1

What are the good statistical methods to study failure rates? For example, let's say I manufacture 100 tire models and I sold 3 million units that hit on road. Let's say I have the manufacture date, sales date and failure date at a tire level.

  1. How can I do an early detection of a bad model? (Important goal)

    • So I can stop putting these tires on road

  2. How can I compare failure rates of tires by car models? For example, is the tire likely to fail in SUVs compared to sedans? How do I prove this statistically?

I'm looking for statistical approaches that can be used to answer these two questions

Prasanth Regupathy
  • 119
  • 1
  • Seems like you're interested in prediction. Consider trying a basic linear regression or logistic regression (if failure is a binary variable). For comparing failure rates by type of cars, research linear contrasts (e.g. emmeans library in R). – Parseltongue Sep 22 '20 at 00:25
  • Maybe hazard proportional rates (cox model). – igorkf Sep 22 '20 at 01:43

1 Answers1

0

My recommendation would be survival analysis. I say survival analysis as opposed to linear or logistic regression (as mention in the comments) because you can make use of tires which have NOT failed via censoring.

"Early detection" is a nebulous term. What you could possibly due is collect data on failing tires and compute a Kaplan Meir estimator for the survival curve stratified my tire model (assuming you have sufficient data to do this). To control for variables, you could perform a cox regression. As for statistical "proof", you will never "prove" anything with statistics but you can make very convincing arguments with appropriate models, adequate and quality data, and precise estimates.

For more on survival analysis I recommend Frank Harrell's treatment of the material in his "Regression Modelling Strategies" book.

Demetri Pananos
  • 24,380
  • 1
  • 36
  • 94