Autologistic Actor Attribute Model

A conditional exponential-family model of binary actor outcomes on an observed social network, with statistics that can express outcome dependence along ties.

Version

v1 · 2026-10-31 · History

Domain-specific #

14465

Domain group

Social Sciences

Origin domain

Sociology & Anthropology

Subdomains

Social Network Analysis, Statistical Network Models → Sociology & Anthropology

Aliases

ALAAM, Autologistic Actor Attribute Models

Core Idea

An autologistic actor attribute model (ALAAM) fixes an observed network of actors and assigns a joint probability to possible patterns of a binary actor outcome on that network.^[1][2] It is a family of conditional exponential-family models: weighted statistics can represent baseline outcome prevalence, actor covariates, and outcome patterns along ties.^[2] A particular fitted specification can omit a tie-dependence term as a nested comparison without ceasing to belong to the family.^[2]

The direction of modeling matters. An ordinary exponential random graph model treats social ties as the random outcome; here the ties are held fixed and actor outcomes vary.^[1][3] ALAAM coefficients describe associations under the specified network and covariates.^[3] A cross-sectional fit, by itself, does not establish that a neighbor caused an actor to change behavior.^[3]

The same formal move appears in unlike data settings. Stivala, Wang and Lomi fit a model of smoking among an illustrative group of 50 girls using friendship ties, then fit models of liking jazz among Deezer users in three national online networks.^[3] The outcomes, network sizes and fitted terms differ; the conditional modeling relation remains.

How would you explain it like I'm…

Friend Map Patterns

Imagine a map that shows who is friends with whom in a class. Some kids like jazz and some don't. This tool keeps the friend map the same and figures out how likely different patterns of jazz-liking are, like whether friends tend to match. Finding that friends match doesn't prove one friend made the other like jazz.

Fixed Friendships, Changing Traits

Scientists who study groups of people sometimes draw a map of who is connected to whom, like friendships. An Autologistic Actor Attribute Model keeps that map fixed and looks at a yes-or-no trait each person has, like smoking or liking jazz. It works out how likely different patterns of that trait are across the whole map, using things like how common the trait is, facts about each person, and whether connected people tend to share it. It can show that friends share a trait, but by itself it can't prove one friend caused the other to change.

Fixed-Network Trait Model

An Autologistic Actor Attribute Model (ALAAM) is a statistical model for social networks. It takes the network of ties as fixed and treats a yes/no outcome for each person (say, whether they smoke) as the random thing to be explained. Contrast this with an exponential random graph model, which does the reverse: it treats the ties themselves as random. An ALAAM can include terms for how common the outcome is overall, for individual characteristics, and for whether connected people share the outcome. Its results describe associations given the network and the measured characteristics. A one-time snapshot can't show that a friend caused someone to change behavior. The same method has been used for very different data, from smoking among a small group of girls to liking jazz in large online networks.

 

An autologistic actor attribute model (ALAAM) holds an observed network fixed and assigns a joint probability distribution over possible configurations of a binary actor-level outcome on that network. It is a family of conditional exponential-family models in which weighted statistics capture baseline outcome prevalence, effects of actor covariates, and patterns of the outcome along network ties (e.g., contagion-like configurations). A specific model can drop the tie-dependence term as a nested comparison while still belonging to the family. The key contrast is with exponential random graph models, where ties are the random variable; in ALAAMs the ties are given and actor outcomes vary. Coefficients describe associations conditional on the specified network and covariates. A cross-sectional fit alone does not establish that a neighbor's outcome caused an actor's outcome. The same formal structure has been applied across very different settings, such as smoking among 50 girls in a friendship network and jazz preference among Deezer users in national online networks.

Structural Signature

Sig role-phrases: observed actors — fixed social ties — binary actor outcome — available configuration statistics — parameterized conditional law.

• Observed actors. A population supplies the units whose outcome states are recorded, such as pupils or music-service users.^[3] Without actor-indexed observations there is no attribute vector to model.

• Fixed social ties. The observed network supplies who is connected to whom and which outcome configurations the model can count.^[1][2] If the ties themselves become the dependent variables, the model has crossed into a different network-modeling task.

