1. Hands-tying versus bluster

According to the standard domestic audience costs theory, a state’s threat incurs domestic audience costs ex post, constraining leaders from retracting their commitments (Fearon, Reference Fearon1994; Reference Fearon1997; Tomz, Reference Tomz2007; Chan et al., Reference Chan, Liu and Quek2021). As Quek (Reference Quek2021), 537 writes, tying-hand signaling involves “an action that increases the costs of backing down if the would-be challenger actually challenges, but otherwise entails no costs if no challenge materializes.” Recent studies extend this logic to autocracies, suggesting that autocratic leaders also face pressures to align their actions with their public statements (Weeks, Reference Weeks2008; Weiss, Reference Weiss2013; Dafoe et al., Reference Dafoe, Liu, O’Keefe and Weiss2022). Consequently, the implicit logic is that states tend to adhere to their publicly communicated threats due to domestic audience expectations.

However, recent scholarship on strategic interaction suggests that autocracies often employ rhetoric as “bluster”—that is, the use of implicit threats with no intentions of subsequently conducting military escalation (see Weiss and Dafoe, Reference Weiss and Dafoe2019, 3–4, 6). Existing experimental studies indicate that domestic audiences in autocracies generally approve of such implicit threats, defined as a message merely implying negative consequences without specifying concrete actions or timelines (Quek and Johnston, Reference Quek and Johnston2018, 24–25), even in the absence of actual military escalation subsequently. This domestic approval arises partly because some audiences view tough but vague statements as a sign of strength—often valuing them more than consistency between words and actions or actual military escalation (Weiss and Dafoe, Reference Weiss and Dafoe2019, 4–5). Consequently, by using implicit threats, leaders can “untie their hands,” opting for inaction or later offering an “olive branch” to de-escalate (see Quek and Johnston, Reference Quek and Johnston2018; Mattes and Weeks, Reference Mattes and Weeks2019; for related observational studies, see Wang and Womack, Reference Wang and Womack2019; Wang, Reference Wang2021; Chubb and Wang, Reference Chubb and Wang2023). Accordingly, unlike in the hands-tying mechanism, when a state issues an implicit threat, it may not signal future military escalation but rather inaction or de-escalation.

2. A case of China–Taiwan cross-strait interaction (2016–2022)

To adjudicate the contrasting theoretical predictions regarding the consistency between a state’s words and deeds, I select the cross-strait China–Taiwan relationship from 2016 to 2022 as a within-case study for both methodological and empirical reasons. Methodologically, China’s monopolistic control over the press enables a cleaner and more precise measurement of verbal signaling, which is rarely possible in countries with more complex state-media relations (Quek, Reference Quek2015, 281). The expansive timeframe covering the two terms of Tsai’s administration also avoids the problem of focusing solely on short-term crisis moments. Empirically, Taiwan is often regarded as a critical global hotspot (see Allison, Reference Allison2017; Glaser, Reference Glaser2021; Hass et al., Reference Hass, Glaser and Bush2023; Kang et al., Reference Kang, Wong and Chan2025a). However, much analysis in this area relies on speculation rather than empirical evidence. As Kang (Reference Kang2022), 138–139 states, existing studies often reflect American scholars’ interpretations of Chinese rhetoric, rather than the rhetoric itself. Therefore, an analysis of China’s behavior towards Taiwan can provide fresh empirical evidence for a more nuanced understanding of China’s strategic behavior and contribute new insights to the bluster and hands-tying literature.

To test the competing theoretical predictions by the bluster and hands-tying mechanisms, I propose the following two hypotheses:

H1a: The higher the degrees of implicit threats invoked in China’s rhetoric, the more likely China is to escalate in the future.

H1b: The higher the degrees of implicit threats invoked in China’s rhetoric, the less likely China is to escalate in the future.

H1a is grounded in the standard hands-tying literature, positing that China’s official rhetoric serves as a predictor of its future actions. If H1a is correct, we should observe that implicit threats should be followed by escalations in the case of China–Taiwan interactions. H1b is derived from the recent bluster literature, which argues that autocratic states may use implicit threats in their rhetoric as a means to appease domestic sentiment for subsequent inaction or de-escalation. This hypothesis suggests an inverse relationship between the intensity of China’s rhetoric and the likelihood of its follow-through on such declarations. And thus, if H1b is correct, we should observe that implicit threats are not followed by escalations but de-escalations in the case study.

