[statnet_help] Changes in how standard errors are reported
across ERGM versions
David Kretschmer
dkretsch at mail.uni-mannheim.de
Wed Aug 12 06:57:18 PDT 2020
Dear Pavel,
thank you, of course. Below in this e-mail, you find the summary() results for both model fits (for version ergm_3.10-4 and ergm_3.1-0, respectively). The affected coefficient is edgecov.edge_cov_num2.
Best,
David
############
ergm_3.10-4:
############
==========================
Summary of model fit
==========================
Formula: all_ergm_classes_prepared_code[[i]]$network_youthid ~ edges +
mutual + gwesp(0.25) + edgecov(edge_cov_num1) + edgecov(edge_cov_num2) +
nodematch("girl") + absdiff("rel_pop_std") + absdiff("max_isei_std")
<environment: 0x57a6fd48>
Iterations: 3 out of 4
Monte Carlo MLE Results:
Estimate Std. Error MCMC % z value Pr(>|z|)
edges -3.71337 0.45744 0 -8.118 < 1e-04 ***
mutual 2.33923 0.55391 0 4.223 < 1e-04 ***
gwesp 0.78552 0.27668 0 2.839 0.00452 **
gwesp.decay -0.12137 0.38231 0 -0.317 0.75089
edgecov.edge_cov_num1 -0.01048 0.24068 0 -0.044 0.96526
edgecov.edge_cov_num2 3.07179 NA NA NA NA
nodematch.girl 1.21212 0.39698 0 3.053 0.00226 **
absdiff.rel_pop_std -0.35737 0.19916 0 -1.794 0.07275 .
absdiff.max_isei_std -0.02750 0.15534 0 -0.177 0.85950
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Null Deviance: 0.0 on 380 degrees of freedom
Residual Deviance: -171.2 on 371 degrees of freedom
Note that the null model likelihood and deviance are defined to be 0.
This means that all likelihood-based inference (LRT, Analysis of
Deviance, AIC, BIC, etc.) is only valid between models with the same
reference distribution and constraints.
AIC: -153.2 BIC: -117.7 (Smaller is better.)
###########
ergm_3.1-0:
###########
==========================
Summary of model fit
==========================
Formula: all_ergm_classes_prepared_code[[i]]$network_youthid ~ edges +
mutual + gwesp(0.25) + edgecov(edge_cov_num1) + edgecov(edge_cov_num2) +
nodematch("girl") + absdiff("rel_pop_std") + absdiff("max_isei_std")
<environment: 0x6c247348>
Iterations: 4
Monte Carlo MLE Results:
Estimate Std. Error MCMC % p-value
edges -3.847e+00 4.478e-01 2 < 1e-04 ***
mutual 2.577e+00 5.826e-01 1 < 1e-04 ***
gwesp 8.631e-01 2.996e-01 1 0.00419 **
gwesp.alpha -9.886e-02 4.051e-01 7 0.80733
edgecov.edge_cov_num1 8.447e-03 2.448e-01 0 0.97249
edgecov.edge_cov_num2 -2.250e+00 2.448e-16 NA < 1e-04 ***
nodematch.girl 1.301e+00 3.955e-01 6 0.00110 **
absdiff.rel_pop_std -3.756e-01 2.053e-01 0 0.06817 .
absdiff.max_isei_std -4.686e-02 1.568e-01 0 0.76519
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Warning: The standard errors are suspect due to possible poor convergence.
Null Deviance: 0.0 on 380 degrees of freedom
Residual Deviance: -167.8 on 371 degrees of freedom
Note that the null model likelihood and deviance are defined to be 0.
AIC: -149.8 BIC: -114.4 (Smaller is better.)
--
David Kretschmer
Universität Mannheim
Mannheimer Zentrum für Europäische Sozialforschung (MZES)
A5, 6
68159 Mannheim
Tel.: +49-621-181-2024
> On 12. Aug 2020, at 04:43, Pavel N. Krivitsky <p.krivitsky at unsw.edu.au> wrote:
>
> Dear David,
>
> Could you by any chance provide the coefficient tables for the two fits you are discussing?
>
> Best Regards,
> Pavel
>
> On Mon, 2020-07-27 at 09:39 +0200, David Kretschmer wrote:
>> Dear all,
>>
>> for a replication project, I have worked with an old version of the ERGM package (ergm-3.1.0) and the more recent versions (ergm-3.10.1, in particular) and I have a question on differences between these versions.
>>
>> I have noticed that, for some of the networks I analyse, the ergm() command returns different results for the old and the new version. In particular, this holds true for networks in which some of the network parameters cannot be estimated from the data, as is obvious from the MCMC chain values having the same value across all iterations and the MCMC% in the ERGM output being NA (both in the old and the new version of ergm()).
>>
>> However, the results between versions differ in the fact that, for those parameters that cannot be estimated, the newer ergm versions report the standard error as NA while the old ergm version reports a standard error that is basically indistinguishable from 0 (value of the order of 10^-15- 10^-30). In both versions, regular numeric values for the coefficients themselves are reported.
>>
>> It seems clear to me that the ERGM estimates from these networks, or at least the estimates for these parameters, do not have a meaningful interpretation. Still, I am interested to learn 1) why, in the older versions of the ERGM package, these standard errors close to zero were reported and 2) when this behaviour changed.
>>
>> Any help would be greatly appreciated.
>>
>> Best,
>> David
>>
>>
>> --
>> David Kretschmer
>> University of Mannheim
>> Mannheim Centre for European Social Research (MZES)
>> A5, 6
>> 68159 Mannheim, Germany
>> Tel.: +49-621-181-2024
>>
>> _______________________________________________
>> statnet_help mailing list
>> statnet_help at u.washington.edu
>> <mailto:statnet_help at u.washington.edu>
>> http://mailman13.u.washington.edu/mailman/listinfo/statnet_help
>> <http://mailman13.u.washington.edu/mailman/listinfo/statnet_help>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman13.u.washington.edu/pipermail/statnet_help/attachments/20200812/8a9fd990/attachment.html>
More information about the statnet_help
mailing list