Quality vs. Populism in Short-Video Political Communication: A Multimodal Study of TikTok

Abstract

The article examines how framing and actor identity structure attention in short-video politics using a country-level corpus from Ecuador. It assembles 4612 public TikTok videos from official accounts and politically salient hashtags, extracts multimodal text via automatic speech recognition and on-screen OCR, and constructs two continuous indices: a quality index (programmatic, efficacy-oriented content) and a populism index (antagonistic, people-versus-elite cues). Engagement is modeled as a fractional response (binomial GLM with logit link), with robustness checks using OLS on logit(ER) and Poisson counts with an offset for log(plays + 1). Models include affect (positive sentiment and anger), hour/day controls, and actor fixed effects (leader, creator, institution, party, and media). The indices display construct validity: quality aligns with positive/joyful tone and populism with anger. Net of controls, populism is positively and consistently associated with engagement across estimators; quality is small and often null or negative. Effects are heterogeneous: leaders gain under both frames, creators primarily under populism, and media modestly under populism, while institutions face penalties under both, and parties show limited returns. Monthly series reveal event-linked intensification of populism, and hashtag networks are modular, mapping onto institutional, partisan, and creator ecosystems. A design analysis identifies a non-populist pathway—benefit-first micro-explanations, concise captions, targeted hashtags, and joyful/efficacy affect—that raises engagement without antagonism. The study contributes a reproducible, open-source pipeline for survey-free, multimodal framing measurement and clarifies how persona × frame interactions and meso-level discursive structure jointly organize attention in short-video politics.

1. Introduction

Short-video platforms have redrawn the opportunity structure of political communication in Latin America by compressing messages into seconds, privileging persona-forward storytelling, and optimizing distribution through algorithmic feedback loops. In that environment, framing—the selection and salience of problem definitions, causal attributions, and remedies—remains the core mechanism by which political actors render issues legible and motivating to mass publics (Entman, 1993). A complementary strand of theory shows that populism, understood as a thin-centered ideology that pits a morally pure “people” against a corrupt “elite,” travels well across contexts precisely because of its formal simplicity and affective charge (Pérez-Curiel & Rivas-de-Roca, 2022; Yilmaz & Morieson, 2021). On platforms such as TikTok, those properties interact with design features—brevity, sound–image coupling, and reward structures based on social feedback—so that emotional signals can be reinforced and learned through likes, shares, and comments, amplifying expressive norms such as moral outrage (Brady et al., 2021; Cartes-Barroso et al., 2025). Recent scholarship has begun to trace these dynamics empirically in TikTok’s electoral uses and performative affordances in the region (e.g., Peru’s presidential race and the “politainment” turn: (Cervi et al., 2023); early computational analyses of political TikTok: (Medina Serrano et al., 2020)), while newer Ecuador-specific work documents leader-centric, crisis-framed storytelling in presidential feeds during 2024 (Cervi et al., 2023).

Against that backdrop, the article addresses a substantive and methodological gap. Substantively, it focuses on Ecuador, where 2024–2025 politics were punctuated by public-order crises, a security referendum, and a presidential cycle, conditions under which the payoffs to antagonistic framing plausibly rise and institutional voice risks marginalization in attention markets. Methodologically, most evidence on TikTok politics still relies on manual samples, single-modality text, or campaign-specific case studies; less is known about how multimodal signals (caption, on-screen text, and speech) map onto frame indices at scale and how those indices relate to engagement across heterogeneous actors using models suited to fractional outcomes (Papke & Wooldridge, 1996).

The study therefore assembles a country-level corpus of 4612 public TikTok videos from official accounts and politically salient hashtags in 2024–2025, extracts multimodal text (automatic speech recognition and on-screen OCR), and builds two continuous framing measures: a programmatic quality index (procedural, efficacy-oriented content) and a populism index (antagonistic/people-versus-elite cues). It then tests whether those indices capture substantively meaningful variation and whether they predict engagement once affect, timing, and actor type are held constant while tracing their temporal dynamics and discursive ecosystems. The approach answers four research questions that organize the contribution. First, do the indices exhibit construct validity in this corpus—i.e., coherent distributions by actor category and expected correlations with affect (quality aligning with positive/joy; populism with anger)—as theory would predict from framing and outrage-reinforcement accounts (Brady et al., 2021; Entman, 1993)? Second, are the indices associated with engagement net of controls, using fractional response models, appropriate for rates (Papke & Wooldridge, 1996)? Third, are effects heterogeneous by actor, given TikTok’s affordances for performative authenticity and persona-centric campaigning documented in the region (Cervi et al., 2023; Medina Serrano et al., 2020)? Fourth, how do frames evolve over time and within hashtag communities, and do event-linked windows of intensification emerge in Ecuador’s political calendar?

The justification is both theoretical and practical. Theoretically, specifying how quality and populism travel through short-video affordances in a non-hegemonic media system extends classic framing accounts into a multimodal, high-velocity setting (Wang & Suthers, 2022). Practically, identifying a non-populist path to engagement—if one exists—matters for institutional communication, watchdog journalism, and civic actors who must compete for attention without adopting antagonistic repertoires. By combining open-source extraction with scalable modeling, the article contributes a replicable template for survey-free, country-level analysis and a set of actionable implications for democratic communication on TikTok in Ecuador and comparable Latin American contexts (Angulo Moncayo et al., 2025; Cervi et al., 2023).

Theoretical Framework

This study conceives “framing” as the strategic selection and salience of elements in a message to promote particular problem definitions, causal attributions, moral evaluations, and remedies. The classic definition anchors the construct and justifies operationalizing frame content as measurable text features and tropes extracted from speech and on-screen text in short-video posts. In this sense, the paper follows the mainstream in political communication that treats frames as observable choices in emphasis rather than latent ideologies, enabling quantitative inference from large corpora (Entman, 1993; Jacobi et al., 2016; Mostafa et al., 2025).

The work separates two broad families of frames: programmatic/solution oriented (hereafter quality) and antagonistic/identity mobilizing (hereafter populism). The second follows the ideational tradition in which populism is a thin-centered ideology that opposes a virtuous “people” to a corrupt “elite,” often dismissing pluralism and claiming to channel the volonté Générale (Kidron & Ish-Shalom, 2025; Olivas Osuna, 2021). This approach, widely adopted in comparative research (Dolci & Melli, 2025; Mede, 2024; Tuğal, 2021), allows the study to capture populism as a style of communication manifested in lexical choices and narratives rather than as a fixed party family. It draws on Mudde’s formulation of the “populist zeitgeist” and subsequent elaborations on thin versus “host” ideologies that piggyback on substantive policy positions (Castanho Silva et al., 2023; Mudde, 2004).

Quality is not merely the absence of populism, so it needs to be conceptualized as a positive communicative regime oriented toward public information, policy substance, and problem-solving clarity. Concretely, quality captures issue- and policy-centered messages that foreground proposals, procedural guidance, and accountability/effectiveness claims rather than strategic conflict narratives. This operationalization parallels scholarship distinguishing issue frames from strategic/game frames in political communication (Lawrence, 2000; Dekavalla, 2018) and aligns with deliberative ideals that emphasize reason giving and mutual understanding as benchmarks of civic-oriented public discourse (Russmann, 2021). It also resonates with research on government communication via social media, which highlights the role of institutional accounts in delivering public service information, transparency, and actionable guidance to citizens (Criado & Villodre, 2022).

Two empirical regularities from the platform literature also ground the study’s hypotheses. First, moral–emotional and anger-laden language travels farther in social networks, increasing diffusion probabilities, what the “moral contagion” program documents across domains. Second, high-arousal emotions generally amplify online virality. If populist framings prime anger and moralized ingroup–outgroup boundaries, their diffusion advantage should materialize as higher engagement; conversely, quality frames should correlate with positive or constructive affect and need not benefit from the same amplification. These expectations inform the validation tests linking populism to anger and quality to positive tone (Brady et al., 2017, 2021).

Short-video affordances matter for how frames travel. TikTok’s design centers on a highly personalized “For You” feed, low friction for remix/mimesis, and a creator toolkit that privileges performance and audiovisual cues. The platform thus rewards “performative authenticity” and face-forward narration—traits long noted in the study of mediatized leadership—and pushes political communication toward vernacular genres (skits, duets, trends). This paper adopts the view that TikTok’s affordances shift the marginal returns of frames relative to text-dominant networks: visually legible conflict and identity cues are more discoverable, while dense programmatic content demands scaffolding to compete in the attention economy (Cervi et al., 2023; Kang & Lou, 2022).

