Breast cancer and alcohol

One in eight women in the United States (US) will be diagnosed with invasive breast cancer in their lifetime [1]. Current estimates suggest that 40% of breast cancers may be preventable [2]. One modifiable risk factor is alcohol consumption [3]. Alcohol is a Group I carcinogen [4,5], known to increase risk for breast cancer by increasing endogenous estrogen levels in pre-menopausal females. Estrogen levels are linked to higher breast density and breast cancer development [6–9], and as such, this is one of the biological mechanisms of alcohol’s effect on breast cancer risk. Importantly, the relationship between alcohol and breast cancer risk is dose-dependent with no lower threshold – meaning all levels of alcohol consumption increase breast cancer risk, and risk increases as consumption increases [3,4,10,11]. Pre-menopausal breast cancer risk increases by as much as 34% with one and a half drinks per day [3,10].

Recent cancer statistics indicate that the cancer burden for women has increased. The American Cancer Society estimates that, in the US, younger females have an 82% higher incidence of cancer than males, which is a significant increase from the incidence of 51% in 2002 [12]. This increase is concerning, especially in view of evidence demonstrating that the relationship between alcohol consumption and breast cancer risk is not well known publicly [13–16] and that breast cancer risk is frequently underestimated [17]. Lack of awareness and underestimation of the risk are problematic public health concerns, given that alcohol consumption remains prevalent [18]. While a recent Gallup Poll reported a decline in the prevalence of alcohol use among US adults to 54%, this is consistent with data on the prevalence of past 30-day alcohol consumption from the recent editions of the National Survey on Drug Use and Health [19,20]. Further, findings from a National Alcohol Survey analysis that compared 2019 and 2024 data show an increase in the volume of alcohol consumed among past-year drinkers, suggesting current drinkers are increasing their alcohol use [21]. This finding is particularly relevant to the cancer risk from alcohol given the dose dependency of the risk. Therefore, effective health communication is warranted to increase awareness and address this public health concern.

In January 2025, the US Surgeon General issued an advisory report recommending that health warning labels about cancer and alcohol be included on alcohol packaging [22]. Current US alcohol packages and product warnings do not include nor are they required to include information about cancer risk. Alcohol product warnings in the US remain unchanged since 1988 and use warning themes such as risks related to alcohol consumption during pregnancy, dangers related to operating heavy machinery, and a general statement that alcohol consumption may cause health problems [23]. The themes are not generally relevant, nor do they include potential consequences from the health harms (e.g., in utero effects or long-term injury/death). Other countries, such as South Korea and Canada [24–26], are making concerted efforts to inform the public about the carcinogenic effects of alcohol. Ireland was set to be the first country to require the inclusion of labels about the relationship directly on alcohol products, starting in May 2026; however the mandate has been delayed [27–29].

Health communication and evidence-based health warning messages (HWMs) are crucial for cancer control strategies [30]. Health warning messages that include pictures about tobacco health harms, such as cancer, have been studied and proven successful at reducing the appeal of tobacco products and intentions to use tobacco [31,32]. However, less is known about what type of HWMs regarding cancer risk and alcohol consumption would be the most effective in achieving a reduction in the appeal and intentions to consume alcohol. To date, the few studies about HWM attributes and their effectiveness and thus their potential use for US alcohol product labels have mostly focused on colon, liver, or oral cancers using narrative messages compared to pictorial messages. Narrative health warning messages are defined as messages that include pictures and/or words about people with a lived experience (e.g., a person with a feeding tube as a result of cancer). Pictorial messages, or the non-narrative health warnings in these studies, included graphic images of health consequences (e.g., pictures of cancerous lesions) [33–36]. Other research has examined breast cancer or cancer messaging (text and/or icons) alongside other alcohol health harms [37–39]. Our online experiment and the results we report here are among the first in the US to examine HWMs specifically and solely about alcohol consumption and breast cancer risk [40].

Understanding the potential impact of HWMs addressing alcohol and breast cancer risk is important because cross-sectional survey studies in other countries have demonstrated that support of alcohol control policies increases when awareness of the relationship between alcohol consumption and cancer risk increases [14,18,41,42]. Likewise, in an online experimental study in the United Kingdom, young adults’ interest in alcohol products was shown to decrease after they viewed alcohol health warnings on alcohol packages (including warnings about cancer risk) [43].

