Pictorial Warning Labels and Memory for Cigarette Health-risk Information Over Time

Abstract

Background

Pictorial cigarette warning labels are thought to increase risk knowledge, but experimental research has not examined longer-term effects on memory for health risks named in text.

Purpose

To investigate memory-consolidation predictions that high- versus low-emotion warnings would support better long-term memory for named cigarette health risks and to test a mediational model of warning-label effects through memory on risk perceptions and quit intentions.

Methods

A combined sample of U.S.-representative adult smokers, U.S.-representative teen smokers/vulnerable smokers, and Appalachian-representative adult smokers were randomly assigned to a warning-label condition (High-emotion pictorial, Low-emotion pictorial, Text-only) in which they were exposed four times to nine warning labels and reported emotional reactions and elaboration. Memory of warning-label risk information, smoking risk perceptions, and quit intentions were assessed immediately after exposures or 6 weeks later.

Results

Recall of warning-label text was low across the samples and supported memory-consolidation predictions. Specifically, immediate recall was highest for Low-emotion warnings that elicited the least emotion, but recall also declined the most over time in this condition, leaving its 6-week recall lowest; 6-week recall was similar for High-emotion and Text-only warnings. Greater recall was associated with higher risk perceptions and greater quit intentions and mediated part of warning-label effects on these important smoking outcomes. High-emotion warnings had additional non–memory-related effects on risk perceptions and quit intentions that were superior to text-only warnings.

Conclusions

High- but not Low-emotion pictorial warning labels may support the Food and Drug Administration’s primary goal to “effectively convey the negative health consequences of smoking.”

ClinicalTrials.gov Identifier

NCT03375840

Smoking, Pictorial warning labels, Risk knowledge, Emotion, Memory, Risk perceptions, Quit intentions

Smoking is the leading cause of preventable death in the USA with more than 480,000 smoking-related deaths per year. Mortalities are high because smoking is linked to diseases of most major organs in the body including lung diseases, liver and colorectal cancers, diabetes, rheumatoid arthritis, and cardiovascular disease [1]. On average, smokers live 10 years less than nonsmokers, but smoking cessation has a significant impact on longevity even into a smoker’s 40’s and 50’s [2]. Smokers’ knowledge of tobacco’s health risks, however, is low, and most adults can name only one to four of the many smoking-related diseases [3–5]. In the present study, we examined short-term and longer-term memory for health risks presented in the text of cigarette warning labels among adult smokers, teen smokers/vulnerable smokers, and Appalachian adult smokers to examine generalizability of effects.

Pictorial warning labels on cigarette packages were first implemented in Canada in 2001 and are now required in at least 100 countries [6]. They are intended, at least in part, to increase knowledge of cigarette’s health effects. To measure warning-label knowledge, researchers generally assess memory (free recall, recognition), but experimental research has examined only shorter term or non–disease-specific memory effects as reviewed below. This paper’s primary objective was to test the potential power of negative emotional reactions created by warning labels to increase memory for smoking risks presented in the text immediately and 6 weeks after exposure to them.

Survey data from countries that require pictorial warnings suggest that they generally improve risk knowledge over text-only warnings especially when risks are new and/or the subject of recent warnings [7–9]. These effects may be due to the labels being seen many times a day; a pack-a-day smoker, for example, will be exposed 20 times a day and 7,300 times per year to his or her cigarette packs. Alternatively, warning labels are often part of larger anti-smoking efforts which may also contribute to greater knowledge. Nonetheless, in Canada, which requires pictorial warnings, 84% of smokers cited cigarette packages as a source of information about smoking’s risks versus only 47% of U.S. smokers whose packages contain only text warnings [4]. In this same study, smokers who reported noticing pictorial warning labels were 1.5 to 3.0 times more likely to endorse smoking’s negative health effects, and those who read, thought about, and discussed health warnings were more likely to intend to quit.

Furthermore, pictorial warning labels’ effects on memory may be enhanced by them eliciting negative emotional reactions to smoking [10]. In fact, past psychological research has demonstrated that emotional reactions (both positive and negative) generally increase memory for associated information [11–13]. Emotional reactions also can produce greater cognitive processing and elaboration of information congruent with the reaction, with subsequent effects on decision-making [14], perhaps because it improves the memory of information [15]. Stronger emotional reactions, however, may also be associated with greater incorrect recognition of related information as people confuse information they saw in one context (e.g. the present experiment) with related information seen in some prior context [16].

Consistent with memory research, survey and experimental studies have demonstrated that smokers exposed to pictorial versus text-only warnings scrutinized and elaborated more on associated health risks, and they recalled more health-risk information [3, 17, 18]. Thus, greater warning-label-elicited emotional reaction should increase processing and generally improve memory for health risks.

However, pictorial warning reactions have been linked to less short-term understanding of tobacco-related health information. Kees, Burton, Andrews, and Kozup [19] found that only 60% of participants recalled a health problem related to the warning’s text immediately after the presentation of a more emotional pictorial warning compared with 78% and 70%, respectively, who recalled it when it had been associated with a less emotional pictorial warning or a text-only warning. These immediate-memory results, however, failed to consider emotion’s enduring effects on memory [13, 20]. Long-term memories are not made instantaneously; instead, they consolidate over time. High emotion appears to cause a reverberating memory trace that interferes with immediate memory but produces better long-term memory. Low-emotion memories show forgetting over time, whereas high-emotion memories either improve or decline less as the greater emotion focuses attention, causes rehearsal, and activates hormonal and brain systems that regulate long-term memory storage [21]. In fact, experimental studies support more correct recognitions of entire pictorial warning labels than text-only labels after 20 minutes [10] and greater recall of some piece of information from the text or images 1 week later [22], but neither study tested memory for health effects mentioned in the text. Thus, prior studies have generally been observational, and experimental studies that tested delayed memory did not assess the understanding of specific tobacco-related health risks mentioned in the text (operationalized here as risk recognition and recall), an important component of public health education.

