[statnet_help] temporal ergm with partially observed networks

martina morris morrism at uw.edu
Wed May 13 10:27:46 PDT 2020

Hi Tom,

Identifiable alters are only important if:

1. you are interested in a fixed network on this particular set of
nodes, or

2. the structures you want to model are higher order than dyadic
independent/degree based configurations.

Even in case (2), it may be that these lower order processes place such
constraints on the higher order stats (e.g., the geodesic distn), that you
can reproduce the higher order stats by modeling the lower order ones.
And in your case, you might be able to test that, by modeling the network
egocentrically, simulating from it, and comparing some of the higher order
stats for the observed and simulated data. (with the residual caveat that
the missing data must be accounted for somehow)

You should take a look at our Network Modeling for Epidemics course
materials. We teach that workshop each year. This year there will be a
remote option (in person only if possible).

Course materials/info: http://statnet.org/nme/

Applications still open for this year's course in August, link on the site


On Wed, 13 May 2020, Tom Kraft wrote:

> Hi Martina,

> Thank you for the input! The case I am describing here is in fact part of an epidemic modeling study. The main

> problem I see with the approach of treating the network egocentrically is that I will be forced to sacrifice

> information about identifiable alters (again, if a node is observed then I know with certainty all of the other

> actors they are linked to and their identities, and I also know the full list of people they are not connected

> to). As such, like you say I won't be able to model dependency terms like gwesp. But these are undirected

> contact networks in which if A-B and B-C, then A-C necessarily exists as well.


> To be sure that I am understanding your suggestion: For a given timestep, completely remove all unobserved

> nodes from the network. Estimate parameters using an ego ERGM. Then simulate from the model onto a full network

> containing the observed and unobserved nodes and information about their nodal and edge covariates.


> Is this correct? I am familiar with the tutorial link you sent, but it unfortunately does not cover an

> egocentric application of a stergm. Do you know of any worked examples like that which are available?


> Thank you again,

> Tom


> On Wed, May 13, 2020 at 9:43 AM martina morris <morrism at uw.edu> wrote:

> Hi Tom,


> The alternative approach may be to treat your observed sample

> egocentrically, and use our egocentric inference for ergms/stergms.


> If the terms you anticipate using are restricted to those that can be

> estimated from egocentric data (basically, the dyad-independence terms

> plus degree distributions) these methods would allow you to estimate

> stergms, and simulate networks that reproduce the sufficient statistics in

> expectation.  This is what we are using for our epidemic modeling studies.


> Two refs of interest:


> Theory:

> Krivitsky, P. N., & Morris, M. (2017). Inference for social network models

> from egocentrically sampled data, with application to understanding

> persistent racial disparities in hiv prevalence in the us. Annals of

> Applied Statistics, 11(1), 427-455. doi:10.1214/16-aoas1010


> Tutorial for our software:

> https://statnet.github.io/Workshops/ergm.ego_tutorial.html





> hth,

> Martina



> On Tue, 12 May 2020, Tom Kraft wrote:


> > Hi Zack,

> > Thank you for this helpful reference. I had come across this paper in my literature search but

> should have

> > given it more attention.

> >

> > Is there a place where the source code associated with your paper is available? It remains a bit

> unclear to me

> > how several of the strategies you used (Shat approximation) might be implemented in R.

> >

> > I suppose an alternative is to use multiple imputation as discussed here:

> > Wang, C., Butts, C. T., Hipp, J. R., Jose, R., & Lakon, C. M. (2016). Multiple imputation for

> missing edge

> > data: a predictive evaluation method with application to add health. Social networks, 45, 89-98.

> >

> > Thank you!

> > Tom

> >

> > On Tue, May 12, 2020 at 1:26 PM Zack Almquist <zalmquist at uw.edu> wrote:

> >       Hi Thomas,

> > My paper on missing data for TERGM style analysis might be helpful; take a look at it and see if

> it is

> > addressing your problem of interest:

> >

> > Almquist, Zack W., and Carter T. Butts. "Dynamic network analysis with missing data: theory and

> > methods." Statistica Sinica 28.3 (2018): 1245-1264.

> >

> > Best,

> >

> > Zack

> > ---

> > Zack W. Almquist

> > Assistant Professor

> > Department of Sociology

> > Senior Data Scientist Fellow, eScience Institute

> > University of Washington

> >

> >

> > On Tue, May 12, 2020 at 8:25 AM Tom Kraft <kraft.tom at gmail.com> wrote:

> >       Dear statnet,

> > Thank you for all your wonderful work. I am looking for advice on methods applicable to network

> > data that doesn't fit neatly into the types of data that regularly appear in statnet tutorials. I

> > have a list of sequential discrete time networks involving the same set of actors. At each

> > timepoint, however, I only observe some percent of actors in the whole network (~30%). For all

> > actors observed I have complete knowledge of their ties, including the identity of alters. I

> would

> > like construct a model of these empirical data with relatively basic terms that can be used to

> > simulate full temporal networks with similar network structures.

> >

> > Given the temporal nature of the networks, it seems that a standard tergm could appropriately be

> > used to model the formation and dissolution of ties. Yet given the nature of the sampling, for

> any

> > given network I'm inclined to think that the data are essentially egocentric with known alter

> info

> > because if I have information on a node I can be sure that I know all the connections of that

> > individual. Thus, I am wondering if it is possible to conduct a temporal version of the ergm.ego

> > model. From the materials I have found online, it seems this should be possible:

> >

> > "The principles of egocentric inference can be extended to temporal ERGMs (TERGMs). While we will

> > not cover that in this workshop, an example can be found in another paper that sought to evaluate

> > the network hypothsis for racial disparities in HIV in the US (Morris et al. 2009)."

> > - http://statnet.org/Workshops/ergm.ego_tutorial.html#6_example_analysis

> >

> > However, I am unable to find worked examples to follow up on that reference/approach and it is

> not

> > clear to me how to implement this. Additionally, it seems like ergm.ego is not designed to

> > incorporate information on identifiable alters -- in which case the methods developed

> in Kosikinen

> > and Robins (2010) perhaps would be more appropriate so that this useful information is not

> ignored.

> >

> > Alternatively, I could imagine that this analysis is best conceived of as a tergm with missing

> data

> > at each time step. In this case info on alter ids could be fully utilized I think.

> >

> > I would be very grateful if anyone could comment on whether one of these approaches seems

> feasible,

> > or if there are other options I might consider. Any materials or references to

> > vignettes/tutorials/papers on the topic would also be appreciated. Thank you in advance! Best,

> >

> > Tom

> >

> > Thomas Kraft

> > Postdoctoral Scholar

> > Department of Anthropology

> > University of California, Santa Barbara

> > _______________________________________________

> > statnet_help mailing list

> > statnet_help at u.washington.edu

> > http://mailman13.u.washington.edu/mailman/listinfo/statnet_help

> >

> >

> >


> ****************************************************************

>   Professor Emerita of Sociology and Statistics

>   Box 354322

>   University of Washington

>   Seattle, WA 98195-4322


>   Office:        (206) 685-3402

>   Dept Office:   (206) 543-5882, 543-7237

>   Fax:           (206) 685-7419


> morrism at u.washington.edu

> http://faculty.washington.edu/morrism/




Professor Emerita of Sociology and Statistics
Box 354322
University of Washington
Seattle, WA 98195-4322

Office: (206) 685-3402
Dept Office: (206) 543-5882, 543-7237
Fax: (206) 685-7419

morrism at u.washington.edu

More information about the statnet_help mailing list