• Binary actor outcome. Each actor has an observed 0/1 state for the modeled attribute; other actor characteristics may enter as covariates.^[2][3] The inspected standard form specifies a binary dependent outcome, not an ordinal or continuous one.^[1][2][3]

• Available configuration statistics. The family can count patterns that join outcomes, ties and covariates, such as connected pairs sharing the outcome.^[2][3] A null or nested fitted specification may omit a particular dependence term.^[2][3]

• Parameterized conditional law. Coefficients weight the selected statistics, and a normalizing factor turns their exponentiated sum into a probability distribution over possible binary outcome vectors, conditional on the observed network.^[2][3]

What It Is Not

• A model of tie formation. A standard ERGM makes ties endogenous; the ALAAM keeps the network observed and conditions on it.^[1][3]
• Proof of contagion. A positive shared-outcome tie statistic is a model-conditional association; prior similarity, common exposure and omitted covariates can complicate a causal reading.^[3]
• One software package or estimation algorithm. MPNet, ALAAMEE and Bayesian implementations are ways to fit related models; their availability, diagnostics and data handling are implementation claims, not the model's definition.^[3]
• Every conditional statistical model. Conditioning on a graph alone does not fix the binary actor-outcome space or the ALAAM statistics. An ordinary regression with independent actor observations lacks the modeled network dependence, although it can serve as a nested null case.^[2]

Scope of Application

ALAAMs apply when investigators observe a social network and an actor-level binary outcome and want to evaluate network-dependent outcome configurations while accounting for selected actor covariates.^[2][3] The model may be fitted to directed, undirected or larger network data under implementation-specific terms and algorithms; those variants require their own assumptions and diagnostics.^[3]

The bounded scope here is cross-sectional conditional modeling of a binary outcome. The frozen source suggests ordinal and continuous dependent-outcome versions, but the inspected foundational and worked sources specify binary outcomes.^[1][2][3] The screened extension assertion is excluded from this bounded binary-outcome entry because the inspected definitions specify binary outcomes.^[1][2][3] A separately specified model would require its own primary review.

Clarity

To recognize an ALAAM, ask which variables are treated as observed context and which are assigned a probability law? If the adjacency matrix is fixed and the law ranges over possible binary outcome vectors, the task has the right orientation.^[2] If the law ranges over possible tie matrices, an ERGM is a closer description.^[2]

A second question is which configurations does this specification include? A connected pair sharing the outcome can be a statistic, but a null fit may set its coefficient to zero or omit it.^[2] A missing contagion coefficient in one nested fit therefore does not reclassify the model family.^[2]

Manages Complexity

The model separates three sources of structure that can otherwise be conflated: overall outcome prevalence, actor covariates and dependence associated with network configurations.^[3] By specifying statistics and estimating their weights jointly, an analyst can compare a tie-outcome pattern with the baseline and other modeled features rather than counting matching neighbors alone. The interpretation remains conditional on the chosen terms and observed network.

This compression has a cost. As the model includes richer network-outcome configurations, estimation and goodness-of-fit checks become more demanding. Stivala, Wang and Lomi changed Deezer model terms to achieve convergence; the same feature set did not fit all three national networks.^[3]

Abstract Reasoning

The model's reasoning step is to hold the ties fixed while varying possible 0/1 assignments to actors.^[2][3] A statistic summarizes each assignment's alignment with ties and covariates; its coefficient changes the relative weight of assignments in the conditional law.^[2][3] This makes the role of a connected-pair outcome pattern explicit without declaring the pattern to be a transmission event.

The model can test competing specified patterns. A null model with no tie dependence and an alternative with a tie-outcome statistic ask different questions about the same observed network.^[2] The result is evidence about fit and association under those assumptions.^[3] It is not a temporal account of how the observed actor states arose.^[2][3]

Knowledge Transfer

Within social-network analysis, the same conditional orientation transfers from a school friendship graph with smoking as an outcome to an online music graph with genre preference as an outcome.^[3] Actors, ties, outcomes, covariates and configuration terms must be redefined for each dataset. A smoking association cannot be copied as a jazz-preference coefficient.

The transferable statistical skeleton is an exponential family for configurations conditional on observed structure. The named ALAAM remains domain-specific because actor-indexed social ties, binary actor outcomes and network-attribute configurations do the recognition work. Similar equations in other fields do not by themselves show that practitioners use this same named model there.

