Estimation of smoking prevalence in Canada: implications of survey characteristics in the CCHS and CTUMS/CTADS.
This commentary provides insight into a simple yet difficult-to-answer question: What is the prevalence of smoking in Canada? Many data sources can be used to estimate the extent of smoking behaviour in Canada and its provinces (e.g., a 2015 directory updated by the Ontario Tobacco Research Unit lists more than 60 sources (2)). In contrast to other countries, such as New Zealand, our national census does not inquire into Canadians' smoking status. The assessment of the prevalence of smoking in Canada, therefore, requires specific national surveys. To fulfill the mandate of producing representative estimates of smoking prevalence at the national and regional levels, Statistics Canada operates two main surveys: the Canadian Community Health Survey (CCHS) and the Canadian Tobacco, Alcohol and Drugs Survey (CTADS). Upon closer inspection, understanding the estimation of smoking prevalence requires a critical appreciation of these surveys and their idiosyncrasies.
While a definitive answer (or number) cannot be offered here, this commentary proposes to contribute to this issue by reviewing the two main data sources that are used to study smoking prevalence, examining how smoking prevalence is measured in these surveys, and exploring potential reasons for their noticeable discrepancies.
ESTIMATING SMOKING PREVALENCE: THE CANADIAN COMMUNITY HEALTH SURVEY AND THE CANADIAN TOBACCO, ALCOHOL AND DRUGS SURVEY
The CCHS is a repeated cross-sectional survey created in 2000 that collects annually information related to health status, health care utilization and health determinants for the Canadian population. It collected and published information bi-annually from 2000-2001 through to 2007, whereupon it became an annual survey. (3) The CTADS is a repeated cross-sectional survey created in 2013 to provide bi-annual data on tobacco, alcohol and drug use and related issues, with a focus on 15-24 year olds. (4) The CTADS built upon the longstanding Canadian Tobacco Use Monitoring Survey (CTUMS), which repeated annually from 1999 until 2012, and was combined with Statistics Canada's last alcohol and drugs survey, the Canadian Alcohol and Drugs Use Monitoring Survey (CADUMS).
How do these surveys measure smoking? The most common definition of smoking is "current smoking status", which includes daily and non-daily (occasional) smokers based on responses to the question "At the present time do you smoke cigarettes every day, occasionally, or not at all?". This question was developed in the mid-1990s to standardize measures across Statistics Canada surveys. (5) In order to ensure consistency, this question follows no skip patterns and its wording and ordering within the full questionnaire remains almost identical across surveys. Notably, the CTUMS/CTADS questionnaire adds the following preamble: "I am going to start with questions about cigarette smoking. Include cigarettes that are bought ready-made as well as cigarettes that you make yourself."
Figure 1 presents the prevalence estimates in the CCHS and CTUMS/CTADS between 2001 (the first year of the CCHS) and 2013 (the latest available year for CTADS data when this commentary was written) (full estimates are given in Supplementary Table 1, in the ARTICLE TOOLS section on the journal website). During this period, smoking prevalence changed from 25.9% (95% CI: 25.6-26.3) to 19.3% (95% CI: 18.7-19.9) in the CCHS; and from 21.7% (95% CI: 20.6-22.9) to 14.6% (95% CI: 13.5-15.8) according to the CTUMS/ CTADS. The latter reports a systematically lower smoking prevalence estimate than that of the CCHS. Percentage point differences varied from 2.1 (2003) to 4.7 (2013) with an average difference of 3.4 between surveys. Readers can compare the non-overlapping confidence intervals as a conservative test to confirm that the differences in estimates are statistically significant.
EXPLORING POTENTIAL REASONS FOR DISCREPANCIES BETWEEN SURVEYS
Why do these sources produce different smoking prevalence estimates? Statistics Canada's Health Division addresses this concern by providing the following warning:
"Users should be aware of a number of differences between CCHS and CTADS. CCHS collects information from respondents aged 12 and over, CTADS collects information from respondents aged 15 and over; the two surveys use different sampling frames; the annual sample for CTADS is 20,000 compared to 65,000 for CCHS; in CCHS, smoking questions are asked in the context of a wide range of health-related behaviours whereas in CTADS, all questions are related to the use of multiple products and substances with addictive properties." (6)
Table 1 presents and compares the main survey design elements of both surveys (i.e., coverage, sampling, non-response and measurement). Differences suggested here above can be summarized to age range, sampling frame, sample size and survey context. Some of these differences are unlikely to explain the lower prevalence systematically observed in the CTADS. First, the lower age group included in the CCHS (12+ versus 15+) should not produce a higher estimate because smoking rates at ages 12-14 years are significantly lower than those in the rest of the population. Second, the sample sizes used in both surveys are determined based on their statistical ability to provide precise estimates at the national level; therefore, they are unlikely to be at the base of their consistently different results. On the other hand, differences in sampling frame and questionnaire content may contribute to the discrepancies between CCHS and CTUMS/CTADS.
