GENERAL COMMENTS

Though the paper is generally well written and provides substantiated rationale for the raised research question, there are several concerns about the analytic approach that was used. While the use of data for several consecutive years could definitely be an advantage, change of the response options wording introduces measurement bias, and all the speculations regarding the increases or decreases in particular sources of cigarette purchase become groundless.

Moreover, the Tables 3 and 4 are only devoted to 2012 data, so they do not refer to the research question of whether sources of cigarettes change over time. Overall, authors can be invited to make their line of analysis more logical and targeted to their research questions. Several minor comments: Analysis is insufficiently explained in the Methods section. In Table 1, regular and intermittent smokers are considered as portions among smokers, not among all the surveyed adolescents, which leaves the processes of changing prevalence of smoking aside.

Results presented in Table 2 are shown without pointing to whether they are related to all smokers or regular ones. Second paragraph after Table 2 has the style pertinent to the Discussion section, not the Results. Table 3: (5%CI) is written instead of (95%CI). The Table shows bivariate analysis results, and confounders are not controlled for, if they were some of the associations with the smoking status and the age would be attenuated.

GENERAL COMMENTS

1. The approach of the authors to explore the prevalence of regular smokers among smokers not among all the population is questionable.

2. Because 'getting cigarettes from friends' in 2006-2010 was limited to 'getting cigarettes from friends of same age' in 2011-12, and 'getting cigarettes from older friends' obviously moved to 'other sources', this could obviously account for the changed authors observe and derive their conclusions from.

3. As the question about sources of cigarette supply was not asked as a multiple choice question (sum of answers is 100% as far as I can see) it is not clear why authors apply binary logistic regression while multinomial regression would be more appropriate in this case