Evaluating health warning messages

Before HWMs can be placed on alcohol products, or disseminated in campaigns, they should be tested among the intended audience [31,44]. Testing involves understanding the format and content of the warnings as well as which message attributes elicit reactions, influence attitudes or beliefs, or indicate potential for behavior change [31]. One method of assessing HWMs is exposing participants to hypothetical HWM designs and evaluating their perceptions and reactions via spontaneous reactions provided in open-ended response opportunities. Including open-ended response opportunities or thought-listing tasks alongside closed-ended survey items is an often-used method to evaluate attitudes of participants that may not have been otherwise assessed [45,46]. Open-ended response opportunities allow participants to freely express their immediate perceptions about health warnings, often providing rich insights that may not be captured in survey items on their own terms, without the confines of the researchers’ closed-ended items [45–48]. No studies, to our knowledge, have assessed HWM reactions and perceptions within open-ended and spontaneous responses to HWMs about alcohol consumption and increased breast cancer risk.

The message impact framework

This study was informed by the Message Impact Framework, derived from meta-analyses of tobacco health warning research [44,49]. The Message Impact Framework identifies constructs “sets” that are theorized as antecedents of behavior change. The framework posits that message attributes influence individual message reactions, such as attention to the warnings and emotional reactions. Message reactions are posited to impact attitudes and beliefs, which include participants’ perceptions about the severity of the harm, their susceptibility to the harm, and their efficacy to respond to the health harm. Lastly, attitudes and beliefs are posited to influence behavioral intentions and health behaviors such as reducing or stopping alcohol consumption [44,49]. For this study, we used content analysis techniques to code the open-ended spontaneous responses to HWMs guided by the Message Impact Framework.

Objectives

The open-ended responses (hereafter referred to as responses) were collected during an experimental study evaluating different alcohol and breast cancer HWM designs [40]. Here, we report qualitative directed content analysis and findings, the corresponding quantitative analysis, and results not reported elsewhere. Coding was guided by the variable sets within the Message Impact Framework, which were used for directed content analysis of the responses [50–53]. Our goal was to answer the following questions:

1. 1) What components of a) message reactions, b) attitudes and beliefs, and c) behavioral intentions to reduce or stop drinking alcohol were present in participants’ responses after health warning message exposure?
2. 2) When considering the type of health warning message (text-only or picture-and-text) and the health harm that participants were exposed to, were there similarities or differences in: a) message reactions, b) attitudes and beliefs, and c) behavioral intentions to reduce or stop drinking alcohol present in participants’ responses?
3. 3) When considering the health harm participants were exposed to and select demographics (e.g., age or education), were there similarities or differences in: a) message reactions, b) attitudes and beliefs, and c) behavioral intentions to reduce or stop drinking alcohol present in participants’ responses?

Participants

This study analyzed data from an online experimental study evaluating different alcohol and breast cancer HWM designs and attributes [40]. We evaluated the perceptions of young adult female participants in the US who self-reported alcohol consumption in the last 30 days after exposure to text-only or picture-and-text HWMs about how alcohol consumption can increase risk for breast cancer (described below). Data collection occurred from April 1 to April 24, 2024, and was facilitated by the Qualtrics social marketing research firm. Qualtrics recruited and compensated participants from a panel, and the research team designed all experimental materials. Sample quotas were approximated using US Census data and estimates of the national population of females who drink [54,55]. The University of Missouri Institutional Review Board approved this study (2099603). Our reporting here is guided by the Consolidated Criteria for Reporting Qualitative Research Checklist (COREQ) [56] and content analysis reporting guidelines [53].