Bringing these strands together, the paper treats engagement on TikTok as the outcome of an interaction between (i) message-level framing choices and (ii) platform-specific attention and recommendation dynamics. Framing theory provides the general mechanism—salience, interpretation, and implied remedies—while the two indices operationalize two distinct communicative regimes within that framework: populist framing as antagonistic, identity-mobilizing storytelling (people-versus-elite cues) and quality framing as informational, policy-oriented, problem-solving communication. TikTok’s affordances and vernacular genres shape how these frames are enacted (through face-forward narration, audiovisual cues, trends, and remix) and how they are rewarded because the For You feed translates user reactions into distributional signals. Accordingly, the same framing strategy may yield different engagement payoffs depending on the platform’s logic and the actor’s capacity to perform the frame in a native style, motivating our emphasis on actor heterogeneity and the bounded-rate modeling of engagement.

Internationally, the emerging state of the art on political TikTok shows rapid professionalization by candidates and parties and the migration of “politainment” formats into campaign communication (Cartes-Barroso et al., 2025; Moir, 2023). Early computational audits of political TikTok underline how personalized feeds privilege creator style performance and trend-based content (Li et al., 2025; Orbegozo-Terradillos et al., 2024; Van Remoortere et al., 2024); comparative work documents candidate uptake in multiple countries. This literature motivates distinguishing leaders, parties, institutions, media, and creators in Ecuador because platform-native performance advantages should accrue unevenly across actor types.

Regionally, scholarship in Latin America suggests that TikTok lowers entry barriers for nontraditional actors and rewards leader-centric storytelling (Morejón-Llamas et al., 2024; Quimis Arteaga, 2024; Wilches et al., 2024). Studies highlight how outsider or newly prominent candidates leverage the app’s vernacular to reach youth cohorts, as well as the increasing reliance on influencers as message brokers. These patterns suggest that populist styles may be especially effective when embodied by personalized accounts or creator collaborations, while formal institutional accounts face a disadvantage in the short-video attention market.

In Ecuador, a small but growing literature documents platform-native political communication. Work on the 2024 referendum campaign argues that government messaging used TikTok to project authority and mobilize support; recent peer-reviewed research examines discursive strategies on Instagram and TikTok by President Daniel Noboa, evidencing leadership-centered narratives in hybrid governmental–campaign communication (Angulo Moncayo et al., 2025). Complementary studies on earlier presidential contenders analyze TikTok’s influence on youth perceptions. These sources situate the present study’s empirical focus and justify the emphasis on actor labels in modeling engagement effects (Posligua Quinde & Ramírez Rodríguez, 2024; Quimis Arteaga, 2024).

Finally, the paper conceives “engagement” as a fractional outcome that captures interaction conditional on reach (Saqr & López-Pernas, 2021). Methodologically, modeling such proportions with a binomial GLM and logit link (fractional logit) is appropriate and consistent with econometric best practice for variables bounded in [0,1], a choice that supports the study’s claims about marginal effects of frames after controls. This connects the theoretical expectations above to estimable quantities on the platform.

The framework therefore expects (i) construct validity: quality should align with positive/constructive sentiment, while populism should align with anger/moralized affect; (ii) distributional differences: actor types positioned closer to creator logics (leaders, influencers, and some media) should show higher central tendencies for populist style and larger engagement elasticities to it than institutions and parties; and (iii) temporal dynamics: populist framings should intensify near political shocks or high-conflict windows, consistent with moral–emotional diffusion and platform affordances that favor conflict salience over technical detail. The remainder of the article tests these expectations on the Ecuadorian TikTok corpus assembled for this study.

2. Materials and Methods

This section details the end-to-end pipeline used to assemble the corpus, extract multimodal text, construct framing indices, and estimate the engagement models. All data come from public TikTok posts retrieved from official or public-facing accounts and institutional/partisan hashtags in Ecuador; no private content or surveys were used. The full code and configuration are replicable with open-source software (Python 3.13.5, Tesseract 5.3.4, faster-whisper, pandas, statsmodels, and networkx).

2.1. Corpus Construction and Time Window

The corpus combines two seed sets curated ex ante: (i) public accounts (institutions, parties, leaders/politicians, media outlets, and political creators) and (ii) political/institutional hashtags. Using an automated scroll-and-parse collection routine, we retrieved each video’s URL, TikTok ID, timestamp, caption, author handle, and interaction counters (plays, likes, comments, and shares). The scrape covers 2024–2025, with dense coverage from April 2024 through October 2025 (the span used in monthly panels). After deduplication by video ID, removal of malformed records, and dropping entries without visibility counters, the analytic dataset contains 4612 unique videos. Actor labels were merged from the curated seeds; the labeled subset used in heterogeneity models comprises 4588 videos across five actor types—creator (2453), leader (928), institution (499), party (404), and media (304)—with the remainder classified as other/unknown for overall models.

2.2. Multimodal Text Extraction (ASR and OCR)

Audio was transcribed using an open-source automatic speech recognition pipeline to obtain segment-level Spanish transcripts with timestamps. On-screen text was captured through frame sampling and optical character recognition using a standard OCR engine with Spanish language resources. The OCR stage produced 3823 non-empty text records, reflecting that not all clips contain stable text overlays. For each video, a message blob concatenates caption, ASR, and OCR text after normalization and de-duplication (e.g., lowercasing and removal of URLs/emojis).

2.3. Affect and Auxiliary Features

From the message blob, the pipeline computes affective scores—positive sentiment, anger, joy, fear, and sadness—using open-source Spanish/multilingual models (code in scripts/annotate_text.py). It also derives caption length (characters), hashtag count (number of “#” tokens), and binary flags for has_caption and has_ocr. All continuous features are standardized (z) within the analytic sample to ease interpretation in regression.

2.4. Framing Indices: Quality and Populism

Two continuous indices summarize the dominant frame of each video. The quality index targets programmatic/informational communication (problem/solution statements, procedural details, service delivery, and policy benefits). The populism index targets antagonistic/identity communication (people vs. corrupt elite, blame attribution, ingroup/outgroup signaling). Both indices are produced by supervised scoring over the message blob (pattern features and lexical/semantic projections; implementation in scripts/annotate_text.py) and then rescaled to z-scores.

Operationally, the scoring combines two complementary components. First, it uses interpretable lexical and pattern cues extracted from the multimodal message blob (caption + ASR + OCR), including procedural/policy terms and problem–solution structures for quality, and antagonistic “people-versus-elite” templates, blame attribution, and ingroup/outgroup signaling for populism. Second, it uses lexical/semantic projections to reduce sparsity and capture paraphrases that express the same framing logic beyond exact keyword matches. As brief illustrations, high-quality messages tend to include patterns such as “cómo solicitar/cómo acceder”, “plan/cronograma”, “medidas/reformas”, “beneficios”, “entrega/servicio”, and “solución/implementación”, whereas high-populism messages often include cues such as “los corruptos/los de siempre”, “mafias”, “nos roban”, “traidores”, and explicit “ellos vs. el pueblo” contrasts.

Construct validity is checked in two ways: (i) distributional variation across actor labels (wide dispersion and label shifts rather than ceiling effects) and (ii) affective correlations: quality aligns positively with positive sentiment/joy and negatively with anger, while populism shows the opposite profile.

2.5. Engagement Measures and Exposure

The main outcome is the engagement rate (ER),

$$\mathrm{E}{\mathrm{R}}_{\mathrm{i}} = {\displaystyle \frac{\mathrm{like}{\mathrm{s}}_{\mathrm{i}}+\mathrm{comment}{\mathrm{s}}_{\mathrm{i}}+\mathrm{share}{\mathrm{s}}_{\mathrm{i}}}{\mathrm{play}{\mathrm{s}}_{\mathrm{i}}+1}}$$

(1)

bounded in (0,1) by construction (the +1 prevents division by zero). Robustness analyses vary the outcome—likes + comments only, likes only, or counts—with and without top-tail winsorization.

