The major assumption of PMT is that, once individuals receive information about a threat to their health (e.g., the risk of an infection with COVID-19), two cognitive processes are initiated that ultimately determine the probability of showing an adaptive coping response to the threat (Rogers, 1983). One of these processes, threat appraisal, comprises the following three components: (a) the assessment of the likelihood of being personally affected (perceived vulnerability), (b) the evaluation of magnitude of negative consequences (perceived severity), and (c) emotions of fear that are evoked by the threat (fear arousal). Coping appraisal, on the other hand, subsumes (a) beliefs about the efficacy of the recommended response to effectively avert negative consequences (response efficacy) (b) beliefs about one’s capability to perform the recommended protective behaviors (self-efficacy) and (c) the expected personal costs of adopting these (response costs). High threat perceptions (i.e., high levels of perceived severity, vulnerability, and fear arousal) and positive evaluations of the recommended coping responses (i.e., high levels of response efficacy and self-efficacy; expectation of low response costs) are proposed to increase an individual’s intention to engage in protective behaviors (protection motivation). Through their effect on intention, they are hypothesized to indirectly increase the likelihood that protective measures are actually taken.

Earlier research conducted during past pandemics (e.g., SARS-CoV-1, avian influenza/flu H5N1, swine influenza/flu H1N1) and the current COVID-19 pandemic lends initial support to the proposed bivariate associations of the six appraisal components with intentions and performed protective behaviors. Specifically, cognitive risk perception and the experience of fear or worry due to a specific infectious disease have been shown to be positively related to reported intentions and protective behaviors (Bish and Michie, 2010; Rubin et al., 2010; Dryhurst et al., 2020; Coifman et al., 2021). Furthermore, individuals reported stronger intentions and more thorough adherence when they believed that the different protective measures were effective at reducing their risk of contracting a certain disease (Rubin et al., 2010; Bults et al., 2011; Kaspar, 2020; Kowalski and Black, 2021; Scholz and Freund, 2021; Van Loenhout et al., 2021) and were confident in their own ability to do so in everyday life (Kaspar, 2020; Kowalski and Black, 2021; Scholz and Freund, 2021; Van Loenhout et al., 2021). On the contrary, expecting the measures to be costly, e.g., in terms of time and effort needed for hygiene and distancing practices, was associated with lower intentions and adherence (Al-Rasheed, 2020; Kaspar, 2020). In many of these studies, effects were stronger and more consistent for the coping variables as compared to the threat variables (Farin, 1994; Teasdale et al., 2012; Kaspar, 2020; Kowalski and Black, 2021; Scholz and Freund, 2021; Van Loenhout et al., 2021).

Research in light of pandemic events has oftentimes been carried out as a rapid response to emerging influenzas (Rubin et al., 2010), characterized by profound methodological weaknesses and a lack of a theoretical framework (Leppin and Aro, 2009). For the most part, data has been collected through convenience sampling or by surveying specific subgroups of the population (e.g., inhabitants of a certain region, students, or medical staff) (Al-Rasheed, 2020). This poses problems to the generalizability of results to the general public (Kaspar, 2020; Van Loenhout et al., 2021). An even greater concern is the predominance of cross-sectional surveys casting doubt on causal inferences about the effects of the appraisal processes on subsequently formed intentions or performed behaviors (Brewer et al., 2004; Norman et al., 2015).

