We administered 4 KAP surveys from August 2014 through July 2015 in Sierra Leone (Technical Appendix Figure 1). In all KAP surveys, we applied multistage cluster sampling. KAP 1 was conducted in 9 districts from all 4 regions of Sierra Leone. KAPs 2–4 (October 2014, December 2014, and July 2015) covered all 14 districts in the country. The 2004 Sierra Leone Population and Housing Census List of Enumeration Areas (http://www.sierra-leone.org/Census/ssl_final_results.pdf) served as the sampling frame for random selection of enumeration areas from the districts. Within enumeration areas, we used the random walk method, a form of systematic random sampling, to select households. In each selected household, we conducted 2 interviews: first with the head of the household, and second with either a woman or a young person 15–24 years of age who was randomly selected. All 4 KAP surveys were designed to produce national- and regional- level estimates at a 95% confidence level within a 2.5% margin of error for national estimates and a 3.5% margin of error for regional estimates (Technical Appendix Table 1).

We captured exposure to a source of information in the KAP surveys with the question: “Through what means/ways did you learn about Ebola?” Categories were not mutually exclusive; data collectors ticked “yes” in all boxes that applied. Furthermore, more response options were included in later versions of the KAP surveys. Five categories reflected the different information sources: electronic media (radio and television); print media (newspapers, brochures, and other print materials); new media (mobile phones, text messages, and internet); government (house visits by health workers); and community (religious and traditional leaders, megaphone public announcements, community meetings, friends, and relatives). We categorized number of exposures into 0–1 source, 2, 3, and 4–5 sources. The sample of respondents exposed to 0 or 5 sources was low (n = 38 for 0 sources and n = 265 for 5 sources). Therefore, we combined 0 with 1 and 5 with 4. The category “level of outbreak” was broken down into before the peak, peak, after peak, and no transmission for each region. We consulted World Health Organization data to create this variable for every region in all 4 KAP surveys (1) (Technical Appendix Table 2).

We tested Ebola-specific knowledge as both an outcome of exposure to information and a mediating factor in the association between information exposure and behavior (Technical Appendix Figure 2). We assessed Ebola-specific knowledge in the KAP surveys with 2 open-ended and 5 closed-ended questions, resulting in 2 scores, 1 for knowledge and 1 for misconception (Technical Appendix Table 3). In the knowledge score, a maximum score of 8 points was possible in KAP 1–4. In the misconception score, 12 points was the maximum in KAP 1 and 13 points in KAP 2–4, due to an additional response option. We dichotomized the scores based on the means (22).

We measured Ebola-specific behavior by 2 open-ended questions and created 2 scores, 1 for protective behavior and 1 for risk behavior. For the protective behavior score, respondents could tally a maximum of 7 in KAP 1; in KAP 2–4, the maximum was 9. The risk behavior score had a maximum of 9 points in KAP 1 and 11 in KAP 2–4. We again dichotomized the scores based on the means.

Working with information from previous KAP survey analyses (17,18), we chose the following covariates as potential confounders: gender (male, female); age (13–20, 21–35, 36–49, >50 years); education (no education, primary education, and secondary and above); religion (Islam, Christianity); and region (Northern Province, Eastern Province, Southern Province, Western Area).

The overall response rate of the 4 KAP surveys was 98%. The pooled sample of KAP 1–4 consisted of 10,604 respondents. Because of missing values for some variables, we excluded 95 respondents (0.9%), making the total study sample 10,509. The age distribution was similar in the different regions, with those 21–35 years of age representing the largest group (Table 1). The total sample had an even distribution of men and women (49.1% men, 50.9% women). In all regions apart from the Northern Province, most of the sample had at least primary schooling. Two thirds of the respondents were affiliated with Islam (67.3%). EVD messages were received mostly through electronic media (94.1%) and community sources (59.5%) (Table 2); a substantially smaller portion were exposed to print (8.7%) and new media (15.5%). Government information campaigns reached 47.7% of respondents. Most respondents were exposed to >2 different types of information sources (72.6%).

The results of this study clearly show that exposure to different types of information sources was associated with increased knowledge and protective behaviors during the EVD epidemic in Sierra Leone. A strong dose-response association was seen between information exposure and knowledge and protective behavior. However, exposure to all information sources (apart from electronic and print media) was also significantly associated with misconceptions and risk behavior. For all sources, apart from print media, the association between information exposure and behavior was mediated by knowledge. Knowledge had a significant association with level of outbreak at all stages of the outbreak. Protective behavior was significantly associated with level of outbreak at the peak and after the peak but no longer showed an association at the point of no transmission. Misconceptions showed a decline at first but increased toward the end of the outbreak. Risk behavior was prevalent during the entire course of the outbreak, but with a declining trend.

The results reinforce the importance of radio in Sierra Leone; electronic media (which consists mainly of radio) had the strongest association with knowledge and protective behavior but also with misconceptions. Even though electronic media were not directly linked with risk behavior, there was a mediating effect on risk behavior through misconceptions. In 2013, Sierra Leone had a literacy rate of only 46% (25), which makes radio the preferred and most accessible source of information for a large part of the population (78% have access to radio) (26). Other studies have also reported the importance of radio in promoting knowledge in low-income settings (11,27,28). Nonetheless, the influence of radio on promoting protective behavior in these studies was not as strong (11) or did not have an association at all (28). The setting of our study, an infectious disease outbreak with high death rates, might have provided more urgency and a different risk perception and behavior change pathway than other studies (20). Furthermore, as part of the EVD response, radio was used as a channel to disseminate messages from community and religious leaders (3). An additive effect of different information sources (such as community and government) coming together in radio could explain the strong associations.