Unlike most prior research that exposed participants once and to fewer labels than that mandated by the Food & Drug Administration (e.g. [19]), in the present study, we tested the effect of four exposures to each of nine labels. Participants in each of three representative samples of smokers (U.S. adults, Appalachian adults, and U.S. teen smokers/vulnerable smokers) were randomized to one of three warning-label sets selected a priori to differ in emotional reaction (Text-only, Low-emotion pictorial, High-emotion pictorial). We then examined smoking-risk memory either immediately postexposure or after 6 weeks (participants were also randomized separately in each sample to delay condition). Our operationalization allowed us to model how recognition and recall of key risk concepts in the text of nine warnings (e.g. “death”) related to one another and to recognition of related nontextual risks, risk perceptions, and quit intentions. We used a dimensional approach to characterize emotion, the circumplex model, that conceptualizes emotion as arising from feelings based in valence (good vs. bad) and arousal (calm vs. excited; e.g. [23]). Because negative arousal drives people to prepare for action [24], which may allow them to avoid health and other hazards, we chose only negative-valence warnings and focused on the effects of arousal (called emotional reaction from hereon). We hypothesized that:

 

Hypothesis 1: Health risks in the text of a set of high-emotion pictorial warnings would be remembered less well than those in a low-emotion pictorial set or a text-only set immediately after exposure.

Greater memory consolidation then should emerge over time with more emotion. The observed effect may concur with Kleinsmith and Kaplan’s [20] original findings of memory improvement for high-emotion stimuli and memory decline for low-emotion stimuli. However, a confound existed in their original studies, and subsequent results have pointed to less memory decline over time (rather than memory enhancement) with greater emotion (e.g. [13, 25]).

 

Hypothesis 2: Memory for health risks in the text of more versus less emotional warning-label sets would improve or decline less over time such that the highest emotion set would be remembered better than the lowest emotion set after a 6-week delay.

Finally, to test whether memory for cigarette health risks might affect tobacco usage, we conducted structural equation modeling (SEM). The goal of the model was to examine both recognition memory and free recall as possible results of emotional reaction to and elaboration of the warnings and as potential influences on risk perceptions and quit intentions. Our hypothesized initial SEMs were developed from past research demonstrating that emotional reaction is an important mediator of pictorial warning labels’ impact on elaboration, risk perceptions, and quit intentions [3, 26] and an important causal factor in memory [13]. We then modeled elaboration as leading to recall and recognition. Recognition itself was modeled as a precursor of recall memory because recognition is simpler and thought to be a subset of processes required for free recall and/or to share processes with free recall [27, 28]; free recall, of course, is affected by other self-initiated processing [29]. Because recognition memory is influenced by other processes too (e.g. familiarity, [29]), it was also modeled as potentially important to risk perceptions and quit intentions. As noted earlier, we expected more emotional warning-label sets to drive greater correct and nontextual recognition of tobacco risks (compared with less emotional label sets) as smokers correctly remember tobacco’s risks more often but also confuse tobacco health risks that were presented in the text of the warnings with other smoking-related health risks that were not shown [16].

 

Hypothesis 3: Negative emotion evoked by warnings would mediate effects of warning labels on warning elaboration, memory, and risk perceptions, each of which would mediate, in turn, effects on quit intentions.

Method

Participants and Design

In 2015, nationally representative teen and adult samples and a regionally representative Appalachian sample were recruited through an internet survey company, YouGov (see [26] for additional sample details). The sampling frame for the national adult population was constructed from the 2014 National Health Interview Survey; for Appalachian smokers, the 2011 Tobacco Use Supplement to the Current Population Survey was used due to its larger sample size and representativeness in smaller geographic areas. Adult panelists were 19–65 years old, had smoked 100+ lifetime cigarettes, and currently smoked “every day” or “some days.” The teen sampling frame was constructed from the 2011–2012 National Health and Nutrition Examination Survey and included smokers and vulnerable smokers. Eligible participants were 14–18 years old and had “ever tried or experimented with cigarette smoking, even a few puffs.” Across all samples, 4,424 people completed baseline (N_adult=1,272, N_Appalachian = 1,620, N_teen=1,532); 1,932 completed the entire study (N_adult=736, N_Appalachian = 727, N_teen=469). Sample sizes were determined with a power analysis designed to detect the two-way interaction between warning-set emotionality and time delay, separately for each sample (see [26]). The teen sample did not reach the intended size.

Separately within each sample, participants were randomly assigned to one cell of a 3 (warning-label set: Text-only, Low-emotion pictorial, High-emotion pictorial) × 2 (measures delay: immediate, 6-weeks) completely between-participants design (participant flow chart in Supplementary Figure 1). Participants viewed the same set of nine warnings four times over 2 weeks: Exposure 1 occurred at baseline (Time 1, they rated emotional reaction to and credibility of each label). Exposures 2 and 3 occurred 1 week after baseline in the same session (Time 2, they evaluated perceived matching of picture and text—or were simply exposed in the text-only condition—and warning relevance); Exposure 4 occurred 2 weeks after baseline (Time 3, they evaluated emotional reactions and credibility). Either immediately after the last exposure (i.e. Time 3) or 6 weeks after the last exposure (Time 4), participants responded to dependent measures including free recall memory, recognition memory, risk perceptions, and quit intentions (for a list of all measures, correlations, and descriptive statistics, see Figure 1 and Table 1, respectively). Participants viewed warnings and completed all measures over the Internet on their own electronic devices. Some measures not examined here are studied in other published papers (see [26, 30]).

Fig. 1.

Overview of study procedure. Participants viewed each warning label from their experimental condition four times in total. At exposure three, participants in the image conditions rated the extent to which the images matched with the warning text. Participants in the text-only condition were also exposed to their warnings, but were not asked to answer the question about the image.

