suest - variance matrix is nonsymmetric or highly singular

[Sunita Rao]

I fit the full multinomial logit model for all four alternatives of my response variable and four restricted multinomial models in which one alternative is excluded:

estimates dir

---------------------------------------------------------------------------------------
name | command depvar npar title
-------------+-------------------------------------------------------------------------
m1 | mlogit IntlFW 280 all four
m2 | mlogit IntlFW 280 IntlFW != No Intl FW :IntlFW
m3 | mlogit IntlFW 280 IntlFW != AA1000 only :IntlFW
m4 | mlogit IntlFW 280 IntlFW != ISAE3000 only :IntlFW
m5 | mlogit IntlFW 280 IntlFW != Both AA1000 & ISAE3000 :IntlFW
---------------------------------------------------------------------------------------


I then increased the matsize to 800 (I have Stata IC) and the used the following code to conduct the suest test:

suest m1 m2, vce(cluster CompanyName) noomitted

However, I am getting the following warning and an empty results table: "Warning: variance matrix is nonsymmetric or highly singular"

I also used dataex to include some of my data with this question, but I Stata is giving me the following response:
dataex IntlFW i.NumTypeofAP GeoRevSeg ROA LMarketCap DisclosureIndex MandatedAssurance i.FSaud
> itorCode i.SVCGrouped i.NumHQ i.Fiscalyear, count(150)
factor variables and time-series operators not allowed
r(101);

Do you know what the warning in red above means and what I can do to get a proper response for my suest test? I would be very grateful for some help.

[Carlo Lazzaro]

Sunita:
I do hope that the following (old) thread can be helpful: https://www.stata.com/statalist/arch.../msg00980.html.

[Sunita Rao]

Hi Carlo,

It is very kind of you to send a link to that thread.

So, I should remove the "problem" predictor variables? I have an idea which ones they are.

Warm Regards,
Sunita

[Carlo Lazzaro]

Sunita:
I'm sorry, no.
You should check your dataset and spot the sparse ones (usually they're categorical variables).

[Sunita Rao]

Hi Carlo,

Thank you so much for this suggestion.

I have not done that before. Could point me in the right direction?

Warm Regards,
Sunita

[Carlo Lazzaro]

Sunita:
for instance, a categorical variables that can take on 0/1 values but scores 1 for a too limited proportion of the sample (summed up from https://www.stata.com/statalist/arch.../msg00980.html).

[Sunita Rao]

Hi Carlo,

This is very useful. I will examine all my indicator variables asap.

Warm Regards,
Sunita

[Sunita Rao]

Hi Carlo,

Thank you so much for your kindness!

I examined all my indicator variables and none of them seem sparse to me.

I have a total of 4438 observations. If there are 217 missing observations from a total of 4438 (as in the variable FSauditorCode below), it that considered sparse?

I would be very grateful for your guidance and opinion.


codebook IntlFW MandatedAssurance FSauditorCode SVCGrouped NumHQ Fiscalyear

------------------------------------------------------------------------------------------------
IntlFW
------------------------------------------------------------------------------------------------

type: numeric (byte)
label: Intl

range: [0,3] units: 1
unique values: 4 missing .: 66/4,438

tabulation: Freq. Numeric Label
1,445 0 No Intl FW
892 1 AA1000 only
1,625 2 ISAE3000 only
410 3 Both AA1000 & ISAE3000
66 .

------------------------------------------------------------------------------------------------
MandatedAssurance (unlabeled)
------------------------------------------------------------------------------------------------

type: numeric (byte)

range: [0,1] units: 1
unique values: 2 missing .: 0/4,438

tabulation: Freq. Value
3,972 0
466 1

------------------------------------------------------------------------------------------------
FSauditorCode FSauditorCode
------------------------------------------------------------------------------------------------

type: numeric (byte)
label: FSauditorCode

range: [0,2] units: 1
unique values: 3 missing .: 66/4,438
unique mv codes: 2 missing .*: 151/4,438

tabulation: Freq. Numeric Label
457 0 All auditors non-Big4
155 1 Mix of Big4 and non-Big4
3,609 2 All Big4
66 .
151 .n

------------------------------------------------------------------------------------------------
SVCGrouped RECODE of svcInd
------------------------------------------------------------------------------------------------

type: numeric (byte)
label: SVCInd

range: [1,5] units: 1
unique values: 5 missing .: 66/4,438

tabulation: Freq. Numeric Label
306 1 Mining
1,860 2 Manufacturing
402 3 Utilities
789 4 Finance
1,015 5 Other
66 .

------------------------------------------------------------------------------------------------
NumHQ HQ Country
------------------------------------------------------------------------------------------------

type: numeric (byte)
label: NumHQ

range: [1,45] units: 1
unique values: 45 missing .: 66/4,438

examples: 12 France
21 Japan
35 South Africa
41 Taiwan

------------------------------------------------------------------------------------------------
Fiscalyear Fiscal year
------------------------------------------------------------------------------------------------

type: numeric (int)

range: [2008,2015] units: 1
unique values: 8 missing .: 0/4,438

tabulation: Freq. Value
208 2008
281 2009
439 2010
581 2011
583 2012
814 2013
883 2014
649 2015

[Carlo Lazzaro]

Sunita:
my temptative answer puts the blame on:

MandatedAssurance (unlabeled)
------------------------------------------------------------------------------------------------

type: numeric (byte)

range: [0,1] units: 1
unique values: 2 missing .: 0/4,438

tabulation: Freq. Value
3,972 0
466 1

to the prevalence of 0s.

[Sunita Rao]

Hi Carlo,

Thank you so much for your help. I greatly appreciate your kindness.

Would you recommend dropping the variable MandatedAssurance and running all my models again?

Warm Regards,
Sunita

[Carlo Lazzaro]

Sunita:
yes, I would give it a try.

[Sunita Rao]

Hi Carlo,

Thank you very much for your suggestion. I will do that.

Warm Regards,
Sunita Rao

[Rabab Al hasni]

Dear Sunita,

I got the same problem with the multinomial regression. I would like to ask please how did you solve it? I tried to apply Carlo advice of MandatedAssurance but it did not work with me. Thus, Can I ask you please what is the command syntax that needs to be run it in Stata?


Many thanks for your kind help in advance

Best regards,
Rabab

[Carlo Lazzaro]

Duplicate post. See reply at: https://www.statalist.org/forums/for...ighly-singular

[Sunita Rao]

Hi Carlo,

I apologize profusely for taking so long to get back to you.

It discussed this issue with my co-author. It turns out that we need to keep MandatedAssurance in our model for theoretical reasons.

A Big Thank you for giving me advice. I will certainly keep it in mind if I do multinomial regression.

Warm Regards,
Sunita