sofiejulie.vanneste
- All Messages By This Member
#3960
Hi all,
I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
Thanks much,
Kind regards
Sofie
Richard DeJordy
- All Messages By This Member
#3965
Hopefully someone with more statistical background will chime in here, but in the meantime, here are a couple thoughts.
If you only have five different (and importantly INDEPENDENT) networks, I suspect you will have difficult getting any meaningful significances for whole network measures.
One thing I have played with is using "Average Degree" as a measure, by including all the nodes' Degree centrality with an indicator of the network and then using T-Tests (for two networks) or ANOVA (for more than two) to see if the "Average" between groups is significantly different; HOWEVER, I am not certain that this would meet all the prerequisite conditions for such tests, as the "within network" observations are not truly independent of each other (and, for example, are sensitive to network size, etc.) My view is, if you don't find any statistical differences using this approach, there probably aren't any, but if you do, it may be an artifact of some other phenomenon. (Which may or may not be useful information for you.)
But, again, these are just a couple immediate thoughts so hopefully someone with more statistical insight will chime in.
Rich
toggle quoted messageShow quoted text
On May 9, 2011, at 3:27 PM, sofiejulie.vanneste wrote:
Hi all,
I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
Thanks much,
Kind regards
Sofie
Mark Temple (Public Health Wales - Health Protection)
- All Messages By This Member
#3969
Rich,
I’m not a statistician but as a Public Health Medic I live and work in a word drenched in statistics, I wonder if the real quest we need to answer is not the purely statistical one is there a significant difference between two networks, which I interpret as meaning are the differences so large that they are unlikely to be due to chance, that is essentially a value judgement however we dress it up as we select the rejection level of our hypothesis in advance.
In my view the important question is really Are the network so different that this has important implications. Again a judgement call but at least it is out in the open and clear. And the interpretation is very much based on your prior beliefs as to what is important and how much this new data alters that belief you hold.
To illustrate my point many new drugs are tested on very large numbers of subjects and can be shown to have “significant” advantages over tried and tested alternatives, but in reality the benefit is so small that the cost of the improvement is so large that actually both the patient and the payer will benefit from using the better known less expensive medication. There are always exceptions but the general point is that significant changes may be unimportant and non-significant ones (insulin for diabetics for example tested in a trial of one patient, who lived to be 75) can be very important!
Not sure if this helps or hinders but I could not resist offering my two ha’pennyworth.
Mark
Dr Mark Temple
GMC Reg Number: 2488589
Ymgynghorydd mewn Meddygaeth Iechyd Cyhoeddus, Teml Heddwch ac Iechyd,Caerdydd CF10 3NW
Consultant in Public Health Medicine, Communicable Disease Surveillance Centre, Public Health Wales,Temple of Peace and Health,Cathays,Cardiff, CF10 3NW
Ffôn/Tel: 029 2040 2471
Symudol/Mobile: 07850 313365 Ffacs/Fax: 029 2040 2529
Ebost/Email: Mark.Temple@...
Rhyngrwyd/Internet: www.iechydcyhoedduscymru.wales.nhs.uk/ www.publichealthwales.org/
Mewnrwyd/Intranet: nww.publichealthwales.wales.nhs.uk
toggle quoted messageShow quoted text
From: ucinet@... [mailto:ucinet@...] On Behalf Of Rich Dejordy
Sent: 10 May 2011 20:15
To: ucinet@...
Subject: Re: [UCINET] Test for significant differences between network measures
Hopefully someone with more statistical background will chime in here, but in the meantime, here are a couple thoughts.
If you only have five different (and importantly INDEPENDENT) networks, I suspect you will have difficult getting any meaningful significances for whole network measures.
One thing I have played with is using "Average Degree" as a measure, by including all the nodes' Degree centrality with an indicator of the network and then using T-Tests (for two networks) or ANOVA (for more than two) to see if the "Average" between groups is significantly different; HOWEVER, I am not certain that this would meet all the prerequisite conditions for such tests, as the "within network" observations are not truly independent of each other (and, for example, are sensitive to network size, etc.) My view is, if you don't find any statistical differences using this approach, there probably aren't any, but if you do, it may be an artifact of some other phenomenon. (Which may or may not be useful information for you.)