2.6. Baseline Regression for Fractional Outcomes

Because ER is a fractional response, the baseline specification uses a GLM with binomial family and logit link, following Papke and Wooldridge (1996) for proportions:

$$logit\left({\mu }_i\right) = {\mathsf{\beta }}_0\hspace{0.25em}+\hspace{0.25em}{\mathsf{\beta }}_Q\hspace{0.17em}{Q}_i\hspace{0.25em}+\hspace{0.25em}{\mathsf{\beta }}_P\hspace{0.17em}{P}_i\hspace{0.25em}+\hspace{0.25em}{\mathsf{\gamma }}^{\mathrm{\top }{S}_i}\hspace{0.25em}+\hspace{0.25em}{\mathsf{\delta }}^{\mathrm{\top }{L}_i}\hspace{0.25em}+\hspace{0.25em}{\mathsf{\lambda }}_{h\left(i\right)}\hspace{0.25em}+\hspace{0.25em}{\mathsf{\omega }}_{d\left(i\right)},$$

(2)

where ${\mu }_i =\mathrm{E}\left[\mathrm{E}{\mathrm{R}}_{\mathrm{i}}\right|\cdot ],$ Q_i and P_i are the quality and populism z-scores, S_i is the vector of affect controls (positive sentiment, anger; optionally joy/fear in sensitivity checks), L_i encodes actor fixed effects (creator, leader, institution, party, and media), and λ_h(i) and ω_d(i) are hour-of-day and day-of-week fixed effects. HC3 robust standard errors are used throughout. This model is estimated on the full analytic set (N = 4612); the by-label heterogeneity version replaces β_Q, β_P with label-specific slopes by interacting Q_i and P_i with actor indicators.

To verify that results are not artifacts of the rate metric, the pipeline also models interaction counts. Negative-binomial variants probe over-dispersion; signs and relative magnitudes are reported and match the fractional results.

2.7. Temporal Aggregation and Event Alignment

For temporal analyses, the pipeline collapses to monthly panels overall and by actor label (monthly_overall.csv, monthly_by_label.csv) by averaging the quality and populism indices and counting videos per month. Trends are interpreted against Ecuador’s political calendar (e.g., 2024 security referendum; 2025 electoral cycle) to discuss windows of intensification.

2.8. Hashtag Ecosystems and Modularity

Captions are tokenized to extract hashtags, and then two graphs are built: a co-occurrence network (hashtags as nodes; weighted edges when two tags appear in the same video) and an author × hashtag bipartite network. Using Louvain community detection, the analysis reports modularity, community counts, and top tags per community and relates community composition to actor labels.

2.9. Coordination/Replication Screen

Potential coordination is screened by comparing pairs of videos across accounts using textual similarity on captions and OCR/ASR snippets (cosine over sentence embeddings) and near-contemporaneity thresholds. The output is a shortlist of high-similarity pairs with time deltas and URLs; the goal is descriptive (triangulation), not attributive.

2.10. Quality Control and Missing Data

All counters are retrieved from TikTok at the collection time; heavy-tail influence is addressed by winsorization in sensitivity checks. Videos with missing plays or corrupted counters are dropped before modeling. Both ASR and OCR coverage are reported descriptively; the engagement models do not condition on having OCR/ASR to avoid conditioning on measurement.

2.11. Ethical Considerations

Only public content and public-facing accounts are included. No face recognition, demographic inference, or personal data extraction is performed. The study focuses on aggregate patterns and hyperlinks only to already public posts.

2.12. Software and Reproducibility

All steps run on Windows (Anaconda env), with Tesseract 5.3.4 for OCR and faster-whisper for ASR. Scripts referenced above produce the intermediate CSV/JSONL files cited in the Results. The modeling is implemented in statsmodels with reproducible seeds where applicable; tables and figures are written to analysis/subfolders. A replication package includes seeds, intermediate files, and notebooks to regenerate all tables and plots from the 4612-video analytic dataset.

This methodological design—multimodal text extraction, weakly supervised framing indices, and fractional/offset GLMs with affect and schedule controls—aims to maximize construct validity, statistical identification for fractional outcomes, and practical reproducibility under computational constraints.

3. Results

3.1. Framing Measurement: Quality and Populism (Validity and Distribution)

Two video-level indices are analyzed: quality (informational/constructive content) and populism (people-centric/antagonistic cues), each standardized as z-scores. Before modeling engagement, the section tests whether the indices exhibit substantive variation and whether distributions shift across actor types.

3.1.1. Distributions by Label

Figure 1a,b display violin plots for quality and populism by actor type; Figure 1c,d show ridgeline densities for the most frequent labels.

Table 1. Descriptives (z-scores) by label.

Label	n	quality_mean	quality_sd	populism_mean	populism_sd
other	1544	−0.166	1.315	−0.028	1.330
leader	1146	0.165	0.494	0.144	0.600
creator	993	−0.260	0.960	0.059	1.025
institution	499	0.562	0.451	−0.333	0.297
party	406	0.111	0.891	−0.036	0.825
Overall	4588	−0.000	1.000	−0.000	1.000

Note. Means are z-scores (overall mean ≈ 0; sd ≈ 1 by construction), but between-label means depart from zero, validating the indices’ discriminatory power. Sample sizes reflect the auto-label mapping described above.

reports means, standard deviations, and sample sizes. Wide dispersion around zero in both indices indicates meaningful heterogeneity in framing intensity across videos (Figure 1a,b). Institutions center well above zero on quality (mean = 0.562) and below zero on populism (mean = −0.333), consistent with informational, programmatic messaging. Leaders tilt upward on both dimensions (quality = 0.165; populism = 0.144), evidencing a hybrid style that mixes constructive cues with mobilizing/antagonistic appeals. Creators skew lower on quality (−0.260) and slightly positive on populism (0.059), aligning with attention-seeking, identity-forward delivery. Parties sit near zero, slightly positive on quality (0.111) and neutral/slightly negative on populism (−0.036), suggesting a heterogeneous mix (see Table 1). Ridgeline panels (Figure 1c,d) show rightward mass for leaders on populism and for institutions on quality, highlighting actor-specific framing regimes rather than pure noise.

Despite an overall mean of zero by construction, between-group means differ meaningfully, and the shapes are wide, confirming that the measures capture real, interpretable variation in political communication styles on TikTok.

3.1.2. Validity Checks

If the indices capture recognizably different communicative styles, quality should align with a more constructive/positive tone, whereas populism should align with mobilizing anger (and possibly lower positivity). The section correlates both indices with affective cues available in the dataset (sent_pos, anger, joy, and fear).

Table 2. Pearson correlations among framing indices and affect (video level).

	quality_index	populism_rate	sent_pos	anger	joy	fear
quality_index	1.00	−0.21	0.71	−0.89	0.50	−0.02
populism_rate	−0.21	1.00	−0.11	0.34	−0.06	−0.04
sent_pos	0.71	−0.11	1.00	−0.43	0.73	0.01
anger	−0.89	0.34	−0.43	1.00	−0.30	0.01
joy	0.50	−0.06	0.73	−0.30	1.00	−0.05
fear	−0.02	−0.04	0.01	0.01	−0.05	1.00

Note. Entries are Pearson r. All key pairs have N = 4588 complete cases. The 95% CIs for the dot plot use the Fisher z approximation.

and Figure 2 show patterns that match theoretical expectations:

• Quality ↔ positive tone: The quality index correlates positively with overall positive sentiment (r = 0.71, N = 4588) and with joy (r = 0.50, N = 4588). It correlates negatively with anger (r = −0.89). These are large associations in social media text/audio settings and indicate that high-quality videos systematically present more constructive affect and far less anger.
• Populism ↔ anger: The populism rate correlates positively with anger (r = 0.34) and slightly negatively with positive sentiment (r = −0.11), consistent with a mobilizing, adversarial style. Associations with fear are near zero.
• Because N ≈ 4.6k videos for all pairs, even moderate effect sizes are estimated precisely (95% CIs in the dot plot).

The pattern strongly supports construct validity. The indices are not arbitrary normalizations: they sort videos along dimensions that coincide with widely studied affective signatures: constructive/joyful content for quality and angry/mobilizing content for populism. These validity checks show that the indices cleanly separate programmatic/constructive from mobilizing/antagonistic styles in TikTok political videos. This justifies treating them as substantive frames.