Protection motivation theory provides a comprehensive description of factors that might influence the motivational process of deliberation, choice of goals and corresponding actions (Norman et al., 2015). Intentions are proposed to be the main driver and sole predictor of performed protective behaviors (Norman et al., 2015). Although meta-analytic work shows that there is a strong relationship between intentions and subsequent health behavior (r = 0.53), this means that 70% of variance are typically not explained (Sheeran, 2002). An individual’s motivation, therefore, does not seem to sufficiently explain whether or not he or she is taking health protective action (Sheeran, 2002). We propose that, once an intention to adopt preventive measures is set, the two PMT-constructs self-efficacy and perceived costs continue to play a role in determining its successful implementation. As the authors of other prominent social cognitive theories [e.g., the theory of planned behavior/reasoned action approach (Ajzen, 1991; Schwarzer, 2008; Fishbein and Ajzen, 2009)] point out, the belief in one’s capability of performing a behavior (i.e., self-efficacy) is likely to remain crucial for behavioral initiation and maintenance after an intention has been set, especially when difficulties arise and setbacks have to be overcome. Relatedly, perceived costs can reflect those difficulties [i.e., inconvenience, discomfort, financial expense, disruption of everyday life (Rogers, 1983)] that might be subjectively experienced in the very moment a protective behavior needs to be carried out. Visceral drives–e.g., aversive experiences such as hunger, fatigue, and negative emotions–play an important role in explaining impulsive risk behaviors that are in conflict with individuals’ long-term health goals (Loewenstein, 1996, 2005; Hofmann et al., 2008; Nordgren et al., 2008). As powerfully demonstrated by Nordgren et al. (2008), those temporary states affect individuals’ health cognitions and behavioral intentions: Hungry dieters and smokers with a momentary cigarette craving follow less ambitious weight loss goals and intentions to quit smoking, respectively, than did their satiated or noncraving counterparts. Furthermore, the latter were more confident in their ability to reach their goals. Their greater optimism reflects the well-documented phenomenon of the cold-to-hot-empathy gap: When in a “cold state” or not affectively aroused (e.g., not hungry), people seem to systematically underestimate the intensity of future aversive states and their impact on their own preferences and behavior in the future (Loewenstein, 1996, 2005). Likewise, people who are currently unaffected by the burden of COVID-19 protective behaviors (e.g., the discomfort of wearing a facemask) might underestimate the motivational force of those costs when reporting their intentions to adhere, yet be strongly influenced by them in situations critical for infection protection. Thus, perceived costs might have a strong influence on performed behavior, whereas their influence on intentions might be less pronounced.

Besides their psychological effects, both self-efficacy and perceived costs might reflect the amount of actual control people possess over performing a behavior based on their actual abilities, opportunities, and the existence of actual barriers to action (Ajzen, 1991; Fishbein and Ajzen, 2009). Such non-motivational factors may affect behavioral performance independently from a person’s intentions (Fishbein and Ajzen, 2009). In fact, several PMT-based studies found self-efficacy to be directly predictive of health behavior along with intention (Conner and Norman, 2015; Dowd et al., 2016). The direct effect of costs on behavior can possibly be inferred from nudging interventions that proved successful in improving health behavior (Sniehotta et al., 2014; Bucher et al., 2016). In this domain, nudging refers to a strategic change in the choice architecture aimed at making healthy choices easier, e.g., by making the respective option psychologically salient or appealing or reducing the option-related effort (Bucher et al., 2016; Hansen et al., 2016). For example, placing fruit within consumers’ reach has been shown to positively affect its intake (Bucher et al., 2016). Low costs─favorable environmental conditions and the absence of barriers to actions─may therefore contribute to health protective behaviors without affecting intention or self-efficacy (Sniehotta et al., 2014). Based on the aforementioned arguments and empirical findings, a modification of the PMT model seems to be necessary: Besides intention, self-efficacy and perceived costs might also directly predict performed protective behavior (Fishbein and Ajzen, 2009; Conner and Norman, 2015).

The main aim of this study was to test the utility of Rogers’ PMT for better understanding protective behaviors in the COVID-19 pandemic. Expanding on past work, we present the results of a longitudinal survey in a sample representative for age and gender of the German adult population. The longitudinal survey design and the representative sample (stratified by age and gender of the German adult population) address main weaknesses of existing research and reduce ambiguity in causal interpretation of the findings (Brewer et al., 2004; Norman et al., 2015; Al-Rasheed, 2020; Kaspar, 2020; Van Loenhout et al., 2021).

Existing studies during pandemics investigated one particular prevention measure as outcome variable (e.g., either hand washing or facemask wearing) or combined multiple different behaviors into one scale measuring overall adherence (Bish and Michie, 2010; Hagger et al., 2020). Here, we aim to explain the adoption of the three “DHM-measures”–keeping physical distance (“D”), proper hand hygiene (“H”) and wearing a face mask (“M”)–which were promoted by the German Federal Centre for Health Education since the beginning of the COVID-19-pandemic (Federal Centre for Health Education, 2020). Since these behaviors differ in many aspects (e.g., in terms of discomfort, frequency, novelty, and impact on others) it is to be expected that different cognitions determine their performance (Bish and Michie, 2010). By specifying three SEMs for each of these behaviors, we aim to identify more precisely which of the PMT constructs might be specifically promising target variables to promote each of them individually.

