Course outlines

Three outlines, each built on the reading paths of Getting started. A week assumes one lecture and one computer lab of about 90 minutes each; a workshop day assumes about six hours. For each week or day the outline gives the chapters, the core idea, a demonstration with the data set and the commands the chapter uses, and exercises to assign. Exercises are cited as “Ch. 8, Ex. 3”, numbered as in the book. The chapter notes list all exercises of each chapter by level, so you can swap one for another.

The demonstrations are taken from the chapters’ own examples. The do-files of the book contain every printed command in order and can be run in class as they are.

  1. A. Descriptive and inferential network analysis (12 weeks)
    1. Week 1. Relations and their vocabulary
    2. Week 2. From a question to network data in Stata
    3. Week 3. Pictures of networks
    4. Week 4. Describing a whole network
    5. Week 5. Paths and reachability
    6. Week 6. Centrality and brokerage
    7. Week 7. Cohesive groups and communities
    8. Week 8. Positions and roles
    9. Week 9. Ego networks
    10. Week 10. Baselines
    11. Week 11. Permutation tests and QAP regression
    12. Week 12. One complete analysis, and reporting it
  2. B. Statistical models for networks (12 weeks)
    1. Week 1. Random graphs and generators
    2. Week 2. Networks from attributes
    3. Week 3. Null models
    4. Week 4. Permutation tests and QAP regression
    5. Week 5. ERGM I: dependence and change statistics
    6. Week 6. ERGM II: closure, comparison and fit
    7. Week 7. ERGM III: simulation, degeneracy and missing data
    8. Week 8. Change between waves
    9. Week 9. Stochastic actor-oriented models
    10. Week 10. SAOM: adequacy, estimators and several groups
    11. Week 11. Event models
    12. Week 12. One complete analysis, and reporting it
  3. C. Workshop or summer school (5 days)
    1. Day 1. Networks as data
    2. Day 2. Describing networks and positions in them
    3. Day 3. Groups, positions and ego networks
    4. Day 4. Inference for one network
    5. Day 5. A first network model

A. Descriptive and inferential network analysis (12 weeks)

For students new to network analysis. The course follows the common ground (Chapters 1 to 6), the path Describing a network up to Chapter 11, and the first part of the path Statistical models for one observed network (Chapters 15, 18, 19 and 20). It ends with the parts of the capstone that do not need an ERGM.

Each week has a deck with speaker notes, a hands-on slide for each block and the book’s exercises, as PowerPoint with a PDF; the lecture and the lab of a week are in one deck, and instructors are free to use, adapt and share them.

Week 1. Relations and their vocabulary

  • Slides: PowerPoint · PDF
  • Chapters: 1 Why networks?; 2 What is a network?
  • Core idea: a relational question needs data on pairs. A tie is directed or undirected, binary or valued, signed or not, and each choice keeps some information and discards other information. A network can be a cause, an outcome or the context of another process.
  • Demonstration: the four-person network mini, typed with input and declared with nwset; then nwwebuse lazega, nwclear, nwsummarize lazega_adv and nwsym, which turns the 892 advice arcs into 700 edges; the eight-lawyer subset with nwsubset and its edge list with nwtoedge.
  • Exercises: Ch. 1, Ex. 2 (attribute or relation); Ch. 1, Ex. 6 (influence and its rivals for a correlation among friends); Ch. 2, Ex. 3 (density of the work network by hand, before and after symmetrizing); Ch. 2, Ex. 7 (mutual pairs from the numbers of arcs and edges).

Week 2. From a question to network data in Stata

  • Slides: PowerPoint · PDF
  • Chapters: 3 From a research question to a network and a method; 4 Working with networks in Stata
  • Core idea: a network is constructed, not found. Boundary, tie definition, name generator, the rule for one-sided reports and missing data each shape what it can answer, and each belongs in the report. In Stata a network is declared once and then found by name.
  • Demonstration: the OR and AND rules on the lawyers’ friendship reports (501 against 74 edges) with nwsym lazega_fr, mode(max) and mode(min), and nwdyads; the survey example of Chapter 4, from a wide export through reshape long to nwfromedge; nwset, mat(), nwimport, wildcards over several networks, nwsave, nwuse, nwsubset and nwsymmetrize.
  • Exercises: Ch. 3, Ex. 2 (AND rule, OR rule and arcs by hand); Ch. 3, Ex. 3 (the two rules on the advice network); Ch. 4, Ex. 1 (declare a five-person network from a matrix); Ch. 4, Ex. 6 (directed, union and intersection versions of the survey network).

