[Rsiena-help] Standard Error Estimates

[Tom Snijders]

Dear Nate,

Such differences are very unpleasant indeed, and unacceptable for practical purposes.  I assume that your estimates were obtained with the Method of Moments (the default estimation method).

The parameter for increasing precision of standard errors is the number of runs in Phase 3; this is set by the n3 parameters in sienaModelCreate.

However, such a large instability of s.e.s obtained with n3 = 1000 suggest that something is not right. My guess is that perhaps this has to do with multicollinearity between parameter estimates, which can be diagnosed by looking at the estimates for the inter-parameter correlations given under the heading "Covariance matrix of estimates (correlations below diagonal)" in the output file, and in the summary(...) of the estimation results. If this is the case, then the thing to do is drop some of the effects highly correlated with others. If this is not the case, then there is reason for worry and it would be good to try and find out the cause of this instability - because it is unacceptable.

By the way, just 2 weeks ago I encountered another case of such instability, where indeed the reason was multicollinearity, which led me to add some remarks to Chapter 7 (p. 91) of the manual. This is in the manual which is (since a few minutes ago) available at http://www.stats.ox.ac.uk/~snijders/siena/RSiena_Manual.pdf, not yet the manual that is included in the package.

By the way (2nd), this topic really is more appropriate for the Stocnet/Siena yahoo discussion group (it's a fine line indeed, but if my diagnosis is correct this is not about programming), therefore I copy my reply to the Stocnet/Siena group.

Please let us know what you'll find out.

Best wishes,

Tom

================================================================
Tom A.B. Snijders
Professor of Statistics in the Social Sciences
Department of Politics and Department of Statistics
Nuffield College
University of Oxford
tel. +44-01865-278599

From: rsiena-help-bounces at r-forge.wu-wien.ac.at [mailto:rsiena-help-bounces at r-forge.wu-wien.ac.at] On Behalf Of Nate Doogan
Sent: 18 September 2011 14:41
To: rsiena-help at r-forge.wu-wien.ac.at
Subject: [Rsiena-help] Standard Error Estimates

Hi group.


I will paste two result excerpts. Both are results of the same model on the same data with the same precision parameters (at least those that I know of). I chose to use five phase two sub phases, and 1000 phase three iterations in both estimations.

Here's number one:

  11. eval RSB alter                               -1.1657  ( 0.6275   )  0.0023
  12. eval RSB ego                                 -2.5752  ( 1.2052   )  0.0188
  13. eval same RSB                                -0.8655  ( 0.7055   )  0.0462

Behavior Dynamics
  14. rate rate RSB period 1                        1.4035  ( 0.1899   ) -0.0589
  15. eval behavior RSB linear shape               -1.0667  ( 0.1923   ) -0.0408
  16. eval behavior RSB average similarity          0.8539  ( 0.3881   )  0.0015

Here's number two:

  11. eval RSB alter                               -1.1412  ( 1.9588   )  0.0015
  12. eval RSB ego                                 -2.6017  ( 2.6454   ) -0.0682
  13. eval same RSB                                -0.8478  ( 2.1887   )  0.0587

Behavior Dynamics
  14. rate rate RSB period 1                        1.4049  ( 0.1531   ) -0.0373
  15. eval behavior RSB linear shape               -1.0670  ( 0.2532   )  0.0160
  16. eval behavior RSB average similarity          0.8468  ( 0.7681   ) -0.0453

The only difference is that I asked for the simulated nets in number two (returnDeps=T). In my opinion, the parameter estimates of these two fits are acceptably close. I don't believe that the standard errors are, however. My decision to reject would vary in these estimates for effects 11, 12, and 16. Is it likely that my data is odd in that way? Or is there a software parameter that will allow me to perhaps trade some time for some precision in the estimation of standard errors?


Nathan Doogan
Doctoral Student
The Ohio State University


________________________________
Nuffield College is a Registered Charity No. 1137506. Registered Office: Nuffield College, New Road, Oxford, OX1 1NF
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/rsiena-help/attachments/20110919/a7d4ca69/attachment.htm>