Dario Schiraldi : How do I create and interpret regression models?

Question

Hi everyone,

I am Dario Schiraldi, CEO of Travel Works, starting to work with regression models in JMP and would love some advice.

What’s your approach to creating a solid regression model, and how do you interpret the key outputs like coefficients, p-values, and diagnostics?

Are there any tips or common pitfalls I should watch out for? Also, which JMP features or visualizations do you find most helpful for understanding your model?

Regards

Dario Schiraldi CEO of Travel Works

Dario Schiraldi

statman · Answer

Dario, IMHO, you are asking the wrong questions.  You should start with: What questions you want to answer? (e.g., What factors contribute to demand forecasting? or What factors affect travel delays?)  What hypotheses do you want insight into?  (e.g., When travel tickets are purchased (day of week, proximity to holidays...) affects pricing due to dynamic pricing models used by suppliers (e.g., airlines).

Understand the situation and then select the tools to assist in this endeavor.  Pardon my analogy, but starting your investigation with regression is like; here is a hammer, let's see what I can hit with it.

Applying regression to a data set without understanding the context is very dangerous.  The ability to model the past may be useless for predicting the future (e.g., history may not predict the future).  You need to understand data sources, data types, how the data was gathered, measurement error, etc. before applying regression methodologies.

I agree with both Victor's and Pete's suggestions.

P_Bartell · Answer

I agree with @Victor_G 's reply. But to add just a bit more general advice...when I was tasked with doing any data analysis work with an eye towards modeling at some point in my problem solving process, I always started with plotting the data before any modeling. Which plots depends on the data for both responses and predictors, continuous, nominal, ordinal, etc. But this much I can tell you I always started with the Distribution platform so I had a sense of the following for both predictors and responses:

1. Where's the middle?

2. How spread out is the data?

3. What's the shape look like?

4. Is there anything funny, suspicious, or unusual that might make subsequent analysis problematic?

This last question can lead to all sorts of data clean up, recoding, outlier decisions, etc. One small example, suppose Traveler Citizenship is one of the predictors,,,and you pulled data from different databases...maybe one database has a Traveler's Citizenship as 'USA' another database, 'American'...well JMP is gonna treat this scenario as two different levels for citizenship...which is not how you want it treated I suspect?

So once my data is clean and good to go...then it was always off to at least some sort of scatter plotting of predictors vs. responses...multiple platforms in JMP for this, with lots of flexibility for visualization modification within. Graph Builder, Fit Y by X, Multivariate platforms all have useful features within.

Then looking at these plots and asking the following type questions:

1. Do these make physical sense?

2. Any relationships that look odd or suspicious?

3. Are there outliers that need further investigation?

4. What type of regression techniques might work?

5. Is there evidence of multicollinearity among the predictors?

Then based on the answers to these questions as long as the data still looks clean and reasonable...it's finally off to modeling.

Hope this helps?

Victor_G · Answer

Hi @darioschiraldi1,

Welcome in the Community !

I would encourage you to first watch and read some ressources to get you started about these vast questions, too general to be answered concisely on this forum :

STIPS Module 1: Statistical Thinking and Problem Solving

STIPS Module 2: Exploratory Data Analysis

STIPS Module 4: Decision Making with Data

STIPS Module 5: Correlation and Regression

STIPS Module 7: Predictive Modeling and Text Mining

And visit the Statistics Knowledge Portal | Introduction to Statistics | JMP to better understand assumptions and pitfalls for regression models.

Hope these first few ressources may help,