Dear Jeoffrey, There isn’t a one-size-fits-all answer to which function is more appropriate; it often depends on the specifics of your data and research questions. However, understanding the underlying assumptions, methodologies, and how each function handles standard errors can guide you to make a more informed decision. Comparing model fits and running sensitivity analyses can also provide additional insights. Discussion Points:

1. Understanding the Differences:

• The polr function from the MASS package and the clm function from the ordinal package use different algorithms and methods for estimating the parameters and standard errors. It’s important to understand these differences when interpreting the results.
• polr typically uses a frequentist approach with maximum likelihood estimation, while clm can also incorporate a variety of link functions and threshold structures, possibly explaining the differences in standard errors for intercepts.

2. Standard Error Calculations:

• The discrepancy in the significance of the intercepts likely stems from the way each function calculates the standard errors. This can be influenced by the method of estimation and the assumptions made about the distribution of errors.

3. Model Fit and Assumptions:

• Check the assumptions of each model. The clm function allows for more flexibility in specifying the threshold structure, which might better suit your data depending on the distribution of responses and the nature of the predictors.
• Consider conducting a goodness-of-fit test or comparing information criteria (AIC, BIC) between the models to see which model provides a better fit for your data.

4. Interpreting Intercepts in Ordinal Regression:

• In ordinal regression, intercepts (or thresholds) are less interpretable on their own compared to the coefficients of the predictors. Focus on the consistency and significance of the predictor coefficients.
• If the interpretation of intercepts is crucial, ensure that the chosen model aligns well with the theoretical underpinnings of your research question.

5. Simulation and Sensitivity Analysis:

• Consider running a simulation study to see how both models behave under known conditions. This can help understand the robustness of each model’s estimates.
• Perform sensitivity analysis to check how small changes in data or model specification affect the results from both functions.

Many thanks for this very detailled answer Joseph. Highly helpful indeed. Following your suggestions, I will look at every criteria you mentionned before choosing. By reading papers about both of the functions, I thought assumptions on which clm is based were less restrictive than polr. That's why I have more inclined to choose clm as a first guess. Does it make any sense ? Or am I confusing ?

Once again, thank you very much for your help. Really appreciated.

Best Regards

/Jeoffrey

Jeoffrey Dehez , your thanks should not go to Joseph - he just used GPT and pasted the answer into RG. There is no intellectual work associated with that. And I am a bit sorry that you find that answer useful, since it is very generic and nowhere does is give any specific answer.

I would say sometimes it is useful to go back to basics. Especially when, like me, you're not a specialist. Thank you for your comment. It encourages me to be more vigilant.

But where does GPT get it's information? It's trained on a collection of primary sources (i.e. texts) composed by people. If the people who composed those sources are incorrect, so too will GPT's answers be incorrect.