This cross-sectional study was conducted among undergraduate students of Hawassa University in southern Ethiopia. At the time of data collection, the university had over 17,000 regular undergraduate students. More description of the university and the study population is available elsewhere [20].

Because we also had a plan to determine khat chewing practice and its associated factors among the student population, we restricted the sample population to only students who were in their second year of university and above. Those in their first year were not included because they were considered to be in a buffer period transitioning from being a high school student to a university student [20].

Initially, the statistical method proposed to analyse the data was a Poisson regression; hence, the sample size required for a Poisson regression analysis was estimated using the software G*Power version 3.1 (Franz Faul, Universität Kiel, Kiel, Germany) [21]. However, Poisson regression was not performed, and the data were analysed using a negative binomial regression. The following assumptions were used in this calculation: a to-be-detected effect size (exp [β₁]) of 1.5, alpha-value of 0.05 (two-tailed), power of 80%, disease rate in non-chewers (base rate exp [β₀]) of one (based on our best guess since there was no prior information available), fixed mean exposure time of one month and variance accounted for by other predictors (R² other X) of 0.5. Since there was no prior information on the variance, this value was used instead of the default zero to maximize the sample size. In addition, a binomial distribution for the exposure variable, long-term regular khat chewing (i.e., X distribution=binomial), was used. The parameter of the exposure variable (X-param π) was set at 28% based on the current prevalence of khat chewing among students of Axum University, Ethiopia [22]. This percentage was used as a proxy for the prevalence of long-term regular khat use since it has not been previously estimated. With these assumptions, the required sample size was found to be 377. Considering a design effect of 1.5 and a non-response rate of 5%, the final estimate was 595. As this study was part of a larger study designed to address other objectives that require a sample size as large as 1,290, the sample size requirement to address our objective for this paper was fulfilled.

The sampling was accomplished using a two-stage stratified cluster sampling design. All undergraduate study programmes within the university constituted the primary sampling units. From each college/institute, two primary sampling units were selected using simple random sampling. Within each primary sampling unit, sections/classes formed the secondary sampling units. Then, the secondary sampling units were stratified by university year of study (second year, third year, fourth year and above). From each year of study, one secondary sampling unit was selected using simple random sampling. All of the students in the selected secondary sampling unit were included in the study since students within a university are assumed to be relatively homogenous.

The data collection was accomplished using a pre-tested self-administered questionnaire. The questionnaire contained items to assess khat chewing behaviour and other relevant variables. It also contained nineteen items about oral symptoms such as toothache, sores/irritations in the mouth, gingival bleeding, oral malodour, dry mouth, teeth sensitive to heat/cold, and temporo-mandibular joint pain. These items were adapted from the study of Locker and Miller [23]. Such subjective oral health status assessment techniques have been used in different populations worldwide and have been found useful for epidemiological studies on the oral-health status of populations where clinical assessments of oral health are costly [23,24].

Data were collected in a class after the necessary information on privacy and confidentiality and a comprehensive instruction on how to complete the questionnaire were provided to the study participants. While an Amharic-version of the questionnaire was mainly used for the data collection, the English version of the questionnaire was made available for those who reported difficulty understanding Amharic. All data collection sessions were facilitated by facilitators who were not part of the respective campus community. In addition, no one outside of the study was allowed to enter the classrooms at the time of data collection to ensure maximum privacy.

The dependent variable was oral symptom count. Respondents were asked a series of nineteen questions on whether they had specific oral symptoms within the four weeks preceding the study. Then, a new variable was computed for the number of symptoms each respondent reported, and this new variable was used for subsequent analyses in this study as a count variable.

The major explanatory variable was long-term regular khat chewing defined as khat chewing for at least three years with a frequency of at least once a week. Regular chewing was categorized as yes or no. Other independent variables were also considered mainly to control for their potential confounding effects on the association between khat chewing and oral symptom count. These variables included cigarette smoking in the past 12 months (yes/no), usually using sugar in hot drinks (yes/no), frequency of soft drink consumption (less than once a week or never vs. one or more times per week), self-reported dental fluorosis (yes/no), tooth brushing (using a brush and tooth paste) frequency (not at all, less than once a day, one or more times per day), alcohol consumption in the past 12 months (yes/no), amount of monthly pocket money (in quintiles) and sex (female vs. male). Monthly pocket money was used as a proxy indicator of socioeconomic position.

To facilitate detection of data-entry errors, all data were entered twice using the software EPI Info version 3.5 (Centers for Disease Control and Prevention, Atlanta, GA, USA). Data entry errors were corrected by comparing the two independently entered datasets. Further data analyses were accomplished using Stata version 12 (StataCorp., College Station, TX, USA).

As is commonly encountered in self-administered questionnaires, we had some variables with item-missing data. Multiple imputation using chained equations [25] was performed in Stata to replace missing items and to retain as many of the respondents as possible in the analysis. Twenty imputations (M=20) were created for each missing item using a burn-in period of 10 after checking the convergence of the imputation models using trace plots, as recommended in the Stata manual. Augmented regression was used to handle cases of perfect prediction when using categorical predictors [26]. Most of the important variables with item-missing data were successfully imputed. However, because the imputation models failed to converge, some variables were not imputed. Sensitivity analyses were employed by comparing the descriptive summaries of the imputed variables before and after imputation, and the analyses showed comparable results confirming the absence of any aberrations during the imputation [26].

The analysis was performed considering the complex sampling design. Hence, the stratification and clustering both in the first and second stages of sampling were taken into account. In addition, an analysis weight was employed that accounted for the unequal probability of selection in the first and second stage sampling units as well as a post-stratification adjustment factor based on the sex of the respondents [27].

To determine the association between long-term regular khat chewing and oral symptom count, a negative binomial regression was used instead of a Poisson regression since the assumption was not fulfilled due to the overdispersion of the data [28]. Univariable negative binomial regression using the major explanatory variable and each of the possible confounders was undertaken as the initial step. Then, variables with p-values of less than or equal to 0.25 in the univariable regression and those deemed important based on the literature were selected for inclusion into the initial multivariable model [27]. Serial refining and refitting of the initial multivariable model was also carried out. Variables with insignificant coefficients whose removal did not significantly affect the model as a whole and individual variables remaining in the model were deleted. Any improvement in the model was checked by including relevant interactions and higher-order terms one at a time, but none was found significant. Thus, the final model was a main-effects model containing the major explanatory variable (long-term regular khat chewing) and four other covariates (cigarette smoking, self-reported dental fluorosis, frequency of soft drink consumption and pocket money). After running the final model, the average adjusted predictions of oral symptom count by khat chewing status were estimated using Stata’s user written command “mimrgns” [29]. The reproducibility of the results, which were based on multiply-imputed data, was assessed using Monte Carlo error estimates as per White et al.’s [25] guidelines. The Monte Carlo error estimates suggested that the results are quite reproducible. In other words, our results would be stable “across repeated uses of the same imputation procedure” [26].

This study was approved by the institutional review board (IRB) of the College of Medicine and Health Sciences of Hawassa University, Hawassa, Ethiopia (IRB No. IRB/165/05). The respondents did not provide any personal identifiers on the questionnaires to ensure the anonymity of their responses. Respondents took part in the study voluntarily after a verbal informed consent that was endorsed by the aforementioned IRB.