Table 1

Correlations of condition variables, emotional reactions, risk elaboration, memory variables, risk perceptions, quit intentions, and demographic variables used as covariates in the full sample

		1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16
1	Delay
2	PWL: High vs. Low	.07**
3	PWL: Text vs. image	−.02	−.03
4	Emot. react.	−.01	.17**	.04
5	Risk elaboration	.02	.08**	−.02	.47**
6	Correct recog.	.05*	.11**	−.05*	.22**	.14**
7	Nontextual recog	.23**	.06**	−.06**	.21**	.30**	.44**
8	Incorrect recog: non-risks	.09**	.07**	.03	.13**	.29**	.25**	.67**
9	Risk perceptions	.02	.01	.00	.41**	.28**	.17**	.16**	.06**
10	Quit intentions	−.06**	.03	.00	.40**	.35**	.08*	.06*	.03	.36**
11	Correct recall	−.26**	−.04	−.03	.09*	−.02	.20**	−.22**	−.21**	.12**	.12**
12	Age	.03	.05*	−.02	−.02	−.15**	.09**	−.01	−.09**	−.01	−.21**	−.08**
13	Gender	−.02	.02	.02	.10**	−.03	−.02	−.09**	−.14**	.04	.02	.08**	.00
14	Race	−.01	−.01	.05*	−.07**	−.20**	.06*	−.04	−.09**	−.01	−.12**	.03	.13**	.11**
15	Educ.	.00	.00	−.01	.01	−.06**	.01	.01	.01	.04	−.05*	.00	.27**	−.06*	−.01
16	Adults vs. other	.02	.02	−.02	−.04	−.09**	.02	.00	−.03	.06**	−.08**	.01	.31**	.00	−.06**	.28**
17	Teens vs. other	−.02	.01	−.01	.08*	.11**	−.02	−.01	.03	−.04	.26**	.09**	−.72**	.00	−.17**	−.38**	−.44**
Descriptive statistics
Minimum					1.00	1.00	0.00	0.00	0.00	−3.00	−1.99	0.00
Maximum					5.00	5.00	1.00	1.00	1.00	3.00	1.59	0.89
Mean					3.17	2.68	0.71	0.21	0.09	1.18	0.00	0.25
SD					0.95	1.34	0.21	0.27	0.22	1.28	0.76	0.19
Cronbach’s α					.96	.89	.81	.77	.87	.90	.64	.54

Descriptive statistics for indices are included at the bottom of the table. Delay condition is coded 0 = immediate, 1 = delay; PWL condition is High- versus Low-emotion (−1 = Low emotion, 0 = Text, 1 = High emotion) and Text versus image (1 = Low emotion, 1 = High emotion, −2 = Text). Gender is 0 = male, 1 = female, race is 0 = nonwhite, 1 = non-Hispanic white, education is 0 = high school or less, 1 = more than high school, adult versus other is 1 = adult sample, 0 = teen sample, 0 = Appalachian sample, teens versus other is teens = 1, adults = 0, Appalachian = 0. *p < .05, **p < .01.

Materials and Procedure (Clinical Trial Registration # NCT03375840)

The study’s nine text warnings were from the 2009 Family Smoking Prevention and Tobacco Act and originally mandated by the Food & Drug Administration, but are not currently in use in the USA (e.g. “WARNING: Cigarettes cause stroke and heart disease”). We did not use the nine Food & Drug Administration-mandated images; instead in the High-emotion condition, the text warnings were paired with matching images pre-tested to elicit strong negative emotion to smoking from smokers. In the Low-emotion condition, the matching images paired with the text warnings were pre-tested to elicit less negative emotion. Thus, all warnings were negative in valence but differed in arousal. See Figure 2 for example stimuli and Supplementary Text 1 for pretesting details.

Fig. 2.

Sample warning labels. Example warning labels by experimental condition. The same text warnings were used in all three conditions and shown on a white background, without cigarette packages. The text warnings were formatted to mimic black-and-white text warnings on cigarette packages; text was sized comparably across conditions (approximately 375 × 120 pixels). Warnings appeared at approximately 375 × 368 pixels on participants’ monitors. The total size of warnings viewed by participants in the text-only condition was smaller than that of participants in the pictorial image conditions. Center image purchased via iStockphoto.com/Dmytro Sobko; Right image courtesy of Food & Drug Administration Center for Tobacco Products.

Emotional reaction

On the first and fourth warning exposures (Times 1 and 3), participants rated emotional reaction to each warning label using the Self-Assessment Manikin ([31]; see [26], using the same adult and teen samples). Participants selected one of five stick figures to describe how they felt about each warning from “calm, drowsy and peaceful on the left to excited, energized and alert on the right.” These reactions were coded from 1=calm to 5=excited (Cronbach’s α_{Exposure 1}= .95; Cronbach’s α_{Exposure 4}=.95).

Warning elaboration

All participants were asked three questions at Time 3 (modified from 17): “In the past week, how often have you thought about the health warnings we showed you?,” “In the past week, how often have you noticed the warning labels on cigarette packages?,” and “In the past week, have health warning labels made you think about smoking’s health risks?” The items were measured on a scale from 1 (not often) to 5 (very often) and averaged to form an index of warning elaboration (Cronbach’s α = .89).

Risk perception (assessed at the final time point)

Participants were asked five questions about risk perceptions of smoking [32]. Specifically, participants were asked three relative risk questions and two absolute risk questions. For relative risks, they responded to: “Compared to the average nonsmoker your age, gender, and race, how would you rate your chances of”: 1) “getting a life threatening illness because of smoking,” 2) “getting lung cancer,” and 3) “dying at a younger age than average” (−3 = Much lower, +3 = Much higher). For absolute risk, they were asked to rate their agreement to the statement “If I don’t stop smoking, I would feel very vulnerable to dying at a younger age because of smoking” (−3=strongly disagree, +3=strongly agree) and “If I continue to smoke, likelihood of getting a life-threatening illness because of smoking are:” (0=Almost zero, 6=Almost certain). This last item was recoded so that −3 represented the lowest risk perception and +3 indicated the highest risk perception. The five items were averaged to form an index of risk perceptions (Cronbach’s α = .90).

