Assessing Responses to Cannabis Health Warnings among Adults in US Recreational States

Abstract

Introduction:

Guided by the Extended Parallel Process Model (EPPM), this study used BERTopic modeling and manual coding to analyze reactions to cannabis health warnings and to assess how those responses were associated with message reactions, health beliefs, and intentions.

Methods:

In 2022, adults (≥ 21; N=1,078) living in recreational cannabis-use approved US states who reported using cannabis in the past year were randomly assigned to view text-only or pictorial cannabis health warnings. Participants wrote free-response reactions to the warnings, which were analyzed in 2025 using BERTopic, a clustering approach of topical modeling. Participant responses were manually coded into EPPM response categories (maladaptive or adaptive). Three linear regression models were conducted with the main outcome measures (i.e., reactance, health beliefs, and intentions to prevent harms from using cannabis) as the dependent variables and frequency of cannabis use (e.g., past 30-day use and problematic cannabis use [PCU] risk), cannabis risk perceptions, and EPPM responses as predictors, and condition as a control variable.

Results.

BERTopic identified 11 thematic subtopics that were coded for EPPM responses. Responses were more maladaptive (624; 57.9%) than adaptive (454; 42.1%). Multiple linear regression results showed maladaptive EPPM response was significantly associated with higher levels of reactance (β = 0.38), less accurate health beliefs about cannabis harms (β = −0.32), and lower intentions to prevent harms from using cannabis (β = −0.31).

Conclusions:

Results provide data-driven evidence on how US cannabis consumers evaluate health warnings from established cannabis control systems. Participant responses reveal areas of resistance and acceptance to warnings, identifying potential ways to improve future messaging.

Introduction

The majority of states in the United States (US) have legalized access to medical or recreational cannabis.¹ Although cannabis may have limited benefits, such as alleviating chronic pain,^2–5 cannabis risks include impaired driving,⁶ onset of psychosis and psychotic symptoms,^7,8 and exposure to smoke.⁹ High-frequency cannabis use is a risk factor for developing dependence and cannabis use disorder (CUD).^10,11 Cannabis use prevalence has increased among adults in the US alongside policy liberalization.^11–14 Public polling indicates that US consumers generally overestimate the benefits of cannabis and underestimate its risks.^15,16 Decreased risk perceptions were associated with increased prevalence of cannabis use,^16–18 highlighting the need to promote awareness about health effects as cannabis becomes available in legal markets.^19,20

Health warning messages (i.e., health warnings) are interventions delivered on products where consumers are likely to view them repeatedly.^21–23 Health warnings are key components of regulatory systems for tobacco^21,24 and alcohol²⁵ promoting risk awareness during purchase and use. Tobacco research showed that warnings with pictures demonstrating health effects (i.e., pictorial warnings) increased the salience of the warnings²⁶ and intentions to quit smoking.^27–29 Reviews evaluating “real world” impacts of warnings on alcohol products found exposure to warnings increased risk perceptions about harms, although effects decayed over time due to repeated exposure.³⁰ A systematic review of experimental research found that exposure to several unique “rotating” warnings across products had a greater impact on alcohol consumption than viewing a single warning repeatedly.²⁵ Rotating multiple warnings to prevent habituation is a best practice in tobacco control, along with pictorial warnings.³¹

In Canada, where cannabis is nationally legal, policies^32–34 informed by tobacco research³⁵ require cannabis products to be sold in packaging with rotating warnings. In the US, cannabis remains federally illegal but legally available in most states, where warnings are often mandated before research has evaluated the designs.^36,37 Surveys sampling respondents from Canada and US legal and illegal cannabis states found that exposure to Canadian cannabis health warnings mandated on products increased notice, recall, and recognition of cannabis health effects (compared to US states with or without legal cannabis policies).^38,39 A review of cannabis health warning research found that Canadian warnings were generally more effective than US warnings for improving knowledge about cannabis health effects.⁴⁰ Canadian warnings provide benchmarks for less established cannabis control systems in the US. Moreover, in the US, public health agencies have called for research into effective cannabis health messaging.⁴¹ This study addresses this call by analyzing the reactions of US consumers to cannabis warnings adapted from Canadian designs and assessing the theoretical mechanisms behind their effectiveness.