The finding that both knowledge and misconceptions and both protective and risk behavior were linked to most of the information sources could be explained by the presence of contradicting messages during the outbreak. Previous studies have reported a high prevalence of misconceptions, which have led to rejection of interventions during the EVD epidemic (12,29,30). Furthermore, correct and incorrect explanations of a disease might coexist in a community. In Guinea, for instance, 83% of respondents believed that a virus caused EVD, but at the same time, 36% believed that higher powers were involved in causing the outbreak (17). This finding could explain, in part, some of the results in our study, in which community sources were most strongly associated with risk behavior. Community sources were employed mostly in areas with high EVD transmission. Selection bias could, therefore, be an explanation of the observed associations. Furthermore, the temporal relationship between community sources and the reported risk behaviors cannot be ascertained.

Print media showed a different pattern from other information sources in the models. Exposure to print media was associated with both knowledge and protective behavior but not with misconceptions and risk behavior. A reason for this finding might be that those who read newspapers and brochures differ from the general population of Sierra Leone in that they are literate and have the means to buy a newspaper. Only 11% of the population reads newspapers on a weekly basis (26). Educational level was taken into account in the analyses, but residual confounding cannot be excluded.

We found a clear dose-response association for knowledge and protective behavior. Other studies highlight this finding as well; exposure to multiple information sources increases the chances of demonstrating higher knowledge and protective behavior (31–34).

Previous studies show conflicting results as to whether knowledge is needed for behavioral change (19). From the results of our study, it seems that behavioral change can be achieved through directly influencing behavior as well as through a mediation effect of improvement in knowledge. Even though having knowledge appears to be beneficial for adopting protective behaviors, misconceptions mediate the effect of risk behavior. This connection was also seen in reality during the EVD outbreak in West Africa, where misconceptions about the virus led to sometimes violent resistance to public health measures (35).

In the analysis of the association between level of outbreak and knowledge and behavior, we found that risk behavior was prevalent during the entire course of the outbreak. In Sierra Leone, concerns were expressed about the response system, such as fears of calling the national hotline and of chlorine spraying (36). Ebola-specific information decreased in intensity after the outbreak, which might have resulted in loss of protective behavior and an increase in misconceptions. It can be assumed that repetition of messages is crucial.

In general, information dissemination during the Ebola epidemic was not unidirectional. For instance, radio channels encouraged listeners to call in and share questions. Communication channels should therefore be placed in a broader context, acknowledging the complex interactions among societal, community, and individual features, which in turn affect knowledge and behavior (37).

Major strengths of this study are the random sampling method, the large sample size, and the timing of the data collection. Even though the KAP survey instrument was not validated because of the urgency of the setting, KAPs are an established way of gathering data (15), and some of the adapted items have been validated in other contexts (38). In the surveys, some questions were asked in an open-ended manner, reducing the likelihood of bias because respondents would name only answers they actually knew. The response rate was high, which also reduced the risk of bias. Clear dose-response associations were found, strengthening the hypothesis that exposure to information sources can influence Ebola-specific knowledge and behavior. Further strengths include the possibility to adjust for potential confounders.

Limitations of the study include the fact that the data came from cross-sectional studies and that, therefore, temporality cannot be established. Consequently, directions of the found associations have to be interpreted with caution, as reverse causality cannot be ruled out. Performing mediation analyses with cross-sectional data can seem controversial because there is no way to properly measure if whether the exposure and mediator happen before the outcome. However, because EVD was a new disease in this region of Africa, Ebola-specific knowledge was low at the start of the outbreak (29). Furthermore, various information sources were used to disseminate EVD messages at the beginning of the epidemic. It can therefore be assumed that both the exposure and the mediator came before the outcome.

Another limitation is the inherent high correlation between time and district in an infectious disease outbreak, which was addressed by applying multilevel modeling. A further drawback is that, in contrast to KAP 2–4, KAP 1 contained only 9 out of 14 districts. The results were adjusted for region and district, somewhat mitigating this limitation. Furthermore, as previous studies have shown, there is a risk that self-reported behavior might not reflect actual behavior, but in the given outbreak situation, the survey was deemed to be the most feasible method (39). Closed-ended questions could have introduced a bias for the risk behavior outcome, owing to the risk of socially desirable answers. To minimize this factor, respondents were encouraged to give honest opinions.

Misclassification of the exposure to information sources could have influenced the results. The number and types of interventions during the EVD epidemic might have made it difficult for respondents to distinguish between information exposures. For instance, lay health workers and community mobilizers might have been viewed as health workers (which was classified as government exposure), overestimating the exposure to that information source (and because of the nondifferential nature of this potential bias, underestimating the results). Last, because exposures were not mutually exclusive during the data collection, a respondent could have been exposed to more than one source of information. To address this factor, for subsequent analyses, mutually exclusive media categories were created.

In conclusion, the results of this study show the importance of information sources in influencing knowledge and behavior—in both positive and negative directions—and could be used for communication strategies in emergency preparedness and disease outbreaks in low-income settings. Our findings underscore the value of risk communication for rapid disease control efforts and, therefore, for global health security.