Quit intentions (assessed at the final time point)

Participants answered three questions about quit intentions: a contemplation ladder in which participants chose a number indicating their current status regarding smoking from 0 = no thought of quitting to 10 = taking action to quit (e.g. cutting down, enrolling in a program) (this item was also completed at baseline), the likelihood that they would try to quit in the next 30 days (−3 = very unlikely, 3 = very likely), and “in the next week, do you expect your tobacco use to” (−3 = decrease a lot, 3 = increase a lot). The latter item was reverse coded. Because of differences in response scales, each item first was standardized and then standardized ratings were averaged (Cronbach’s α = .64).

Free recall memory (assessed at the final time point)

Following Kees et al. [19], participants were provided nine text boxes and asked to recall each warning label’s text. Responses were coded by two independent coders for the label’s exact words and equivalent words. Coder agreement was high, with Krippendorf’s α of .91 to .98. The primary memory variable used in analyses was memory of the focal concepts (hereafter called focal-concept recall) through either exact or equivalent words. For example, for the “Cigarettes are addictive” warning, focal-concept recall was considered correct if the participant mentioned the exact word “addictive” or equivalent words such as “addicting,” “can’t quit,” or “habit” (see Supplementary Table 1 for coding details). Participants received a score for each label (1=correct recall, 0=incorrect); these binary variables were used as repeated measures to test the primary memory hypotheses. A focal-concept recall index, the proportion of labels correctly recalled, was used only in the SEM (Cronbach’s α = .54 based on the nine binary correct/incorrect responses).

Recognition memory (assessed at the final time point)

Participants were subsequently shown 26 negative health effects and asked “Which of the following health effects were mentioned in the warnings we showed you?” and responded “Yes, I saw this health effect” or “No, I did not see this health effect.” Fourteen risks were mentioned in the warnings (e.g. cancer, harms children, emphysema); six were risks of smoking not presented in the warning text (e.g. wrinkles, diabetes); and six were not risks of smoking (e.g. acne, thyroid problems). We then calculated the proportion of correct recognition (Cronbach’s α = .81), of nontextual recognition (Cronbach’s α = .77) and one for nonsmoking risks (Cronbach’s α = .87).

Preliminary Analyses and Analysis Strategy

The three samples were combined in analyses, and sample effects were examined for generalizability. All analyses used weighted data (based on gender, age, race, education, and region of country) unless otherwise indicated; follow-up analyses with unweighted data revealed similar results. The primary hypotheses (Hypotheses 1 and 2) concerning the memory impact of warning-label condition immediately and over time were tested using generalized estimating equations (GEE) with an identity link to estimate experimental differences in probabilities directly; interactions including warning-label set, time, and sample were examined. This model appropriately allows correlation between binary outcomes measured on the same individual. Analysis was conducted separately for recognition memory and focal-concept recall.

To test Hypothesis 3, we first estimated two alternative SEMs using Mplus [33] and compared model fits to determine whether delay should be included as a covariate or effect moderator. Thus, in one model, we controlled for possible effects of delay on all measures assessed at the two time points (i.e. memory variables, risk perceptions, and quit intentions). Because memory should change with delay, the second model allowed for moderation by delay on each path that involved any measure assessed at the two time points.

To compare effects of experimental conditions, we created two planned contrasts. In one, pictorial warnings (High-emotion = +1, Low-emotion = +1) were contrasted with Text-only warnings (coded as −2); the other contrast compared Low-emotion warnings (−1) with High-emotion warnings (+1; Text-only warnings = 0). In the initial SEMs, these contrasts were used to predict emotional reaction; to test Hypothesis 3, we assessed their indirect effects on risk perceptions and quit intentions via emotional reaction and via emotional reaction, elaboration, and memory.

To test whether warning-label condition was related to these important tobacco outcomes through memory in the SEMs, we first conducted preliminary analyses using STATA [34]. We examined the consistency of emotional reaction across exposures. GEE revealed that emotional reaction to warnings did not differ between the first and fourth exposures (see Supplementary Text 2). Thus, we averaged across these two indices of average emotional reaction to the nine warning labels (Cronbach’s α = .96, M = 3.17, SD = 0.95) to create a measure of emotional reactions.

Because emotional reaction, warning elaboration, risk perceptions, memory variables, and quit intentions were nonnormally distributed, maximum-likelihood parameter estimates with standard errors and a chi-square test robust to nonnormality were used [35]. This approach handles missing data on dependent measures by using all available data and selecting the set of values of model parameters that maximizes the likelihood function. Cases with missing data on independent and control variables were deleted list-wise. Preliminary SEMs also controlled for demographic variables related to sampling frame and that might affect memory, risk perceptions, or quit intentions. These variables were sample (i.e. teen, adult, or Appalachian), education (i.e. high school or less vs. some college or more), race/ethnicity (non-Hispanic white vs. nonwhite), gender, and age. Sample was recoded as two contrast variables to compare teens with both adult samples and to compare the nationally representative sample of adults with those of teens and regionally representative Appalachians. To control for guessing, we included incorrect recognition of risks unrelated to tobacco use (e.g. incorrect recognition of “non-risks” in Table 1) as a predictor of correct recognition of tobacco risks that were shown in the warnings and nontextual recognition (i.e. of tobacco risks not shown).

To evaluate model fit, we used multiple fit indices with recommended cutoff values (e.g. [36]; Root Mean Square Error of Approximation [RMSEA] ≤ 0.05 with 90% Confidence Intervals = 0.00–0.08, Comparative Fit Index [CFI] ≥ 0.95, Standardized Root Mean Square Residual [SRMR] ≤ 0.08. We then chose the model with a better fit to the data. Estimated indirect effects were calculated using robust standard errors; simulations show that such estimates yield accurate estimates of sampling variability when the distribution of model parameters is nonnormal [37]. We used bootstrapping with 5,000 resamples to evaluate significance of indirect effects [38]. For the best-fitting model, we then removed variables nonpredictive of memory, warning elaboration, risk perceptions, or quit intentions one at a time, starting with the least predictive variable; the model was rerun after each deletion.