Week 3. Pictures of networks

  • Slides: PowerPoint · PDF
  • Chapters: 5 Seeing networks
  • Core idea: in most layouts the ties are data and the node positions are not. A claim about structure needs a measure defined on the ties. Fixed coordinates are needed to compare pictures; a matrix plot avoids the coordinate question.
  • Demonstration: nwplot flomarriage, label(_nwnode) in the default, layout(circle) and layout(frucht) layouts, and with size(wealth); nwplot lazega_adv, color(office); nwplotmatrix lazega_adv, group(office); the busiest US airports on a map with nodexy() and basemap(usmap); the Chesapeake Bay food web with edgesize() on log flows; an interactive page with generate(x y).
  • Exercises: Ch. 5, Ex. 1 (one network, three layouts); Ch. 5, Ex. 5 (matrix plot by status); Ch. 5, Ex. 10 (which display for which question).

Week 4. Describing a whole network

  • Slides: PowerPoint · PDF
  • Chapters: 6 Describing networks
  • Core idea: density is hard to compare across sizes, so report mean degree with it. Reciprocity, transitivity and the triad census describe closure; the E-I index and assortativity are read against the value the group sizes alone would produce, not against zero.
  • Demonstration: nwdyads lazega_adv, nwtriads lazega_adv, nwmixing lazega_adv, attribute(office) and attribute(status), which print the E-I index with a permutation p-value, and nwassortativity; the gang’s birthplace groups with nwmixing gang, attribute(Birthplace); the Glasgow girls as a contrast in size.
  • Exercises: Ch. 6, Ex. 2 (reciprocity of pairs and of arcs); Ch. 6, Ex. 5 (the work network beside advice and friendship); Ch. 6, Ex. 7 (gender and the expected E-I index); Ch. 6, Ex. 11 (density or mean degree).

Week 5. Paths and reachability

  • Slides: PowerPoint · PDF
  • Chapters: 7 Paths, distance, and reachability
  • Core idea: reachability asks whether anything can flow; distance asks how far. A network that is not connected needs a stated convention for unreachable pairs. In a directed network distance need not be symmetric. Short distances alone do not make a small world.
  • Demonstration: nwpath lazega_adv between two lawyers, nwgeodesic lazega_adv; nwreach flomarriage and nwgeodesic flomarriage, unconnected(16) for the isolated Pucci family; nwgeodesic glasgow1 on a network with unreachable pairs.
  • Exercises: Ch. 7, Ex. 1 (the reachability matrix and Pucci); Ch. 7, Ex. 3 (distances among the Glasgow girls); Ch. 7, Ex. 6 (one pair of lawyers in both directions); Ch. 7, Ex. 10 (what an average distance says about a rumor).

Week 6. Centrality and brokerage

  • Slides: PowerPoint · PDF
  • Chapters: 8 Centrality, brokerage, and structural holes
  • Core idea: name the mechanism (activity, access, control, prestige) before choosing the measure. Say whether ties are counted in their direction. A score measures position, not power or influence.
  • Demonstration: the Florentine table of degree, closeness, betweenness, effective size and constraint; on the advice network nwdegree, generate(adv_out adv_in), nwcloseness with and without nosym, nwbetween, nwevcent, nwburt, nwbrokerage, group(officenum), nwbridges and nwsimmelian.
  • Exercises: Ch. 8, Ex. 3 (a star network by hand); Ch. 8, Ex. 4 (closeness under two penalties); Ch. 8, Ex. 7 (tenure as a rival source of a high in-degree); Ch. 8, Ex. 11 (whom to brief first).

Week 7. Cohesive groups and communities

  • Slides: PowerPoint · PDF
  • Chapters: 9 Cohesion, subgroups, and communities
  • Core idea: “hard to separate” has three meanings: local density (cliques, k-plexes, n-cliques, k-cores), robustness (k-components, the cohesion hierarchy) and separation (communities). A detected community is not by itself evidence of a cohesive social group.
  • Demonstration: the reciprocated advice network from nwsym lazega_adv, mode(min); nwclique, nwkplex, nwnclique, nwkcore, nwkcomponents and nwcohesion; nwcommunity lazega_adv, algorithm(louvain) seed(12345) symmetrize, the same with leiden and infomap for comparison, and nwmodularity lazega_adv, group(officenum) symmetrize; the gang and the macaque cortex as further cases.
  • Exercises: Ch. 9, Ex. 1 (k-components against k-cores); Ch. 9, Ex. 4 (communities of the co-work network); Ch. 9, Ex. 5 (label propagation under four seeds); Ch. 9, Ex. 7 (what a cohesion result stands for).