In line with PMT, we hypothesized that higher levels of perceived severity of the disease, of perceived personal vulnerability to an infection, and of experienced fear arousal would have a positive effect on an individuals’ intention to adopt protective measures at the baseline measurement. Second, both perceived response efficacy and self-efficacy beliefs concerning the performance of the behaviors in everyday life were expected to have a positive effect, whereas perceived costs were hypothesized to reduce the respective behavioral intentions. We also hypothesized that these factors would have an indirect effect on actual behavior performed in the following month through their influence on intentions. That is, severity, vulnerability, fear arousal, self-efficacy and response efficacy were expected to increase the actual adoption of protective behavior via increased behavioral intentions. Costs, on the other hand, were assumed to be negatively associated with performed protective behavior via its inhibiting effect on intentions. In an additional exploratory investigation, we tested a second, integrated social cognition model designed to illuminate the implementation of adherence intentions more comprehensively. Our proposed model alternative incorporated two additional direct effects of both self-efficacy and costs on behavior.

Data were collected in two consecutive surveys waves in mid November and mid December 2020. The study thus was conducted at the beginning of the second wave of the COVID-19 pandemic in Germany, after an exponential surge of confirmed cases had been observed in October 2020 (Schuppert et al., 2021; Robert-Koch-Institute, 2022). Starting from November, new containment measures of partial lockdown were implemented by the German government that primarily aimed at restricting leisure time activities (i.e., closing of restaurants, cinemas, and theaters) and contacts (i.e., banning gatherings of members of more than two households in public) (Schuppert et al., 2021). On the day the first assessment (Time 1 [T1]; 13.11.─19.11.2020) started, 751,095 cases and 12,200 deaths due to COVID-19 had been reported in Germany (Robert-Koch-Institute, 2022). Since growth of infections remained strong in November and December, lockdown measures were extended by the closure of schools, day care centers and retail shops on December 16, 2020 (Schuppert et al., 2021). On the first day of the second survey wave (Time 2 [T2]; 14.12─20.12.2020), there had been 1,351,510 confirmed cases and 22,475 deaths related to COVID-19 in Germany (Robert-Koch-Institute, 2022).

A nationally representative sample of the German adult population (stratified by age and gender) was recruited by the panel provider Respondi¹ and invited to complete the survey hosted on Qualtrics.² Out of N = 467 individuals who started the survey, n = 8 refused to give their consent. n = 30 dropped out before completing the questionnaire, and n = 97 were excluded due to a failed attention check (“Please mark the response option 1”). In a series of items presented to detect careless responses (Meade and Craig, 2012), n = 3 individuals indicated that we should not use their data for analyses. We removed their records as well as the second record of one participant whose ID indicated duplicate submission. When the final sample at T1 (N = 328) was contacted again one month later at T2, n = 285 individuals (87%) opened the follow-up questionnaire, out of which N = 278 gave their consent and fully completed the survey. The mean age of this final sample was 45.83 years (SD = 14.57), and 50.7 percent were female. Detailed demographic characteristics of the T1 and T2 samples are displayed in Supplementary Table A1 in the Supplementary material for this article. Note that merging data into a longitudinal data file revealed eight persons with inconsistent or implausibly altered indicator variables across waves (age, gender). We decided to keep their records in the sample as no obvious signs of poor-quality responses (e. g. speeding, response sets) could be detected.

Item examples and scale statistics can be found in Table 1 (see Supplementary Tables A2–A4, for a translated version of all measures). The Time 1 questionnaire measured threat (severity, vulnerability, fear) and coping (response efficacy, costs) appraisals as well as behavioral intentions. Four weeks later, performed behavior was assessed. Measures were based on items used in previous work on PMT (Milne et al., 2000) and adapted to the adoption of protective behavior during the pandemic. For each of the three DHM-behaviors, separate sets of items were presented to assess coping appraisal (response efficacy, self-efficacy, perceived costs), intentions, and behavioral frequency. Within these sets, each item referred to the adoption of either distancing, hand hygiene, or mask-wearing behavior in specific everyday life situations that have been emphasized to be especially important for prevention in public health campaigns (Federal Centre for Health Education, 2020). For example, looking at perceived response efficacy of wearing a mask, participants would indicate whether they believed that it would reduce their risk of an infection while (a) using public transport, (b) spending time in public places, (c) meeting other people and (d) entering supermarkets or other shops.

