Recruitment and data collection
The was a cross-sectional study of adolescents and young adults aged 15-25 attending high schools, colleges and universities. In Kenya students normally attend colleges and universities from 19-23 year of age, but we added two years to take account of delayed entry. However, because there was a larger number of mature students than expected, the number who were over the age of 25 was only 3.4% and we decided to include all respondent in the analysis.
The participants were all Kenyans, who were recruited from Nairobi County and three counties in South Eastern Kenya: Machakos, Kitui and Makueni. We were already working with these four countries on other project and had gained acceptance from local people, so they seemed a natural choice for this project. The high school studies were all from these areas and the college and university students came from across Kenya as admission are handled centrally. The study comprised a convenience sample of following:- (1) one administrative location in which there are several high schools and because of school closure we only accomodated those students who could make it to the data collection points (2) all colleges located within the four counties and were in session at the time of the data collection and (3) one public university located in one of the four counties. It was conducted from July 2016 to October 2018. The subjects were recruited from the colleges and universities and the high school students were recruited from their local communities, because the schools were closed during the study period as the result of a lengthy teachers’ strike. The college and university students were approached in their classrooms after lectures, once the various institutions provided permission for their students to take part in the study. Permission to approach the high school students was obtained from local community administrators.The local administrators contacted the students and asked them to come along to local community centres at a specific time and date. Of all the 9,742 participants:- 6648(68.6%) were university students and 1534(15.8%) were college students who were approached in their classrooms agreed to take part in the study. Of the 1506(15.5%) high school students who presented themselves at the data collection points, agreed to participate. It is a Government requirement to seek permission for any activities taking place in the community so that logistical support can be provided, such as suitable meeting venues and security. The research assistants were informed of the schedules for the college and university students and the high school students were directed to specific public meeting areas, within walking distance, where they assessed them with the help of local community leaders. Participants were only included in the study if they were able to speak, read and write in English and had voluntarily agreed to participate in the study by signing the informed consent form. Consent was obtained from parents and guardians if the participants were under 18 years of age.
We worked with the colleges and universities to make sure there was mental health support available if the students needed help because of the issues raised by the survey. Trained staff from local health center facilities, who had received training on the World Health Organization (WHO) Mental Health Gap Action Programme Intervention Guide  were available to support the high school students, if needed. This instrument was developed for use by non-mental health specialists to identify common mental health issues and suggest interventions that they could provide. We also informed the participants where they could seek help at institutional and community levels.
We used several instruments
Socio-demographic characteristics. The questionnaire included questions about socio-demographic variables, including age, gender, whether they were attending high school, college or university, marital status and birth order.
Economic indicators. The respondents provided details of their household, including what items were in their home and how they accessed water and toilet facilities and cooked. They were put into one of five wealth index categories, as a reflection of their economic status. The wealth index we used was based on the World Bank recommendation for low-income and and middle-income countries and has been adopted by the Kenyan Government. It contains five levels, with one representing the lowest level of wealth and one indicating the highest level.
Psychiatric conditions. The Psychiatric Diagnostic Screening Questionnaire (PDSQ) was used to assess the respondents. It comprises 126 questions that assess the symptoms of 13 Axis I disorders in the Diagnostic and Statistical Manual of Mental Disorders, fifth edition . These are: eating disorders (bulimia/binge-eating disorder), mood disorders (major depressive disorder), anxiety disorders (panic disorder, agoraphobia, post-traumatic stress disorder, obsessive-compulsive disorder, generalized anxiety disorder and social phobia), substance use disorders (alcohol abuse/dependence and drug abuse/dependence) and somatoform disorders (somatization disorder and hypochondriasis). It also contains a six-item psychosis screen. The disorders that were selected were the most prevalent in community-based epidemiological surveys[40,41] and the most frequently reported in large clinical samples[42–44]. In a validity study of 994 psychiatric outpatients, the 13 PDSQ subscales demonstrated good to excellent internal consistency. Cronbach’s alpha exceeded .80 for 12 of the 13 subscales and the mean of the alpha coefficients was .86. Test-retest reliability was examined in 185 subjects who completed the PDSQ twice in one week. The test-retest reliability coefficients exceeded .80 for nine subscales and the mean of the test-retest correlation coefficients was .83. The convergent and discriminant validity of the PDSQ subscales were examined in 361 patients who completed a package of questionnaires at home less than a week after completing the PDSQ. The last six questions from the PDSQ major depressive episode domain are used to measure suicidal ideation. These are related to people frequently thinking of dying in passive ways, like going to sleep and not waking up, wanting to be dead, thinking they would be better off dead, thinking about suicide, seriously considering taking their life and thinking about specific ways to take their life. The questions are coded as no (zero points) or yes (one point).