Week 8. Positions and roles

  • Slides: PowerPoint · PDF
  • Chapters: 10 Positions, roles, and structural equivalence
  • Core idea: two actors can occupy the same position without being tied. Similar tie profiles, blocks, regular equivalence and core and periphery are different answers to “who could replace whom?”
  • Demonstration: the small office typed with input and nwset, with nwsimindex and nwregular; on the advice network nwsimindex lazega_adv, measure(jaccard), nwconcor lazega_adv, splits(2), nwcollapse, nwhierarchy with single and Ward linkage, and nwcoreperiphery with nwplotmatrix lazega_adv, sortby(core) group(core); nwdendrogram on the Florentine families. CONCOR needs a number of blocks in advance; Chapter 23’s stochastic block model chooses it from the data.
  • Exercises: Ch. 10, Ex. 3 (structural equivalence by hand); Ch. 10, Ex. 4 (Ward and CONCOR as two-group splits); Ch. 10, Ex. 7 (is the core just the busiest lawyers?); Ch. 10, Ex. 8 (who could replace lawyer n17).

Week 9. Ego networks

  • Slides: PowerPoint · PDF
  • Chapters: 11 Ego networks. Chapters 12 (two-mode), 13 (valued, signed, multiplex) or 14 (over time) can replace it, or Chapter 10, if the students’ own data call for them.
  • Core idea: an ego network is a unit of analysis in its own right. Size, density, effective size and constraint are largely redundant. Composition is read against what the setting makes available, and exposure before an outcome is a covariate, not proof of influence.
  • Pointer: the exposure regression treats the actors as independent. The model that does not is the autologistic actor attribute model (nwalaam; Chapter 24, Section 24.4), taught in Course B, Week 7. Chapter 24 also takes up the question with a hazard model and a spatial lag.
  • Demonstration: nwego lazega_adv, sizevar(egosize) densvar(egodens), nwegoplot, and nwaltergen with proportion() and diversity(); exposure to smoking friends in the Glasgow data with nwaltergen and regress; the same measures from survey data with collapse.
  • Exercises: Ch. 11, Ex. 1 (ego density and effective size by hand); Ch. 11, Ex. 4 (ego density against constraint); Ch. 11, Ex. 6 (the exposure regression one wave later); Ch. 11, Ex. 9 (which ego measure for a question about promotion).

Week 10. Baselines

  • Slides: PowerPoint · PDF (the deck is shared with Course B, Week 3)
  • Chapters: 15 Random and structured network generators (the sections on density and random graphs); 18 Random networks and null models
  • Core idea: size and density alone set the chance level of many statistics. A null model holds some features fixed, and that choice defines the question a test answers. Dyads share nodes, so ordinary tests do not apply.
  • Demonstration: the density sweep with nwrandom; nwcug lazega_adv, stat(nwclustering) rname(cluster_global) reps(1000) seed(12345) condition(density) plot, then the same test of reciprocity with stat(nwdyads) rname(reciprocity) under density, dyad-census and degree conditioning.
  • Exercises: Ch. 15, Ex. 3 (the transitivity chance leads you to expect); Ch. 15, Ex. 7 (how often a small random graph has no closed triads); Ch. 18, Ex. 3 (the friendship network’s null mean); Ch. 18, Ex. 5 (clustering inside the Boston office).

Week 11. Permutation tests and QAP regression

  • Slides: PowerPoint · PDF (the deck is shared with Course B, Week 4)
  • Chapters: 19 Permutation inference and comparing networks; 20 QAP regression
  • Core idea: relabeling the nodes keeps the structure and breaks the link between identity and position. QAP keeps a regression’s point estimates and replaces its inference with permutation p-values. Significance and predictive accuracy are different claims.
  • Demonstration: nwpermute lazega_fr, nwcorrelate lazega_adv lazega_fr, permutations(1000), nwmixing lazega_adv, attribute(office) permutations(1000); nwqap lazega_adv lazega_fr office practice status, permutations(500) plot, the same with qapspp, and predict(); nwqap glasgow2 glasgow1 smoke1 sport1.
  • Exercises: Ch. 19, Ex. 2 (the smallest possible p-value); Ch. 19, Ex. 4 (advice and co-work); Ch. 20, Ex. 1 (fitted tie probabilities by hand); Ch. 20, Ex. 5 (the advice model without friendship).