Health warnings use threat appeals to motivate behavior by highlighting susceptibility to serious threats.⁴² The Extended Parallel Process Model (EPPM) is a threat-appeal theory predicting three responses to health warnings: 1) maladaptive response, 2) adaptive response, or 3) no response. A maladaptive response is posited when a threat is deemed relevant, but one’s perceived efficacy is low, resulting in a tendency to resist warning recommendations.^43–45 Adaptive response occurs when perceived threat and efficacy are high, resulting in motivation to follow warning recommendations.^43,44,46 No response occurs when a threat is perceived as irrelevant. This paper examined maladaptive and adaptive responses, as these represent engagement with relevant health threats and inform intervention strategies.

One way to assess EPPM responses is by displaying health warnings to individuals and collecting free-response feedback, which can be coded for EPPM categories (referred to as EPPM responses).⁴⁷ This content analysis approach^48,49 provides rich insights into public reactions to health communications.⁴⁷ For example, a content analysis of public Tweets about the TIPS anti-smoking campaign found that maladaptive responses theorized in EPPM (e.g., reactance; “Every time I see an anti-smoking commercial, I light one up”) emerged naturally in responses to the campaign.⁴⁷ Another experimental study exposed participants to varied cigarette health warnings and asked for free-response written reactions. Responses were coded for EPPM categories,⁵⁰ and adaptive responses (e.g., “You have to be crazy to smoke”) correlated with greater odds of smoking quit intentions, consistent with EPPM predictions.⁵⁰

While EPPM human coding effectively identifies motivational stances (maladaptive or adaptive), it provides limited insights into topical nuances that drive responses. Conversely, recent transformer-based text analysis employing BERT (Bidirectional Encoder Representations from Transformers)⁵¹ and its derivative BERTopic⁵² excel at clustering semantically related content into coherent themes, identifying fine-grained topical structures that human coding might miss. However, BERTopic cannot differentiate between motivational stances when similar language appears across responses (e.g., skepticism versus cautious acceptance). Integrating both methods creates a complementary approach that captures motivational stances through EPPM and the rich thematic content through BERTopic, providing a comprehensive understanding of reactions to health warnings.

Despite studies analyzing EPPM responses to tobacco warnings,^47,50,53 the authors are unaware of any studies that have applied the model to cannabis health warnings. This paper reports on a national experiment where US adult cannabis consumers viewed different cannabis health warnings and provided free-response written reactions. These free responses were analyzed using BERTopic to identify thematic subtopics and manually coded to assess EPPM motivational stances. This paper specifically examined how EPPM responses related to three key measures of health warning efficacy: Reactance (i.e., resistance to persuasive messages),^54–56 health beliefs,⁵⁷ and behavioral intentions.⁵⁸ EPPM posits that maladaptive responses increase resistance and lower acceptance of health warnings, while adaptive responses promote acceptance. The authors hypothesized that maladaptive EPPM responses (vs. adaptive) will predict higher resistance to cannabis health warnings (indicated by reactance), less acceptance (characterized by inaccurate health beliefs about cannabis harm), and lower intentions to mitigate harm from cannabis use (H1).

Method

Sample

This study was part of a larger project evaluating cannabis health warnings.⁵⁹ Participants were 1,078 adults (≥ 21) living in legal recreational cannabis use-approved states in the US who reported past 12-month cannabis use. Sample quotas were based on gender, age, race, and ethnicity for people who use cannabis in the US.⁶⁰ The research firm Qualtrics recruited participants from research panels. Inclusion criteria were being 21 years or older, past 12-month cannabis use, and living in a recreational US cannabis use-approved state. The study team designed experimental materials. The University of Missouri Review Board approved the protocol (2080622).