Results

Demographics

In unweighted data, participants (N = 1,932) averaged 39.7 (SD = 16.8) years and were 42.5% male and 79.2% White; 54.2% had more than a high school education. They smoked 12.0 (SD = 9.9) mean cigarettes per day, with 71.0% smoking daily; 38.3% made at least one past-year quit attempt (Table 2; for data in the separate samples, see Supplementary Table 2). We conducted separate logistic regressions predicting study completion from warning-label condition contrasts, delay condition, study-sample contrasts, demographics and smoking behaviors (i.e. age, gender (male/female), race (white/nonwhite), education (high school or less/more than high school), cigarettes per day, Time 1 quit contemplation ladder, past quit attempts, and smoking frequency). National adults were more likely to complete the study than Appalachian adults or national teens (58%, 45%, and 31%, respectively, OR = 1.69, Wald χ² (1) = 47.78, p < .001); teens were noncompleters more than adults (OR = .54, Wald χ² (1) = 67.39, p < .001). People were more likely to complete if they were in the immediate versus delayed post-test (OR = 1.96, Wald χ² (1) = 119.63, p < .001), were older (OR = 1.03, Wald χ² (1) = 256.04, p < .001), were white (OR = 1.29, Wald χ² (1) = 12.43, p < .001), had more than a high school education (OR = 1.98, Wald χ² (1) = 122.63, p < .001), had lower baseline quit intentions (OR = 1.02, Wald χ² (1) = 5.43, p = .020), were heavier smokers (OR = 1.03, Wald χ² (1) = 68.75, p < .001), and had no past quit attempts (OR = 1.37, Wald χ² (1) = 24.12, p < .001). Warning-label condition did not predict completion.

Table 2

Self-reported demographics and smoking behaviors for participants who did not complete the study (noncompleters) and those who did complete (completers)

	Noncompleters (N = 2,492)	Unweighted completers (N = 1,932)	Weighted completers (N = 1,932)
Age	31.4 (16.1)	39.7 (16.8)	35.8 (15.1)
Gender
Male	41.0%	42.5%	52.8%
Female	58.9%	57.3%	47.0%
Race
White	74.9%	79.2%	75.9%
Black	10.8%	10.3%	9.6%
Hispanic	6.6%	4.5%	9.5%
Other	7.6%	5.7%	4.8%
Education
≤High school	62.4%	45.7%	68.8%
Some college	24.1%	33.0%	21.8%
≥Bachelor’s degree	13.3%	21.2%	9.4%
Cigarettes per day	9.4 (10.0)	12.0 (9.9)	11.9 (9.9)
Time 1 quit contemplation	6.0 (3.5)	5.7 (3.3)	5.5 (3.3)
Past quit attempts
0	44.5%	53.7%	53.3%
1+	43.5%	38.3%	38.0%
Smoking frequency
Every day	55.7%	71.0%	70.5%
Some days	10.4%	9.4%	10.1%
Occasionally (teens)	5.1%	3.2%	3.8%
Rarely (teens)	13.8%	7.2%	6.8%
Not at all (teens)	15.0%	9.2%	8.7%
Warning label
Text-only	34.6%	33.3%	31.9%
Low-arousal	34.4%	34.4%	36.1%
High-arousal	31.0%	32.3%	31.9%
Delay condition
Immediate	40.4%	57.1%	57.1%
Delay	59.6%	42.9%	42.9%
Sample
Adult	21.5%	38.1%	38.1%
Teen	42.7%	24.3%	24.3%
Appalachian adult	35.8%	37.6%	37.6%

For those who completed, both unweighted and weighted averages are reported. Weighted data are used in all analyses. Standard deviations of continuous measures are in parentheses.

Tests of Primary Memory Hypotheses 1 and 2

Our primary hypotheses (Hypothesis 1: that health risks in high-emotion warning text would be remembered less well immediately and Hypothesis 2: that memory for the health risks in the text of more versus less emotional warnings would improve over time or decline less over time such that the highest emotion warning set would be remembered better than the lowest emotion warning set after 6 weeks) were tested in the combined sample (see Supplementary Text 3 for descriptive results). Based on pilot data (Supplementary Text 1) and earlier analyses of these adult and teen data, we expected emotional reaction to be strongest in the High-emotion condition and weakest in the Low-emotion condition, with the Text-only condition falling in between [26]. As hypothesized, the two-way interaction of label condition and time on focal recall was significant (Wald χ² (2) = 12.59, p =.002; see Table 3); the three-way interaction with sample did not approach significance (Wald χ² (2) = 1.90, p = .75). In the Immediate condition and consistent with Hypothesis 1, participants recalled the least in the High-emotion condition (26%) compared to Low-emotion (31%) and Text-only (30.1%) conditions; High:Low, b = −.06, SE = .02, z = 2.49, p = .013; High:Text-only, b = −.04, SE = .02, z = 2.02, p = .043; Low:Text, b = .01, SE = .02, z = 0.51, p = .610). The data also supported Hypothesis 2; focal-concept recall was greater with delay in the High- versus Low-emotion set. In particular, recall declined the most in the Low-emotion condition which elicited the least emotional reaction; it declined least in the High-emotion condition which elicited the most emotional reaction (Immediate:Delay in Low, b = −.15, SE = .02, z = 6.87, p < .001; Text-only, b = −.10, SE =.02, z = 4.91, p < .001; High, b = −.05, SE = .02, z = 2.21, p = .027). Memory-decay differences were pairwise significant in unadjusted analyses only for High versus Low (b = .11, SE = .03, z = 3.55, p < .0001; High vs. Text, b = .05, SE = .03, z = 1.78, p = .075; Low vs. Text, b = −.06, SE = .03, z = 1.90, p = .058). Adjusting for demographics made little difference to results (two-way interaction, Wald χ² (2) = 13.30, p = .001; three-way interaction, p = .79). The nonsignificant three-way interaction suggests sample effects are similar. Results with exact recall were similar to focal-concept recall, but average exact recall was at floor by 6 weeks (5.5–6.5% recall).