3.2. Engagement Effects

If frames are substantively meaningful, they should also predict audience response. The analysis models the engagement rate (ER)—likes + comments + shares divided by plays—as a fractional outcome with a binomial GLM and logit link (the canonical fractional logit), which keeps predictions in [0,1] and supports robust inference even with many observations near zero. This specification is standard for rates and proportions and suits social media ERs, a core metric in digital communication studies (Sanches & Ramos, 2025).

TikTok and short-video platforms have become central to electoral communication in the country—most visibly during and after the 2023 presidential cycle—so testing which frames move engagement is substantively relevant for Ecuador’s contemporary political discourse (Noboa, 2025; Primicias, 2025).

3.2.1. Baseline Fractional Logit

The dependent variable is ER at the video level. The main regressors are quality and populism (both z-scores). Controls include hour and day-of-week fixed effects and affective covariates (sent_pos, anger) to absorb valence. Estimation uses a binomial GLM with logit link and HC3 robust SEs (N = 4588). Methodologically, this follows fractional response best practice (Papke & Wooldridge, 2008).

Table 3. Fractional logit model for engagement rate.

Term	coef	se_robust	z	p_Value	OR	OR_lo	OR_hi
Intercept	−2.659	0.042	−64.13	0	0.07	0.065	0.076
quality_index	−0.075	0.044	−1.71	0.088	0.927	0.85	1.011
populism_rate	0.111	0.013	8.42	0	1.117	1.089	1.146
sent_pos	0.058	0.021	2.79	0.005	1.060	1.017	1.103
anger	0.018	0.038	0.49	0.625	1.019	0.946	1.097

Note. Odds ratios (ORs) translate logit coefficients into multiplicative changes in the odds of engagement.

reports coefficients and odds ratios; the calibration plot (Figure 3) compares predicted vs. observed ERs by deciles.

Populism increases engagement. A one-SD increase in populism is associated with an 11.7% increase in the odds of engagement (OR = 1.117, z = 8.42, p < 0.001). In ER units, the average marginal effect is +0.007 (≈+10% relative to the sample mean ER = 0.068).

Quality is (weakly) negative in the aggregate. A one-SD increase in quality is associated with a 7.3% decrease in the odds of engagement (OR = 0.927, z = −1.71, p = 0.088), AME = −0.0048. This is modest and borderline significant, suggesting that programmatic/informational content is not rewarded on average when competing in the same feed against more mobilizing appeals.

Valence matters but does not subsume framing. Positive sentiment is independently positive (OR = 1.060, p = 0.005), consistent with constructive tone drawing interaction, while anger is not significant once populism is in the model, indicating that the frame (people-centric/antagonistic) captures more than raw affect.

Model fit and calibration: The McFadden-style pseudo-R² is 0.048, typical for micro-level ER models. The calibration curve tracks the 45° line reasonably well across most deciles; the model slightly under-predicts in upper-middle deciles and slightly over-predicts in the top decile, which is acceptable for an overall baseline.

Overall, the baseline model indicates that higher levels of populist framing are associated with higher engagement rates, whereas quality-oriented framing shows smaller and statistically weaker (often negative) aggregate associations once timing and affect controls are included (Table 3). We interpret the implications of these patterns for platformized political communication in the Discussion section.

3.2.2. Heterogeneity by Actor

Engagement is modeled as a fractional response with a binomial GLM and logit link, including interactions between the framing indices and actor labels. Controls are positive sentiment, anger, and hour/day fixed effects. This follows best practice for fractional outcomes; actor-specific interaction results are visualized in the forest plots (Figure 4a,b), the marginal effect curves (Figure 5a,b), and the joint 2D response surfaces (Figure 6).

As shown in the forest plots (Figure 4a,b) and summarized in

Table 4. Slopes by actor. Fractional logit with robust (HC3) SEs; hour and day-of-week fixed effects included; and N by label in last column. RR = exp(β) per +1 SD.

Label	N	β_Quality	SE_q	RR_q	RR_q_lo	RR_q_hi	β_Populism	SE_p	RR_p	RR_p_lo	RR_p_hi
creator	2453	−0.044	0.045	0.957	0.876	1.046	0.090	0.015	1.094	1.062	1.127
institution	499	−0.607	0.079	0.545	0.467	0.636	−0.559	0.224	0.572	0.368	0.888
leader	928	0.551	0.089	1.734	1.458	2.063	0.489	0.038	1.630	1.513	1.757
media	304	−0.075	0.058	0.928	0.829	1.039	0.059	0.025	1.060	1.010	1.113
party	404	−0.253	0.073	0.776	0.673	0.895	0.004	0.056	1.004	0.900	1.121

, the actor-specific slopes (table x), converted to rate ratios (RRs) per +1 SD change, show a clear stratification: (i) leaders gain from both frames: quality RR = 1.734 and SE = 0.089; populism RR = 1.630 and SE = 0.038 (see the leader lines in Figure 4a,b; the monotonic leader curves in Figure 5a,b; and the warm ridge in the leader surface in Figure 6—leader). In TikTok’s personalized attention market, the leader persona—already salient in Ecuador—converts programmatic content and antagonistic frames into higher interaction. This matches comparative work on TikTok as a vehicle for “performative authenticity” by incumbents and contenders in Latin America and beyond (Bergengruer, 2024; Grantham et al., 2025). (ii) Creators benefit mainly from populism RR = 1.094 and SE = 0.015; quality is roughly neutral/slightly negative, with RR ≈ 0.96 (see creator points in Figure 4b; the steep creator curve in Figure 5b; and the creator surface in Figure 6—creator, where high populism raises ER even at low quality). This is consistent with the logic of creator economies in which conflict-coded content (outrage, in-group signaling) pulls algorithmic exposure and social feedback. Evidence that angry posts diffuse better reinforces the mechanism (Brady et al., 2017; Gerbaudo et al., 2023; Han et al., 2023). (iii) Parties show no reliable populism payoff, with RR ≈ 1.00 and a penalty for quality RR = 0.776 (flat/neutral party points in Figure 4a,b; near-flat party curves in Figure 5a,b). Organizational branding looks less competitive than leader-centered or influencer content in short-video feeds, echoing Ecuadorian coverage of parties’ struggles to match leaders’ digital reach and the use of influencers to compensate (Páez, 2023). (iv) Institutions are penalized by both: quality RR = 0.54 and populism RR = 0.572 (institution points <1 in Figure 4a,b; declining institutional curves in Figure 5a,b; the institutional surface in Figure 6—institution sits mostly below 1 across the (Q,P) plane). In Ecuador’s environment—where presidential and first-lady accounts dominate social narratives while institutional vocerus lag—formal accounts appear disadvantaged in the attention economy, and populist pivoting does not help (Primicias, 2024). (v) Media has a modest populism dividend RR = 1.060 and nearly neutral quality RR = 0.928 (see media points in Figure 4b and the slight upward media curve in Figure 5b), suggesting that conflict-framed coverage marginally boosts interaction without the strong returns leaders obtain.

The pattern is coherent with local dynamics: leaders invest heavily in platform-native storytelling and advertising; TikTok is a key youth-reach channel; and regulation of digital campaigning remains patchy. The result is a pecking order in which (i) leaders are payoff dominant under both frames, (ii) creators win with populism, (iii) parties capture little added value from either frame, and (iv) institutions lose engagement as they become either more technical or more combative, mirroring reporting that institutional communication struggles to convert attention into positive interaction in Ecuador’s feeds. The joint response surfaces (Figure 6) reinforce this division: the creator and leader planes show wide areas above parity, whereas the institutional plane lies largely below 1. Finally, the pattern also squares with experimental/observational evidence that anger-laden messaging diffuses efficiently online, helping personalized actors monetize conflict more effectively than organizations (Brady et al., 2021).

3.2.3. Engagement Rate Robustness

This subsection tests whether the framing–engagement relationship is stable to alternative summaries of the ER and alternative estimators. First, the ER is summarized nonparametrically by quality/populism quartiles within actor groups; second, the baseline fractional logit is re-estimated alongside an OLS on the logit transform of the ER (a common robustness check for bounded outcomes). Signs and relative magnitudes are expected to hold if the baseline result is not an artifact of functional form.