Procedure and Design

Participants provided consent before pre-exposure measures (i.e., administered before exposure to warnings), including demographic information, frequency of cannabis use, risk for problematic cannabis use (PCU), recalled exposure to cannabis health warnings, and cannabis risk perceptions. Participants were then randomly assigned to view either three text-only or pictorial cannabis health warnings (described below) presented as stand-alone images and counter-balanced across topics. After viewing warnings (i.e., post-exposure), participants wrote free-response reactions, completed main outcome measures, and received a debrief with resources for help with substance use.

Stimuli

In terms of stimuli, the experiment tested two types of cannabis health warnings: text-only vs. pictorial. The warning text was from the mandated warnings in Canada, describing the risks of impaired driving, mental health, and smoke exposure. We used these warnings since Canada has a well-established national cannabis control system,¹⁹ including guidelines for warnings based on extensive consumer testing.⁶¹ The text-only condition showed three cannabis health warnings with no pictures, and the pictorial condition showed the warnings with pictures (Appendix Figure 7). Draft warnings were pilot-tested with a national sample; details are reported elsewhere.⁵⁹

Pre-exposure measures

Participants provided their age, gender, race, ethnicity, and education. We measured the frequency of cannabis use (e.g., monthly use vs. less than monthly use).⁶² PCU risk was measured with the ASSIST questionnaire.⁶³ Participants were asked if they had “a strong desire or urge to use cannabis” in the last 3 months (0=Never; 6=Daily or almost daily); “how often use of cannabis led to health, legal, or financial problems” (0=Never; 7=Daily or almost daily); “how often have you failed to do what was normally expected of you because of your use of cannabis” (0=Never; 8=Daily or almost daily); “Has a friend or relative or anyone else expressed concern about your use of cannabis?” (0=No, never; 6=Yes, in the past 3 months; 3=Yes, but not in the last 3 months) and “Have you ever tried and failed to control, cut down or stop using cannabis?” 0=No, never; 6=Yes, in the past 3 months; 3=Yes, but not in the last 3 months). Using ASSIST criteria, PCU risk categories were low (0–3), moderate (4–26), and higher (27+). Participants reported recall of cannabis health warnings in the past 12 months.⁶² Cannabis risk perceptions were measured with six items (“If you experience health harms from using cannabis, how severe will it be?”; How serious are the health harms of cannabis? How significant are the health harms of cannabis? How at risk are you to experience health harms from using cannabis? How likely is it for you to experience health harms from using cannabis? How possible is it you will experience health harms from using cannabis? (1=not at all; 9 =extremely).^56,64

Free-response reactions

Participants were asked to provide free-response reactions after viewing cannabis health warnings (“Please type in every thought that came to mind while looking at the messages. Don’t worry about spelling or punctuation. Please list as many thoughts as possible that you can recall thinking while looking at the messages”). The text of participants’ responses was content analyzed⁴⁹ using BERTopic and human coding (described below) to identify thematic subtopics and EPPM responses underlying subtopics. Each participant response was scored for the presence (coded as 1) or absence (coded as 0) of EPPM response category and subtopic code. All codes were mutually exclusive.

Main outcome measures

The main outcome measures (defined in Table 3) were reactance toward the cannabis health warnings,^65,66 health beliefs about cannabis health harms,⁶² and intentions to prevent harm from using cannabis.⁶²

Table 3.

Measures and definitions of main outcome measures variables.

Main outcome measures	Response Options	M (SD)a	Reliability
Reactanceb - The messages tried to manipulate me. - The information in the messages was overblown. - The messages annoyed me.	1 (not at all) – 9 (extremely)	4.47 (2.64)	α = .87
Health Beliefs Based on what you know or believe… - How harmful is cannabis smoke? - Does cannabis increase the risk of mental health problems? - How dangerous is driving or operating machinery after using cannabis?	1 (not at all) – 9 (extremely)	5.27 (2.15)	α = .79
Intentions to prevent harm from using cannabis - How likely are you to avoid cannabis smoke as much as possible? - What is the likelihood you will limit daily or near-daily cannabis use? - What is the likelihood you will reduce cannabis use to avoid mental health problems?	1 (not at all) – 9 (extremely)	5.13 (2.22)	α = .78

^aM=mean; SD=standard deviation, and α = Cronbach’s alpha.