Week 12. One complete analysis, and reporting it

  • Slides: PowerPoint · PDF (the deck is shared with Course B, Week 12)
  • Chapters: 29 A complete network analysis (theory, data, description and QAP); 30 Reproducible and responsible network research
  • Core idea: state the expectations, and what would count against them, before any model is fitted. Agreement of analyses on the same arcs is one piece of evidence, not three. A report owes its readers the construction decisions, and its subjects a judgment about disclosure risk.
  • Demonstration: advice density by the status of sender and receiver; activity and seniority by status; the QAP model of Chapter 29. The ERGMs that follow are easier after Chapter 21, so present their results here (associates do not favor partners; being asked follows seniority) and leave the code to course B. From Chapter 30, stored results after nwcug (return list) and nwqap (ereturn list).
  • Exercises: Ch. 29, Ex. 1 (density ratios from the status table); Ch. 29, Ex. 7 (could an all-partner core be chance?); Ch. 30, Ex. 1 (the construction decisions of Chapter 29); Ch. 30, Ex. 4 (how many lawyers their degrees alone identify).

B. Statistical models for networks (12 weeks)

For students who know the material of Chapters 1 to 8, from course A or from another course. It can be taught as a second course. The course follows the path Statistical models for one observed network to its end (Chapters 15, 18 to 22) and then the path Change and events (Chapters 14, 25 to 28), and closes with the capstone. Chapter 16 is added because its logit on the edge list leads directly to dyadic models. Week 7 carries two optional blocks. Chapter 23 (Latent space and block models, nwsbm and nwlsm) gives a second and a third account of the advice network that the ERGMs explain, with hidden groups and hidden positions in place of counted tendencies. Chapter 24 (Influence and contagion) has a section on autologistic actor attribute models (nwalaam), for questions in which an attribute of the actors, not the ties, is the outcome. The rest of Chapter 24 is reading after Week 9, because the co-evolution model of Week 9 answers the question that Chapter 24 leaves open.

Each week has a deck with speaker notes, a hands-on slide for each block and the book’s exercises, as PowerPoint with a PDF; Weeks 3, 4 and 12 use the same decks as Course A, and instructors are free to use, adapt and share them.

Week 1. Random graphs and generators

  • Slides: PowerPoint · PDF
  • Chapters: 15 Random and structured network generators
  • Core idea: a random graph fixes only size and density; other generators add one idea each (geometry, shortcuts, growth). A resemblance shows that a mechanism could produce a pattern, not that it did. One draw is not the model’s typical outcome.
  • Demonstration: nwrandom 71, density(.179) with set seed, nwring, nwlattice, nwsmall and nwpref; the same generators through nwgen; a small-world sweep over rewiring probabilities with nwgeodesic.
  • Exercises: Ch. 15, Ex. 2 (predict transitivity and reciprocity, then draw); Ch. 15, Ex. 6 (degree distributions of a random and a preferential-attachment network); Ch. 15, Ex. 8 (evidence against preferential attachment); Ch. 15, Ex. 9 (what a small-world claim needs).

Week 2. Networks from attributes

  • Slides: PowerPoint · PDF
  • Chapters: 16 Generating networks from attributes and tie probabilities
  • Core idea: tie probabilities can come from group membership or from a homophily coefficient, which is a log-odds ratio. Group sizes and opportunity can produce the same sorting as preference. A logit on the edge list turns several attributes into a generator; what it fails to reproduce shows what attributes alone cannot explain.
  • Demonstration: nwexpand for same-group networks, nwgen for tie probabilities and nwdyadprob to draw; nwhomophily with nwmixing to check recovery; nwtoedge, logit, predict and nwfromedge for the estimated generator.
  • Exercises: Ch. 16, Ex. 2 (the odds factor of a homophily coefficient); Ch. 16, Ex. 3 (the E-I index that ignores status); Ch. 16, Ex. 4 (the homophily sweep for office); Ch. 16, Ex. 9 (preference or opportunity).

