Determinants of Knowledge About and Use of Dietary Supplements

1. MEASURES ASSESSED IN THE STUDY:

The selection of measures to be evaluated in the study was based on the scientific literature review and personal interest of the researchers:

1. Measures related to dietary supplements:
2. knowledge about dietary supplements (KaDS) - assessed with a 17-item KaDS questionnaire, which was developed as a Polish version by Karbownik et al. (2019). A respondent will be asked to rate each of the 17 statements concerning dietary supplements as "true" or "false". KaDS will be primarily coded as a sum of both General and Specific questionnaire subscales. KaDS will be operationalized as an 18-level ordinal variable.
3. self-rated sources of KaDS - assessed separately for 5 categories: "medical doctors", "pharmacists", "dietitians", "friends (with no medical education)", "media (magazines, TV, radio, Internet)". A single-item question will be asked for each category: "to what extent do you get knowledge about dietary supplements from...?". Each category will be operationalized as a 4-level ordinal variable (from "not at all" to "to a large extent"). TV - television.
4. use of dietary supplements (UoDS) - assessed with a single-item question: "have you used any dietary supplements within the past 30 days?". UoDS will be operationalized as a 2-level categorical variable ("no" and "yes").
5. positive personal experience with dietary supplements - assessed with a single-item question: "if you take dietary supplement, do you feel it helps you?". The measure will be operationalized as a 3-level categorical variable ("no", "yes" and "not applicable, I don't use dietary supplements"). Cases presenting "not applicable..." response, will be coded in the analyses as "no".
6. negative personal experience with dietary supplements - assessed with a single-item question: "if you take dietary supplement, do you feel it hurts you?". The measure will be operationalized as a 3-level categorical variable ("no", "yes" and "not applicable, I don't use dietary supplements"). Cases presenting "not applicable..." response, will be coded in the analyses as "no".
7. interest in dietary supplements - assessed with a single-item 5-point semantic differential: from "No. This is completely indifferent to me." to "Yes! Every day I look for information on this topic.". The measure will be operationalized as a 5-level ordinal variable.
8. trust in advertising dietary supplements (TiADS) - assessed with an 8-item TiADS questionnaire, which was developed as a Polish version by Karbownik et al. (2019). A respondent will be asked to express her/his opinion about the information conveyed by the advertisements of dietary supplements using a 5-point semantic differential scale. TiADS will be primarily considered as a sum of Reliability, Intelligibility and Affect questionnaire subscales. TiADS will be operationalized as a 33-level ordinal variable.
9. having contact with dietary supplement advertisements - assessed with a single-item question: "have you had any contact with dietary supplement advertisements within the past week?". The measure will be operationalized as a 2-level ordinal variable ("no" and "yes").
10. Measures related to other health issues:
11. general beliefs about medicines (BMQ-General) - assessed with an 8-item General part of the Beliefs about Medicines Questionnaire, which was developed by Horne et al. (1999), and adapted to Polish language and validated by Karbownik et al. (2019) (paper currently under peer-review). A respondent will be asked to express her/his agreement with 8 statements concerning medicines using a 5-point Likert scale: from "completely disagree" to "completely agree". BMQ-General will be coded as 2 separate variables: BMQ-General-Overuse subscale and BMQ-General-Harm subscale. Both the subscales will be operationalized as 17-level ordinal variables.
12. self-rated health - assessed with a single-item question: "how do you assess your health?". The measure will be operationalized as a 4-level ordinal variable ("poor", "fair", "good", "excellent").
13. self-rated diet - assessed with a single-item 5-point semantic differential: from "I eat a lot of fast food, chips, sweets, etc." to "I only eat healthy, balanced meals". The measure will be operationalized as a 5-level ordinal variable.
14. self-rated physical activity - assessed with a single-item 5-point semantic differential: from "I have no physical activity at all" to "I play sport intensively 5 times a week". The measure will be operationalized as a 5-level ordinal variable.
15. conventional cigarette smoking - assessed with a single-item question: "do you smoke conventional cigarettes?". The measure will be operationalized as a 3-level categorical variable ("never", "no, but I smoked in the past", "yes").
16. electronic cigarette use - assessed with a single-item question: "do you use electronic cigarettes?". The measure will be operationalized as a 3-level categorical variable ("never", "no, but I used in the past", "yes").
17. Measures of sociodemographic data:
18. age - operationalized as a continuous variable rounded to whole years. The allowed values to be reported will be: "below 18", "18", "19", "20", ..., "113". The oldest Pole currently (October 23, 2019) living is 113-year-old.
19. sex - operationalized as a 2-level categorical variable ("male" and "female").
20. education level - operationalized as a 5-level ordinal variable ("primary", "secondary or vocational", "higher-bachelor", "higher-master", "higher-doctorate").
21. medical education - operationalized as a 2-level categorical variable ("no medical education" and "medical education"). This measure will be used as auxiliary to check whether a respondent meets exclusion criteria. Missing values in this variable will be coded as "medical education".
22. number of inhabitants in a place of residence - operationalized as a 4-level ordinal variable ("below 5,000", "5,000-50,000", "50,000-500,000", "over 500,000").
23. monthly net household earnings per family member - operationalized as a 4-level ordinal variable ("below 1,000 PLN", "1,000-2,000 PLN", "2,000-3,000 PLN", "over 3,000 PLN"). PLN - Polish currency (zloty).

