An independent study reference written by Dr Phuc V. Nguyen. It is not official subject material — for assessment requirements always follow your subject outline and vUWS.

Networksadvanced

Running a network study

Q: An analyst achieves a 70 per cent response rate on a 60-person network survey, records a tie only when both actors name each other, and reports betweenness scores as final. What is the strongest objection?

Roughly half the pairs cannot be observed, and betweenness depends on paths through actors who may be absent. At a response rate near 0.7, with a tie needing confirmation from both ends, only about 0.49 of pairs can be assessed. Betweenness is built from paths, so a missing actor removes routes and can shift brokerage onto whoever remains. That response rate is not unusual for survey work in general, and random sampling is the wrong fix, because a network needs coverage of a defined population rather than a representative sample of it.

Q: A study of a 30-person team publishes its network diagram with names replaced by codes. A team member reads the report. What is the realistic privacy risk?

A reader can identify individuals from their position, such as the only person linking two sub-groups. Structure is itself an identifier. Anyone who knows the team can match a distinctive position to a person, and small networks make that easy. This is why network studies need consent that covers being named by others, careful decisions about what is published, and often aggregated reporting rather than a full diagram.

A network study succeeds or fails at the design stage. Four decisions do most of the work. The question, which determines everything after it. The boundary, meaning exactly who is inside the network and who is not. The relation, which has to be one specific tie definition rather than a vague sense of connection. And the collection method, whether a roster, a name generator or digital trace data. Missing actors do far more damage here than missing rows do in ordinary survey work, because every tie incident to the missing actor goes with them, along with every path that ran through them.

Why it matters

Ordinary survey work loses a respondent and loses one row. A network study loses a respondent and loses the ties incident to that person, plus every path that ran through them, so two people you did collect can look further apart than they are. That is why design comes first. You cannot patch a network dataset the way you patch a spreadsheet, because what is missing is not the cell, it is the connection.

Before you read on — recall

An analyst achieves a 70 per cent response rate on a 60-person network survey, records a tie only when both actors name each other, and reports betweenness scores as final. What is the strongest objection?

Formulas

Share of dyads with both ends responding

P(\text{both ends respond}) \approx r^{2}

This is complete-case dyad coverage. Under a design that accepts a tie only when both actors confirm it, and with actors responding independently at rate

r

, roughly

r^2

of pairs can be assessed. At a response rate of 0.8 that is about 64 per cent, even though only a fifth of people are missing. If a single nomination is enough to record a tie, a pair is covered whenever either end responds, which is a much larger share, so state the collection rule before quoting any coverage figure.

Pairs implied by each boundary choice

N_{\text{pairs}} = \frac{n(n-1)}{2}

Widening a boundary from 50 actors to 150 lifts the pairs to consider from 1,225 to 11,175. Roster methods, which ask every person about every other person, become unworkable long before the actor count feels large.

Worked examples

Scenario

A hospital asks for a study of the network behind handover quality. The obvious boundary is the ward roster of 46 nurses.

Solution

The roster is a nominalist boundary, drawn by the researcher from an administrative list. It is defensible and it will miss the two agency staff who cover most night shifts and the pharmacist everyone phones. A realist alternative asks participants who they actually rely on and includes whoever is named often enough. Each approach produces a different network and a different answer about fragility. Neither is wrong. What is wrong is failing to state which one was used, because the reader cannot otherwise judge the density or the bridges.

Scenario

A firm has three years of internal messaging logs and wants a collaboration network built from them without running a survey.

Solution

Trace data is complete, cheap and measures something different from what people report. A message log records contact rather than reliance, over-counts scheduling chatter, and misses conversations held in person. It also leaves a consent question that a survey answers explicitly. A workable design builds the network from the logs, validates it against a short survey on a sample of staff, reports the agreement rate, and defines a tie by a stated threshold such as a minimum of two-way exchanges per month.

Common mistakes

✗A network study is a survey with extra questions. The dependency structure changes the statistics. Standard significance tests assume independent observations, and network data violates that by construction, which is why permutation approaches such as the quadratic assignment procedure are used instead of ordinary tests.
✗Removing names protects participants. Removing names does not remove position. Someone who is the only bridge between two departments is identifiable from the structure alone, and small networks are effectively re-identifiable from the pattern of degrees. Protection has to come from aggregation, access control and informed consent, not from relabelling.
✗A high response rate makes the network safe. Coverage of pairs falls faster than coverage of people. Under mutual confirmation, even at 90 per cent response roughly a fifth of pairs have a missing end, and the actors most likely to be missing, such as contractors and part-timers, are often the ones holding the bridges.
✗Digital trace data removes the measurement problem. Trace data measures the channel it was logged on and nothing else. It records contact rather than reliance, it stops at the platform boundary, and it silently excludes anyone who works mostly offline. It is a different instrument, not a neutral one.

Revision bullets

•Design order: question, then boundary, then relation, then collection method
•Nominalist boundaries come from a list; realist boundaries come from the actors
•One relation per network; never merge advice, friendship and reporting lines
•Under mutual confirmation, dyad coverage is roughly the square of the response rate
•Structural position re-identifies people even after names are removed
•Trace data and self-report measure different relations, so validate one against the other

Quick check

A study of a 30-person team publishes its network diagram with names replaced by codes. A team member reads the report. What is the realistic privacy risk?

Connected topics

Sampling Bias What SNA Is Actors and Ties Betweenness Density CRISP-DM Why Ethics Matters

More in Networks

Weighted Ties Degree Centrality Closeness Bridges and Cliques

Sources

Laumann, Marsden & Prensky (1983)
Laumann, E. O., Marsden, P. V., & Prensky, D. "The Boundary Specification Problem in Network Analysis." In R. S. Burt & M. J. Minor (eds), Applied Network Analysis: A Methodological Introduction, SAGE, 1983, pp. 18-34.
The standard statement of the nominalist and realist approaches to drawing a network boundary.
Marsden (1990)
Marsden, P. V. "Network Data and Measurement." Annual Review of Sociology, 16, 435-463, 1990.
Reviews name generators, rosters and informant accuracy, and what each instrument actually measures.
Kossinets (2006)
Kossinets, G. "Effects of Missing Data in Social Networks." Social Networks, 28(3), 247-268, 2006.
Quantifies how non-response and boundary error distort network measures.
Kadushin (2005)
Kadushin, C. "Who Benefits from Network Analysis: Ethics of Social Network Research." Social Networks, 27(2), 139-153, 2005.
Sets out why consent and anonymity work differently when respondents name other people.

How to cite this page

Dr. Phil's Quant Lab. (2026). Running a network study. Derivatives Atlas. https://phucnguyenvan.com/concept/ba-running-a-network-study