Assess Missing Network Data and Boundary Sensitivity
Diagnose how nonresponse, fixed-choice nominations, and boundary decisions alter a network, then report which structural conclusions remain stable across plausible alternatives.
Method sources
By the end of this tutorial
- 1Distinguish an observed absence of a tie from an unobserved actor, dyad, or interaction context.
- 2Construct sensitivity scenarios that match plausible boundary and missing-data mechanisms.
- 3Report robust and fragile network conclusions without filling gaps as if they were observed facts.
SNA
Network specification
Analysis scenario
A district maps advice seeking among instructional staff in six schools before allocating coaching support. Eight invited educators did not respond, two schools supplied incomplete staff rosters, and the survey allowed at most five nominations. The team must learn which findings survive plausible missingness and boundary choices before acting.
Nodes
All instructional staff employed in the six participating schools on the census date, with roster inclusion, response status, school, and role documented for every eligible actor.
Ties
A directed advice-seeking nomination from educator A to educator B for help used during the previous eight weeks, capped at five outgoing nominations by the survey design.
Network type
One-mode, directed, unweighted whole network with a six-school organizational boundary, one recall period, a fixed-choice cap, and missing actors and dyads retained as unknown rather than coded as no tie.
Step-by-step tutorial
Audit coverage and missingness
Join the invitation roster, response log, and edge list before calculating network measures. Mark eligible actors who are absent, distinguish item nonresponse from a reported zero, record the five-nomination cap, and summarize coverage by school and role rather than only for the network overall.
Checkpoint
Every eligible actor has a known roster and response status, and no unknown dyad has been silently converted into an observed absence of a tie.
Define boundary alternatives
Write the primary six-school boundary and at least two defensible alternatives before viewing centrality results. For example, compare current employees with staff active during the full recall window, and compare the six participating schools with a restricted boundary that excludes schools whose rosters are incomplete.
Checkpoint
Each boundary has a substantive inclusion rule, a time reference, and a documented reason; none was chosen because it produces a preferred network picture.
Run mechanism-matched sensitivity checks
Recalculate density, components, reciprocity, degree, betweenness, and community membership under the pre-specified boundaries. Then repeat the analysis after plausible actor and tie omissions that reflect observed nonresponse or fixed-choice censoring, keeping random and targeted omissions as separate scenarios.
Checkpoint
The comparison table shows how each conclusion changes by boundary and missingness mechanism, including rank stability and uncertainty rather than one replacement estimate.
Report what is robust and fragile
Classify findings as stable, magnitude-sensitive, rank-sensitive, or unsupported across scenarios. Describe which missingness assumptions drive each result, show more than one network view when needed, and recommend additional data collection before any decision that depends on a fragile actor label or subgroup boundary.
Checkpoint
The final claim names its boundary and coverage, separates observed data from sensitivity scenarios, and does not treat imputed or simulated ties as facts about people.
Interpret with care
A stable conclusion is one that keeps its substantive meaning across defensible boundaries and plausible missingness mechanisms. Similar-looking overall density can coexist with unstable components, centrality rankings, or community assignments, so robustness must be checked for the exact claim being made.
Nonresponse is rarely neutral. Missing peripheral actors can inflate apparent cohesion, while missing central actors can fragment paths and reorder betweenness. A fixed-choice cap may censor highly active respondents differently from others, so random deletion alone is not an adequate sensitivity model.