3. Explanatory variable: degrees of implicit threats

In this study, the explanatory variable is the degrees of implicit threats invoked in rhetoric. Implicit threat refers to the verbal warnings of negative consequences aimed at an adversary without explicitly mentioning concrete actions and timeline (Quek and Johnston, Reference Quek and Johnston2018, 24–25).Footnote ² As suggested in the former section, experimental studies suggest that implicit threats can both trigger and mitigate audience costs (Quek and Johnston, Reference Quek and Johnston2018, 24–25; Weiss and Dafoe, Reference Weiss and Dafoe2019, 6). Therefore, this rhetorical frame serves as the most appropriate proxy for directly testing the two competing hypotheses.

3.1. Measuring degrees of implicit threats: a word embedding approach

To measure the levels of implicit threats made in an article, I developed a lexicon of implicit threats through a word embedding method. Prevailing methodologies of examining the rhetoric of China’s foreign affairs have focused on quantifying a few key phrases (Quek, Reference Quek2015, 299; Johnston, Reference Johnston2013, 42) and tallying headlines before and after distinct crises (Chan and Zhong, Reference Chan and Zhong2019; Wang, Reference Wang2021, 18). While valuable, these methods may not comprehensively capture the full spectrum of sentiment regarding China’s foreign affairs, given the unique usage of phrases in Chinese foreign affair contexts (see Johnston, Reference Johnston2022). Moreover, even more advanced semi-supervised models, such as Latent Semantic Scaling (LSS), require a quality list of seed words to scale the ideological position of a corpus (Watanabe, Reference Watanabe2021, 5–10).

To build a lexicon that fully captures words representing implicit threats, I collected more than 2 million People’s Daily news articles published from 1949 to 2023. I chose the People’s Daily because it is China’s flagship official channel used to signal foreign policy intentions (Shambaugh, Reference Shambaugh2007). Moreover, when the People’s Daily, representing the will of the Central Committee of the Chinese Communist Party (CCP)—the organ of the party-state—together with Xinhua News Agency, sets the tone on a particular issue, all other channels need to align with its stance (See Stockman Reference Stockmann2012, 180–200; Godwin and Miller, Reference Godwin and Miller2013, 39–45). In this regard, the People’s Daily is the most authoritative proxy for China to signal its behaviors to foreign and domestic audiences because it sets the tone for other outward- and inward-facing channels and directly represents the reputation of the party.Footnote ³ To reduce noise, I first classified the text corpus to include only articles related to foreign affairs using a lasso classifier.Footnote ⁴

After refining the corpus, I employed the Word2Vec model, a neural network-based word embedding technique developed by Google, to analyze semantic relationships within the text. Unlike traditional bag-of-words approaches such as Latent Dirichlet allocation or Structural Topic Modeling, which often treat words as independent entities and risk losing contextual meaning, Word2Vec captures the semantic associations between words by representing them as dense vectors in a continuous vector space. This method positions words with similar meanings closer together, reflecting their contextual usage in the corpus. Recent findings also showed that in sentiment analysis, lexicons derived from Word2Vec embeddings have been shown to outperform off-the-shelf, general-purpose dictionaries (Widmann and Wich, Reference Widmann and Wich2023; see also, Liang et al., Reference Liang, Ng and Tsang2023).

To select the seed words for “implicit threats,” I drew on Quek and Johnston (Reference Quek and Johnston2018), supplementary material, 4 survey experiment vignette, which was used to test public approval after a state issued an implicit threat. In the original text, the vignette states:

“The Chinese leader warned that the Japanese Government must be held fully responsible for the consequences arising therefrom if Japan failed to remove its structures on the territory.”

Based on this, I used the word “consequence (后果)” as the primary seed word, along with two synonyms—“all consequences (一切后果)” and “serious consequences (严重后果)”—as the seed words for the skip-gram models. To ensure consistency, I trained the skip-gram model using both 100- and 300-dimensional word representations. I then collected the top 100 words most closely associated with the 3 seed words and each of the 5 leadershipsFootnote ⁵ in both the 100- and 300-dimensional models, producing a total of 6000 words.Footnote ⁶ After removing duplicates, the dataset was reduced to 1160 words.Footnote ⁷

With the assistance of 2 research assistants familiar with China’s foreign affairs, I identified 19 phrases associated with implicit threats. To ensure consistency in word selection, I operationalized the concept using a coding rule stipulating that the phrases must convey the notion of negative consequences without referencing explicit actions or specific timelines. This approach distinguishes implicit threats from explicit threats (e.g., “respond with force” [军事回应]) and from broader expressions of negative attitudes and implied actions and negative attitudes (e.g., “not tolerate” [不姑息]). By doing so, the selection process ensures that the identified phrases reflect implied, rather than overtly stated, punitive outcomes. The 19-word lexicon is presented in Table 1. Table 2 reports the People’s Daily article on Taiwan that exhibits the highest level of implicit threat sentiment.