Table 3

Percent focal-concept recall, correct recognition, and nontextual recognition by warning-label condition and time across three samples

	Low emotion (n = 699)	Text-only (n = 617)	High emotion (n = 617)
% focal recall
Immediate (n = 1,104)	31.4 (1.2)	30.1 (1.2)	26.0 (1.0)
6-week delay (n = 829)	15.9 (1.0)	20.7 (1.1)	21.1 (1.2)
% correct recognition
Immediate (n = 1,104)	69.4 (1.0)	72.0 (1.0)	76.3 (0.9)
6-week delay (n = 829)	71.0 (1.2)	76.6 (1.0)	76.2 (1.0)
% nontextual recognition
Immediate (n = 1,104)	10.6 (0.9)	16.1 (1.3)	20.8 (1.5)
6-week delay (n = 829)	29.4 (2.0)	32.3 (1.6)	22.5 (1.5)

Standard errors of the proportions estimated from the weighted sample are in parentheses.

The primary hypotheses were tested again using correct recognition memory as the dependent measure (see Table 3). However, Hypothesis 1 and 2 were not supported (two-way warning-label condition × Time interaction, Wald χ² (2) = 2.13, p = .34; three-way interaction with sample, Wald χ² (4) = 5.57, p = .23). A main effect of condition did emerge, with participants in the High-emotion and Text-only warning conditions recognizing more of tobacco’s health risks than those in the Low-emotion condition (condition: Wald χ² (2) = 9.20, p = .010; High:Low, b = .06, SE = .02, z = 3.03, p = .002; Low:Text-only, b = −.04, SE = .02, z = 2.08, p = .037; High:Text-only, b = .02, SE = .02, z = 1.33, p = .180). Adjusting for demographics made little difference to the results (two-way interaction, Wald χ² (2) = 2.48, p = .29; three-way interaction, Wald χ² (4) = 6.55, p = .16). We had no hypotheses for nontextual recognition, but participants did not recognize these risks more in the High-emotion condition over time (Table 3).

Test of Hypothesis 3: Does Negative Emotion Evoked by Warnings Mediate Warning-label Effects on Warning Memory, Risk Perceptions, and Quit Intentions?

Our primary interest in this section concerned Hypothesis 3’s hypothesized indirect effects of warning-label condition through negative emotion on warning elaboration, memory, risk perceptions, and quit intentions. For completeness, separate analyses were used to investigate the total effects of label exposure on key outcome variables (see Supplementary Text 4).

Structural equation model

Because memory is known to be affected by delay, we initially tested two theory-based SEMs, one that included delay as a covariate and another that included it as a moderator of paths to post-test variables (i.e. recognition variables, recall, risk perceptions, and quit intentions). The initial model with delay as a covariate (N = 1,929, see Supplementary Figure 2 and Supplementary Table 5 for initial model results and Table 1 for simple correlations) provided an adequate fit to the data (χ² [35] = 241.95, p < .01; RMSEA = .055 [90% CI: .05 to .06]; CFI = .89; SRMR = .04, BIC = 18,365.07). The alternative model with moderation by delay did not fit the data well (χ² [60] = 3,936.22, p < .001; RMSEA = .18; CFI =.27; SRMR = .07, BIC = 18,386.48, see Supplementary Figure 3 and Supplementary Table 6). Thus, we continued with the model with delay as a covariate. Nonsignificant paths were dropped one at a time based on the highest p-value, with the exception of the path from the contrast comparing pictorial labels with text-only labels to emotional reaction, which was retained in the model. The final model fit the data well (χ² [43] = 217.68, p < .001; RMSEA = .05 [90% CI: .04 to .05]; CFI = .91; SRMR = .05, BIC = 18,201.95, see Figure 3, Supplementary Table 7 for effects of covariates, and Supplementary Table 8 for indirect effects). SEM results controlling for baseline quit intentions and conducted in each separate sample were largely similar; see Supplementary Figures 4 and 5. The model was similar using unweighted data (χ² [43] = 481.82, p < .001; RMSEA = .07 [90% CI: .07 to .08]; CFI = .89; SRMR = .05, BIC = 18,028.15). All paths and indirect effects remained significant.

Fig. 3.

Final structural equation model for antecedents and consequences of memory in full data. Path coefficients are unstandardized, *p < .05. Effects of covariates (i.e. demographics and incorrect recognition of risks unrelated to tobacco) are not shown in this figure and can be found in Supplementary Table 7; indirect effects are reported in Supplementary Table 8.

Participants in High- versus Low-emotion warning-label conditions reported greater emotional reaction (b = .20, p <.001; see Figure 3). Consistent with prior research [3, 26], greater emotional reaction acted directly as information to increase risk perceptions (b = .52, p <.001) and to motivate quit intentions (b = .21, p <.001). The greater emotional reaction was associated with greater correct recognition of tobacco risks in the warning text, greater nontextual recognition (i.e. of tobacco risks not presented), and greater focal-concept recall of health risks in the warning text (b = .04, p <.001, b = .03, p =.003, and b = .02, p =.014, respectively). In addition, greater emotional reaction was related to more warning elaboration (b = .66, p <.001). Although we expected greater elaboration to be associated with more recognition of tobacco’s health risks (both correct and nontextual), elaboration only predicted nontextual recognition (b = .02, p = .027). Greater correct recognition and greater nontextual recognition were associated, respectively, with more (b = .32, p < .001) and less recall of focal concepts (b = −.25, p < .001).

In response to our question about the relation of memory to risk perceptions and quit intentions), more recall and nontextual recognition both were associated with greater risk perceptions (b = .78, p < .001 and b = .48, p = .001, respectively). Delay reduced recall (b = −.08, p < .001) and quit intentions (b = −.09, p = .035), but increased nontextual recognition (b = .09, p < .001). Finally, greater risk perceptions led to higher quit intentions (b = .16, p < .001). No direct paths were retained from memory variables to quit intentions.

