CSSS Notes

_csss_uw_2024.md

Center for Statistics & the Social Sciences (CSSS) 25th Anniversary

• Text as Data Short Course
• Advances in Social Networks Analysis
• Bias, Fairness and Inequality in an Algorithmic Age

bias_fairness.md

Bias, Fairness and Inequality in an Algorithmic Age

Chair: Sasha Johfre (UW) Moderator: Kosuke Imai (Harvard)

Algorithmic Reinterpretations: College Rankings and Socioeconomic Self-Shorting

James Chu (Columbia)

• ways that people interpret metrics may drive self-selection
• problems:
  • cannot capture all that is valuable
  • exacerbate racial, gender and class inequalities
  • school rankings: do not adequately capture career and social mobility
  • recidivism: do not capture historic injustices
• if algorithms reduce dimensions to a metric, people "decode" or reinterpret these metrics back to the input domain (and importantly, are biased in how they decode)
• Algorithmic reinterpretation = inconsistent decoding
• decoding of a metric is not 1:1
• hypothesis: interpretation of college ranking metrics is clustered by socioeconomic status (SES)
• impressions findings (n=2k):
  • high SES: interpret high score for exclusivity, rigor, safety, stress
  • low SES: interpret high score as signal for exclusion
• differences in adolescent perceptions of college tuition price

Advancing Algorithmic Fairness: A Statistical Learning Approach with Causal Constraints

Razieh Nabi (Emory)

• legal framing: "would the employer have taken the same action if the employee had a different race and everything else was the same?" maps to causal DAG
• is unfairness always about the direct effect of X on Y? definitions should be context specific
• mitigation approaches:
  • pre-process the observed data
  • post-process the statistical output
  • re-train subject to fairness constraints (constrained optimization)
• intuition: lagrange multiplier indexes a constraint-specific path encoding constrained model space where influence of the sensitive attribute is 0
• observation: formalizations of fairness only codify legal and political desiderata (no substitute for discourse, debate & policymaking)

Fair inference in multilevel data analysis

Peter Hoff (Duke)

• goal: infer group-specific parameters from group-specific samples
• hierarchical modeling enable data-sharing between groups (perhaps small sample size within group) but are biased & lack group-specific error control
• Smaller p-values via indirect information FAB Inference Slides linear FAB R package

social_networks.md

Advances in Social Network Analysis

Chair: Tyler McCormick (UW) Moderator: Yuan Hsiao (UW)

Social Networks & Health

Weihua An (Emory)

• Combine peer-reports and self-reports for measurement of sensitive behaviors
  • Smoking, bullying etc.
  • You said, they said: A framework on informant accuracy with application to studying self-reports and peer-reports
• Weighting peer-reports by peer network centrality, correct for biased self-reports
• Challenges to social contagion modeling
• Social network-based smoking prevention intervention

Mosuo social networks do not support universalist theories of gendered social relationships

Siobhan Mattison (U New Mexico)

• Gender believed in the literature to constrain activities & social relationships

• Evolutionary theorists link constraints to gendered differences in payoffs from childcare vs politicking & securing reproductive partners

• “WEIRD” societies: Western Educated Industrialized Rich & Democratic

  • Non-“naturalistic” societies
  • Academic framing of WEIRD societies universalize their cultural norms about gender through the literature

• Setting: Ethnic Mosuo of China

  • Matrilineal & patrilineal subgroups
  • Lots in common (language, art, etc.) but gendered kinship is different (monogomy vs “walking” marriage, paternal vs maternal authority & inheritance)
  • Stat. sig. impacts (reversals) in health, social graph sparsity between men & women in matrilineal & patrilineal groups
  • Conjecture: differences in local ecology motivates differences in gendered social relation? (Kind of a “nothing burger”)

• Universalism is boring. Essentialist thinking about gender is not doing the work that scientifically meaningful categories should do. Need to be more specific to social/cultural context in describing paths from sex to gender

Population-Level Balance in Signed Networks

Beijing Tang (CMU)