But, again, these are just a couple immediate thoughts so hopefully someone with more statistical insight will chime in.
Rich
On May 9, 2011, at 3:27 PM, sofiejulie.vanneste wrote:
Hi all,
I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
Thanks much,
Kind regards
Sofie
John Skvoretz
- All Messages By This Member
#3970
Snijders and Borgatti 1999 in Connections 22: 161-170 have some suggestions using bootstrap or jackknife.
John Skvoretz
5309 Ambrose Ct
Tampa FL 33647-1010
To: ucinet@...
From: fieke_vanneste@...
Date: Mon, 9 May 2011 19:27:49 +0000
Subject: [UCINET] Test for significant differences between network measures
Hi all,
I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
Thanks much,
Kind regards
Sofie
Ray Paquin
- All Messages By This Member
#3972
Rich, Mark & Sofie
I ditto what you`ve said below except that I use non-parametric equivalents of the t-test and anova, the Mann-Whitney and Kruskall-Wallace tests, since the network data is non-parametric. That said, I too, would be interested in more voices here.
Best,
Ray
toggle quoted messageShow quoted text
--- In ucinet@..., "Mark Temple (Public Health Wales - Health Protection)" <mark.temple@...> wrote:
Rich,
I'm not a statistician but as a Public Health Medic I live and work in a word drenched in statistics, I wonder if the real quest we need to answer is not the purely statistical one is there a significant difference between two networks, which I interpret as meaning are the differences so large that they are unlikely to be due to chance, that is essentially a value judgement however we dress it up as we select the rejection level of our hypothesis in advance.In my view the important question is really Are the network so different that this has important implications. Again a judgement call but at least it is out in the open and clear. And the interpretation is very much based on your prior beliefs as to what is important and how much this new data alters that belief you hold.
To illustrate my point many new drugs are tested on very large numbers of subjects and can be shown to have "significant" advantages over tried and tested alternatives, but in reality the benefit is so small that the cost of the improvement is so large that actually both the patient and the payer will benefit from using the better known less expensive medication. There are always exceptions but the general point is that significant changes may be unimportant and non-significant ones (insulin for diabetics for example tested in a trial of one patient, who lived to be 75) can be very important!
Not sure if this helps or hinders but I could not resist offering my two ha'pennyworth.
Mark
Dr Mark Temple
GMC Reg Number: 2488589
Ymgynghorydd mewn Meddygaeth Iechyd Cyhoeddus, Teml Heddwch ac Iechyd, Caerdydd CF10 3NW
Consultant in Public Health Medicine, Communicable Disease Surveillance Centre, Public Health Wales, Temple of Peace and Health, Cathays, Cardiff, CF10 3NWFfôn/Tel: 029 2040 2471
Symudol/Mobile: 07850 313365 Ffacs/Fax: 029 2040 2529
Ebost/Email: Mark.Temple@...<mailto:Mark.Temple@...>
Rhyngrwyd/Internet: www.iechydcyhoedduscymru.wales.nhs.uk<http://www.iechydcyhoedduscymru.wales.nhs.uk/>/ www.publichealthwales.org/<http://www.publichealthwales.org/>
Mewnrwyd/Intranet: nww.publichealthwales.wales.nhs.uk<http://howis.wales.nhs.uk/sitesplus/888>From: ucinet@... [mailto:ucinet@...] On Behalf Of Rich Dejordy
Sent: 10 May 2011 20:15
To: ucinet@...
Subject: Re: [UCINET] Test for significant differences between network measuresHopefully someone with more statistical background will chime in here, but in the meantime, here are a couple thoughts.
If you only have five different (and importantly INDEPENDENT) networks, I suspect you will have difficult getting any meaningful significances for whole network measures.