Table 1. Implicit threat lexicon

3.3. Sentiment analysis: a dictionary approach

After validation, I used the 19-word lexicon to calculate the implicit threat sentiment score per article (indexed as $i$) using the formula below. I multiplied the score by 100 for better visualization:

\begin{equation*} \text{Implicit threat score}_{i} = \left( \frac{\text{Number of implicit threat phrases}_{i}}{\text{Total number of words}_{i}} \right) \times 100 \end{equation*}

To conduct the analysis, I identified 1111 People’s Daily news articles that mention “Taiwan” and its acronyms in the titles, excluding sports and commercial news unrelated to cross-strait relations, from 2016 to 2022. These exclusions ensured that only relevant political content was included in the analysis.

The implicit threat score ranges from 0 to 2.06, where 0 indicates no implicit threat sentiment in the article, and 2.06 represents that 2.06% of the total words in an article were classified as implicit threat phrases.

Table 2. Example of an article with the highest implicit threat score

4. Outcome of interest: China’s military escalation

To measure China’s military escalation, defined as “the use of force to seize land or coerce an opponent” (Fravel, Reference Fravel2008, 4–5), I utilized event data from the Integrated Crisis Early Warning System (ICEWS), a comprehensive platform that tracks and analyzes daily state-to-state interactions (Boschee et al., Reference Boschee, Lautenschlager, O’Brien, Shellman and Starz2018). The analysis focuses exclusively on events with the highest severity intensity ( $-$10), corresponding to the category of “fight,” which includes the “use of force.” Events were selected where “China” is identified as the source country and “Taiwan” as the target country from 2016 to 2022.

The data are coded as binary, with 1 indicating an escalation and 0 indicating no escalation at the daily level. This binary measure was chosen for two methodological reasons. First, while ICEWS draws on news reports to classify events—a process that may conflate verbal threats with material escalations—I have explicitly excluded events coded as “threaten,” which include “verbal threats” (intensity level $-$6). Only events categorized as “fight,” reflecting the use of force (intensity level $-$10), were retained to ensure that the outcome of interest captures material escalations distinct from implicit verbal threats, which constitute the EV. Second, although ICEWS provides a spectrum of hostility intensity scores ranging from $-$1 to $-$10, only events coded as $-$10 qualify as escalations under Fravel’s definition of the use of military force. Other high-severity events, such as mass killings ( $-$9) or the imposition of embargoes ( $-$8), do not meet the criteria for escalation in the context of interstate conflict nor necessarily reflect the direct use of military force.

5. Quantitative results

To test the two competing hypotheses, I use the following model:

\begin{equation*} \text{Escalation}_{t} = \beta_0 + \sum_{k=1}^{d} \beta_k \, \overline{\text{Implicit Threat}}_{t-k} + \alpha_y + \epsilon_{t}. \end{equation*}

The model estimates escalation ( $\text{Escalation}_{t}$) as a function of daily average implicit-threat levels observed over the preceding $d$ days ( $\overline{\text{Implicit Threat}}_{t-1}, \overline{\text{Implicit Threat}}_{t-2}, \ldots, \\ \overline{\text{Implicit Threat}}_{t-d}$), while controlling for year fixed effects to account for differences across Tsai Ing-wen’s two presidential terms. In this specification, $t$ denotes the current day for which the likelihood of escalation is estimated, and $k$ indexes the number of days prior to $t$ from which implicit-threat scores are drawn. The inclusion of multiple lagged implicit threats allows for a more comprehensive analysis of how daily rhetoric influences the likelihood of escalation events over time, thereby capturing the temporal relationship between China’s signaling and subsequent actions. By incorporating these lags, the model addresses potential autocorrelation, leading to more reliable and unbiased estimates.

To account for days when China did not publish any articles, which result in missing values, I replaced these missing values with zeros. This approach assumes that the absence of publications indicates a lack of implicit threats on those days, aligning with observed patterns in China’s official media communication. Over the entire period of analysis ( $ N = 2557 $ days). Lastly, I aggregated the implicit threat scores by calculating the daily average.

As shown in Table 3, all models strongly support H1a, indicating that China’s rhetoric, particularly implicit threats, is significantly associated with subsequent escalation events. Specifically, Model 1 reveals that a one-unit increase in the one-day lagged implicit threat score (representing the percentage of implicit threat-related words at a daily level) corresponds to a 2.7 times increase in the odds of escalation.

