The 2021 German federal election was unique due to the absence of an incumbent chancellor, a three-way competition for the chancellorship, and three televised debates (Triells) with significant viewership (4-11 million). This study focuses on the second Triell, a pivotal moment in the campaign where public opinion remained malleable. The research question centers on understanding what caused viewers to alter their perceptions of who won the debate. Previous research suggests that debate winner perception strongly influences electoral decisions, particularly in countries with weak partisan identification like Germany. Existing studies identify pre-existing candidate preferences, perceived debate performance, and party identification as key factors. However, they often lack detailed analysis of the impact of individual candidate statements. This study addresses this gap by analyzing real-time response (RTR) data, enabling a detailed examination of how specific statements influence perception shifts, considering political predispositions simultaneously. By using machine learning, the study aims to identify which combination of political predispositions and candidate statements are most effective in changing viewers' perceptions of the debate winner, offering a novel approach to evaluating political debate performance and expanding the empirical debate research toolbox.

Existing research highlights the significant role of pre-existing candidate preferences as cognitive filters influencing debate perception, aligning with cognitive dissonance theory. Viewers tend to favor their pre-debate preferred candidate as the winner, even after the debate. Candidate images (credibility, competence) also significantly impact post-debate judgments. Selective exposure and selective perception are proposed mechanisms explaining how viewers process information aligning with their pre-existing beliefs. However, the literature debates the extent to which candidate statements influence viewer perceptions, with some suggesting a 'tipping point' where even strong partisans adjust their views based on compelling new information, particularly when party loyalty is weak. Party identification is considered a key heuristic in processing political information. Pre-debate expectations about the debate outcome also play a role, though generally a weaker one. Candidates can utilize rhetorical strategies (attacks, acclaims, defenses), but the systematic effects of these strategies on audience perception remain debated. Real-time response (RTR) measures are crucial for assessing audience evaluations of candidate statements, though previous studies often focus on average evaluations across the entire debate rather than individual statements. This study aims to address the existing gap by examining the influence of specific statements and their interaction with political predispositions on the dynamic process of perception change.

This study uses a large-N field study design, analyzing data from over 4600 participants who used a web application (Debat-O-Meter) to rate candidate statements in real-time during the second Triell of the 2021 German federal election. The Debat-O-Meter allowed participants to rate each candidate's statements on a five-point scale (+2 to -2). A pre-survey collected data on socio-demographic characteristics, political attitudes, candidate preferences (as future chancellor and debate winner), and evaluations of candidate images (credibility, likability, leadership, competence). A post-survey collected data on post-debate winner perception. The study utilized a convenience sample recruited through newspapers and a university panel, acknowledging potential limitations regarding generalizability. The data were analyzed using random forest and decision tree models. Random forest models were used to identify the most influential variables in predicting changes in debate winner perception for each candidate. Decision tree models were then used to visualize and understand the relationships between these variables in more detail, mapping out pathways of responses that led to changes in winner perception. The study coded the debate into 293 speech phases to examine the impact of individual statements, addressing the limitation of previous studies that typically considered average ratings across the entire debate. Robustness checks were conducted using a subsample of only TV viewers, excluding online streamers, to assess the stability of the findings.

The study's findings support the hypotheses that pre-debate dispositions and agreement/disagreement with key speech moments significantly determine changes in debate winner perception. Random forest models revealed that pre-debate winner expectation and pre-debate chancellor preference were highly important in predicting perception changes. Candidate images (credibility and competence) also played significant roles, especially in determining if a participant would shift their perception towards a different candidate. Party identification showed surprisingly little influence on perception change. Analyzing specific candidate statements, decision trees identified several key moments that significantly influenced perception shifts. For Annalena Baerbock (Greens), successful attacks on Armin Laschet (CDU) proved effective. For Armin Laschet, a patriotic address at the end, attacks on Olaf Scholz (SPD), and a defense of his climate policy were pivotal. For Olaf Scholz, a humorous defense against a critical question and a positive statement on pension policy were effective, but had comparatively less impact than the statements of the other candidates. The models demonstrated high accuracy in predicting perception changes, indicating that both political predispositions and specific debate moments played key roles in shaping viewer opinions.

The study's findings address the research question by demonstrating the interplay of pre-existing beliefs and specific candidate statements in shaping debate winner perceptions. Pre-debate expectations significantly filtered information processing, but they were not deterministic. Candidate statements on specific issues had a demonstrable impact, indicating that candidates can effectively influence audience perceptions through targeted messaging. The relative lack of influence of party identification is noteworthy, suggesting that in a context of weak partisan attachment, candidate performance and messaging can significantly shape opinions even among those with pre-existing preferences. The use of machine learning provided a nuanced understanding of the interaction between political predisposition and candidate statements, offering a novel method for analyzing debate effects. The granular approach to real-time responses allowed for the identification of specific speech moments that proved highly influential in shifting opinions.

This study makes several key contributions. It provides a novel method for analyzing debate impact using machine learning and real-time response data, offering a more granular analysis of perception shifts. The findings highlight the importance of both pre-existing political predispositions and the content of candidate statements on viewer perceptions. Future research could explore emotional responses to statements and examine whether the findings are generalizable to other contexts, particularly those with stronger partisan attachments. Further investigation is needed into the interaction between different rhetorical strategies and viewer responses across diverse demographic groups.

The study’s reliance on a convenience sample limits the generalizability of the findings to the broader German population. The study’s focus on a single debate also limits the generalizability of the findings to other debates and election contexts. The use of RTR data can be influenced by factors beyond candidate statements. While the study controlled for several relevant variables, the possibility of omitted variable bias cannot be entirely ruled out. The use of machine learning techniques brings its own methodological limitations: the ability of the algorithm to detect patterns depends on the balance of categories in the data.