Error message: Cannot compute an improvement -- discontinuous region encountered (for panel ordered logit/probit model)

[Sean Bilic]

I have panel data for OECD countries between the years 2010 and 2021.

I categorize GINI into 4 groups as follows

This is named as d_gini, which is my dependent variable. And I regress it on World Pandemic Discussion Index (WPDI) as follows:

But, regardless of probit or logit model, I encounter with this error. How can I fix it?

Thank you.


PS: I find such a solution. https://www.statalist.org/forums/for...on-encountered

But I cannot adopt this solution to my analysis.

[Sean Bilic]

Note: When I try this panel ordered logit regression with different explanatory variable, for example I use log_GDP in stead of WPDI, this regression works. I don't see any error. However, I need to use this explanatory variable WPDI. How can I fix?

[Andrew Musau]

I have panel data for OECD countries between the years 2010 and 2021.

I categorize GINI into 4 groups

You start off with a variable that is continuous over the unit interval and then you categorize it. Is there a reason why you do this instead of analyzing it in its original scale?

Note: When I try this panel ordered logit regression with different explanatory variable, for example I use log_GDP in stead of WPDI, this regression works. I don't see any error. However, I need to use this explanatory variable WPDI. How can I fix?

Can you show the result of

sum WPDI

[Sean Bilic]

You start off with a variable that is continuous over the unit interval and then you categorize it. Is there a reason why you do this instead of analyzing it in its original scale?



Can you show the result of

sum WPDI

All variables are like that. (Note: two LFP variables in the bottom of the table are not used. )

note: dgini and d_gini are the same, I have changed for a trail.

[Andrew Musau]

My guess is that you have too many zero values for this variable. Maybe you can also show the following figure:

hist WPDI, discrete

[Sean Bilic]

My guess is that you have too many zero values for this variable. Maybe you can also show the following figure:

hist WPDI, discrete

Yes you are right. For each country, only 10 of 12 observations are zero. How can I fix this problem?

[Andrew Musau]

I do not think that there is anything to fix, the variable just does not have much variation. From the link you referenced in #1, you can try gllamm with the -adapt- option as below to see if it helps:

ssc install gllamm, replace
gllamm d_gini WPDI, i(Country_id) link(ologit) adapt

But, you still did not answer my question in #3:

You start off with a variable that is continuous over the unit interval and then you categorize it. Is there a reason why you do this instead of analyzing it in its original scale?

It is poor statistical practice to categorize continuous variables if there is no rationale behind doing so as you are throwing away information. In that case, you may solve your problem by analyzing the outcome in its original scale.

[Carlo Lazzaro]

Sean:
Andrew's excellent advice is well covered in statistical literature (see, free of charge, http://citeseerx.ist.psu.edu/viewdoc...=rep1&type=pdf).

[Nick Cox]

Please note our policy on cross-posting, which is that you should tell us about it. Similar points are being made to your postings on Reddit.

FAQ Advice #8

[Jared Greathouse]

Also note that we (mainly I) ask a lot to show data examples using dataex and to put your results in

Code delimiters.

screenshots are unclear (usually) and are just not optimal in 99.9% of cases. You can literally just put your output in between code delimiters, and it shows us exactly how it looks on your screen, no uploading or taking pictures needed. i get that pictures and descriptions are well intentioned, but they're just not helpful to us helping you.

[Nick Cox]

Correction to #9: Stack Overflow not Reddit.