^bThis is the “brief” version of the reactance scale.⁶⁶

Statistical Analysis

Free response reactions were analyzed with each participant’s written response as the unit of analysis. Using responses as input, BERTopic was performed after removing 17 responses (1.6%) that were missing, gibberish, or too short (e.g., “nothing”), leaving 1,078 valid responses. The embedding process used the pre-trained Sentence Transformer “all-MiniLM-L6-v2” model to generate embeddings in a 358-dimensional space. Next, UMAP dimension reduction⁶⁷ and HDBSCAN clustering⁶⁸ fine-tuning and model evaluation steps followed to extract thematic topics from the data (Appendix Figures 1 through 6). Compared to alternative bag-of-word methods, such as LDA topic modeling, BERTopic leverages context-aware transformer embeddings that capture meaning beyond exact word overlap, yielding topics that are both more coherent and less redundant for downstream analysis (see Appendix Table 1 and 2 for sensitivity analysis).

Second, each participant’s response was deductively coded on EPPM categories.^45,50 EPPM theory⁴⁴ and meta-analysis⁴³ posit maladaptive and adaptive responses are exclusive, so EPPM coding was exclusive, creating a binary variable (1=maladaptive; 0=adaptive response). Maladaptive responses were defensive reactions against cannabis health warnings. Adaptive responses expressed beliefs, attitudes, and intentions or behaviors in line with warning recommendations. Responses with adaptive and maladaptive content were coded adaptive if they expressed intentions to take protective action aligned with warning recommendations and maladaptive if they did not express protective intentions. Two coders independently coded 10% of responses. Interrater reliability was assessed^48,49 using Krippendorff’s alpha,⁶⁹ with a threshold of .70. Any responses that fell below this threshold were discussed and disagreements were resolved. Another sample consisting of 5% of the responses was coded, followed by another reliability assessment. The remaining responses were then divided and coded based on the acceptable reliability results.

Third, since demographic and psychosocial factors influence reactions to threat appeals,⁴⁴ we examined descriptives and bivariate correlations between pre-exposure measures, experimental conditions, EPPM responses, and main outcome measures to identify covariates for regression analyses. Any variables correlated with main outcome measures at r ≥ .3 were retained as covariates in linear regression.⁷⁰ Fourth, to test H1, three separate multiple linear regression models were conducted with the main outcome measures (i.e., reactance, health beliefs, and intentions to prevent harm from using cannabis) as criterion variables, frequency of cannabis use and PCU risk, cannabis risk perceptions, and EPPM response (maladaptive vs. adaptive) as predictors and experimental condition as a control variable. All analyses were conducted using BERTopic and SPSS v29.

Results

Table 1 presents characteristics for the sample of 1,078 cannabis consumers aged 21–78 years (M=40.2, SD=15.5). Of these, 49.4% identified as male, 78% as White, and 13.0% as Hispanic, Latino/a/x or Spanish origin. Most reported using cannabis “every day or almost every day” (40.4%) and “moderate” risk for PCU (69.4%). The average cannabis risk perceptions were M=2.6 on a 9-point scale.

Table 1.

Participant characteristics for adults (21+) reporting past 12-month cannabis use.

