Separate Diffusion from Homophily
Use time ordering and a defensible identification strategy before claiming that connected students transmitted an attitude, behavior, or outcome.
Method sources
By the end of this tutorial
- 1Define diffusion, influence, and homophily confounding from an explicit node, tie, boundary, direction, weight, and observation-window specification.
- 2Apply and audit this procedure: Draw a temporal causal diagram, distinguish prior similarity from later convergence, measure shared contexts, specify an identification strategy or randomized encouragement where feasible, test negative controls, and report effects only under stated assumptions.
- 3Interpret the result with a sensitivity check and the following evidence boundary: Connected people can resemble one another because of selection, shared context, measurement, simultaneous change, or influence; ordinary regression cannot generally separate these explanations.
SNA
Network specification
Analysis scenario
Students with connected study partners become more similar in attendance over a year, and a dashboard labels the pattern as peer influence.
Nodes
Students eligible throughout the longitudinal observation period, with entry, exit, exposure opportunities, and baseline attendance recorded.
Ties
Time-stamped study-partner nominations observed before and during repeated attendance measurements.
Network type
Longitudinal directed network and repeated behavior data with changing ties, shared environments, and measured plus unmeasured confounders.
Step-by-step tutorial
Freeze the relational question
Write the decision the analysis must inform, then lock the eligible node roster, tie-generating event, direction, weight, self-tie rule, observation window, and missing-data code. Preserve a read-only source copy and record why this specification represents the stated question.
Checkpoint
A second analyst can reconstruct the same node set and edge table from the written rules without guessing what an absent record means.
Compute the focal structure
Work on a versioned analysis copy and carry out the focal method exactly as specified: Draw a temporal causal diagram, distinguish prior similarity from later convergence, measure shared contexts, specify an identification strategy or randomized encouragement where feasible, test negative controls, and report effects only under stated assumptions. Save software and package versions, every threshold, normalization, seed, and intermediate count needed to reproduce the result.
Checkpoint
The output is tied to one named data version and includes the denominator, parameter settings, and a reproducible calculation record.
Run a structural sensitivity check
Repeat the analysis under at least one defensible alternative boundary, missingness rule, tie threshold, weight transformation, or model setting. Compare membership and substantive conclusions, not only a single coefficient, and investigate every change large enough to alter a decision.
Checkpoint
The audit states which patterns persist, which actors or groups change classification, and which conclusion depends on an analyst choice.
Report for responsible action
Pair the numerical result with a table or structure-preserving visual, document excluded and missing actors, and explain uncertainty in plain language. Convert the finding into a reversible support question, not an automatic ranking, while stating this boundary: Connected people can resemble one another because of selection, shared context, measurement, simultaneous change, or influence; ordinary regression cannot generally separate these explanations.
Checkpoint
The final note contains the question, specification, result, sensitivity evidence, uncertainty, privacy controls, and a proportionate next step.
Interpret with care
diffusion, influence, and homophily confounding describes a property of the specified relation and network boundary. It does not transfer automatically to another relation, time period, class, platform, or population.
Connected people can resemble one another because of selection, shared context, measurement, simultaneous change, or influence; ordinary regression cannot generally separate these explanations. Compare the result with raw counts, missingness, plausible alternative specifications, and contextual evidence before acting.