In the quartile patterns (Figure 7a,b), the median ER rises monotonically across populism quartiles for leaders and creators, flattens for media, and falls for institutions. For quality, creators show a flat/slightly negative slope, parties trend downward, and leaders exhibit a mild upward tilt (consistent with dual payoff). These shape-invariant patterns by label indicate that the sign results are not driven by a handful of outliers: higher populism generally lifts the median ER where personalization is strongest (leaders/creators), whereas quality does not pay off for organizations, especially institutions.

Alternative estimators (

Table 5. Robustness models (HC3 robust SEs; hour and day-of-week fixed effects included; coefficients reported for key regressors. “GLM-binomial” = fractional logit; “OLS logit(ER + ε)” = linear regression on the logit transform).

Term	GLM-Binomial	OLS logit(ER + eps)
quality_index	−0.075	−0.044
populism_rate	0.111	0.130
sent_pos	0.058	0.036
anger	0.018	0.063

Note: Signs for quality (negative) and populism (positive) match across both estimators, reinforcing construct stability. Coefficients are in log-odds units for the GLM and logit-ER units for OLS; magnitudes are not directly comparable, but the direction and ranking of effects are.

): Re-estimating ERs with (i) a binomial GLM (logit link) and (ii) OLS on logit(ER + ε) yields consistent signs on the framing indices. Populism remains positive and statistically clear in both models; quality remains negative (weaker in OLS but same sign), while positive sentiment is positive, and anger is small. These checks support the conclusion that the direction of effects is robust to the modeling choice.

3.3. Temporal Dynamics of Frames

The monthly series shows a gentle upward drift in populism and a late decline in quality in the last part of the window. The two lines are relatively flat through mid-2024 and then diverge in 2025: populism edges up from early 2025 and ticks higher around mid-year; quality decelerates and falls in late 2025. (See Figure 8).

Alignment with the political calendar: Three windows plausibly structure the series:

• Security shock and constitutional referendum (Jan–May 2024): Ecuador entered 2024 under a state of emergency and an unprecedented declaration of “internal armed conflict” after gangs stormed a TV station on 9 Jan 2024. The government framed policy as wartime and then advanced a popular consultation on security/justice/investment approved in April 2024. A calm in our series after May is consistent with a “consolidation” phase rather than escalation of antagonistic framing (Collins, 2024).
• Electoral cycle (Feb–Apr 2025): The 2025 general election cycle—first round in February and second round in April—re-activates competitive incentives. Campaign coverage documented strategic emphasis on digital media (TikTok/Instagram) by the incumbent and allies. Our line for populism nudges upward from early 2025, consistent with electoral mobilization and “performative authenticity” practices on short-video platforms (Robertson, 2025).
• Referendum push and protests over fuel subsidies (Aug–Oct 2025): In August 2025, the executive floated seven referendum questions to advance far-reaching reforms; by late September, large demonstrations erupted after the removal of the diesel subsidy, including a viral video of military abuse of a dying protester that triggered national outrage. In our data, populism rises and video volume spikes in September 2025, followed by a sharp October drop, which aligns with a short, intense contentious cycle.

The macro pattern is consistent with the micro estimates: populist framing intensifies around political shocks and mobilizing events, while quality weakens late in the period as communication becomes conflict centered. This sits in a broader context of persistent insecurity and militarization debates in Ecuador in 2024–2025.

Trends by Label/Actor

By-label time series indicate who drives the temporal movement. Leaders and creators are the main amplifiers of populist framing around political windows; institutions contribute relatively more quality during calmer months but converge toward neutral or conflict-coded discourse near major events. (See Figure 9a–c).

First, the populism-by-label line (Figure 9a) climbs for leader/creator accounts around the 2025 campaign and again ahead of the late-September protest wave, while institutional series remain flat or dip, consistent with the penalties for institutions and the payoffs for leaders/creators. Second, the quality-by-label line (Figure 9b) shows institutions sustaining higher relative quality during mid-2024, with convergence downward in late-2025 as the agenda shifts to conflict and protest management. Finally, a stacked area of monthly video volume (Figure 9c) confirms that the September 2025 surge is a composition shock (more content), which mechanically increases the chance of conflict-framed content surfacing in feeds.

This actor timing structure makes sense in light of (i) the documented centrality of the presidency’s digital strategy and heavy ad spend in 2024, (ii) the election year incentives to personalize and dramatize conflict on TikTok, and (iii) the late-2025 policy shock (subsidy removal) that precipitated protests and viral grievance frames. Institutional accounts, which already underperform in engagement, are structurally constrained: technical messaging seldom scales on short video, and pivoting to antagonism appears neither credible nor rewarded.

Framing does evolve with Ecuador’s political calendar. Populism intensifies around campaigns, referendum signaling, and protest cycles, and quality declines late in the period as conflict frames dominate. Leaders and creators drive the peaks; institutions briefly hold the quality line but do not convert it into engagement during contentious windows. The macro time series thus triangulates the micro evidence, reinforcing the claim that platform-native, personalized actors monetize conflict better than organizations in Ecuador’s TikTok public sphere.

3.4. Hashtag Ecosystems (Communities and Topics)

A co-occurrence network of hashtags is built from the TikTok captions/OCR text; Louvain community detection identifies modular topical clusters. The resulting hashtag network contains 526 nodes and 3976 edges with Louvain modularity = 0.368 and 35 communities, indicating meaningful but permeable topical segmentation. A companion author–hashtag bipartite graph (who uses which hashtag) contains 652 nodes and 897 edges and is used to profile the communities.

A modularity of ~0.37 is typical of thematically organized, event-reactive networks: clusters are distinct yet connected through generalist or trending tags. In this corpus, five thematic basins stand out:

• Pro-government leadership cluster (C0): Tags such as #ecuador, #DanielNoboa, #ElNuevoEcuador, #DanielNoboaPresidente, and #EcuadorParaAdelante dominate the largest community. The degree rankings confirm that these act as hubs—e.g., #ecuador (degree 370) and #danielnoboa (231)—which increases their bridging capacity into adjacent topics and For You exposure. This is consistent with the leader-centered engagement advantage documented.
• Generic trend/For You basin (C14, C11): Tags like #fyp, #viral, and #parati (and variants such as paratiiiii…) form a creator economy cluster that co-occurs with security-coded terms (#seguridad and #seguridadprivada) and with political tags near mobilizing events. These high-degree generic tags (e.g., #fyp degree 232) function as bridges that can amplify visibility across communities when paired with political content, aligning with the ER gains creators obtain under populist framing.
• Correísmo/opposition basin (C18): Tags such as #RafaelCorrea, #LuisaGonzalez, #LuisaPresidenta, and #RC5 structure a compact cluster that becomes more active during electoral windows. Its topical cohesion suggests an audience niche with strong identity signaling.
• Indigenous movement and territorial cluster (C3): #Pachakutik, #LeonidasIza, and #CONAIE, alongside #Quito/#Guayaquil, reveal a mobilization-centric community consistent with protest cycles and territory-rooted frames. Its location next to city tags hints at issue–territory coupling during contentious episodes.
• Institutional election administration (C16): Tags like #CNE, #CNECumplió, #LaDemocraciaNosUne, and #EleccionesEcuador2025 map an institutional basin. Institutional actors face engagement penalties; here, their hashtag repertoire is coherent but relatively peripheral (small community and modest degree), which may limit organic spread in short-video feeds.

The 20 × 20 author–hashtag heatmap (Figure 10) shows a block-diagonal structure typical of topical specialization: high-volume creators cluster with trend/For You tags; leadership-adjacent accounts saturate the Noboa cluster; opposition accounts concentrate on Correísmo tags; and institutional accounts reuse CNE tags. Cross-blocks mostly appear where authors blend a political tag with a generic bridge (#fyp/#viral/#parati), a strategy that plausibly pulls algorithmic exposure beyond the base audience, again consistent with the creator/leader populism dividend in engagement.

Using node metadata, top hubs by degree—#ecuador (370), #fyp (232), #danielnoboa (231), #viral (181), and #parati (105)—are precisely the tags that either (i) carry national scope (#ecuador) or (ii) speak the platform’s discovery grammar (#fyp/#viral/#parati). Their presence across communities indicates likely diffusion conduits: pairing a partisan or institutional tag with one of these hubs increases the chance of cross-community reach, which helps explain why leaders/creators—already advantaged in personalization—convert populist framing into ER gains while institutions remain siloed.