Week 3. Null models

  • Slides: PowerPoint · PDF (the same deck as Course A, Week 10)
  • Chapters: 18 Random networks and null models
  • Core idea: the conditional uniform graph test compares an observed statistic with networks that share some of its features. Conditioning on density, the dyad census or the degrees answers different questions. Rejecting a null leaves every mechanism that could produce the pattern in play.
  • Demonstration: nwcug lazega_adv for transitivity and reciprocity under condition(density), condition(census) and condition(degrees), with plot.
  • Exercises: Ch. 18, Ex. 1 (distance from the null mean in standard deviations); Ch. 18, Ex. 2 (why the census condition is refused for an undirected network); Ch. 18, Ex. 4 (Glasgow and the lawyers under two conditions); Ch. 18, Ex. 6 (questions to ask of a “significant” clustering).

Week 4. Permutation tests and QAP regression

  • Slides: PowerPoint · PDF (the same deck as Course A, Week 11)
  • Chapters: 19 Permutation inference and comparing networks; 20 QAP regression
  • Core idea: permutation keeps the observed structure and relabels it; the p-value concerns the data in hand. QAP regression adds several predictors, and turns logit coefficients into odds ratios and fitted probabilities. It does not model dependence between ties.
  • Demonstration: nwcorrelate and nwmixing with permutations(); nwqap lazega_adv lazega_fr office practice status, permutations(500) plot, qapspp, mode(receiver absdist), type(regress) on a valued outcome, and predict(); the Glasgow model with an earlier wave as predictor.
  • Exercises: Ch. 19, Ex. 5 (is advice sorted by gender?); Ch. 20, Ex. 2 (odds ratio and the p-value floor); Ch. 20, Ex. 4 (the Glasgow model with fewer predictors); Ch. 20, Ex. 9 (what is wrong with ordinary standard errors).

Week 5. ERGM I: dependence and change statistics

  • Slides: PowerPoint · PDF
  • Chapters: 21 Exponential random graph models, from the start through the reciprocity step and the check of the MCMC simulation
  • Core idea: a model of a whole network lets ties depend on other ties. An ERGM coefficient is a change in the conditional log-odds of one tie, given the rest of the network, read through its change statistic. A specification is built in stages.
  • Demonstration: nwergm lazega_adv, edges seed(12345); then nodematch(office practice), nodeofactor(status) and nodeifactor(status); then mutual; estat mcmcdiag after the fit.
  • Exercises: Ch. 21, Ex. 1 (what the office coefficient says about a pair); Ch. 21, Ex. 2 (why the normalizing constant cannot be computed); Ch. 21, Ex. 3 (homophily on gender); Ch. 21, Ex. 7 (two MCMC intervals).

Week 6. ERGM II: closure, comparison and fit

  • Slides: PowerPoint · PDF
  • Chapters: 21, from the transitivity step to the end
  • Core idea: homophily, status, reciprocity and closure compete for the same ties. Convergence of MCMLE is not a verdict on the model. Fit is judged on statistics the model does not contain.
  • Demonstration: the final model with gwesp(.5); estat mcmcdiag, plot; estimates store and lrtest; estat gof, gof(model distance odegree idegree espartners) seed(12345) plot; predict, frame(pr) pr.
  • Exercises: Ch. 21, Ex. 4 (goodness of fit with 300 draws); Ch. 21, Ex. 6 (other gwesp decays and gwodegree); Ch. 21, Ex. 9 (co-work as a rival to closure); Ch. 21, Ex. 10 (a statistic for unreachable pairs).