Other Measures: The Washington Early Recognition Center Affectivity and Psychosis (WERCAP) screen[47,48] was used to quantitatively assesses psychosis-risk symptoms and bipolar-risk symptoms (affectivity) based on the frequency of symptoms and their effects on functioning. It has high test-retest reliability and validity, with affectivity of sensitivity of .91, specificity of .71, psychosis sensitivity .88 and specificity of .82. We also used the WERC stress screen, a self-report questionnaire, to assess total stress burden and the severity of individual stressors [47,48]
Data management and statistical analysis
The coded data were checked, cleaned and exported into SPSS,version 21 IBM Corp, NY, USA
Creation of Suicidal Index scores
Data reduction techniques were used to summarize the observed suicidal ideation variables, namely the last six questions of PDSQ on depression subscale, into a few dimensions by Rasch analysis through latent variable modelling using the eRm, ltm and difR R packages (R Foundation, Vienna, Austria). Component internal consistency and reliability were used to compute the suicidal ideation scores, by calculating Cronbach's alpha and this was high (.776).
Before we performed the Rasch factor analysis, the correlation matrix was inspected to check for the strength of the correlation. Then the factorability was tested using exploratory factor analysis using the Kaiser-Meyer-Olkin measure of sampling adequacy and Bartlett's test of sphericity. Exploratory factor analysis, with varimax rotation, was carried out to determine the dimensional structure suicidality using the following criteria: (a) eigenvalue >1, (b) variables should load >.50 on only one factor and less than 0.40 on other factors, (c) the interpretation of the factor structure should be meaningful and (d) the Scree plot should be accurate when the means of communalities are above .60 . Computations were based on a covariance matrix, as all the variables received values from the same measurement scale . Bartlett's test of sphericity with p <.05 and a Kaiser-Meyer-Olkin measure of sampling adequacy of .6 were used when performing this factor analysis. A factor was considered as important if its eigenvalue exceeded 1.0 . Pairwise associations between the six items, corresponding to two-by-two contingency tables for all possible pairs, were computed. The Cronbach's alpha of the six items was.776, with only one component loading. Factor scores were then generated because loadings were all similar. The scores had bimodal negatively skewed distribution, suggesting there were two groups. Respondents scoring less than zero were classified as not having suicidal ideation while those with more than zero were classified as having suicidal ideation.
The results of the exploratory and statistical data analysis are presented in the tables. We employed descriptive statistics to estimate the prevalence of suicidality as well as the participant’s characteristics. Mean prevalence rates were estimated. The outcome variable of suicidality scores were grouped into those with, and those without, suicidal ideation. Univariate associations between suicidal ideation and other variables were estimated using bivariate logistic regression, after they were fitted to identify potential confounding factors. Variables with a P-value of less than <.05 were entered into generalized linear models using logit link to identify independent predictors of suicidality. Adjusted odds ratios (aOR) with 95% confidence intervals (CI) were calculated to assess the strength and significance of the association. All tests were two-sided and statistical significance was set at p <.05. We did not include depression in the analysis of psychiatric disorders, because of high collinearity between depression and suicidality.
Overlap between suicidal ideation, depression and the wealth index
We triangulated overlaps of the poorest and highest wealth indicator (quintiles 1 and 5) and the least and most potent suicidal ideations, which were numbers seven and one on Table 3. We then further triangulated this with depression, which was the mental disorder most significantly associated with suicidal ideation in the literature. Chi-square tests were used to test whether there were significant associations between the wealth index, suicidal ideation and depression.