**3. Data and Methods**

### *3.1. Study Area and Data Collection Method*

The study area covers five intersections in Yaoundé, Cameroon: (1) Central Post O ffice, (2) Carrefour Bata, (3) Carrefour Mokolo, (4) Vog-Ada, and (5) Etam Bafia (Figure 1). These hubs were selected on the basis of criteria such as: compliance with CPTED principles, the presence of a nearby police station, and relatively high density. At each intersection, a spatial delimitation of the area considered in the study was carried out as shown in Figure 1.

For contextual and e fficiency reasons, we opted for data collection by the paper-based method, entitled Interview Paper and Pencil (IPC). The on-field interview in the study areas was conducted for three weekdays (8, 9, and 11) on July 2019. To control for potential impacts of brightness level on FoC, the interview was conducted only during the daytime (from 10 a.m. to 6 p.m.). To carry out the data collection in the field, a team of 5 interviewers consisting of 2 male urban planners and 3 female social science students were involved, and they were divided into two teams with one urban planner each.

Because the target population of this study was pedestrians on the five intersections, it was di fficult to apply any pre-designed systematic sampling techniques. In addition, due to the severely congested local context, probability sampling method like stratified or random sampling could not be applied either. Accordingly, we had to adopt a non-probability sampling method, the so-called 'street corner sampling'. That is, each interviewer team approached the pedestrians present on the sidewalk (within red lines in Figure 1), asking for their consent before submitting the questionnaire. Although we did not count the exact number of people we approached, a significant number of them did not wish to respond to the questionnaire. Great care was taken to avoid any form of discrimination on the basis of age, ethnicity or any other discriminatory characteristics of the respondents, as the objective of the interviewers was to obtain the opinion of as many people as possible. Finally, a total of 186 pedestrians were surveyed. However, only 185 samples were used in the analysis because one respondent did not want to continue with the survey after completing the first part of the questionnaire. Apart from this, to preserve as many observations as possible, a mean imputation approach was

applied when treated minor missing values in two variables: Income Level and Sense of Community (Table 1).

**Figure 1.** *Cont.*

**Figure 1.** Intersection map: (**a**) Etam-Bafia (**b**) Bata (**c**) Vog-Mbi (**d**) Central Post (**e**) Mokolo. Note: We used several photos for each intersection to count people. The images presented above only aim to show the whole area of each intersection we are interested in.


**Table 1.** Definitions and descriptive statistics of the variables.


**Table 2.** Interviewees' responses on the two questions regarding Fear of Crime (FoC).

> Note: Bold cases were defined as people who felt FoC.

Each survey had an average duration of 30 min including the time taken to explain the purpose of the survey to the respondent. The questionnaire was structured into two main parts: (1) perceived FoC and (2) general socio-demographics, respectively served as dependent and control variables in the analysis. Details are described in Section 3.2. (Variables and Model Specification).

To measure intersection-level key test variable (i.e., 'pedestrian density'), we used aerial photographs (snapshot) taken by drone on the same day with the interview. Because a single still image cannot capture whole area of each intersection, we took 4 (Bata and Etam-Bafia) to 17 (Central Post) photos for each area and stitched them to count pedestrians on the sidewalk. Every photo had a resolution high enough to distinguish a person. We calculated the surface area of the sidewalk using the city map produced as part of the Yaoundé City Master Plan and available in AutoCAD format.

In our multi-level analysis, the number of clusters were very small, just five. However, Austin's Monte Carlo study suggested that five or less clusters can be considered, "as long as the number of subjects per cluster exceeds approximately 30" [48] (p. 18), and this study met this minimum requirement.

### *3.2. Variables and Model Specification*

As shown above, the data of this study are nested (i.e., pedestrians nested within intersections). While FoC and other socio-demographics are measured in individual level, pedestrian density is measured at the intersection level. Therefore, we applied hierarchical linear modelling (particularly, multi-level binary logistic regression analysis) using the IBM SPSS 25 software.

Level-1 variables include dependent variable and control variables. The dependent variable, 'fear of crime', was defined using two dichotomous choice questions: (1) "are you frightened to cross the intersection?" and (2) "do you feel like making a detour if approaching the intersection?" (see Table 2). Both questions were selected based on the results of previous study [49], in which they concluded that an individual's level of FoC could be observed and understood at di fferent psychological levels such as perception, cognition, and behaviour. Namely, the two questions were designed to capture these three dimensions of individual perception of FoC. By our operational definition, the pedestrians who answered "yes" to any of the two questions were defined as feeling FoC. Because respondents were likely to say "yes" to the above two questions due to the fear of tra ffic accidents, we asked them to answer the questions, considering only fear of crime before asking.

The other level-1 variables include various control variables deduced from the five theories of FoC summarised in the work of Vilalta [50] and other literatures suggested below.

(1) *Victimisation theory*: this theory is based on the assumption that people who have previously been victimised by crime are likely to su ffer from a higher level of FoC than those who have not [35,51]. Regarding this, we applied the 'victimisation experience' variables, which was measured by the following question: "have you ever been a victim of a crime, personally witnessed a crime, or heard of a crime in your surroundings?".


In addition to them, marital status, religion, stance on CCTV, time of the day, and weather condition were applied as level-1 control variables.

As explained above, level-2 (i.e., intersection level) variables include our key test variable: 'pedestrian density'. This was defined as the number of people on the sidewalk captured by drone photographs divided by its area (person/m2). To test the potential non-linearity discussed in Section 1, we took quadratic regression forms by applying both density and density squared variables. Table 1 shows the definitions and descriptive statistics of the variables.