Characteristics	Overall N = 1,078(100%) n (%)	Maladaptive n = 624 (57.88%) n (%)	Adaptive n = 454 (42.11%) n (%)
Age
21–29	438 (40.6)	270 (43.2)	168 (37.0)
30–44	226 (21.0)	132 (21.1)	94 (20.7)
45–59	247 (22.9)	135 (21.6)	112 (24.6)
60+	167 (15.5)	87 (13.9)	80 (17.6)
Gender
Male	533 (49.4)	310 (49.7)	223 (49.1)
Female	511 (47.4)	296 (47.4)	215 (47.4)
Transgender	24 (2.2)	11 (1.8)	13 (2.9)
Do not identify as female,male, or transgender	10 (0.9)	7 (1.1)	3 (0.7)
Race identification
White	841 (78.0)	496 (79.5)	345 (76.0)
Black or African American	134 (12.4)	70 (11.2)	64 (14.1)
Asian or Pacific islander	35 (3.2)	11 (1.8)	24 (5.3)
Middle Eastern	27 (2.5)	19 (3.0)	8 (1.8)
American Indian or Alaska Native	10 (0.9)	6 (1.0)	4 (0.9)
Other	30 (2.8)	21 (3.4)	9 (2.0)
Ethnicity
Hispanic, Latino, Latina, Latinx, or Spanish origin?
Yes	140 (13.0)	82 (13.1)	58 (12.8)
Education
Less than high school	30 (2.8)	20 (3.2)	10 (2.2)
High School or GED	300 (27.8)	188 (30.1)	112 (24.7)
Some college/technical school	412 (38.2)	240 (38.5)	173 (38.1)
Associate’s degree or higher	335 (31.1)	176 (28.2)	159 (35)
Frequency of cannabis use a
Less than once per month	210 (19.5)	104 (16.7)	106 (23.3)
One or more times per month	204 (18.9)	88 (14.1)	116 (25.6)
One or more times per week	224 (20.8)	121 (19.4)	103 (22.7)
Every day or almost everyday	436 (40.4)	309 (49.5)	127 (28.0)
Problematic Cannabis Use (PCU) Risk
Low	272 (25.2)	127 (20.4)	145 (31.9)
Moderate	748 (69.4)	471 (75.5)	277 (61.0)
High	58 (5.4)	26 (4.2)	32 (7.0)
Recall seeing cannabis health warnings on products of packaging in the past 12 months?
Yes	504 (46.8)	287 (46.0)	217 (47.8)
Cannabis risk perceptions
M (SD)	2.6 (1.7)	2.2 (1.5)	3.0 (1.8)

^aSome columns may not add up to 100% due to missingness: 4 participants did not provide frequency of cannabis use (overall), 2 participants in the maladaptive group and 2 participants in the adaptive group.

Regarding EPPM coding, intercoder reliability^48,49 was acceptable at .90 for the second round.⁶⁹ EPPM responses were primarily maladaptive (57.9%). Table 2 shows EPPM response categories by subtopics in BERTopic analyses. Subtopics that were majority maladaptive included skepticism (i.e., 82.3% of skeptical responses were coded maladaptive), downplay severity (66.4%), uncertainty (65.0%), distrust (100%), indifference (89.2%), and negative emotions (67.9%). Mostly adaptive subtopics were mental health harm (75.8%), general harm (72.4%), risk recognition (81.5%), and responsible use (68.7%). The subtopic mixed response was approximately mixed for adaptive (49.4%) and maladaptive (50.6) codes.

Table 2.

Examples of EPPM responses about cannabis health warnings.