Examples

Canonical: friendship ties and smoking

In the ALAAMEE paper, a public excerpt of 50 girls from the Teenage Friends and Lifestyle Study provides a small-network illustration.^[3] The authors model smoking as a binary actor outcome on friendship ties and include terms for network-outcome patterns and actor characteristics.^[3] Their fitted example reports positive smoking co-occurrence along ties and an alcohol-consumption association; it also performs a goodness-of-fit check.^[3] The excerpt is illustrative, and the cross-sectional association does not by itself establish peer transmission.^[3]

Mapped back: the pupils are actors, friendship nominations are observed ties, smoking is the binary outcome, connected smokers are a possible configuration, and fitted weights compare that configuration with other included terms.^[3]

Applied/practice: music preferences in online networks

The same authors fit ALAAMs to Deezer friendship networks in Croatia, Hungary and Romania with liking jazz as the binary outcome.^[3] The specifications account for network configurations and numbers of genres liked.^[3] The contagion term is positive and statistically significant in Hungary and Romania, while the Croatia estimate is not significant; convergence required removing triangle terms for Croatia and Hungary.^[3]

Mapped back: Deezer users are actors, online friendship links are fixed ties, liking jazz is the modeled 0/1 state, and network/covariate statistics are weighted in a conditional law.^[3] The country-specific fits show reuse of the model family alongside limits on transferring a fitted effect.^[3]

Structural Tensions

T1: Fixed network vs. coevolving ties. Conditioning on observed ties isolates actor-outcome variation for a defined task, but it does not model friendship formation or change. Diagnostic: Could the outcome itself have influenced who became tied?^[1][3]

T2: Rich configurations vs. stable estimation. Additional patterns can express competing network hypotheses, while highly parameterized specifications may be harder to fit and diagnose. Deezer term removals make that tension concrete. Diagnostic: Which term can be removed without losing the question being tested?^[3]

T3: Association vs. causal influence. A tie-outcome coefficient can distinguish one modeled association from included covariates, but a single cross-section cannot establish the temporal direction of influence. Diagnostic: What independent design or temporal evidence would support a causal claim?^[3]

T4: Comparable model family vs. local specification. The same conditional form supports school and online data, but outcome coding, network boundaries, covariates and convergence differ. Diagnostic: Are two coefficients answering the same specified question before comparing them?^[3]

T5: Statistical reduction vs. social interpretation. A compact probability law makes a complex actor network analyzable, while its coefficients remain tied to how people, relations and outcomes were measured. Diagnostic: Has the substantive network construction been made explicit enough to interpret the model terms?^[2]

Structural–Framed Character

The identity leans structural within statistical social-network analysis. Once the observed graph, actor outcome space, statistic set and parameters are specified, the conditional law assigns probabilities by a formal rule.^[2] The same rule form is applied in the paper's school friendship and online music-network studies.^[3] A fitted dependence coefficient reports an association under model assumptions, not a demonstrated peer influence.^[3] On that basis I assign the entry low evaluative weight: the model itself does not label a social outcome good or bad or certify an intervention's success.

Its human-practice dependence lies in drawing the actor boundary, recording ties, coding the binary outcome and choosing model terms.^[2][3] Those decisions shape what a coefficient means; the law's calculation is then formal. Its institutional origin is statistical social-network research, not a single software vendor or a mandatory organization type.^[1][2][3] The source-defined ALAAM treats actor ties as observed and assigns a conditional law to binary actor outcomes.^[1][2][3] I therefore judge that its vocabulary travels literally across actor-network studies with binary outcomes, while a merely similar exponential equation outside that setting does not inherit the name. For import versus recognition, a new case qualifies by identifying fixed actor ties, a binary actor-outcome vector and the relevant conditional model family, rather than by borrowing the term for any spread-like pattern.

The thin portable skeleton is statistical representation of dependent data conditional on observed structure; Statistical Model is its independently reviewed strict genus. A substrate-neutral version beyond actor networks is a future-Prime question requiring independent instances and nonduplication proof. Its character: a formally defined, evaluatively neutral model family whose recognizability is strong inside social-network analysis and whose named scope remains domain-bound.

Structural Core vs. Domain Accent

The skeletal relation is a restricted family of probability laws for binary configurations conditional on observed structure, with parameters weighting available statistics.^[2] That family-of-laws structure is expressed by the live Statistical Model entry and supports the reviewed subsumption edge. Conditional Probability explains part of the notation, but its whole Prime signature is not established as the ALAAM's direct parent by writing \(P(Y\mid X)\).

