization of the linear model. The choice is mathematically arbitrary but each way has its advantages. JMP uses effect parameterization. That way is supposed to make the parameter estimates more interpretable. They sum to zero, so the intercept is the mean response at the origin. The parameter for a given level is the change from the intercept. So an estimate of -0.55112 for COO - USA = 0 means that the mean for that level is 5.8482381 - 0.55112 = 5.2941181. The plots are meant to help interpret the selected model.

Since you are using numeric values with the correct nominal modeling type, JMP is correctly estimating the effect of these categorical predictors. The numeric code, though, might be interpreted by Excel as a continuous variable, since it has nothing like modeling type that is fundamental to statistical modeling and testing. That model could lead to a very different estimate, test, and interpretation.

See Help > Books > Fitting Linear Models > Appendix A: Statistical Details > The Factor Models.