Multinomial logistic regression - RRR is too big

[Gurpreet Singh]

Hi,

I am running a regression where a firm performance measure ( a categorical variable) is regressed on certain predictors using multinomial regression model.

- Q4 is the firm performance variable that has 4 categories. Rest of the variables are explanatory variables.
- The estimated coefficients in Stata are relative risk ratios (RRR).

One of the comments from my reviewer is that the RRR coefficient of import capital intensity is TOO LARGE. I estimated the model again by:
a. Using winsorized values of import capital intensity at 1% and 5% level
b. Using winsorized values of all variables at 1% and 5% level

However, nothing seem to work and the coefficient is still too large. I am looking for suggestions.

Thanks!

[Richard Williams]

How is the variable in question scaled? What is the basis for saying the effect is too large? Coefficients are affected by how variables are measured, e.g. the effect of one dollar will be less than the effect of a thousand dollars. You can't really say whether an effect is too large without knowing how it is measured.

[Gurpreet Singh]

1. The variable is defined as a ratio -

Import capital intensity = Imported capital/Operating Expense

2. Below is the summary statistics for this variable:

So certainly outliers are not influencing the RRR coefficient of the regression


(It is a panel data - the data is given for firms and years.)

How is the variable in question scaled? What is the basis for saying the effect is too large? Coefficients are affected by how variables are measured, e.g. the effect of one dollar will be less than the effect of a thousand dollars. You can't really say whether an effect is too large without knowing how it is measured.

[Gurpreet Singh]

3. Also, if i take a log (import capital ratio), then these coefficients do not look that big. But, my guide suggests that it is not ideal to take a log transformation of a ratio.

[Richard Williams]

A coefficient tells you the effect of a one unit increase in an independent variable. In this case a 1 unit increase would be a super huge increase. I suggest multiplying the variable by 100, maybe even 1,000.

Based on what you have said, I don't see any basis for saying the RRR is too big. Unless maybe other studies have measured variables the same way and found far smaller effects.

[Gurpreet Singh]

A coefficient tells you the effect of a one unit increase in an independent variable. In this case a 1 unit increase would be a super huge increase. I suggest multiplying the variable by 100, maybe even 1,000.

Based on what you have said, I don't see any basis for saying the RRR is too big. Unless maybe other studies have measured variables the same way and found far smaller effects.

I see.

Multiplying it by 100 or 1000 resolves the issue. Thanks for your help

[Gurpreet Singh]

A coefficient tells you the effect of a one unit increase in an independent variable. In this case a 1 unit increase would be a super huge increase. I suggest multiplying the variable by 100, maybe even 1,000.

Based on what you have said, I don't see any basis for saying the RRR is too big. Unless maybe other studies have measured variables the same way and found far smaller effects.

I looked at percentiles of this variable. The median of this variable across different quartiles in really small which could explain why the estimated coefficient is so big.

[Bilal Caliskan]

Hello,
I have panel data which was observed one week. I have run multinomial logit regression with xtmlogit command with stating repeated observation. I obtained the following outcome. But, I have a question and I am confused some extreme coefficient and RRR values. My distance1 distance2 and distance3 (they are binary variables) variables seem to have very high coefficient values as you can see attached file. When I converted them into RRR values, one of them (distance1) is over 1000. SO, I would like to ask you that it is normal ? and do you have any suggestions to fix this problem. By the way, I only could attach the results for second category due to large table size. Thanks in advance.

[Clyde Schechter]

No, it is not usual to see RRRs that large for discrete variables--it is rare for anything to be that strongly associated to anything else in real life. My best guess, though it is only a guess without seeing the data, is that your base outcome 1 is a very rare event in your data. If outcome 1 almost never occurs in the estimation sample, then the coefficients of other variables can be very large. I suggest you take a look at the output of -tab decison distance1 if e(sample)- (and similarly with distance 2 and 3 to get a feel for what is going on. I suspect you are going to find some very tiny cells in the decision = 1 rows.

[smith Jason]

I see.

Multiplying it by 100 or 1000 resolves the issue. Thanks for your help

I don't understand why multiplying 100 or 1000. Also, multiply with the very big RRR value variable?
Thank you for your explanation.