Response subtopics	Description	Examples responses	Subtopic counts	EPPM response categories
			N (%) (n/1078)	Maladaptive	Adaptive
Skepticism	Expresses some concerns about potential harms but strong defense of cannabis and skepticism of the warnings.	• “I don’t think cannabis is as dangerous as these messages are putting off. I’m sure it can cause some type of anxiety, but I don’t believe it is that harmful.” • “I can understand the cancer-causing particles as it is burnt plant matter but [the cannabis health warning] is completely undermining the health qualities.”	164 (15.2)	135 (82.3)	29 (17.7)
Downplay severity	Recognition of potential risks with cannabis, but downplaying severity or personal impact, often comparing to other harmful substances.	• “I personally do not believe overuse of cannabis causes mental problems. I do however agree with the operating of heavy machinery. I honestly feel alcohol is the worst instead of cannabis.” • “Yes, I believe excessive use of weed may lead to mental problems. I doubt the physical harm from smoking weed is the same as smoking cigs.”	143 (13.3)	95 (66.4)	48 (33.6)
Uncertainty	Uncertainty toward the truthfulness and accuracy of the cannabis health warnings.	• “Concerning. Scary. Not sure if it’s accurate.” • “Confused, not sure if I believe it all completely.”	120 (11.1)	78 (65.0)	42 (35)
Distrust	Distrust and rejection of cannabis warnings as inaccurate and misleading, e.g. calling them “propaganda.”	• “These ads are inaccurate and not factual at all.” • “To be honest, all of that just looked like fear propaganda.”	76 (7.1)	76 (100)	0 (0)
Indifference	Neutral or mixed reaction to cannabis warning ads, acknowledging some potential risks but not expressing strong concern or agreement.	• “The messages seemed mostly fair. Smoke is carcinogenic, you shouldn’t drive while high, and cannabis can have negative mental health effects on people prone to mental illness.” • “They [warnings] don’t really bother me much since I’m not a heavy user.”	74 (6.9)	66 (89.2)	8 (10.8)
Negative emotion	Emotional impact of cannabis warning ads focused on fear, worry, and concern.	• “I felt scared and nervous. I don’t want to experience any negative side effect.” • “Concerned, cynical, and inquisitive thoughts raced through my mind as I read each message. It just seemed too hard to believe.”	53 (4.9)	36 (67.9)	17 (32.1)
Mental health harm	Perception of significant harm from cannabis use, with emphasis on mental health and addiction concerns.	• “I agree excessive use can contribute to mental health problems. People should not be doing tasks that require coordination when high.” • “It would be a good idea to overcome my marijuana addiction before it creates more harm to my mental health.”	124 (11.5)	30 (24.2)	94 (75.8)
General harm	Perception of significant potential harm from cannabis use, highlighted by the warnings.	• “Cannabis use can be very harmful in that it can contribute to underlying issues, and as you get older can cause health problems.” • “Harmful, dangerous, not as safe as I thought.”	58 (5.4)	16 (27.6)	42 (72.4)
Risk recognition	Recognition of specific risks, primarily focused on avoiding driving high and operating machinery, with mixed views on other warnings (e.g., mental health, anxiety, depression).	• “I thought that these messages gave some valuable insight as to how cannabis can impair your senses especially if you are driving which is why it is not safe to use cannabis when driving.” • “It is common sense that you should not drive or operate heavy machinery while using cannabis, but I didn’t know that the smoke can cause cancer or that prolonged use can worsen anxiety or depression.”	54 (5.0)	10 (18.5)	44 (81.5)
Responsible use	Recognition of potential risks but focus on responsible use, emphasizing avoiding driving high and negative physical health effects.	• “You should never smoke and drive. Stay home or get situated before you smoke.” • “I already knew that cannabis use will cause drowsiness which is why I don’t use and drive just like I wouldn’t drink and drive.”	131 (12.2)	41 (31.3)	90 (68.7)
Mixed	Mixed reactions and planning for potential behavior change (i.e., moderate use), but also counterarguments and reasoning.	• “I agree in not driving while using it. I worry about younger generations who are starting to use it. They may not practice moderation. I think in my 20s I used it a lot. But as I grew older, I learned to use it in moderation.” • “I believe one does become impaired and should not be driving just like alcohol. Not sure about the harmful effects mentioned but believe anything in moderation should be minimal.”	81 (7.5)	41 (50.6)	40 (49.4)

Table 3 presents information on the main outcome measures (see Appendix Table 3 for bivariate correlation results). It was hypothesized (H1) that maladaptive EPPM responses to be associated with more resistance to cannabis health warnings (i.e., reactance) and less acceptance of health warnings (i.e., less accurate health beliefs and lower intentions to prevent harm from using cannabis). Table 4 displays linear regression results. Adjusting for covariates, maladaptive EPPM response (vs. adaptive) was associated with higher levels of reactance (β=0.38, Model 1), less accurate health beliefs about cannabis harms (β=−0.32, Model 2), and lower intentions to prevent harm from using cannabis (β=−0.31, Model 3), supporting H1.

