modelling "indirect" social networks

[Paula de Souza Leao Spinola]

I have a dataset of physicians working in different hospitals in different teams. A physician can work in different hospitals and different teams. Consider the simple example below where I have teams of physicians in two different hospitals. Physicians A and B work together in team 101 in hospital i while Physicians B and C work together in team 210 in hospital j.

clear
input str1 phy_id str1 hosp_id team_id
"A"          "i" 101
"B"          "i" 101
"B"          "j" 210
"C"          "j" 210
end

label var phy_id "physician id"
label var hosp_id "hospital id"
label var team_id "team id"

I want to construct a variable (phy_iv) for each physician that tells me with whom other physicians that work together with him/her have worked in other teams. For instance, for the first row I would like to know with whom physicians that work together with physician A in team 101 (e.g. physician B) have worked in other teams. Hence, the "value" for this variable should be "C" for row 1 and missing for every other row.

My dataset is much more complex than that (physicians may work in different teams within a hospital and teams may consist of more than 2 people). I am stuck on how to deal with this (I will have to construct more than one variable or add rows once I may have more than one physician who will have worked with other physicians who work with a given physician in a given team). So I guess it will be helpful to begin with a simple case.

Many thanks
Paula

[William Lisowski]

You may find this a useful source of information and programs for doing social network analysis in Stata.

https://nwcommands.wordpress.com

[Paula de Souza Leao Spinola]

Many thanks William. But I guess in this simple example a -bys- command should be enough. Would you know how to go about it?

[Mike Lacy]

-nwcommands- is a great resource, as William says, but yes, it's possible your problem has an easy solution with other Stata tools. However, I find it difficult to follow your verbal description. My first suggestion would be for you to create and post a more extensive set of example data that shows 1) what your source data set is like now; 2) what you would like your result data set to look like. In doing this, I'd encourage you to use the -dataex- command, as described in the StataList FAQ. (By the way, I don't see how the hospital id matters for your problem as you have presented it. If one is on a team, that defines membership on that team regardless of hospital location.)

I *think* you might have in mind something like this: "For each subject physician, find all the people s/he has worked with on a team. Among those fellow team members, count how many of them have ever worked together on at least one other team. Assign that value to the subject physician as a measure of her/his ego network density." (Note, by the way, that there could be issues of double counting here. Let say A is the subject, who has worked with B and C on a team. B and C have worked together on three other teams. Does the current team with A count? Does B and C's work on 3 other teams count as just 1, or 3? What if B and C worked on some other team, not the current one with A--does that count?) The rules here would need to be precise, I think.

My gut feeling is that your problem will be most easily handled with social network tools, and I'd guess that the aspect of physicians' ego networks that you want to measure has been thought through and well-developed in the social network discipline before, with all those complexities of double-counting etc. already worked out.

[Paula de Souza Leao Spinola]

Mike, thanks for your helpful suggestions.

It is true that the hospital id does not count in the example above. This is because I generated the variable team_id as egen team_id = group(hosp_id).

Instead of counting, I would like to identify the physicians who have worked with their current colleagues in different hospitals. In other words, I would like to identify the colleagues of the physician´ colleagues who were never their colleagues.

I posted a much more clear post here explaining exactly how my dataset is structured (I also have a time dimension). In the new post all variables matter.
https://www.statalist.org/forums/for...-relationships

I would be great if you could give any hints!