The domain-bound mechanism is actor-network specific: ties link social actors, the outcome is recorded per actor, and connected-actor configurations can enter the statistic set. Removing that setting leaves a general conditional statistical model but not this named family. The inspected examples supply literal reuse among different social networks, not evidence that researchers call every analogous physical or biological conditional model an ALAAM. The possible substrate-neutral remainder is marked as a future-Prime question, with a required comparison to existing Representation, Statistical Model and Conditional Probability identities before any broader admission.

Instantiates / Related Primes

This entry presupposes Social networks and is a kind of Statistical Model.

An ALAAM supplies a restricted family of probability laws for a declared outcome space conditional on observed network data.^[2] The live Statistical Model entry is its reviewed strict genus. The model also presupposes Social Networks as its actor-and-tie carrier; a social network is not itself the probability-model family. Conditional Probability is relevant to the notation (P(Y=y\mid X=x)), but the notation alone does not prove that Prime's full re-normalization signature is a direct DAG parent.

Contagion names one phenomenon an ALAAM can investigate, not a guaranteed inference from fitting it. Random Graph concerns stochastic graph structure, which this bounded model holds fixed. These distinctions prevent a shared word or equation component from substituting for a typed parent proof.

Relationships to Other Abstractions

Current abstraction Autologistic Actor Attribute Model Domain-specific

Parents (2) — outgoing typed relations

• Autologistic Actor Attribute Model is a kind of Statistical Model Domain-specific

  A conditional exponential-family statistical model for binary actor outcomes on a fixed social network.

• Autologistic Actor Attribute Model presupposes Social networks Domain-specific

  The model requires a social network of actors and ties as its fixed structural carrier.

Hierarchy paths (5) — routes to 4 parentless roots

• Autologistic Actor Attribute Model → is a kind of Statistical Model → is a kind of Representation → presupposes Abstraction

Show alternative paths (4)

Neighborhood in Abstraction Space

Autologistic Actor Attribute Model sits in a sparse region of the domain-specific corpus (82^nd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2835 abstractions)

Nearest neighbors

• Generalized blockmodeling of binary networks — 0.82
• Information Entropy — 0.82
• Ethnographic process mapping — 0.81
• Machine-Learning Model — 0.81
• Latent Dirichlet Allocation — 0.81

Computed from structural-signature embeddings · 2026-10-31

Not to Be Confused With

• ERGM or network-selection model: models edges as the random object rather than a binary actor outcome conditional on edges.^[1][3]

• Network autocorrelation model: can also analyze linked outcomes, but the cited ALAAM papers emphasize configuration statistics for multiple network-outcome patterns.^[3]

• Social contagion itself: an empirical process would require evidence beyond a model-conditional fitted association.^[3]

• ALAAMEE or MPNet: software implements estimation; package capabilities can change without changing the model-family identity.^[3]

• A continuous-outcome extension: the original screened article mentions one, but the inspected standard definitions and worked applications use binary endogenous outcomes.^[1][2] The screened extension assertion is excluded from this bounded model because the inspected definitions specify binary endogenous outcomes; any separately specified extension requires its own primary review.^[1][2]

References

[1] Galina Daraganova and Garry Robins, “Autologistic Actor Attribute Models,” in Exponential Random Graph Models for Social Networks: Theory, Methods, and Applications (Cambridge University Press, online 2013), chapter 9, pp. 102–114, publisher chapter PDF. The accessible chapter summary and opening section support fixed ties, binary outcomes and the cross-sectional model orientation; other chapter claims require direct page inspection. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m

[2] Andrew Parker, Francesca Pallotti and Alessandro Lomi, “New Network Models for the Analysis of Social Contagion in Organizations: An Introduction to Autologistic Actor Attribute Models,” Organizational Research Methods 25, no. 3 (2022): 513–540, publisher PDF, especially pp. 517–518, equation 1 and the nested comparison, and pp. 522–530 for the advice-network illustration. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z ↩27 ↩28 ↩29 ↩30

[3] Alex Stivala, Peng Wang and Alessandro Lomi, “ALAAMEE: Open-source software for fitting autologistic actor attribute models,” PLOS Complex Systems 1, no. 4 (2024): e0000021, publisher PDF, especially pp. 2–3, equation 1; pp. 18–20, Figs. 6–8; and pp. 20–23, Table 4. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z ↩27 ↩28 ↩29 ↩30 ↩31 ↩32 ↩33 ↩34 ↩35 ↩36 ↩37 ↩38 ↩39 ↩40 ↩41 ↩42 ↩43 ↩44 ↩45 ↩46 ↩47