Ecuador’s TikTok hashtag ecosystem is modular but bridgeable: leader and creator communities anchor partisan or protest frames, institutional tags form a coherent but peripheral basin, and high-degree generic hubs (#ecuador/#fyp/#viral/#parati) likely act as spread conduits, an architecture that mirrors the engagement asymmetries.

3.5. Coordinated or Replicated Messaging

A lightweight coordination audit inspects nearest-neighbor pairs of highly similar videos and the small clusters they form in a caption/OCR–similarity graph. The exported “top pairs” list contains 25 near-duplicates (cosine similarity sim) and a graph summary with 4590 clusters. Replication exists but is sparse (Figure 11).

The right-skewed profile indicates a thin tail of near-identical content (several pairs with sim ≥ 0.85) atop a background of moderate similarity. Because the export masks author handles and does not include timestamps for these videos, Δt cannot be reported directly. Still, the very small gaps in TikTok video IDs within pairs are consistent with near-contemporaneous uploads, a typical signature of templated or copied messaging. This ecology is coherent with Ecuador’s attention market: leader- and creator-centric narratives travel well and are easy to replicate, whereas institutional communication remains more siloed.

In the Ecuadorian context, such micro-clusters plausibly arise from (i) surrogacy/influencer seeding around campaign pushes, (ii) leader-centered slogan reuse by allied accounts, and (iii) event-triggered scripting (elections and security/protest shocks) that standardizes language for a few days. This minimal but non-trivial replication dovetails with the engagement asymmetries documented: leaders and creators can monetize conflict-coded templates, while institutions derive little benefit from copying.

3.6. Qualitative Verification (Triangulation)

To validate the quantitative signals, a curated set of videos is examined along the two framing indices (

Table 6. Curated examples (top/bottom quality; illustrative populism).

Author	Date	Quality (z)	Populism (z)	Plays	Likes	Comments	Shares	URL
@wilopradoec	5 February 2025	0.50	0.25	884,800	90,300	12,000	5671	https://www.tiktok.com/@wilopradoec/video/7467726142074588421
@rc5oficial	15 September 2025	0.38	0.00	1249	36	1	4	https://www.tiktok.com/@rc5oficial/video/7550114732666080518
@jantopicecuador	4 November 2024	0.31	0.08	42,500	2366	143	68	https://www.tiktok.com/@jantopicecuador/video/7433535188984024326
@wilopradoec	11 September 2025	0.30	0.00	1,300,000	46,500	1611	24,100	https://www.tiktok.com/@wilopradoec/video/7548982035897912582
@rc5oficial	4 July 2025	0.30	0.25	1870	369	34	11	https://www.tiktok.com/@rc5oficial/video/7523023572600687878
@rafaelcorrea_2025	22 July 2024	0.29	0.17	27,300	2376	79	96	https://www.tiktok.com/@rafaelcorrea_2025/video/7394511775556783366
@danielnoboaok	10 September 2025	−0.28	0.00	2,300,000	154,000	6192	11,700	https://www.tiktok.com/@danielnoboaok/video/7548610887351799046
@santylucero14	1 October 2025	−0.28	0.08	2654	418	31	34	https://www.tiktok.com/@santylucero14/video/7556392067065220366
@rc5oficial	7 June 2025	−0.28	0.00	13,500	1695	57	322	https://www.tiktok.com/@rc5oficial/video/7513040508973862152
@lacontraec	22 July 2025	−0.28	0.00	1165	74	2	1	https://www.tiktok.com/@lacontraec/video/7530048022021950725
@maulexbros	9 February 2025	−0.29	0.00	479,200	14,900	176	2185	https://www.tiktok.com/@maulexbros/video/7469499555625585950
@yaku.perez	17 April 2024	−0.29	0.00	41,100	1012	4	88	https://www.tiktok.com/@yaku.perez/video/7358643070285417734

Note. The accessed date for all websites is 5 October 2025.

). These materials confirm that the quality index picks up programmatic/informational styles (policy detail, solution frames, and service information), whereas the populism index captures antagonistic cues (accusation, in-group signaling, and conflict scripting).

Interpretation (How the Clips Read On-Platform)

• Programmatic vs. antagonistic styles are visually and rhetorically distinct. High-quality posts consistently center information value: service delivery or concrete benefits (“MITI-MITI… casa propia”), interviews and policy talk (#LuisaEnMedios), or historical/educational frames (“¿cómo Manta logró su cantonización?”). They adopt instructional or explanatory captions, usually without broad rage-bait tags. By contrast, low-quality posts lean on identity and confrontation (“Se les acabó el festejo… #corruptos”), vague slogans, or viral-seeking generic tags (#parati and #fyp), with little actionable content.
• Engagement patterns mirror Section 3.3. Several “quality” clips achieve very high reach (e.g., @wilopradoec’s housing program explainer at ~1.3 M plays), but antagonistic or conflict-coded low-quality clips also attract substantial interaction (@danielnoboaok’s denunciatory post with ~2.3 M plays). This aligns with the model finding that leaders and creators monetize populist cues, while institutions do not. The curated set shows creators mixing topical hooks with emotive cues, whereas institutional accounts stick to formal updates that convert less efficiently.
• Text layers match the indices. Where OCR captured on-screen text, higher-quality posts carry titles or labels that scaffold comprehension (program names and interview identifiers), while lower-quality posts foreground accusatory or rally phrases. Even when the captions are brief, the visual text in high-quality videos tends to anchor the message (program logos and interview chyron), while the low-quality set relies more on affective cadence and creator performance.
• Topical contrasts are consistent with Ecuador’s context. In 2024–2025, the platform was saturated with content about security, elections, and economic relief. The high-quality examples frequently explain policies or ritualize institutional communication (media interviews and service programs), whereas the low-quality set clusters on accusation and polarizing cues (e.g., corruption labeling and calls against protest). This qualitative texture is exactly what the quality and populism scores are designed to separate, giving confidence in their construct validity within the Ecuadorian ecosystem.
• What this triangulation adds: The qualitative read of 10–20 concrete clips demonstrates that the same directional patterns observed in the models—quality benefits institutions less than leaders, while populism pays off for leaders/creators—are recognizable by human judgment. The clips also clarify mechanisms: quality communicates benefits and procedures; populism simplifies blame and rallies identity, enabling scalable replication in short-video feeds.

3.7. Robustness and Sensitivity

This section probes whether the engagement–framing relationships reflect measurement artifacts. It varies (i) the engagement rate (ER) definition—full ER vs. likes + comments only, with and without winsorization—and (ii) the estimator, comparing a fractional logit for the ER, an OLS on the logit of the ER, and Poisson count models with log(plays + 1) as an offset. All models retain the same basic controls (positive sentiment, anger, hour/day fixed effects, and actor fixed effects).

Signs are stable across specifications. The populism coefficient is positive and precisely estimated in every model; the quality coefficient is small and generally negative (confidence intervals often straddle zero, but the point estimate is consistently ≤ 0), and it remains so under winsorization. Count-based Poisson models reproduce the pattern: adding quality tends to decrease expected interactions (per given exposure), while adding populism increases them. This strengthens the interpretation that conflict-coded frames travel more efficiently on TikTok, whereas programmatic or informational content is not rewarded by the interaction metric at scale.

Table 7. Alternative engagement rate definitions (descriptives).

Metric	Mean	p99	n
ER_all	0.0681	0.2212	4612
ER_all_w99	0.0676	0.2212	4612
ER_lc	0.0627	0.2059	4612
ER_like	0.0589	0.1944	4612
ER_cs	0.0092	0.0464	4612

compares five ER variants: (i) all interactions, (ii) winsorized at the 99th percentile, (iii) likes + comments, (iv) likes only, and (v) comments + shares. Means and upper tails move only modestly across definitions, indicating limited sensitivity to the precise ER construction.

Table 8. Sensitivity of framing coefficients across estimators.