Table 4.

Multiple linear regression models predicting main outcome measures.

					95% CI
Main outcome measures	B	β	SEβ	p	LL 2.5%	UL 2.5%
Reactance a
Frequency of cannabis use	0.33	0.05	0.22	0.14	−0.11	0.76
PCU risk	0.09	0.02	0.17	0.61	−0.25	0.43
Cannabis risk perceptions	−0.21	−0.14	0.04	<.001	−0.30	−0.12
Condition	0.48	0.09	0.14	<.001	0.20	0.76
EPPM response	2.02	0.38	0.15	<.001	1.72	2.31
Health Beliefs b
Frequency of cannabis use	−0.30	−0.06	0.16	0.07	−0.62	0.02
PCU risk	−0.48	−0.11	0.13	<.001	−0.73	−0.23
Cannabis risk perceptions	0.52	0.41	0.03	<.001	0.45	0.58
Condition	0.06	0.01	0.11	0.60	−0.15	0.26
EPPM response	−1.40	−0.32	0.11	<.001	−1.62	−1.19
Intentions to prevent harm from using cannabis c
Frequency of cannabis use	−0.93	−0.17	0.17	<.001	−1.26	−0.59
PCU risk	−0.79	−0.18	0.13	<.001	−1.05	−0.53
Cannabis risk perceptions	0.38	0.29	0.03	<.001	0.31	0.44
Condition	0.23	0.05	0.11	0.04	0.01	0.44
EPPM response	−1.39	−0.31	0.12	<.001	−1.62	−1.16

Notes. Boldface indicates statistical significance (p<0.05). B=unstandardized coefficient, β=standardized coefficient, SE_β=Standard Error of β, CI=confidence intervals, LL=lower level, UL=upper level, EPPM=Extended Parallel Process Model. All significance tests were two-sided. Frequency of cannabis use was coded as 0= less than monthly user, 1 = monthly user, PCU risk was coded as low (1), moderate (2), and higher (3). Condition was coded as 0 = text-only warning and 1 = pictorial warning. EPPM was coded as 1=maladaptive, 0=adaptive.

^aModel results were as follows: Model 1, reactance (R² = .20, R²_adj = .20, F[5, 1068] = 54.45, p < .001)

^bModel 2, health beliefs (R² = .36, R²_adj = .35, F[5, 1068] = 118.85, p < .001)

^cModel 3, intentions to prevent harm from using cannabis (R² = .33, R²_adj = .32, F[5, 1068] = 103.39, p < .001).

Discussion

This study employed an EPPM-response approach to cannabis health warnings using BERTopic and manual coding. Results provide data-driven and human-validated evidence about how US cannabis consumers evaluate health warnings from established cannabis control systems, offering timely data to authorities regulating cannabis products in US markets.

More than half of responses were coded as maladaptive. Subtopics with the largest counts of maladaptive codes were distrust (100% maladaptive codes), indifference, and skepticism. These results indicate considerable resistance to cannabis health warnings. This comports with trends in the US, where risk perceptions about cannabis have declined as market availability has expanded.¹⁸ Notably, the sample reported low averages for pre-exposure cannabis risk perceptions. Together, pre- and post-exposure measures indicated many in this sample underestimated the risk of cannabis harm and resisted warnings describing those risks.

There were less adaptive responses in the sample. Subtopics with the highest counts of adaptive codes were risk recognition, mental health harm, and general harm. Risk recognition focused on operating machinery (“It makes you drowsy and may impair the operation of heavy equipment”). Mental health harm was specific (“It [cannabis] can harm your mental health”), while general harm described general effects (“It [cannabis] can be dangerous and lead to long-term effects”). Thus, many participants acknowledged the possible harms of cannabis use.

