Framing War — Partisan Identity & Support for Military Intervention

// 01 — the question

Classical realism treats public opinion on war as a sideshow to strategic interest. A large body of political-science research disagrees: in democracies, public sentiment constrains foreign policy, and that sentiment is shaped by partisan identity, elite discourse, and framing. The literature is unsettled on a key point — whether humanitarian or national-security justifications win broader support, and whether that depends on which party's leaders do the framing. I designed an experiment to isolate these forces causally, rather than infer them from observational polling that confounds framing with real-world events.

// 02 — design

Respondents completed a pre-treatment battery (baseline attitudes toward force, diplomacy-vs-coercion preferences, political knowledge, party ID), then were randomly assigned to one of five vignettes about a hypothetical crisis in "Country A." The design crosses the partisan source of the endorsement with the rationale, against a neutral control:

• Primary outcome — 5-point support for intervention (ordinal).
• Secondary outcome — the single concern driving the stance: civilian casualties, long-term entanglement, regional stability, or international norms.

// 03 — the instrument: qualtrics engineering

The experiment lives or dies in the survey build. This one ran inside a multi-study omnibus, which means the assignment machinery had to be airtight: treatment is assigned by a block randomizer over embedded-data setters (not over blocks), so the assignment is written into the response record before any branch fires — the same field the analysis pipeline keys on.

• Panel integration — respondent demographics (rid, age, gender, hhi, …) arrive as URL/panel embedded data, merged into the export.
• Consent gate — a branch on the consent choice routes decliners to an EndSurvey terminal before any content.
• Randomization — BlockRandomizer(SubSet=1, Evenly Present) draws one of five ali_treat values; five downstream branches on the embedded field route to the matching vignette block.
• Geo capture — country = ${loc://CountryName} feeds the US-residency screen.
• Quality gates — image attention check, honesty pledge, and Q_RecaptchaScore; the R pipeline requires 2-of-4 to pass.
• Post probe — per-arm follow-ups with display logic on ali_treat, asking which part of the vignette moved the respondent.

// 04 — method & modeling

The primary outcome is ordered, so a linear model would invent distances between "somewhat" and "extremely" likely. I used an ordinal proportional-odds logit, which posits a latent support index sliced by ordered cutpoints:

Rather than call a package, I coded the censored log-likelihood and maximized it with Newton–Raphson and a backtracking line search, parameterizing cutpoints as positive increments so they stay ordered. Two specifications — additive (treatment + party ID) and interaction (treatment × party ID) — adjudicated with a likelihood-ratio test. The secondary outcome uses a multinomial logit over the four concern categories.

# Proportional-odds ordinal logit, fit from scratch (no MASS::polr)
def negll(theta):
    beta, thetas = unpack(theta)        # thetas kept ordered: θj = θj-1 + e^d
    cum = sigmoid(thetas - X @ beta)    # Pr(Y <= j)
    p   = diff_to_category_probs(cum)   # Pr(Y = k)
    return -sum(log(p[arange(n), y-1]))
# maximized by Newton-Raphson + backtracking line search; SEs from the
# observed information. Additive vs. treatment×party compared by an LR test.

// 05 — estimates

Odds ratios from the additive model (reference = control condition, Democratic respondents). Values above one mean higher odds of greater support.

// 06 — the partisan interaction

The averages hide who is moving. Adding the treatment × party interaction improves fit decisively — the effect of a cue depends on the respondent's own identity:

Co-partisan appeals are the reliable channel: a Democratic humanitarian cue lifts Democrats from ~38% to ~50% predicted support; the same message from a Republican source moves them barely at all — and vice versa.

// 07 — interactive: predict a respondent

The widget runs the fitted interaction model in your browser. Choose an identity and the cue they hear; the predicted 5-point support distribution recomputes live.

// 08 — what drives opposition

A multinomial model of each respondent's single greatest concern shows the mechanism behind the support numbers — and how framing reshuffles what people worry about.

// 09 — why it matters

Substantively, the result speaks to how democracies are talked into war: legitimacy is manufactured less by the objective stakes than by the alignment of moral framing with partisan identity.

Methodologically, the project runs the full experimental stack end-to-end: instrument engineering in Qualtrics (randomization, branching, embedded-data pipelines, quality gates), pre-registration, and hand-built estimators for ordered and categorical outcomes — the same toolkit that underpins rigorous A/B testing and product experimentation.