One thing I have played with is using "Average Degree" as a measure, by including all the nodes' Degree centrality with an indicator of the network and then using T-Tests (for two networks) or ANOVA (for more than two) to see if the "Average" between groups is significantly different; HOWEVER, I am not certain that this would meet all the prerequisite conditions for such tests, as the "within network" observations are not truly independent of each other (and, for example, are sensitive to network size, etc.) My view is, if you don't find any statistical differences using this approach, there probably aren't any, but if you do, it may be an artifact of some other phenomenon. (Which may or may not be useful information for you.)
But, again, these are just a couple immediate thoughts so hopefully someone with more statistical insight will chime in.
Rich
On May 9, 2011, at 3:27 PM, sofiejulie.vanneste wrote:
Hi all,
I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
Thanks much,
Kind regards
Sofie
sofiejulie.vanneste
- All Messages By This Member
#3973
Dear John,
Thank you very much for the reference. However, as a layman in that field of research, I wonder how I should proceed to do a bootstrap.
I assume you can program this in some software programs? Because you cannot manually draw 5000 bootstrap samples, no?
Could anyone give me any tips how to proceed?
Thank you very much!
Sofie
toggle quoted messageShow quoted text
--- In ucinet@..., John Skvoretz <skvoretzj@...> wrote:
Snijders and Borgatti 1999 in Connections 22: 161-170 have some suggestions using bootstrap or jackknife.
John Skvoretz
5309 Ambrose Ct
Tampa FL 33647-1010To: ucinet@...
From: fieke_vanneste@...
Date: Mon, 9 May 2011 19:27:49 +0000
Subject: [UCINET] Test for significant differences between network measuresHi all,
I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
Thanks much,
Kind regards
Sofie
Steve Borgatti
- All Messages By This Member
#3974
Ucinet does have that (networks|compare densities) but it will only compare two relations.
steve
Stephen P. Borgatti
Paul Chellgren Endowed Chair of Management
Gatton College of Business and Economics
University of Kentucky
Lexington, KY 40508-0034 USA
E-mail: sborgatti@...; steve.borgatti@...
Tel: +1 859 257-2257 (O); +1 (512) 843-2674 (Google Voice)
Skype: steve.borgatti
toggle quoted messageShow quoted text
From: ucinet@... [mailto:ucinet@...] On Behalf Of sofiejulie.vanneste
Sent: Thursday, May 12, 2011 9:40 AM
To: ucinet@...
Subject: [UCINET] Re: Test for significant differences between network measures
Dear John,
Thank you very much for the reference. However, as a layman in that field of research, I wonder how I should proceed to do a bootstrap.
I assume you can program this in some software programs? Because you cannot manually draw 5000 bootstrap samples, no?
Could anyone give me any tips how to proceed?
Thank you very much!
Sofie
--- In ucinet@..., John Skvoretz <skvoretzj@...> wrote:
>
>
> Snijders and Borgatti 1999 in Connections 22: 161-170 have some suggestions using bootstrap or jackknife.
>
> John Skvoretz
> 5309 Ambrose Ct
> Tampa FL 33647-1010
>
>
>
>
>
>
> To: ucinet@...
> From: fieke_vanneste@...
> Date: Mon, 9 May 2011 19:27:49 +0000
> Subject: [UCINET] Test for significant differences between network measures
>
>
>
>
>
>
> Hi all,
>
> I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
>
> Thanks much,
>
> Kind regards
> Sofie
>
sofiejulie.vanneste
- All Messages By This Member
#3976
Dear Steve,
Thank you very much for your answer, but I would also like to know if differences in network centralities/average geodesic distance/... are significant.
When I want to compare densities through UCINET the network also needs to be composed of the same actors and in my case this requirement is not always fulfilled.