We next considered indirect effects of warning-label condition on memory (see Supplementary Table 8 for detailed information about each estimated Indirect Effect [IE]). The indirect effect of the contrast comparing High- versus Low-emotion pictorial warning conditions on focal-concept recall through self-reported emotional reaction was significant and positive (IE = 0.003, 95% CI: 0.002 to 0.005), as was the contrast’s indirect effect through emotional reaction on correct recognition (IE = 0.003, 95% CI: 0.002 to 0.004). A negative indirect effect of condition on focal-concept recall existed through emotional reaction and then nontextual recognition (IE = −.001, 95% CI: −0.002 to −0.001). The path though emotional reaction, elaboration, and nontextual recognition was ns. Thus and consistent with Hypothesis 3, higher emotional reaction (from High- vs. Low-emotion pictorial warning) mediated effects of warning-label condition on memory, both increasing and decreasing correct recall (total IE = 0.004, 95% CI: 0.002 to 0.006).

We also tested hypothesized indirect effects of warning labels on risk perceptions and quit intentions (Hypothesis 3). Indirect effects of High- versus Low-emotion warning-label conditions on risk perceptions supported Hypothesis 3 (total IE = 0.11, 95% CI: 0.08 to 0.14), although not all indirect effects were significant. Positive indirect effects existed through: (1) emotional reaction (IE = 0.10, 95% CI: 0.07 to 0.13), (2) emotional reaction then correct recall (IE = 0.002; 95% CI: 0.001 to 0.004), (3) emotional reaction, then correct recognition, then correct recall (IE = 0.002, 95% CI: 0.001 to 0.003), and (4) emotional reaction then nontextual recognition (IE = 0.003, 95% CI: 0.001 to 0.004). A negative indirect effect of this contrast existed for warning condition through emotional reaction, nontextual recognition, then correct recall (IE = −0.001, 95% CI: −0.002 to −0.001). The remaining paths did not attain significance (see Supplementary Table 8). Consistent with Hypothesis 3, High- versus Low-emotion warning-label condition had significant indirect effects on quit intentions (total IE = 0.06, 95% CI: 0.04 to 0.08). Positive indirect effects existed through: (1) emotional reaction (IE = 0.04, 95% CI: 0.03 to 0.06) and (2) emotional reaction then risk perceptions (IE = 0.02, 95% CI: 0.01 to 0.02). The remaining paths were ns (see Supplementary Table 8).

Discussion

Supporting memory-consolidation hypotheses (Hypotheses 1 and 2), focal-concept recall was worst in the High-emotion pictorial warning condition when assessed immediately after warning exposure, but recall declined less over time in more versus less emotional warning conditions. Thus, at 6 weeks, recall of Low-emotion pictorial warnings was significantly lower than that for High-emotion pictorial warnings and Text-only warnings, which were similar. These results did not differ across the three samples, supporting generalizability of effects. Recognition memory also was similar across the three samples. It was higher for more than less emotional warning-label sets and was unaffected by delay, perhaps due to ceiling or familiarity effects [29].

The GEE analyses indicated that recall and recognition results were most different between the Low-emotion pictorial warnings as compared to both Text-only and High-emotion pictorial warnings. The High-emotion condition, however, might have demonstrated superior memory to the Text-only condition if the delay was longer than 6 weeks given emotion’s role in supporting memory over time [13]. Indeed, survey and experimental studies indicate greater knowledge when pictorial warnings versus text-only warnings are used [7, 22]. The lack of difference between Text-only and High-emotion conditions could be due to a floor effect at 6 weeks; however, recall in the Low-emotion condition starts the highest and ends up significantly lower than the other two conditions by 6 weeks, making floor effects less likely. Instead and compared to the Text-only condition, the High-emotion condition may have elicited more beliefs in smoking myths [30] and/or reactance (i.e. anger and negative cognitive responses to the message, [21, 39, 40]), that suppressed memory, risk perceptions, and quit intentions (e.g. [30]). In one study, however, stronger emotional reactions provoked by pictorial warnings increased risk perceptions and quit intentions more than reactance reduced them [40]. Overall, the present memory data most strongly supported not using Low-emotion pictorial warnings. A further exploratory analysis of the effects of emotional reaction within each warning-label set revealed that the more emotional the label was within its own set, the better it was recalled and recognized (see Supplementary Text 5). Given overall low focal-concept recall and positive effects on nontextual recognition, these results suggest that regulators interested in increasing public-health knowledge should choose uniformly high-emotion warnings that involve text only or text plus images.

Consistent with Hypothesis 3, our SEM results demonstrated that emotional reactions mediated the effects of warning-label condition on recall and recognition memory. They also pointed towards the increased recognition and recall of smoking’s health risks as potentially important to risk perceptions. In particular and with a control for guessing, our experimental warning-label manipulation of emotional reaction increased memory (focal-concept recall, correct and nontextual recognition). It also was related to more thinking about the warnings, which appeared to spread and activate other important smoking risks stored in memory. We expected greater elaboration to cause more recognition of tobacco’s health risks (correct and nontextual), but only the path from elaboration to nontextual recognition was significant, and no indirect effects of warning conditions through elaboration were significant. Similar-sized effects of emotion on correct and nontextual recognition further suggested that more emotional labels may cause confusion between presented smoking-risk information and related smoking health risks that were not presented (alternatively, nontextual recognition may index memories of images rather than text, but nontextual recognition was not higher in pictorial than text-only conditions; see Table 3). Nontextual recognition then appeared to suppress freely recalling risks from the text. Although this effect could be interpreted as negative, both types of recognition memory concerned actual smoking risks and were “correct” in that sense. Both types also may increase overall smoking knowledge and motivate risk perceptions and quit intentions as people correctly recognize and recall the risks they read and correctly draw parallels to other related risks. Indeed, all three memory variables related directly or indirectly to greater risk perceptions and quit intentions. Finally, the manipulation’s indirect effects on risk perceptions and quit intentions that were independent of memory (i.e. through emotional reaction) were larger than those dependent on memory. Thus and consistent with prior research [3, 17, 18, 41], our data also support the use of High-emotion pictorial warnings over Text-only warnings.

