[Rsiena-help] SIENA: rate function with large s.e.

Wed Sep 5 15:20:31 CEST 2012

Hello Philippe,

you are right about formula 29 in the 2007 book chapter, indeed the 
conditions need to be switched to match what is written (correctly) in 
the text.

We wrote 'left-continuous' because we thought until the decision to 
change something is actually taken, including this decision time point, 
the state remains the same, and it changes in consequence of the 
decision, infinitesimally after that time point. It doesn't actually 
matter, one could just as well argue for a right-continuous process, I 
guess - the difference between the two is on a null set for the Markov 
process anyway. But the left-continuous version makes it more convenient 
to write down equations like (9) and (10), depending on the state at the 
decision moment.

Best regards,
Ch

Am 05/09/2012 14:59, schrieb Philippe Sulger:
> Hi Christian
>
> Ok. Thank you very much for your reply. I will follow your suggestions.
>
> Btw.: I was/am reading Snijders et al. (2007) 
> <http://www.stats.ox.ac.uk/%7Esnijders/siena/chapter_coevol.pdf>. I 
> have already dropped a line to Tom, but I do not want to bother with 
> (misplaced and unjustfied) pickiness, or redundant details: I 
> certainly miss something/do not understand something as I should, but 
> up to now, in equation (29), the "if" - conditions should be switched 
> so that they are consistent with the previous text making sense to me, 
> shouldn't they? In any case, I am doubting that my guess is right, 
> since in the book 
> <http://books.google.ch/books?id=RS_tDqROnacC&pg=PA41&lpg=PA41&dq=Longitudinal+models+in+the+behavioral+and+related+sciences&source=bl&ots=aYjhHBuWu7&sig=7paHd8Dp8QnaCyH2xQ2a9DbgZZg&hl=de#v=onepage&q=Longitudinal%20models%20in%20the%20behavioral%20and%20related%20sciences&f=false> 
> they are as in the paper...
>
> And on page 7: shouldn't the process be a /right/-continuous function 
> of time? I am sorry about these questions but I just want to make sure 
> that I understand it correctly.
>
>
> Best,
> Philippe
>
> On 05.09.12 12:39, Christian Steglich wrote:
>> Hi Philippe,
>>
>> your result suggests that behaviour change is very quick compared to 
>> network change, i.e., that according to the given model specification 
>> (a very sparse one), the first behaviour measure is actually 
>> non-informative for explaining the evolution to the second behaviour 
>> measure.
>>
>> This could be due to several things: actor heterogeneity when the 
>> model assumes homogeneity, some other type of model misspecification, 
>> use of an inflated behaviour scale (your parameters are very small, 
>> suggesting as much), or true independence in your behaviour 
>> responses. I'd suggest you check the micro step distribution first 
>> (who makes how many steps in which direction) for any anomalies.
>>
>> Best,
>> Ch
>>
>>
>>
>> Am 14/07/2012 18:06, schrieb Philippe Sulger:
>>> Dear RSiena-Users
>>>
>>> I jointly analyze network dynamics and behavioral dynamics. 
>>> Something that I am concerned about are the large standard errors on 
>>> the estimate of the behavioral rate parameter. Here an example:
>>>
>>>                                                   Estimate 
>>> Standard   Convergence
>>> Error      t-ratio
>>> Network Dynamics
>>>    1. rate  basic rate parameter friends           7.8441 (  1.3870  
>>> )    0.0131
>>>    2. eval  outdegree (density)                   -1.1287 (  0.1991  
>>> )   -0.0392
>>>    3. eval  reciprocity                            1.8075 (  1.1237  
>>> )    0.0220
>>>    4. endow reciprocity                           -1.7207 (  1.9371  
>>> )   -0.0050
>>>    5. eval  balance                                0.2603 (  0.0565  
>>> )    0.0264
>>>    6. eval  same covariate                         0.2877 (  0.2220  
>>> )    0.0135
>>>
>>> Behavior Dynamics
>>>    7. rate  rate behav period 1                   34.9277  ( 
>>> 36.6589  )   -0.0177
>>>    8. eval  behavior behav linear shape           -0.0327 (  0.0592  
>>> )   -0.0194
>>>    9. eval  behavior behav quadratic shape         0.0003 (  0.0048  
>>> )   -0.0220
>>>   10. eval  behavior behav: effect from covariate -0.0096 (  0.1383  
>>> )   -0.0588
>>>
>>> As you see, convergence is good. Even if I exclude the 10th effect 
>>> (effect from gender, in many networks I get large s.e. for the rate 
>>> on aggressive behavior). I also made the rate of behav dependent on 
>>> covariates with no essential effect.
>>>
>>> Are these large s.e.'s of concern?
>>>
>>> Thanks for your help.
>>>
>>> Best,
>>> Philippe
>>>
>>>
>>> _______________________________________________
>>> Rsiena-help mailing list
>>> Rsiena-help at lists.r-forge.r-project.org
>>> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rsiena-help
>>
>>
>> -- 
>> _ __ ___ ____ ___ __ _ __ ___ ____ ___ __ _
>>
>> Christian Steglich, researcher
>> Faculty of Behavioural and Social Sciences
>> University of Groningen
>> Grote Rozenstraat 31
>> 9712 TG Groningen
>> The Netherlands
>>
>> fon +31-(0)50-363 6189
>> fax +31-(0)50-363 6304
>>
>> http://www.gmw.rug.nl/~steglich/
>> _ __ ___ ____ ___ __ _ __ ___ ____ ___ __ _
>>
>

-- 
_ __ ___ ____ ___ __ _ __ ___ ____ ___ __ _

Christian Steglich, researcher
Faculty of Behavioural and Social Sciences
University of Groningen
Grote Rozenstraat 31
9712 TG Groningen
The Netherlands

fon +31-(0)50-363 6189
fax +31-(0)50-363 6304

http://www.gmw.rug.nl/~steglich/
_ __ ___ ____ ___ __ _ __ ___ ____ ___ __ _

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/rsiena-help/attachments/20120905/429a2f67/attachment-0001.html>