Week 7. ERGM III: simulation, degeneracy and missing data

  • Slides: PowerPoint · PDF
  • Chapters: 22 Advanced exponential random graph models (Section 22.7, a Bayesian ERGM, can be left as reading)
  • Core idea: simulating from a model shows what it implies. A coefficient does not move only its own statistic. A raw triangle count can make a model degenerate; geometrically weighted terms avoid it. A missing dyad is not a non-tie.
  • Demonstration: nwergm simulate from hand-chosen coefficients and from the fitted advice model with one coefficient changed; nwergm mesa, edges nodematch(grade race) gwesp(.25) seed(12345) and the same model with triangle, which fails; a version of mesa with missing dyads built with nwtomata.
  • Optional extra 1, another account of the same network (Chapter 23): the ERGM explains ties by counted tendencies; a stochastic block model and a latent space model explain them by hidden groups and hidden positions. Demonstration: nwsbm lazega_adv, maxblocks(6) seed(12345) generate(block) replace, the table of ICL values, the blocks against office and status with tabulate; nwsbmplot, type(matrix) sort shade; nwlsm with estat mcmcdiag and nwlsmplot. The chapter opens with what both models share: ties that depend on hidden attributes, an amount of hidden structure that is chosen, and positions and labels known only up to relabeling. Homework: Ch. 23, Ex. 2 and 3.
  • Optional extra 2, an attribute as the outcome (Chapter 24, Section 24.4): nwalaam smoker1 glasgow1, effects(density contagion sender receiver) seed(12345), its pseudolikelihood start with mple, and estat gof; a cross-sectional association, not influence. Homework: Ch. 24, Ex. 4 and 5.
  • Exercises: Ch. 22, Ex. 1 (why mutual has no meaning in an undirected network); Ch. 22, Ex. 2 (why a triangle count degenerates); Ch. 22, Ex. 4 (simulate with the gwesp coefficient at 0.3); Ch. 22, Ex. 6 (ten missing dyads).

Week 8. Change between waves

  • Slides: PowerPoint · PDF
  • Chapters: 14 Networks over time; 25 Stochastic actor-oriented models, through the ministep process, the objective function and the first look at the data
  • Core idea: a network observed once is a snapshot of a process. Ties between waves are stable, formed or dissolved. An SAOM represents the unseen process as small choices by actors, with a rate function (how often an actor may change a tie) and an objective function (which change it makes).
  • Demonstration: nwturnover glasgow1 glasgow2, generate(_turnover) with and without formation; plots at fixed positions with nodexy() and nwmovie; nwsaom describe, waves(glasgow1 glasgow2 glasgow3); nwsaom, wave1(glasgow1) wave2(glasgow2) outdegree reciprocity seed(1).
  • Optional: Section 14.6, adoption over the waves: nwdiffuse curve, exposure and threshold on smoking in the three Glasgow waves, which leads to the co-evolution model of Week 9.
  • Exercises: Ch. 14, Ex. 2 (Jaccard index and persistence by hand); Ch. 14, Ex. 7 (formed ties against all pairs without a tie); Ch. 25, Ex. 1 (the counts of nwsaom describe); Ch. 25, Ex. 7 (what the rate counts).

Week 9. Stochastic actor-oriented models

  • Slides: PowerPoint · PDF
  • Chapters: 25, from building up a model on Glasgow to the end
  • Core idea: convergence t-ratios check the estimation; the z statistics test effects. Goodness of fit shows where a model fails. A co-evolution model separates selection from influence, given its assumptions and a behavior that changes often enough.
  • Demonstration: nwsaom, wave1(glasgow1) wave2(glasgow2) outdegree reciprocity transtrip nodematch(smoke1) seed(12345); three waves with waves(); nwsaom refit; estat gof, nsim(100) seed(1); the co-evolution model with behavior(alcohol1 alcohol2 alcohol3), behsim and avalt.
  • Pointer: with one wave only, nwalaam (the optional block of Week 7) describes how an attribute and the ties around each actor line up, and cannot separate selection from influence (Chapter 24).
  • Exercises: Ch. 25, Ex. 2 (the second period and a restart); Ch. 25, Ex. 3 (homophily in sport); Ch. 25, Ex. 5 (avsim in place of avalt); Ch. 25, Ex. 8 (a teacher’s seating plan).

Week 10. SAOM: adequacy, estimators and several groups

  • Slides: PowerPoint · PDF
  • Chapters: 26 Advanced stochastic actor-oriented models
  • Core idea: the score test says which effect to try next; the time test whether effects stay constant; simulation and the importance table show what a model implies. Method of Moments, maximum likelihood and Bayesian estimation have their own diagnostics. Several groups can be pooled in a joint model or combined in a meta-analysis.
  • Demonstration: estat score, estat timetest, fixrates, estat importance and estat simulate; nwsaom simulate; the ml option and nwsaom bayes, nmain(250) seed(1); nwwebuse baerveldt19, nwstack school*_w1, name(w1) group(school), a joint model with groups(school), and nwsaom meta.
  • Exercises: Ch. 26, Ex. 1 (two estimators in standard errors); Ch. 26, Ex. 2 (Bayesian runs with other seeds); Ch. 26, Ex. 3 (freeing outdegree activity); Ch. 26, Ex. 4 (40 classes: joint model or meta-analysis).