Procedure and design

In an online experiment, participants provided informed consent before completing baseline demographic and substance use questions. Participants were randomly assigned to view one standalone HWM (text-only or picture-and-text) about the relationship between alcohol consumption and breast cancer risk along with a description of one of three breast cancer health harms (mastectomy, hair loss from chemotherapy, and mortality). After exposure to the HWMs, participants were presented with an open-ended thought-listing task eliciting their immediate response to the HWM. Outcome measures, guided by the Message Impact Framework variable groups, were then assessed using additional survey items. The end of the survey included a debrief about the research and contact information for the Substance Abuse and Mental Health Services Administration for resources about alcohol consumption and a website link to healthywomen.org for resources about breast health.

Experimental stimuli – health warning messages

We designed HWMs to include the four components theorized for effective warnings [30]: a signal word (i.e., “Warning!”), identification of the risk (i.e., the connection between alcohol consumption and breast cancer risk) description of the effect (i.e., a single health harm resulting from cancer (plus or minus a picture depicting the harm), and steps that could help avoid the risk (i.e., a self-efficacy statement about alcohol consumption reduction and speaking with a healthcare provider about decreasing risk). The HWMs appeared on the screen alone and all text was the same font size, white, and placed on a black background. The picture-and-text warnings were formatted with the same sized graphic images based on previous research testing tobacco health warning designs [57,58]. All HWMs were formatted to match in number of words, reading level, and design. In our pilot with a national sample, draft HWMs using the topics of mortality, hair loss from chemotherapy, lymphedema, and mortality were tested. The health harms with the highest mean scores for perceived message effectiveness, mastectomy, mortality, and hair loss from chemotherapy, were used for our main experiment. Additional description, examples of the HWMs, and full results of the experiment, including the quantitative assessment of the variable groups from the Message Impact Framework, are reported elsewhere [40].

Measures

Alcohol use status

Alcohol use status was measured using alcohol consumption frequency with an item asking, “How many times have you had at least one drink of alcohol in the past 30 days?” followed by a prompt to select from 1 to 30 times.

Open-ended immediate response

The unit of analysis for this study was the response provided by participants who viewed the HWMs with health harms about mastectomy, hair loss from chemotherapy, and mortality. Participants encountered the response opportunity before completing the remaining survey items related to the Message Impact Framework variable sets. Participants were prompted to: “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.” [59]. The prompt was shown on the same survey screen as its corresponding textbox where free text entry was enabled without word count limits. Participants needed to enter content (even blank spaces) into this area before proceeding with the rest of the survey.

Data preparation

We excluded 21 blank/unusable responses from the HWM health harms (responses such as “I’m” or “gut”). This resulted in a total of 748 responses used for analysis. The number of participants who viewed the different types of messages (i.e., text-only or picture-and-text and health harms) is presented in Table 3. Chi-square tests showed experimental assignment to health harm and HWM type (text-only or picture-and-text) and demographic characteristics remained balanced after data preparation [40].

Approach to analysis

Data were analyzed in three steps. Step 1 included assessing the responses to identify the constructs from the Message Impact Framework using directed content analysis [50–52]. Step 2 included assessing the association between response codes and message types (text-only or picture-and-text) and health harms (mastectomy, hair loss from chemotherapy, and mortality), respectively. Step 3 included assessing the moderating effects of the associations between demographic traits and response codes. All the qualitative analysis was done using Dedoose software, while quantitative analyses were done using SPSS software.

Step 1 directed content analysis

Use of directed content analysis provided a systematic approach to qualitative analysis given the high number of open-ended responses collected as part of a larger quantitative experimental study [53]. In Dedoose, each response was labeled with a code and participant descriptors (i.e., age, race, ethnicity, education, income, health status, and drinking history, and drinking history).

Initial codebook creation and structure

The research team deductively created the codebook before coding was initiated and inductively updated it during analysis as additional patterns in the responses were identified [50–52]. Thus, the majority of our code labels were created using the constructs and operationalizations from the Message Impact Framework [44] (See Table 1). These included the overarching categories (parent code labels) based on the theoretical constructs of the Message Impact Framework, including message reactions, attitudes/beliefs, and behavioral intentions, as well as subcategories of corresponding variables within each of these sets. Example subcategories included attention for message reactions and stop drinking for behavioral intentions. Within the codebook, we documented key words and phrases that related to the parent code labels and corresponding variables as subcategories. The definitions and list of keywords evolved as coding progressed, however initial keywords/definitions always served as guideposts or coding rules [52].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Codebook and Danalysis as described by Hsieh definitions.

