ivreg2 warning covariance matrix of moment conditions not of full rank

[Hend She]

Dear Stata Users,
I was trying to estimate an IV regression (2SLS)

I have few dummies in my model, the dependent is a binary variable of benefit's claim and I am instrumenting the amount. I use weights to correct for sampling issues.
I got this error at the end of my 2nd stage results table

 ivreg2 nclaim age age2 married ib3.edu i.eu_nat i.pgemplst_gen i.howner i.singlep i.dis_dummy  i.female i.east ib2.citysize i.haskids  (ben_amt=  prvtransfers needs) [pw=hweight] if  head==1, first 
-----------------------------------------------------------------------------------------------------------
Underidentification test (Kleibergen-Paap rk LM statistic):            146.799
                                                   Chi-sq(2) P-val =    0.0000
------------------------------------------------------------------------------
Weak identification test (Cragg-Donald Wald F statistic):             1545.588
                         (Kleibergen-Paap rk Wald F statistic):        165.936
Stock-Yogo weak ID test critical values: 10% maximal IV size             19.93
                                         15% maximal IV size             11.59
                                         20% maximal IV size              8.75
                                         25% maximal IV size              7.25
Source: Stock-Yogo (2005).  Reproduced by permission.
NB: Critical values are for Cragg-Donald F statistic and i.i.d. errors.
------------------------------------------------------------------------------
Warning: estimated covariance matrix of moment conditions not of full rank.
         overidentification statistic not reported, and standard errors and
         model tests should be interpreted with caution.
Possible causes:
         singleton dummy variable (dummy with one 1 and N-1 0s or vice versa)
partial option may address problem.
------------------------------------------------------------------------------
Instrumented:         ben_amt
Included instruments: age age2 married 0.edu1 1.edu2
                      2.edu3 4.edu4 1.howner 1.singlep 1.dis_dummy  
                      1.female 1.east 1.citysize 3.citysize 1.haskids
Excluded instruments: prvtransfers needs
------------------------------------------------------------------------------

I don't know what to do here with this error message. The results show the coefficient of the instrumented variable/ ben_amt -.007962 to be be significant (P-value 0.095 ) though weakly significant. The two instruments used are shown to be strongly significant in the first stage results (both p-value 0.000)

[Dung Le]

Hi Hend,

That warning comes from singleton issue as reported by ivreg2 (from SSC)

Possible causes: singleton dummy variable (dummy with one 1 and N-1 0s or vice versa)

ivreg2 also gives you a hint to address that issue, that is the partial option.

partial option may address problem.

That said, you code could be modified as (I added the partial option in red)

 ivreg2 nclaim age age2 married ib3.edu i.eu_nat i.pgemplst_gen i.howner i.singlep i.dis_dummy  i.female i.east ib2.citysize i.haskids (ben_amt=prvtransfers needs) [pw=hweight] if  head==1, first partial(list_of_singleton_covariates)

[Hend She]

Dear Dung,
Thanks a lot for the advice! I was aware that it might be caused by the dummies but didn't know how to handle it. Could you please write the syntax similar to how I should type it in Stata? I got an error when I run it

[Dung Le]

Hi Hend,

Since you did not provide an example of your data and I do not have my own data example here, I wan unable to give you an example of Stata code. Can you post back with an example of your data using dataex command?
Just for self-explored, assume that variable dis_dummy in your model is singleton, so your code could be:

 
 ivreg2 nclaim age age2 married ib3.edu i.eu_nat i.pgemplst_gen i.howner i.singlep i.dis_dummy i.female i.east ib2.citysize i.haskids (ben_amt=prvtransfers needs) [pw=hweight] if  head==1, first partial(dis_dummy)

[Hend She]

Thanks a lot, Dung!
Here is the code, yes the one you included is one of of the few dummies used.

  ivreg2  nclaim  age age2  married ib3.edu i.eu_nat i.pgemplst_gen i.howner  i.singlep i.dis_dummy  i.female i.east ib2.citysize i.haskids
 (ben_amt =   prvtransfers needs) [pw=hweight] if  head==1, partial(married howner singlep dis_dummy east haskids)
Error: howner singlep dis_dummy east haskids listed in partial() but not in list of regressors.
invalid syntax

[Dung Le]

Hi Hend,

Is dis_dummy a category variable? If so, try:

partial(i.dis_dummy)

[Hend She]

It is a dummy variable (0,1) reflecting disability.

[Dung Le]

Okay, I have just looked at your code again, try this:

partial(married i.howner i.singlep i.dis_dummy i.east i.haskids)

[Hend She]

Perfect, thank you! That worked very well.

---------------------------------------------------------------------------------------------------------
Underidentification test (Kleibergen-Paap rk LM statistic):            146.799
                                                   Chi-sq(2) P-val =    0.0000
------------------------------------------------------------------------------
Weak identification test (Cragg-Donald Wald F statistic):             1545.588
                         (Kleibergen-Paap rk Wald F statistic):        165.936
Stock-Yogo weak ID test critical values: 10% maximal IV size             19.93
                                         15% maximal IV size             11.59
                                         20% maximal IV size              8.75
                                         25% maximal IV size              7.25
Source: Stock-Yogo (2005).  Reproduced by permission.
NB: Critical values are for Cragg-Donald F statistic and i.i.d. errors.
------------------------------------------------------------------------------
Hansen J statistic (overidentification test of all instruments):         2.012
                                                   Chi-sq(1) P-val =    0.1560
------------------------------------------------------------------------------

According to the last result, the overidentification, should be valid. I assume, one should make a remark that partialling out has been used to get the estimation?

On another note, I think this partial out option does not work with robust, though robust is recommended in case of binary dependent variables.