Regarding demographics, we collected participants’ age and gender in both survey waves. At Time 1 only, they were asked to report their subjective health status (1 = very bad, 5 = very good), whether a medical condition put them at greater risk of severe illness from COVID-19, and whether the latter applied to a significant other. Additional questions tapped into the endorsement of conspiracy theories, trust in politics and science, as well as the motivation for and actual performance of several non-recommended prevention behaviors (Dohle et al., 2020; Imhoff and Lamberty, 2020). These questions were assessed for exploratory purposes and are not discussed further.

Data was analyzed separately for each of the DHM-behaviors. To test the hypothesized direct and indirect causal effects of coping and threat appraisals on behavior, all variables were analyzed as latent variables using a two-step structural equation modeling (SEM) approach (Anderson and Gerbing, 1988). We first performed confirmatory factor analyses (CFA) to ensure adequate measurement of the latent constructs. Following a partial adjustment of the measurement models, we then examined the proposed structural relationships of the variables.

Reliability and validity of the measurement models were assessed using established criteria including the squared standardized factor loadings (squared multiple correlations, SMCs ≥ 0.4 indicate good indicator reliability; Weiber and Mühlhaus, 2014), Raykov’s rho coefficient (ρ >0.7 indicates good construct reliability; Raykov, 2004; Hair et al., 2013), standardized factor loadings (values >0.5 indicate adequate, values >0.7 good convergent validity; Hair et al., 2013), and estimated factor correlations (absolute values <0.85 indicate adequate discriminant validity in factor measurement; Brown, 2006). Moreover, a variety of fit indices was employed to evaluate model fit. In addition to the inferential Chi-square test (Bollen-Stine p value >0.05 indicates good exact fit) we report the chi-square to degrees of freedom ratio (absolute values 2 and 3 indicate an adequate fit; Schreiber et al., 2006), the Root Mean Square Error of Approximation (RMSEA, values ≤0.05 and those ranging from 0.05─0.08, respectively, indicate an good and adequate fit; Browne and Cudeck, 1992), the Standardized Root Mean Square Residual (SRMR; common cut-off ≤0.08 for model acceptance; Hu and Bentler, 1999) and the Comparative Fit Index (CFI; ≥ 0.90 recommended for the given sample size and high model complexity; Hair et al., 2013). In addition to global fit measures, we examined standardized residuals to identify local areas of misfit at the level of pairs of observed variables (values | ≥ 2.58| indicate a large deviation of model-implied covariances from those obtained empirically; Kline, 2005; Byrne, 2013).

Consistent with our pre-registration,³ path coefficients significant at p < 0.05 were considered to support direct and indirect effects of the appraisal variables as hypothesized by Rogers (1975, 1983). In an exploratory part of our analyses, we estimated alternative models including two additional direct paths (costs ➔ behavior; self-efficacy ➔ behavior). Added paths were trimmed if their estimated coefficients were not significant. The resulting alternative structures to the original PMT models were retained if Chi-square-difference tests and information criteria (AIC; Akaike’s information criterion; and Bayesian information criterion; BIC) suggested their superiority: A significant chi-square difference statistic indicates that the hypothesis of equal model fit needs to be rejected as the addition of free parameters (i.e., the direct paths) significantly improved the correspondence of the model with the data (Kline, 2005). In using AIC and BIC for model comparison, we include model parsimony as an additional criterion to the goodness of fit to the data (Kline, 2005).