Table 3. Logistic regression results with year fixed effects

*** p < 0.001; **p < 0.01; *p < 0.05.

To ensure the robustness of the findings and address potential time-series autocorrelation, I introduced additional controls in subsequent models. In Model 4, I included both the current day’s implicit threat score and multiple lagged scores up to four days prior. It also incorporated the previous day’s escalation as a predictor. Results remained consistent across these models. A Durbin–Watson test on Model 4 yielded a value of 2.01 ( $p$ $=$ 0.64), indicating minimal autocorrelation after including all lagged variables. Figure 1 visualizes the predicted probability of China’s escalation against Taiwan as a function of its one-day lagged implicit threat levels (Model 4).

Figure 1. Predicted probability of escalation.

To further explore the potential accumulation of implicit threats in the past several days, I used 3-day and 5-day rolling averages of the implicit threat scores in Models 5 and 6. The coefficients suggest that cumulative implicit threats over multiple days increase the likelihood of escalation, capturing the cumulative effect of China’s rhetoric while smoothing short-term fluctuations and mitigating autocorrelation in the time-series data.

I probed the robustness of the results using five alternative methods: (1) recalculating the implicit threat scores using an LSS scaling approach (Watanabe, Reference Watanabe2021); (2) penalizing the regression models by excluding zero entries from the independent variables when employing a rolling average; and (3) replacing the models with ordinary least squares (OLS), rare event logistic regression (King and Zeng, Reference King and Zeng2001), and vector autoregression. As presented in the appendices, the results remained consistent.

6. Implications and forward

While this research focuses on China–Taiwan strategic interaction, the dynamics uncovered here may also arise in other maritime and territorial disputes that China designates as core interests. Indeed, Taiwan occupies a central place in China’s domestic political discourse.Footnote ⁸ For example, Sino–Japanese tensions over Taiwan have intensified as Japanese leader Sanae Takaichi publicly linked Taiwan’s security to Japan’s defense posture. Beijing views this framing as foreign interference in a core sovereignty issue and has responded with sharp rhetorical condemnation followed by economic and military actions (seeMinistry of Foreign Affairs of the People’s Republic of China, 2025). Such rhetoric raises domestic audience costs for Chinese leaders by narrowing the political space for compromise. However, it also risks provoking strong military responses—from Japan, and particularly from the United States, if Beijing pushes too far.

This push–pull dynamic between domestic imperatives and external deterrence pressures is likewise evident in other disputes, including those over the Diaoyu Islands and the South China Sea, where opposing parties refuse to compromise and benefit from US security commitments. For example, Japan’s Ministry of Foreign Affairs maintains that there is “no issue of territorial sovereignty to be resolved” regarding the Diaoyu/Senkaku Islands (Ministry of Foreign Affairs of Japan, 2025), while the Philippines has vowed “not to yield any inch of its rightful territory” in the South China Sea (Reuters, 2023). This dynamic is not new. Historical conflicts such as the Sino–Soviet border clash in 1969, the Sino–Indian conflict in 1959–1962, and the Sino–Vietnamese War in 1979 illustrate similar patterns. In these cases, China’s threats collided with opposing parties unwilling to back down, resulting in full-blown escalation (see Godwin and Miller, Reference Godwin and Miller2013).

This research opens two key avenues for further investigation. First, while this research focuses solely on a single rhetorical frame—implicit threats, recent experimental studies suggest that other rhetorical frames, such as territorial ownership (Fang and Li, Reference Fang and Li2020), historical memories (Weiss and Dafoe, Reference Weiss and Dafoe2019), provocation (Dafoe et al., Reference Dafoe, Liu, O’Keefe and Weiss2022), and explicit threats (Quek and Johnston, Reference Quek and Johnston2018) can also generate audience costs and incentivize China to escalate if opposing parties do not back down. Nonetheless, these experimental studies have yet to establish whether these rhetorical frames correlate with actual escalation. Future research should explore which rhetorical frames, beyond implicit threats, are most indicative of China’s intentions and which constitute noise, thereby advancing the literature on China’s signaling behaviors and crisis forecasting.

Finally, beyond traditional media, China is capitalizing on platforms like WeChat, Twitter, and Facebook, each catering to distinct domestic and international audiences. However, outlets like the Global Times (English) do not hold the same authoritative weight in China as the People’s Daily does.Footnote ⁹ Future studies should delve into how China’s varying media channels play distinct roles. For example, lower-ranked propaganda channels may in fact engage more in bluster than more authoritative channels, given that they may not carry the reputation of the top leadership.