Week 11. Event models

  • Slides: PowerPoint · PDF
  • Chapters: 27 Relational event models; 28 The Dynamic Network Actor Model
  • Core idea: with a time-stamped log the unit is the event. The relational event model compares, at every event, the pair that fired with all pairs at risk, so it has no intercept. Frequency and recency of past contact are different claims. The DyNAM separates who acts from whom the actor chooses.
  • Demonstration: the ten-event toy log, nwset sender receiver, eventtime(t) name(chat) and nwrem chat, nodsnd nidrec; the company emails (keep if t < 14) with nwrem mail and frpsndsnd, rsndsnd and covevent(samesup) after nwload mail, xvars and nwexpand supervisor; nwdynam mail, submodel(rate), nwdynam mail, inertia recip indeg same(branch) and submodel(rate) intercept.
  • Exercises: Ch. 27, Ex. 2 (the risk set and the log likelihood of equal weights); Ch. 27, Ex. 4 (recency of receipt); Ch. 28, Ex. 3 (an expected waiting time); Ch. 28, Ex. 8 (which of five models for which question).

Week 12. One complete analysis, and reporting it

  • Slides: PowerPoint · PDF (the same deck as Course A, Week 12)
  • Chapters: 29 A complete network analysis; 30 Reproducible and responsible network research
  • Core idea: one question about the law firm, whether advice runs up the status order, is carried from theory and stated expectations through description, QAP regression and a sequence of ERGMs to a synthesis, with sensitivity runs chosen in advance. The report states the construction decisions and confronts the disclosure risk of relational data.
  • Demonstration: the QAP model; the ERGM with friendship as edgecov(); nodemix(status) and lincom for whose choices carry the status effect; seniority rank next to partnership (nodeicov(), nodeocov()); the sensitivity run without the two most active lawyers (nwdrop); stored results from Chapter 30.
  • Exercises: Ch. 29, Ex. 2 (odds ratios of the ERGM, and why QAP and ERGM are not two confirmations); Ch. 29, Ex. 3 (the mixing model and the status-match model are the same model); Ch. 29, Ex. 6 (a sensitivity check of your own choice); Ch. 29, Ex. 9 (deference or competence); Ch. 30, Ex. 5 (a do-file that collects results without retyping them).

C. Workshop or summer school (5 days)

Each day has a deck of 60 to 75 slides, with speaker notes, a hands-on slide for each block and the book’s exercises; the decks are PowerPoint files with PDFs, and instructors are free to use, adapt and share them.

For participants who know Stata and want to apply network analysis to their own data. The first three days cover the common ground and the main descriptive tools, the fourth the inference for one network, and the fifth the first model, with a choice for the afternoon. A morning can take one chapter as a lecture and a lab; an afternoon the next.

Day 1. Networks as data

  • Slides: PowerPoint · PDF
  • Chapters: 1 Why networks?; 2 What is a network?; 3 From a research question to a network and a method (the reporting checklist); 4 Working with networks in Stata; 5 Seeing networks
  • Core idea: a relational question needs data on pairs, and every construction decision changes the network. A network is declared once in Stata and found by name. Node positions in a picture are not data.
  • Demonstration: nwwebuse lazega, nwclear, nwsummarize lazega_*, nwsym and nwdyads; the survey example from reshape long to nwfromedge; nwplot flomarriage in three layouts; nwplot lazega_adv, color(office); an interactive page.
  • Exercises: Ch. 1, Ex. 5 (cause, outcome or context); Ch. 2, Ex. 3 (density by hand before and after symmetrizing); Ch. 4, Ex. 4 (wildcards, nwsave and nwuse); Ch. 5, Ex. 1 (one network, three layouts).