https://doi.org/10.1371/journal.pone.0338687.t001

We initiated directed content analysis as described by Hsieh and Shannon (2005) using the predefined code labels within the codebook. Each response was coded independently and according to the codebook by a minimum of two research team members who were also involved in study design, data preparation, and the development of the codebook. Responses that warranted additional code labels that were not listed in the initial codebook were flagged and discussed (e.g., responses to be labeled later as sad, added as a subcategory of the parent code message reactions). Interim results and discrepancies were discussed during analysis-specific team meetings. We maintained an audit trail within the codebook as code label definitions evolved or required clarifications/discussion [51]. Parent and subcategory code labels were collapsed when discerning between two categories or corresponding Message Impact Framework variables became less discrete or more complex. For example, the code labels related to the variables of perceived susceptibility and perceived threat were adjusted to perceived threat and first person and/or perceived efficacy. Similarly, we added subcategories to message reactions such as sad and shock or surprise when the team noticed their frequency. Table 1 is the final codebook.

Verification of the appropriateness of code labels and their application to responses was ongoing as the research team reviewed each flagged response to reach consensus. To validate rigor and consistency, we viewed responses in groups by code label, confirming reliability and identifying patterns of language within the grouped responses. Patterns of language contributed to objectivity and trustworthiness as the research team verified that the responses were grouped appropriately by code label. We then assessed whether a certain element or variable set of the Message Impact Framework appeared more or less frequently in certain participants’ responses.

Step 2 Chi-square tests of independence

Once each response was labeled, we were able to group responses by Message Impact Framework constructs and subcategories. For example, we identified the number of responses that were labeled with the attention subcategory of message reactions. This allowed us to the code labels (i.e., response types such as attention) as dependent categorical variables and conduct analysis to answer research question 2, to examine whether the message types (text-only or picture-and-text) were associated with participants’ response types (e.g., attention). We analyzed the four parent code labels (message reactions, attitudes and beliefs, behavioral intentions, and new information) and their highest frequency subcodes (attention, first person and/or perceived efficacy, and stop drinking). We used SPSS to conduct Chi-square tests of independence.

Step 3 Moderation analysis

To address research question 3, we used moderation analyses (ANOVA and regression) to test whether demographic characteristics, including age, race, ethnicity, education, income, health status, and drinking frequency, would have a moderating effect on participants’ response types (again analyzed as the dependent categorical variable). For race and ethnicity, a categorical variable, we used two-way factorial ANOVA to examine whether the effect of health harm HWM on response type (e.g., responses labeled with attention) was moderated by race and ethnicity respectively. For age, income, education, health status, and drinking frequency, which were treated as dummy coded continuous variables, we used hierarchical multiple linear regression to test moderation effects. For the analysis for research questions 2 and 3, all statistical significance was evaluated at the 0.01 level to minimize the risk of Type I error.

Sample and dataset characteristics

Table 2 shows all participant characteristics. We included open-ended, substantive responses from 748 female young adult participants between 21 and 29 years of age (M = 25.3, SD = 2.6) who reported consuming alcohol in the past 30-days and viewed the health harm HWMs. Most participants identified as White (71.7%), followed by Black or African American (11.5%), Asian (4.5%), American Indian or Alaskan Native (2.0%), Native Hawaiian or Other Pacific Islander (0.4%), more than 1 race (4.8%), or another race not listed (5.1%). Most did not identify as Hispanic, Latina, Latinx, or of Spanish origin (81.8%). In terms of education, 26.7% of participants were high school graduates, while 24.7% had some college education, followed by bachelor’s degree (19.1%), associate’s degree (10.6%), master’s degree (9.4%), less than high school (5.7%), and professional or doctoral degree (3.7%). 35.8% of participants reported their annual household income as less than $50,000, while 35.0% of people reported it between $50,000 and $100,00. Most participants perceived their health as good (43.2%)

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Sample characteristics overall and by health harm HWM. All participants (n = 748) reported currently drinking alcohol.