Specification	β_Quality	95% CI (Quality)	β_Populism	95% CI (Populism)	N
FracLogit ER (all)	−0.072	[−0.158, 0.014]	0.110	[0.084, 0.136]	4612
FracLogit ER (winsor.99)	−0.051	[−0.130, 0.028]	0.103	[0.080, 0.125]	4612
OLS logit(ER)	−0.049	[−0.229, 0.132]	0.136	[0.095, 0.177]	4612
Poisson counts, offset log(plays + 1)	−0.155	[−0.313, 0.004]	0.108	[0.048, 0.169]	4612
Poisson counts (winsor.99)	0.119	[−0.022, 0.260]	0.066	[0.009, 0.123]	4612
FracLogit ER (likes + comments)	−0.076	[−0.162, 0.010]	0.113	[0.087, 0.140]	4612

shows that the populism coefficient remains positive and significant across all estimators and outcome definitions; the quality coefficient is small and non-positive, with intervals that span zero in most specs and a borderline negative effect in count models with exposure. The winsorized count model slightly attenuates magnitudes, as expected when top-tail influence is down-weighted.

Figure 12 plots β and 95% CIs for quality and populism across all robustness models; the vertical zero line aids visual assessment. The populism intervals sit entirely to the right of zero in every ER-based model and remain positive in count models, while quality hovers near or below zero.

These checks indicate that the paper’s core findings do not hinge on the measurement of engagement or on a particular estimator. Because TikTok interactions are heavy tailed, winsorization reduces extreme influence but does not erase the populism premium. Likewise, switching from a rate (fractional logit) to a count model with exposure yields the same qualitative reading: for a given viewing base, videos with higher populism scores attract more interactions, whereas those with higher quality scores attract no more (and often fewer). Substantively, this aligns with a feed economy where conflict and identity cues are more “algorithmically legible” than informational/problem-solving content. As a result, the signs and relative magnitudes are stable under reasonable perturbations, supporting the claim that framing effects are not artifacts of modeling choices.

3.8. Evidence-Based Design for Non-Populist Engagement

The engagement rate (ER) is modeled as the logit of the ER, regressed on standardized quality_index, populism_rate, affect (sent_pos, anger, joy, and fear), and textual features computed from captions (caption_len, n_hashtags), plus actor fixed effects. For the recommendation curves, populism is fixed to −1 SD (low) (

Table 9. Levers for non-populist engagement (model-based) (OLS on logit(ER), robust HC3 SEs, and actor fixed effects. ΔER is the change in predicted ER when the lever moves from −1 SD to +1 SD, holding populism low at −1 SD).

Lever	Beta (log-odds)	p(HC3)	ΔER (+1 vs. −1 SD)
quality_index	0.005	0.935	0.05 pp
sent_pos	−0.016	0.622	−0.14 pp
joy	0.052	0.009	+0.47 pp
anger	0.107	0.052	+0.96 pp
caption_len	−0.104	<0.001	−0.93 pp
n_hashtags	0.032	0.028	+0.28 pp

).

3.8.1. What Works Without Populism (Interpretation Anchored in Ecuador’s Context)

(i)

Quality helps—but modestly—when the message is tight and concrete. The partial-dependence curve shows a small but positive slope for quality_index at low populism (Figure 13). In practice, this means clips that condense procedures, benefits, and “how it works” micro-explanations (e.g., trámites, programas, and plazos) can raise ER without antagonism, especially for institutional or media actors who already trade in information.

(ii)

“Joy” outperforms “anger” as a non-polarizing energizer. In the multivariate model, joy is positively associated with engagement (ΔER ≈ +0.47 pp, p ≈ 0.009) while anger is the largest pro-engagement driver overall (ΔER ≈ +0.96 pp, p ≈ 0.052) but is precisely the pathway that tends to co-travel with populist frames. For non-populist strategy, the data favor joy/efficacy cues—celebratory outcomes, beneficiaries on camera, and problem-solving progress—over anger. This squares with the platform’s reward to emotions that feel shareable without necessarily attacking out-groups.

(iii)

Be concise: shorter captions systematically win. Each +1 SD in caption length is associated with ≈ −0.93 pp lower ER (p < 0.001), controlling for everything else. High-performing non-populist clips keep captions short and functional (title + 1 key noun phrase), shifting explanation to the video layer (visual steps, lower thirds) rather than long prose in the caption.

(iv)

Use a few targeted hashtags (not a cloud). n_hashtags shows a small but significant positive association (ΔER ≈ +0.28 pp, p ≈ 0.028). The sweet spot is a handful of specific tags (program name, municipality, and policy area) rather than generic #fyp/#parati clouds. This helps contextual discovery without pushing conflict.

(v)

Who says it still matters? Actor fixed effects absorb systematic differences: leaders and creators begin higher, institutions lower. The design levers above are therefore net of actor type and remain useful to ministries, municipalities, and watchdogs seeking engagement with informational rather than antagonistic content.

Figure 13. Partial-dependence curves.

Figure 13. Partial-dependence curves.

3.8.2. Design Recipe

The evidence suggests a non-populist playbook that remains performant on TikTok Ecuador:

• Lead with benefits, not blame. Open in the first seconds with the specific benefit (e.g., “Ahora puedes inscribirte en … en 2 pasos.”) and then show a crisp two-step visual. This aligns with the small but positive quality slope when populism is low (Figure 13).
• Show progress with positive affect. Use joy/efficacy as the affective wrapper: cutaways to successful outcomes, smiling beneficiaries, “before/after” vignettes, or milestone counters. Joy’s positive, significant association (Figure 13) offers a clean, non-polarizing path to interaction.
• Write captions like headlines. Keep caption length down (Figure 13). Title + one concrete noun phrase (“Entrega de 120 kits en La Argelia”) beats paragraph-length descriptions. Move explanation to on-screen text and voiceover.
• Hashtag with intent. Use 1–3 precise tags: program acronym, locality, and policy area (Figure 13). Avoid generic reach tags; they add noise without informational value.
• Treat anger as a constraint. Anger measurably boosts ER (Figure 13), but it is the closest correlate of the populist pathway you are explicitly avoiding. When accountability is necessary, switch to fact-first exposure (documents on screen, timelines, and source labels) and efficacy framing (“qué hacer ahora”), not antagonistic labeling.
• Institutional accounts can still win. Because actor effects are controlled, these levers remain usable by institutions and media, which tend to underperform leaders/creators in raw interaction. The gains are smaller per post, but repeatable and reputationally safer.

4. Discussion

This study asks whether two framing indices—programmatic quality and antagonistic/populist tone—capture substantively meaningful variation in TikTok content from Ecuador and whether such variation relates to engagement net of controls and in ways that differ by actor type, how these frames evolve over time, and whether discursive communities and coordinated content help explain observed patterns. The evidence indicates that both indices behave as theoretically expected in distribution and affective correlates and that they exhibit distinct engagement returns by actor, with leaders and creators reaping the largest dividends from populist framing, while institutions are, on average, penalized under both frames. Temporal patterns show event-linked intensification of populism, and hashtag networks reveal modular ecosystems consistent with institutional, partisan, and creator clusters. Together, these findings support the claim that framing, actor identity, and meso-level discursive structure jointly shape attention in short-video politics.

Before interpreting these patterns, it is important to draw a clear distinction between observed engagement and broader political or democratic outcomes. In this paper, engagement rates capture platform-visible interaction and attention under short-video affordances, but higher engagement should not be equated with political effectiveness (e.g., persuasion, vote choice, or offline mobilization) nor with democratic value (e.g., deliberative quality, informational benefit, or civic accountability). In algorithmically ranked environments, attention can be amplified by entertainment logics, emotional arousal, or conflict cues that may not translate into meaningful political learning or pro-democratic effects. Accordingly, our discussion treats engagement primarily as an indicator of platformed attention dynamics and message performance, while normative and causal claims about democratic quality or political impact remain outside the scope of the present design.

Within Ecuador’s short-video arena, who speaks and how they frame both matters. The analysis shows that the populist path to engagement is powerful but uneven, concentrating among leaders and creators, while institutions can recover ground by translating policy into tight, benefit-first micro-stories with joyful/efficacy affect. This extends framing theory into a platform-specific, multimodal context, aligns with evidence on affective diffusion, and provides a practical, non-populist design space for public communication, one that can be tested and refined in future causal and comparative work (Brady et al., 2021; Cervi, 2021; Entman, 1993; Medina Serrano et al., 2020; Mudde, 2004).