Day 2. Describing networks and positions in them

  • Slides: PowerPoint · PDF
  • Chapters: 6 Describing networks; 7 Paths, distance, and reachability; 8 Centrality, brokerage, and structural holes
  • Core idea: each statistic describes one named relation in one network and is read against what chance and group sizes would produce. Reachability comes before distance. A centrality measure follows a mechanism.
  • Demonstration: nwdyads, nwtriads and nwmixing on lazega_adv; nwgeodesic and nwpath; nwdegree, nwcloseness, nwbetween, nwburt and nwbrokerage, group(officenum).
  • Exercises: Ch. 6, Ex. 7 (gender and the expected E-I index); Ch. 7, Ex. 3 (distances among the Glasgow girls); Ch. 8, Ex. 4 (closeness under two penalties); Ch. 8, Ex. 6 (sought or broker?).

Day 3. Groups, positions and ego networks

  • Slides: PowerPoint · PDF
  • Chapters: 9 Cohesion, subgroups, and communities; 10 Positions, roles, and structural equivalence; 11 Ego networks. Participants with affiliation or valued data can take Chapter 12 or 13 in place of 10 or 11.
  • Core idea: a cohesive group, a community, a position and an ego network answer different questions about the same ties. A partition proposed by an algorithm is not by itself a social group.
  • Demonstration: nwkcore, nwkcomponents and nwcommunity, algorithm(louvain) with nwmodularity; nwconcor and nwhierarchy with tab against office and status; nwego, nwegoplot and nwaltergen.
  • Exercises: Ch. 9, Ex. 4 (communities of the co-work network); Ch. 10, Ex. 4 (Ward and CONCOR as two-group splits); Ch. 11, Ex. 4 (ego density against constraint).

Day 4. Inference for one network

  • Slides: PowerPoint · PDF
  • Chapters: 15 Random and structured network generators (the sections on density and random graphs); 18 Random networks and null models; 19 Permutation inference and comparing networks; 20 QAP regression
  • Core idea: dyads are not independent observations. A conditional uniform graph test compares a statistic with random networks; permutation relabels the observed one; QAP regression adds several predictors. Each rests on its own assumption about what may be exchanged.
  • Demonstration: nwrandom; nwcug lazega_adv with condition(density) and condition(census); nwcorrelate lazega_adv lazega_fr, permutations(1000); nwqap lazega_adv lazega_fr office practice status, permutations(500) plot.
  • Exercises: Ch. 18, Ex. 3 (the friendship network’s null mean); Ch. 19, Ex. 4 (advice and co-work); Ch. 20, Ex. 5 (the advice model without friendship); Ch. 20, Ex. 7 (homophily or shared office).

Day 5. A first network model

  • Slides: morning (ERGM): PowerPoint · PDF; afternoon, panel option (Chapters 14 and 25): PowerPoint · PDF; afternoon, capstone option (Chapter 29): PowerPoint · PDF
  • Chapters: morning: 21 Exponential random graph models. Afternoon, one of two: 14 Networks over time and 25 Stochastic actor-oriented models, for participants with panel data; or 29 A complete network analysis, which brings the week together on the law firm.
  • Core idea: an ERGM makes the dependence among ties the object of the model; its coefficients are read as conditional log-odds through change statistics, and its fit is checked on statistics it does not contain. An SAOM models change between waves as choices of actors.
  • Demonstration: the staged advice model with nwergm, estat mcmcdiag and estat gof; for the panel option, nwturnover, nwsaom describe and nwsaom on the Glasgow waves; for the capstone option, the QAP and ERGM models of Chapter 29.
  • Exercises: Ch. 21, Ex. 1 (what the office coefficient says about a pair); Ch. 21, Ex. 3 (homophily on gender); then Ch. 25, Ex. 3 (homophily in sport) for the panel option, or Ch. 29, Ex. 4 (a step or a slope: seniority in odds ratios) for the capstone option.
  • Optional: participants whose outcome is an attribute of the actors (smoking, adoption, a rating) and who have a single observation of the network can read Chapter 24, Section 24.4 on nwalaam, which is the one-wave counterpart of the panel option; it needs no further preparation than the morning’s ERGM.

Permutation tests and model fits repeat a computation many times. Appendix A of the book advises setting reps() or permutations() deliberately; in a lab with many students, a smaller number of repetitions than the book’s gives the same lesson in less time.


Back to top

nwcommands is free software under the GNU General Public License, version 3, as are the model packages nwsaom, nwergm, nwdynam, nwrem, nwalaam, nwlsm and nwsbm. Citing and licenses · Get help · GitHub. Issues with the core go to the nwcommands repository; issues with a model package go to that package's own repository.

This site uses Just the Docs, a documentation theme for Jekyll.