https://doi.org/10.1371/journal.pone.0338687.t002

All participants self-reported having consumed alcohol at least one day in the last 30 days. The average drinking frequency was 8.7 times (SD = 7.4) in the prior 30 days. Alcohol consumption frequency distribution is displayed in Fig 1 below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The distribution of participants’ drinking frequency over last 30 days.

https://doi.org/10.1371/journal.pone.0338687.g001

Participants who viewed one of the health harm HWMs took a median time of 10.9 minutes to complete the entire survey. Responses ranged from 2 to 1296 characters in length, including spaces. Mean length was 118 characters (SD = 107) with a median of 93 characters. The average number of code labels applied to a response was 3 (SD = 2, range = 1–11; 11 occurring only once). This average number of code labels includes parent code and subcategory code labels.

Research question 1 – types of responses

Message impact framework constructs and sub-categories

Message reactions.

Message reactions was the most frequently used parent code, applied to 545 (72.9%) of the excerpts, meaning 545 participants included a message reaction in their responses to the HWM. We treated the subcategory code labels of attention and cognitive elaboration as types or gradations of message reactions. The code label attention was applied to excerpts that included simpler message reactions and restated the HWM content, such as “Drinking alcohol increases risk for breast cancer.” We applied the code attention to 165 excerpts (22.1%). The code label cognitive elaboration was applied to excerpts that expanded upon the HWM content, implying some level of attention or comprehension but also the ability to apply the content beyond reiteration of the message. Examples of cognitive elaboration excerpts are, “It made me think more deeply about the effects of alcohol” and “Makes me sad because I lost my ex due to a drinking problem…”. We applied the subcode cognitive elaboration to 148 excerpts (19.8%). Table 3 includes the code labels alongside exemplar responses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Code labels, exemplar responses, and percentage of codes applied in total and by health harm exposure (combined text-only HWM and picture-and-text HWM).

https://doi.org/10.1371/journal.pone.0338687.t003

This last example response also received the code label of sad. We delineated emotions using the subcategory code labels: sad, hope, shock/surprise, and fear/anxiety/worry. The code label fear/anxiety/worry was applied the most frequently with 157 responses (21.0%). The code labels sad and shock/surprise were applied to 47 (6.3%) and 43 (5.7%) of the responses, respectively. The code label of hope was applied to 10 responses (1.3%).

Ninety-one participants (12.2%) included comments about the survey. These responses included words or phrases such as “the survey was” or “it was” such as, “I thought it was very informative.” Twenty-four participants responded only about the survey, meaning their responses were only coded with the survey specific parent code. This also means that 67 (9.0%) participants responded specifically about the survey but also with additional content that fell within the Message Impact Framework construct categories.

Attitudes and beliefs

The parent code label attitudes and beliefs applied to 435 responses (58.2%). The subcategory code labels for attitudes and beliefs evolved during analysis because this was one construct of the Message Impact Framework that was more nuanced or qualitatively subtle. We realized the risk of introducing bias and assumptions when evaluating what the perceptions of participants were when we only had the one written response without follow-up. For these reasons, we included two subcategories for attitudes and beliefs that we, as a research team, could objectively delineate between – first person and/or perceived efficacy and perceived threat.

First person and/or perceived efficacy labeled the I or me responses about the impact of the threat and their ability or lack thereof to address the threat. For instance, “The messages made me feel that maybe I could prevent myself or my family members from getting breast cancer by avoiding alcohol…” This code label was applied to 200 responses (26.7%) where participants acknowledged (or dismissed) the risk and its application to themselves. We used one subcategory label for the responses that dismissed the current threat. Dismissal of current threat was applied to 62 of the responses (8.3%) where participants acknowledged the threat in the HWM but also dismissed or deferred the threat, indicating that they were not at risk because of current drinking habits, age, or health status. An example of a response coded with this label is “I don’t really drink alcohol much so those messages didn’t affect me. I understand the effects of alcohol but it can’t affect me if I don’t drink it hardly ever.”