Thanks much,
Sofie
toggle quoted messageShow quoted text
--- In ucinet@..., "Borgatti, Steve" <sborgatti@...> wrote:
Ucinet does have that (networks|compare densities) but it will only compare two relations.steve
Stephen P. Borgatti
Paul Chellgren Endowed Chair of Management
Gatton College of Business and Economics
University of Kentucky
Lexington, KY 40508-0034 USA
E-mail: sborgatti@...<mailto:sborgatti@...>; steve.borgatti@...<mailto:steve.borgatti@...>
Tel: +1 859 257-2257 (O); +1 (512) 843-2674 (Google Voice)
Skype: steve.borgatti
Web: www.steveborgatti.com<http://www.steveborgatti.com>From: ucinet@... [mailto:ucinet@...] On Behalf Of sofiejulie.vanneste
Sent: Thursday, May 12, 2011 9:40 AM
To: ucinet@...
Subject: [UCINET] Re: Test for significant differences between network measuresDear John,
Thank you very much for the reference. However, as a layman in that field of research, I wonder how I should proceed to do a bootstrap.
I assume you can program this in some software programs? Because you cannot manually draw 5000 bootstrap samples, no?
Could anyone give me any tips how to proceed?Thank you very much!
Sofie--- In ucinet@...<mailto:ucinet%40yahoogroups.com>, John Skvoretz <skvoretzj@<mailto:skvoretzj@>> wrote:
Snijders and Borgatti 1999 in Connections 22: 161-170 have some suggestions using bootstrap or jackknife.
John Skvoretz
5309 Ambrose Ct
Tampa FL 33647-1010To: ucinet@...<mailto:ucinet%40yahoogroups.com>
From: fieke_vanneste@
Date: Mon, 9 May 2011 19:27:49 +0000
Subject: [UCINET] Test for significant differences between network measuresHi all,
I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
Thanks much,
Kind regards
Sofie
Robin
- All Messages By This Member
#4875
Hello Steve,
a question regarding the specific routine you mentioned (Networks>Compare densities): does this function use the parametric or non-parametric variants of the bootstrapping technique in order to generate standard errors (according to Snijders & Borgatti 1999 Connections)?
Many thanks, Robin
________________________
Robin Kubitza, PhD student
Department of Biology, Section of Ecology
University of Turku
FI-20014 Turku
Finland
toggle quoted messageShow quoted text
--- In ucinet@..., "Borgatti, Steve" <sborgatti@...> wrote:
Ucinet does have that (networks|compare densities) but it will only compare two relations.steve
Stephen P. Borgatti
Paul Chellgren Endowed Chair of Management
Gatton College of Business and Economics
University of Kentucky
Lexington, KY 40508-0034 USA
E-mail: sborgatti@...<mailto:sborgatti@...>; steve.borgatti@...<mailto:steve.borgatti@...>
Tel: +1 859 257-2257 (O); +1 (512) 843-2674 (Google Voice)
Skype: steve.borgatti
Web: www.steveborgatti.com<http://www.steveborgatti.com>From: ucinet@... [mailto:ucinet@...] On Behalf Of sofiejulie.vanneste
Sent: Thursday, May 12, 2011 9:40 AM
To: ucinet@...
Subject: [UCINET] Re: Test for significant differences between network measuresDear John,
Thank you very much for the reference. However, as a layman in that field of research, I wonder how I should proceed to do a bootstrap.
I assume you can program this in some software programs? Because you cannot manually draw 5000 bootstrap samples, no?
Could anyone give me any tips how to proceed?Thank you very much!
Sofie--- In ucinet@...<mailto:ucinet%40yahoogroups.com>, John Skvoretz <skvoretzj@<mailto:skvoretzj@>> wrote:
Snijders and Borgatti 1999 in Connections 22: 161-170 have some suggestions using bootstrap or jackknife.
John Skvoretz
5309 Ambrose Ct
Tampa FL 33647-1010To: ucinet@...<mailto:ucinet%40yahoogroups.com>
From: fieke_vanneste@
Date: Mon, 9 May 2011 19:27:49 +0000
Subject: [UCINET] Test for significant differences between network measuresHi all,
I wonder how you can test if -for example- certain obtained network measures (like network centrality, density,...) of five different networks differ significantly from each other (p < .05). Do you have any suggestions?
Thanks much,
Kind regards
Sofie