4.1. Interpretation and Links to Prior Work

The framing constructs align with well-established theoretical lenses. By design, quality approximates programmatic, efficacy-oriented messaging, whereas populism proxies antagonistic, people-versus-elite narratives. This maps onto Entman’s definition of framing as the selection and salience of problem definitions and remedies (Entman, 1993) and Mudde’s conceptualization of populism as a thin-centered ideology that pits a morally pure people against a corrupt elite (Mudde, 2004). Empirically, the validity checks are consistent with expectations from research on affect and diffusion online: anger-like tones are positively associated with interaction and spread, whereas positive/efficacy affects yield smaller but reliable gains (Brady et al., 2021; McLoughlin et al., 2024) The actor-stratified returns—leaders benefiting under both frames, creators primarily under populism, institutions lagging—also fit the emerging literature on performative authenticity and the “politainment” affordances of TikTok, particularly in Latin America (Cervi, 2021; Cervi et al., 2023; Freire Alarcón & Medina Chicaiza, 2025; Medina Serrano et al., 2020). The platform’s personalized, short-form grammar amplifies persona-forward communication, rendering institutional styles comparatively disadvantaged unless they adapt to native, visually guided, benefit-oriented storytelling.

4.2. Answering the Research Questions

The four research questions are answered consistently across the empirical sections. RQ1 (construct validity) is supported by wide dispersion across actor categories and the affect-based correlations linking quality to positive/joyful tone and populism to anger (Section 3.1). RQ2 (baseline engagement returns) shows a robust positive association between populist framing and engagement, whereas quality exhibits smaller and more context-dependent returns net of controls (Section 3.2.1). RQ3 (heterogeneity) demonstrates that framing payoffs differ systematically by actor type, with leaders and creators showing steeper engagement responses to populism and institutions facing thinner or negative returns (Section 3.2.2). Finally, RQ4 (temporal and meso-level structure) is addressed by the monthly series and hashtag network evidence, which reveal event-linked intensification of populism and modular discursive ecosystems organized around actor communities and thematic clusters (Section 3.3 and Section 3.4).

4.3. Contributions and Implications

Methodologically, the study integrates multimodal extraction (ASR/OCR) with frame indices and fractional logit engagement modeling on a country-level corpus. Substantively, it clarifies that “populism pays” is not universal: its advantages concentrate where the persona is already salient (leaders) or natively optimized for attention markets (creators). For public institutions, the findings argue for a non-populist optimization path—compact, benefit-first videos, visual procedural cues, and joyful/efficacy affect—which the robustness checks show can improve ERs even when antagonism is avoided. These implications resonate with regional analyses of TikTok as a youth-reach channel, where incumbents and contenders enact curated authenticity to turn policy into watchable micro-narratives, and also intersect with Ecuador-specific dynamics, including the use of influencers and regulatory grey zones around paid political content on TikTok documented by national media, which help explain why creator-centric networks often outperform formal partisan brands in attention capture.

Platform affordances reward emotional immediacy, personalization, and visual proof. Studies of online diffusion show that outrage and anger function as efficient mobilizers of attention and sharing. Leaders and creators can personalize antagonism in ways organizations cannot, leveraging first-person voice, cinematic B-roll, and para-social cues to translate conflict into engagement. Institutions retain a path to performance by substituting antagonism with efficacy—short, concrete “how-to” demonstrations and milestone vignettes—consistent with the small but positive returns to quality once captions are concise and affect is joyful. Hashtag communities amplify this: creator-anchored clusters and leader-specific ecosystems act as meso-level amplifiers, whereas institutional clusters are denser but less bridge-rich, curbing cross-community spillovers.

4.4. Strengths and Limitations

Strengths include country-specific scope, multimodal text extraction, and consistent model signs across estimators. Limitations warrant caution. First, engagement is not persuasion; the ER summarizes visible interactions, not belief or vote change. Second, the dataset covers public accounts and videos available at collection time; TikTok’s content availability and ranking evolve, introducing selection bias. Third, causal inference is limited in algorithmically mediated environments such as TikTok. Engagement is not only an outcome but also a signal that can feed back into distribution: content that performs well is more likely to receive additional exposure via recommendation systems, which can mechanically increase subsequent interactions. Although we model engagement as a rate (interactions per play) to reduce simple reach effects, differential visibility and audience composition remain partly unobserved, and the estimated frame–engagement relationships should therefore be interpreted as conditional associations within a recommendation-driven attention market rather than as causal effects of framing. Future work using experimental or quasi-experimental designs, impression-level exposure data, or exogenous shocks to ranking would be needed to more cleanly separate framing effects from algorithmic amplification.

Fourth, measurement in multimodal, short-video settings is inevitably subject to computational noise. Automatic speech recognition and OCR can introduce error due to background audio, accents, code switching, slang, fast speech, and low-contrast or transient on-screen text. While combining caption + ASR + OCR into a single message blob mitigates single-modality omissions, residual extraction error may still attenuate estimated relationships or unevenly affect certain content styles (e.g., highly edited clips or text-light formats). Accordingly, the indices should be interpreted as scalable proxies rather than perfect reconstructions of intent.

Fifth, we cannot distinguish organic from paid reach or platform-driven boosting. Promoted content and influencer arrangements can shape both exposure and interaction, potentially confounding engagement patterns. Although modeling engagement as a rate (interactions per play) reduces simple scale effects, paid amplification may still alter audience composition and interaction propensity, and this limitation is especially relevant in contemporary campaign environments.

Sixth, external validity is bounded by the Ecuadorian context and time window (2024–2025), which includes salient security and electoral dynamics that may condition both framing strategies and audience responsiveness. The patterns reported here therefore provide strong within-case evidence of how frames perform under TikTok affordances, but cross-national replication is needed to assess generalizability across institutional settings, media systems, and campaign styles in the region.

4.5. Future Research

Three avenues follow directly. The first is causal tests: field experiments or quasi-experiments that manipulate affect wrappers (joy vs. neutral), caption concision, and visual procedural cues for institutional accounts would assess whether the non-populist playbook scales in persuasion, not just ER. Second, cross-national comparisons in Latin America should examine whether leader/creator advantage generalizes, leveraging shared measurement to compare Peru, Colombia, and Chile. Third, meso-level dynamics deserve attention: longitudinal community detection on author × hashtag graphs to test whether bridges (media or hybrid civic creators) mediate frame diffusion into institutional clusters.

5. Conclusions

This article demonstrates that multimodal, survey-free measurement can recover substantively meaningful frames in TikTok politics and link them to observable behavior at scale. Two indices—one capturing programmatic quality and another capturing antagonistic/populist tone—behave coherently across validity checks, display ample variation across actor types, and predict engagement in stable directions across multiple estimators. Within Ecuador’s short-video ecosystem, who speaks and how they frame jointly organize attention: leaders and creators convert both frames—especially populism—into interaction, whereas institutions and parties face thinner returns and, for institutions, even penalties when they lean either into technical detail or into conflict. Temporal series and hashtag networks indicate that these differences are not noise: populism intensifies around political milestones, and discursive communities are modular and thematically specialized.

The study’s most important implication is normative and practical. TikTok’s attention economy clearly rewards conflict cues, yet the analysis also identifies a non-populist path to competitive engagement: benefit-first micro-explanations, concise captions, targeted hashtags, and joyful/efficacy affect, supported by on-screen text that does the explanatory work. This design space is particularly salient for institutional communication, which cannot and should not emulate personalized antagonism. In other words, the platform’s affordances do not make populism inevitable; they make clarity, concreteness, and positive efficacy necessary if one wishes to avoid it.

Methodologically, the contribution is to show that open-source audio OCR/ASR, lightweight frame scoring, and fractional response modeling can be assembled into a replicable pipeline that scales to country-level corpora and yields interpretable quantities for theory and practice. Substantively, the findings extend framing research to the short-video context by specifying a persona–frame interaction: personalization amplifies frame payoffs, and organizational voice dampens them, conditional on the same topical space and temporal shocks. This helps reconcile why “populism pays” in aggregate yet does not pay uniformly across actors.

The evidence portrays a platform where framing, persona, and meso-level discourse structure co-determine attention. Populism remains an efficient—but uneven—route to interaction; at the same time, there is a viable, ethically preferable, and operationally precise strategy for public communicators who seek reach without antagonism. By making that strategy explicit and empirically grounded, the study contributes both to theory—clarifying how frames perform under short-video affordances—and to practice—offering an actionable repertoire that aligns democratic communication with the realities of contemporary feeds.