Perceived threat was applied to 149 responses (19.9%) where participants acknowledged the health harm for others, but not necessarily themselves. An example of a response coded with perceived threat is “It makes me think how easily alcohol can affect someone’s mind and body. Drinking alcohol deteriorates your mind which is linked to your body and how it responds to the amount liquor.” The subcategory of thinking about family/friends was applied to 73 responses (9.8%) to detail the number of participants who mentioned thinking of or telling their sister, mother, or friends about the relationship between alcohol and breast cancer risk. Responses receiving this code indicated participants’ application of how the health harm might impact their loved ones and/or an intention for future interpersonal communication about alcohol and increased risk for breast cancer.

Participants responded about how the HWMs presented information that was new to them. We applied the new information code label to 151 responses (20.2%). Responses receiving this code label included phrases such as “I didn’t know alcohol increases risk for breast cancer” or “This is new information to me,” and “I did not know drinking could lead to breast cancer….This needs to be told to everyone.” We applied the code seeking information to 49 responses (6.6%). Responses receiving this code label included phrases such as “I want to know the source” or “what is the data behind this message.”

Behavioral intentions

The parent code behavioral intentions was applied to 65 excerpts (8.7%) including the subcategory code labels of reduce drinking (18 responses, 2.4%) and stop drinking (48 responses, 6.4%)

Research question 2 – type of HWM (text-only or picture-and-text) and health harm and Written Response

There were no significant associations between the type of message viewed (text-only or picture-and-text) and the type of response. This included message reactions (χ² (1, N = 748) = 1.49, p = .22) and its subcode attention (χ² (1, N = 748) = 0.18, p = .67), attitudes and beliefs (χ² (1, N = 748) = 1.01, p = .32) and its subcode first person and/or perceived efficacy (χ² (1, N = 748) = 1.85, p = .17), behavioral intentions (χ² (1, N = 748) = 0.55, p = .46) and its subcode stop drinking (χ² (1, N = 748) = 2.46, p = .12).

No significant associations between the health harm viewed (mastectomy, hair loss from chemotherapy, and mortality) and type of response (text-only or picture-and-text) were identified. This included within message reactions (χ² (2, N = 748) = 0.97, p = .62) and its subcode attention (χ² (2, N = 748) = 0.58, p = .75), attitudes and beliefs (χ² (2, N = 748) = 0.40, p = .82) and its subcode first person and/or perceived efficacy (χ² (2, N = 748) = 2.46, p = .29), behavioral intentions (χ² (2, N = 748) = 4.81, p = .09) and its subcode stop drinking (χ² (2, N = 748) = 2.96, p = .23).

Research question 3 – type of health harm, response, and demographic moderators

Age was the only variable tested that was associated with type of response. Although the age of participants was restricted between 20–29 years for our study, age was found to have a significant moderating effect on the hair loss from chemotherapy health harm HWMs (β = −0.04, p = .005) in terms of responses coded as attitudes and beliefs. Specifically, when participants were exposed to the hair loss from chemotherapy condition messages, an increase in age was associated with a lower likelihood of responding with attitude and belief related words or phrases.

Limitations

Within the survey, the open-ended response opportunity came after a series of demographic and other assessments. Participants could have been aware of the time spent on survey completion so far and typed less to finish sooner, thought less thoroughly about the HWM, or typed fewer words to spend less time. Likewise, the cross-sectional design did not allow for follow-up questions. Future studies examining the topic of alcohol consumption and breast cancer risk may consider additional in-depth qualitative inquiry to further assess health warning messaging attributes’ ability to change alcohol consumption behaviors. We used directed content analysis that allows for discussion and adjustments during the analysis process. We took many steps to reduce the risk of subjectivity, and our findings add to the growing body of science about HWMs in the alcohol and cancer space. Our sample did report a wide range of drinking frequencies. Future studies should further examine HWMs about alcohol and breast cancer risk among samples of nondrinkers and moderate to heavy drinkers. Lastly, although we found the moderating effect of age in this study, we restricted the age range to focus on the young adult group. Therefore, we should be cautious when interpreting age’s moderating effects to a broader age group. Future studies should further explore the effect of demographic characteristics, especially age, on participants’ reactions to messages, providing more practical guidance for health campaign message design.