Data preparation and calculation of descriptive statistics were performed using R version 4.0.0 and SPSS version 27. Participants remaining at T2 were significantly older (M = 45.63, SD = 14.28) than those individuals who responded only at T1 (M = 37.18, SD = 13.95), t(317 = 3.86, p < 0.001, d = 0.59). Furthermore, they were slightly less likely to report having a significant other belonging to a vulnerable risk group, X² (1) = 5.54, p = 0.019, Φ = 0.14. Despite these two effects, no differences were found between the initial and final samples in terms of demographics or any of our study variables. Furthermore, Little’s MCAR Test (Little, 1988) was nonsignificant, X² (60) = 76.79, p = 0.071, suggesting that observed missing data patterns can be assumed to be missing completely at random (Hair et al., 2013). Prior to SEM analyses and for each behavior, the dataset was evaluated for multivariate outliers by examining squared Mahalanobis distances (D²). Deleting cases with a distinctively large D² (Byrne, 2013; Weiber and Mühlhaus, 2014) left us with sample sizes of N = 275 (physical distancing and mask wearing) and N = 271 (hand hygiene). SEM was performed with AMOS version 27 using maximum-likelihood estimation. Skewness and kurtosis of a number of indicators and Mardia testing (Mardia, 1970) indicated non-normality in the data. This was addressed in all our subsequent analyses by performing bootstrapping (5,000 samples) to derive bootstrap standard errors as well as bias-corrected confidence intervals for all parameter estimates and Bollen-Stine p-values for model testing.

Based on protection motivation theory, we specified three CFA-models (physical distancing, hand hygiene, wearing a mask) with eight factors each. The number of indicators loading on the six appraisal variables, protection motivation and protective behavior varied from 3─5 (see Table 1). All latent variables were permitted to be correlated, all measurement errors were presumed to be uncorrelated, and all indicators were specified to load onto a single factor only. Latent variables were scaled by imposing unit loading identification constraints, meaning that all unstandardized residual path coefficients were fixed to equal 1.0, while the same was done with the factor loading of one indicator per factor (Kline, 2005). We further established identification by fulfilling the requirements of overidentification (df ≥ 0 in all our models) and having more than two indicators per factor (Kline, 2005).

Even though our initial measurement models demonstrated adequate fit with respect to most metrics (see Supplementary Table A5), modification indices (> 4.0) provided by AMOS prompted us to make several adjustments to improve model fit. All modifications of our initial measurement models are documented in the Supplementary material for this article (Supplementary Tables A2–A4) alongside with our rationale underlying these respecifications. Here, we provide a summary of all indicators used in the final measurement models including a documentation of those that we deleted. Graphical representations (Supplementary Figure A1) depict which error covariances have been added. The different fit indices (also reported in Supplementary Table A5) indicated model improvement, with most of them attesting adequate to good fit following our adjustments. Detailed CFA results including interfactor correlations, standardized and unstandardized parameter estimates can be found in the Supplementary Tables A6–A9. Reliability and validity analyses yielded satisfactory results (factor loadings of the individual items 0.59─0.98, all p ≤ 0.001; SMC 0.35─0.95; construct reliabilities ρ = 0.73─0.93). As an exception, we found the factors self-efficacy and intention to be highly correlated (r = 0.86─0.90, p < 0.001) in two of the models. As this observation is in line with the assumption of self-efficacy being the most important predictor of protection motivation (Milne et al., 2000), we kept measurements of these variables distinct.

In the second part of our analyses, the three final measurement models derived from confirmatory factor analyses were respecified as a fully latent SEMs incorporating direct and indirect effects as proposed by PMT. As fit indices listed in Table 2 indicate, these models demonstrated acceptable to good fit to the data. As a second step, each of them was compared to our proposed alternative models. Results from the modified and finally retained models are depicted in Figure 1.

As novel emerging infectious diseases will pose us with new challenges in the future, it will most likely remain central to identify determinants of the public’s adherence to prevention guidelines and thereby find targets for behavioral interventions. The result of this research suggests that motivation and actual behavior depend on individuals’ self-efficacy, as well as perceived and actually existent barriers to adherence. We conclude that during a pandemic, policymakers and public health institutions should not only provide information about the risks and consequences of viral infection. Instead, they should support individuals in developing self-efficacy and address perceptions of difficulties and costs, while also providing them with resources, opportunities and environmental conditions that actually facilitate the adoption of protective measures (Teasdale et al., 2012; Hagger et al., 2020).