Retention was adequate among adults but low among teens. The final sample had a greater proportion of white versus nonwhite participants, was older, more educated, and comprised of heavier smokers with lower quit intentions and fewer past quit attempts than the baseline sample. However, random assignment to warning-label condition was unrelated to study completion. Differential attrition could have biased results (e.g. memory declines with older age). However, using weighted or unweighted data and including sampling frame variables (e.g. age, gender, race, and education) and baseline quit intentions as covariates did not significantly affect results; critical memory-consolidation and SEM results also were similar in the individual samples (see Hypotheses 1 and 2 results and Supplementary Figure 5), consistent with minimal bias.

These findings may have two major implications for health communication and the U.S. future of pictorial warning labels. First, high-emotion pictorial warning labels support the U.S. Food & Drug Administration’s primary goal to “effectively convey the negative health consequences of smoking on cigarette packages and in advertisements” [42]. That is, more emotional warnings help convey factual information more effectively by supporting long-term retention of their text warnings. The use of more emotional warnings may increase the stock of health-risk knowledge that people have. In turn, they may decrease the well-known educational gradient in smoking, that is, people with more education are less likely to smoke and, conditional on being a smoker, are more likely to quit than are people with less education [43]. Text messages can also be emotional and, as in the present data, can be more emotional than some Low-emotion warnings.

Second and related, the Food & Drug Administration currently does not require the use of pictorial warning labels based on an earlier court decision [42]. The causal impact of emotional reaction on memory for tobacco’s risks, however, argues for the benefits of eliciting emotional reaction. In addition, high-emotion pictorial warnings help participants to perceive warnings as more credible than text-only warnings and they scrutinize them more [3]. Since participants who think about and discuss warnings are more likely to express increased quit intentions in the following 6 months [17], these results may further indicate that high-emotion pictorial warnings will help to increase smokers’ motivation to quit smoking. In fact, a recent clinical trial demonstrated that pictorial warning labels with the new text messages used in the present study decreased smoking over current U.S. text-only messages [41]. Thus, based on smoking’s enormous financial and mortality costs, the Food & Drug Administration may have grounds to argue for a substantial interest in requiring high-emotion pictorial warnings to prevent consumer deception and improve public health.

Memory, of course, improves with more exposures. If people indeed look at the warnings, then, over time, we might expect fewer memory differences between more and less emotional warning labels as everybody learns about and recalls more risks. However, this prediction seems overly optimistic given that most people can name only one to four of smoking’s many diseases [3–5], and the same six diseases have been mentioned in U.S. text-only warnings since 1984. In addition, studies have indicated that smokers attend longer to pictorial than text-only warnings [44], suggesting that positive effects for pictorial over text-only warnings may increase over time. At the end of the present study, however, pictorial and text-only warnings had similar memory effects. Nonetheless, future research should examine these questions.

Of course, our operationalization of risk understanding may not fully capture what pictorial images communicate. Given that emotional reactions to danger have been shown to be critical to quick perceptions of and reactions to risk [10, 45], our measure of emotional reaction to the warnings may capture this aspect of risk understanding [46]. Future research should examine additional operationalizations (e.g. comprehension of disease severity, numeric disease risk, disease experience, and likelihood of cure).

In the present study, we used only negative warnings, but positive arousal has also been shown to improve memory [11] and has been used to some extent in tobacco education (e.g. a drooping cigarette to illustrate impotency; funny public service announcements [PSAs]), as well as advertisements (e.g. brand information, smoking cues, healthy smokers smiling and having fun). Positive arousal has weaker effects than negative arousal on improved long-term memories, but it has greater effects than neutral arousal [11]. At the same time, however, providing positive emotional materials may cause perceptions of greater benefit and lesser risk from tobacco [46]. Thus, individuals shown positive cues with health-risk information may understand those risks somewhat better over time (compared with those shown neutral images) but perceive lower risk for themselves. Negative sources such as the pictorial warnings of the present study should produce perceptions of lower benefit and greater risk, plus health-risk information should be better understood over time as those memories are consolidated. More studies need to be conducted to examine positive emotion’s complex implications for health-risk communication.

Limitations of the present study include that participants only viewed warnings four times over 2 weeks (although the number of exposures in this study was greater than in most prior research). In addition, the warnings were presented on computer screens without the context of their own cigarette packages to which they may have positive reactions. Each participant also saw a single set of nine warnings that could have differed in some manner other than emotional reaction. Future research should replicate this study using different images to rule out possible stimulus-sampling effects. Lower reliability of the SEM’s recall index reliability (Cronbach’s α = .54) likely was due to the binary nature of the items (correct/incorrect) that differed considerably in difficulty (see Supplementary Table 3) and, thus, remains a valid indicator of the number of recalled risk. If the present hypotheses are supported, pictorial warnings could potentially be used to improve the credibility, elaboration, and memory of warnings in other areas of tobacco as well as nontobacco health education such as vaccines and antibiotic use.

The current research adds to existing research on the role of emotional reaction in determining the impact of pictorial warning labels. Understanding the myriad processes by which pictorial warning labels influence smokers and vulnerable smokers is crucial to policy makers’ ability to design effective future warning labels.

Acknowledgements

Research reported in this publication was supported by grant number P50CA180908 from the National Cancer Institute of the National Institutes of Health (NIH) and the Food and Drug Administration (FDA) Center for Tobacco Products. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH or the FDA. We thank Louise Meilleur and the OSU-CERTS External Advisory Board for their assistance on this project.

Authors’ Statement of Conflict of Interest and Adherence to Ethical Standards Authors Ellen Peters, Brittany Shoots-Reinhard, Abigail T. Evans, Abigail Shoben, Elizabeth Klein, Mary Kate Tompkins, Daniel Romer, and Martin Tusler declare that they have no conflict of interest. All procedures, including the informed consent process, were conducted in accordance with the ethical standards of the responsible committee on human experimentation (institutional and national) and with the Helsinki Declaration of 1975, as revised in 2000. Research reported in this publication was supported by grant number P50CA180908 from the National Cancer Institute of the National Institutes of Health (NIH) and the Food and Drug Administration (FDA) Center for Tobacco Products. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH or the FDA.