• Signed networks: positive & negative degree (eg “like” or “dislike” edges)
• Collaboration & competition in trade networks
• Presence of signed edges introduce problems / change each other (“the enemy of my enemy is my friend” / “the friend of my friend is my friend”)
• Empirical evidence for balance theory in social networks, protein interactions & ecology
• A signed network is balanced at the population level if for all triples of edges, the expected product of signs is positive
• Authors propose an estimator in this paradigm and generative model satisfying balanced sign constraints
• Correlates of war dataset (CoW)

Measuring and understanding the dynamics of populations of scholars

Emilio Zagheni (Max Planck)

• measuring migration of scholars based on info on institutional affiliation changes
• The demography of the peripatetic researcher: Evidence on highly mobile scholars from the Web of Science
• Scholarly Migration Database
• provide evidence of relationship between economic development & migration propensity
• migration trends & patterns
  • gender inequality
  • international mobility: expands networks & creates visibility
  • what might this mean for gender disparity? higher gender disparity among internationally mobile scholars vs non-mobile scholars (median .32 vs .47 respectively in '98-'02) also changing over time (.5 vs .54 in '13-'17)
  • impact of policies on migration: how did brexit affect migration of scholars?
    • p(entering UK | EU origin) no change
    • p(leaving UK | EU origin) increased
    • p(leaving UK | UK origin) decreased
    • p(entering UK | UK origin) increased
• internal migration observation: center of gravity within the US "moves west" between 2015-2020; reversal of historical trend
• survival probability (liklihood of continuing to publish): little variation between genders, strong differences between countries and entry year

text_as_data.md

Text as data short course

Benjamin Mako Hill: social scientist studying online communities using NLP

NLP: subfield at intersection of computing & human language Text as data: analyzing text to generate insight

• Text as Data: A New Framework for Machine Learning and the Social Sciences
• Data sources relevant to census: Govt and legal documents, annual reports, internal comms, transcripts, legislation, public records
• What can you do w/ text as data?
  • Discovery: EDA, concept & hypothesis discovery
  • Measurement: operationalizing a concept; measuring phenomena at scale
  • Causal inference: assess effect of counterfactual intervention
  • Text as treatment, outcome or confounder

Case studies

• Emotional manipulation content on Facebook: ~7k participants, sentiment analysis LIWC (Linguistic Inquiry & Word Count), attempt to measure impact on text produced by participants given emotional intervention
• Topic modeling for cultural sociology journalistic coverage of government funding for the arts
• Taboo topics on wikipedia: taboo article classifier; wikipedia’s social process of production is shaped by taboo

Representation:

it’s complicated (but mostly about reducing dimensionality)

• Tokenization: split strings up (lots of ways to do this)
• stop words: (commonly used words filtered out in preprocessing “the”, “is”, “at” etc, not always done, sometimes it maters)
• Stemming: reducing words to stem / root form ({changing, changed, change} —> “chang”)
• Lemmatization: converting word to meaningful base form given context
• Vectorization:
  • bag of words (simplest, ignores syntax / word order)
  • N-grams: capture context and word order
• Representing many documents
  • Term/document matrix: term frequency vectors of multiple documents (often big & sparse)
  • Term frequency inverse document frequency (TF-IDF) relevance of word to a document, compared to other documents
  • Further reducing dimensionality: principal component analysis / singular value decomposition (often work well after TF-IDF weighting)
• Word embeddings
  • “You shall know a word by the company it keeps” - firth 1957
  • Represent a word based on words around it in small context window
  • Examples: word2vec (predict word from context or context from word) & GloVe (dim reduction of co-occurence counts within window)
  • Intuition: words that are semantically similar (roughly) map to similar parts of feature space (be careful w/ this, manifolds are messy)
  • Vector “arithmetic”: “king” - “man” = “queen”
• Transformers:
  • predict words in a sequence, trained on large corpora (that's it)
  • Popular proprietary models create reproducibility problems for social sciences
  • Lack of transparency in training data (also bad)

Tools

• Python: NLTK, spaCy
• R: tm, textTinyR

Analysis

• Dictionary methods: LIWC, VADER, Harvard-IV
• unsupervised methods: grouping like things into partitions (clustering, topic models)
• supervised methods: use a small sample of documents to train, make predictions / classifications of other documents