II. DATA ANALYSIS PLAN

1. A study participant will be asked to fill the online survey (https://www.survio.com/pl/) assessing all the above measures.
2. Any cases of respondents, who declared having less than 18 years of age or being medically educated will be deleted from further analyses (as stated in the inclusion and exclusion criteria).
3. All the variables measured with semantic differential scale will have no missing values, because the online survey software attributes a central value ("3") to a response as a default. Missing values are possible in the other variables within the survey, as there will be no forced answering option and no default responses. The missing values will be managed in the following way:
4. The cases with no data, apart from the default values in semantic differential variables, will be deleted.
5. The fraction of missing values for each variable will be calculated.
6. The fraction of cases with at least one missing value will be calculated.
7. Variables with more than 50% of missing values and cases with more than 50% of missing values will be deleted from the database.
8. The pattern of data missingness will be assessed with Little's test for data missing completely at random (MCAR). In case of significant violation of MCAR assumption, for each variable, the association of missingness with all the other variables will be examined.
9. If the fraction of cases with at least one missing value (see point "c") is less than 5% and the result of Little's MCAR test (see point "e") is non-significant, the cases with at least one missing value will be deleted and the analysis will be performed with complete cases only.
10. Otherwise (to point "f"), the missing values will be imputed with multivariate imputation by chained equations (MICE) before any further data analysis is performed. The summation of the items constructing KaDS, TiADS and BMQ will be done after imputation of missing values in each single item.
11. KaDS will be treated as a continuous variable and modeled with multivariate linear regression, whereas UoDS, being a 2-level categorical variable, will be modeled with logistic regression. Ordinal independent variables will be considered as continuous, while included into regression models. Categorical independent variables with more than 2 levels will be converted to dummy variables before being included into regression models. The dependent variables (DVs), KaDS and UoDS, will be modeled in the following steps:
12. Univariate (unadjusted) associations of DVs with all the remaining measures will be assessed (KaDS will be additionally split into General and Specific subscales for the presentation of the results). In addition, the two-way interaction "TiADS"×"having contact with dietary supplement advertisements" will be included. The associations will be reported as both unadjusted and adjusted for all the tested sociodemographic measures (see point "I. 3"). If the data is imputed (see point "3. g"), sensitivity analysis of univariate (unadjusted) associations between DVs and the remaining measures will be performed in the dataset of complete cases only to test the accuracy of missing data imputation. The further multivariate models will be built with the measures statistically significantly (according to the raw p-values in unadjusted analyses) associated with a DV in these analyses.
13. Predictors to be retained in the final model will be selected according to the following criteria:

• Multicollinearity. Multicollinearity of the variables will be assessed with the Pearson's correlation matrix and exploratory factor analysis. In case of detection of multicollinearity, two solutions will be considered: (1) substantially collinear variables may be linearly combined if they may represent the same construct (e.g. "self-rated diet" plus "self-rated physical activity" plus "no cigarette smoking" may represent the construct of "health self-care" or "conventional cigarette smoking" plus "electronic cigarette use" may represent the construct of "nicotine dependence", etc.) or (2) only one variable from a set of substantially collinear variables may be retained and the other may be deleted.
• "Objectivity" of the retained measure. Highly "objective" measures will be favored. They include the one with more convincing proof of validity: KaDS, TiADS and BMQ, followed by the "objective" measures assessed with a single item question: age and sex, followed by more "subjective", not easily verifiable and possibly biased measures: the rest of the measures.
• Information criteria. The models with lower values of Akaike or Bayesian information criterion will be preferred (best subset selection algorithm).

1. Residual analysis of the obtained models will be performed to check for assumptions of general linear modeling. In case of substantially violated assumptions, the processes of model construction may be repeated.
2. Internal validation will be performed with a k-fold cross validation to test for accuracy and stability of the obtained models.
3. Subgroup analyses may be performed in the samples of respondents critical to public health:
4. elderly people (60 or more years of age),
5. rural areas residents (below 5,000 inhabitants),
6. low income people (below 1,000 PLN of monthly net household earnings per family member), etc.
7. P-values throughout the analyses below 0.05 will considered statistically significant.To account for multiple hypothesis testing, if applicable, significance level correction with Benjamini and Hochberg procedure will be applied (false discovery rate 0.05).
8. The analyses will be performed using STATISTICA Software (Statsoft) or R Software (R Core Team).