Insights from EPPM coding and BERTopic modeling identifying nuanced perceptions among users to inform health warning design. For example, the distrust subtopic was coded entirely maladaptive. Distrust responses rejected warnings as inaccurate or misleading (“Literally a bunch of bullshit. Who’s out here lying like this?”). Some participants requested links to websites to verify warning claims (“Bring all scientific data when making safety claims”). Providing links is an audience-driven recommendation that may alleviate distrust and improve health warning effectiveness, although research should assess whether consumers are willing to follow links provided on health warnings. Using data-driven subtopic identification and theory-driven coding can help analyze responses to different message types (e.g., warnings vs. campaign messages) to identify targetable attitudes to reduce resistance.

In regression analyses, maladaptive responses were associated with greater reactance, less accurate health beliefs, and lower intentions to prevent harm from using cannabis. Positive beliefs about cannabis may be more resistant to change than substances with more established profiles. Consumers in the US have had decades of exposure to health warnings on tobacco and alcohol products or health campaigns about those substances. Since legal cannabis markets are relatively new to US consumers, there may be difficulty achieving early acceptance of cannabis health risks, especially for aspects of cannabis products that have changed over time with increasing potency—such as the level of addictiveness and risk of psychosis—where consumer awareness has not yet caught up.

Pictorial warnings (vs. text-only) were associated with greater reactance and greater intentions to prevent harm from using cannabis, but not health beliefs. This suggests adding imagery increased both reactivity to warnings and intentions to act in accordance with warnings, which aligns with research on reactance to cigarette warnings.⁷¹ Regulatory authorities should consider tradeoffs between attention-grabbing visuals and potential message rejection. Future research should explore designs that capture attention while minimizing defensiveness.

The EPPM does not theorize mixed responses, so they were not hypothesized. However, topic modeling revealed nearly balanced counts of maladaptive and adaptive responses, aligning with the human coding. The coding rule distinguished between intentions to protect against health threats (i.e., orienting toward protective action) and mixed responses without intentional language (i.e., indicating ongoing debate). Given the nuanced findings, studies should investigate when mixed EPPM responses occur to advance theoretical insights. Moreover, attitude change can lead to changes in behavior. While this classification of maladaptive response aligns with the EPPM, those maladaptive responses may be precursors of adaptive behaviors, especially since cannabis health harms may be novel to many consumers. The inclusion of human and machine coding revealed nuance in responses, and future research should evaluate how EPPM responses to cannabis health warnings affect behaviors over time.

Limitations and Future Directions

Limitations included self-report measures that may be influenced by social desirability bias. Responses were measured after exposure to multiple warnings. Future research should explore how specific health warnings influence EPPM responses and differentiate by specific risks. The study was cross-sectional, capturing a snapshot within an evolving policy landscape. Research should examine how attitudes toward cannabis warnings evolve as they become integrated into established public health messaging. Our sample excluded youth and those who had not used cannabis in the past year; future studies should include these groups to understand how they respond to warnings. Future studies could also test cannabis health warnings with text directing users to resources that enhance self-efficacy, creating EPPM interventions that combine educational materials with language supportive of adaptive responses. Finally, EPPM is one of many health-behavior frameworks that could assess reactions to cannabis health warnings, and research should continue to test and refine EPPM proposition alongside other theories.

Conclusions

As cannabis products become available in legal retail markets, there is a growing need to promote awareness about the health effects of cannabis. This study identified areas of consumer resistance (e.g., distrust) and acceptance (e.g., recognizing specific risks) to cannabis health warnings, including potential suggestions for improving messaging (e.g., including links to health claims). The goals of this study were to add to the literature on cannabis health warnings and inform health warning designers toward the goal of informing the public and reducing adverse cannabis health effects.

Data Availability

The data underlying this article will be shared on reasonable request to the corresponding author.