model fitted on these data fails to meet the asymptotic assumptions of the Hausman test

[Tiffany Volken]

Hi guys, I am a current economics student and am struggling with Stata and would be very grateful for any guidance.

My study aims to understand the relationship between the decline in male regular employment caused by the lost decade and how this is affecting female non regular labor force participation.
The dependent variable is male reg employment, and the independent variable is female non reg employment.

I am using secondary and tertiary sector share as control variables to make sure the relationship between the decline in male employment and female non regular employment is to do with the Japanese labour market fundamentally changing - nothing to do with deindustrialization, for example.

I have data from 1982 to 2022 in five year intervals for each prefecture of Japan with female and male non regular and reg employment figure and the sectoral employment breakdown as well. The sum of non regular and regular do not account for all employment though - I have not included family workers and self employed ppl etc.

my question is:
I started off with a pooled OLS with an equation ln(Fnonregular)=beta0 + beta1ln(Mregular)it + beta2SecondaryShareit + beta3TertiaryShare +eit

To then check whether i should use the RM or pooled, I ran a an LM test which rejected the null hypothesis.
I also ran the F test which favoured the FM.

Then to choose between the two, I ran a Hausman test but I keep getting the same error:
chi2(3) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= -219.03

Warning: chi2 < 0 ==> model fitted on these data
fails to meet the asymptotic assumptions
of the Hausman test; see suest for a
generalized test.
I tried doing the xtoverid but that doesnt work either stating some installation issue.

I'm wondering if this issue is arising because I don thave enough control variables or data errors.

I am not that clued up on Stata and would love some guidance.

I have attached screenshots from what I have managed to run.

[Carlo Lazzaro]

Tiffany:
some comments about your post:
1) as per FAQ, you are kindly asked to clarify the "installation issues" that make the community-contributed module -xtoverid- inserviceable;
2) I am not clear why you created two categorical variables (-secondary- and -tertiary-) instead on an unique, three-level categorical variable;
3) With 47 panels, you should go -vce(cluster panelid)- standard error (or -robust- that does the very same job in -xtreg-);
4) After 3), run -xttest0- again.

As a sidelight, I would also investigate the potential misspecification of the functional form of the regressand replicating, by hand, the -linktest- procedure.

[Tiffany Volken]

Hi Carlo,

Thank you for your reply! I managed to fix the hausman issue by putting sigmamore at the end of my command which favoured the Fixed Effects model. I also ran the linktest to obtain (p = 0.274) to find that the squared prediction term insignificant is insignificant. Am I correct in assuming that the model is correctly specified? (I have attached the results)

I ran these:
gen femregshare = (femaleregualr / (femalenonregular + femaleregualr)) * 100
gen maleregshare = (maleregular / (maleregular + malenonregular)) * 100
gen secshare = (secondary / (primary + secondary + teriarty)) * 100
gen tertshare = (teriarty / (primary + secondary + teriarty)) * 100

to get:

gen lnfnr = ln(femnonregshare)
gen lnfr = ln(femregshare)
gen lnmr = ln(maleregshare)

and ran my FE model using robust and this time with a time fixed variable and I have attached the results below. In terms of variables, I decided to enter secondary, tertiary, marriage, over65 as control variables and didnt know you could add them all into one variable. I ran
. xtreg lnfnr lnmr secshare tertshare over65 marriage i.year , fe robust and
. xtreg lnfr lnmr secshare tertshare over65 marriage i.year , fe robust
to see how female non reg would compare to female reg in relation to male reg

Looking at these new results, I see that a lot of these are statistically insignificant and just wanted to ask if this means my regressions are wrong.

Also wondering if there is a command where I can compare specific prefectures to see if these results hold for areas that are more rural, for example.

[Carlo Lazzaro]

Tiffany:
1) your first -xtreg,fe- model does not show evidence of misspecification;
2) to choose between the two -xtreg,fe- models, just compare the adjusted Rsq:

di e(r2_a)

3) in your future posts, please use the CODE delimiters (using the -#- toggle) to post what you typed and what Stata gave you back (as per FAQ). Thanks.