Statistics Canada has been aware of the potential for producing differences in smoking prevalence estimates across its different surveys since the development of the CTUMS and its first results in 2000-2001. A working group was formed to tackle this issue and a summary of findings from an unpublished internal report was made available online. (5) This report explored several hypotheses related to sampling frame, use of proxy respondents, questions' wording and ordering, mode of administration and survey context. Among these factors, two were deemed to be potential sources of bias. First, it was proposed that CCHS used a warmer contact approach: participants were contacted in advance for face-to-face interviews, and questions regarding smoking status were asked only after a bond of confidence was established. On the other hand, CTUMS resorted to phone interviews with no advance contacts, and surveyed smoking status earlier in their questionnaire. Second, the report referenced a study documenting discrepancies in the measurement of disabilities (i.e., an outcome liable to social desirability bias and under-reporting) and proposed that respondents could be more open to report their characteristics when it was done in a general health survey (i.e., CCHS) that did not label them as smokers.
The importance of these two factors may have changed since CTUMS' inception. In particular, the new format of the CTADS may further influence the responses of respondents with the inclusion of alcohol- and drug-related issues. Drugs and alcohol surveys tend to have higher non-response rates, which in turn are known to be associated with smoking behaviour. (7,8) This issue is especially problematic given the relatively strong association between cigarette smoking, alcohol and drug use among Canadians. (9) This inclusion may have resulted in survey avoidance by potential respondents because of their alcohol or drug use. The extent of this influence is, however, unknown. Zhao et al. (2009) compared non-respondents in the 2004 Canadian Addictions Survey (CAS, a past iteration of the CADUMS) with those in the 2002 CCHS and found that their samples were not significantly different in terms of educational attainment, household income and marital status. (10)
Other factors have also changed in importance over time as well. For instance, the CCHS and CTADS use a slightly different sampling frame: the CCHS used a combination of an area frame (40%-50%), lists of telephone numbers (40%-50%) and randomdigit dialing (RDD) ([+ or -] 1%) to cover their sampling frame while the CTUMS/CTADS used an RDD-only strategy. RDD cannot reach potential respondents without landline phones (i.e., those without phones or with only cellphones). This sampling frame can translate into a potential selection bias because these individuals without landline are often younger, have lower incomes and also higher smoking rates. (11) In 2002, the Statistics Canada report comparing CCHS and CTUMS argued that these differences were marginal given that only a small percentage of the population did not have landline phones. This argument is however untenable today: in 2013, 21% of Canadian households and 60% of individuals aged 35 years or younger had a cellphone, but no landline phones. (12) Fortunately, CTADS updated their sampling frame in 2015 to include telephone lists that include cellphones.
Canadian public health policy and practice face continued challenges as a result of increasing social inequalities in smoking. (1,13) Notably, this behaviour is increasingly concentrated within specific subpopulations, such as young adults and those who are socio-economically disadvantaged. This necessitates the continual adaptation of surveillance infrastructures, including the use of new survey methodologies, statistical approaches, and information technologies in order to provide the best evidence. There is considerable Canadian research devoted to finding solutions to survey-related issues that arise within the broader context of decreasing response rates and new information technologies. (14) The application of new approaches in survey methodology nonetheless constitutes a considerable hurdle for longstanding surveys, as it hampers their ability to attribute changes in estimates to true behavioural changes rather than to changes in methodology themselves.
Beyond comparisons between CCHS and CTUMS/CTADS, another legitimate question is the extent to which the current utilization of these surveys produces unbiased estimates. The standard approach used to produce prevalence estimates (e.g., descriptive statistics with survey weighting) can also benefit from recent innovations that adjust for sources of error such as nonresponse and under-reporting in order to obtain stronger estimates of smoking prevalence. For instance, Zhao et al. (2009) used predictive models to correct for non-response associated with socio-demographic characteristics (e.g., marital status, educational attainment, household income) and obtained higher prevalence estimates of alcohol and drug use in the Canadian Addiction Survey, with percentage point differences ranging from 0.9 to 2.6.10 Likewise, Land et al. (2014) examined predictors of smoking status under-reporting during pregnancy and used a predictive model integrating socio-demographic correlates to produce more accurate smoking prevalence estimates among pregnant women (in their case, under-reporting lowered self-report estimates by values ranging from 1.1 to 1.9 percentage points). (15)
This commentary aimed to present how smoking prevalence is reported and measured in Canada and to introduce some of the challenges associated with its estimation. Researchers should be more critical of prevalence estimates taken at face value, especially when comparing sources with different methodologies and/or from other countries. Although these challenges influence single-point prevalence estimates, the trends produced by both surveys are extremely consistent over time. Rather than comparing smoking rates produced from the two surveys, Statistics Canada advises users to choose a single source, based on their objectives, and to use that source consistently.
Researchers need to acknowledge the uncertainty over the true value of smoking prevalence in order to address it. To this end, tobacco surveillance experts are encouraged to use multiple data sources and more precise estimation approaches whenever applicable, to systematically discuss their limitations when reporting on prevalence and its change over time, and to continue to explore how potential sources of error should guide the development and utilization of available data sources.
(1.) Corsi DJ, Boyle MH, Lear SA, Chow CK, Teo KK, Subramanian SV. Trends in smoking in Canada from 1950 to 2011: Progression of the tobacco epidemic according to socioeconomic status and geography. Cancer Causes Control 2014; 25(1):45-57. PMID: 24158778. doi: 10.1007/s10552-013-0307-9.
(2.) Ottawa Tobacco Research Unit (OTRU). Directory of Public Use Data on Tobacco Use in Canada. 2015. Available at: http://otru.org/wp-content/uploads/2015/ 08/DirectoryJune2015.pdf (Accessed January 31, 2017).
(3.) Statistics Canada. Canadian Community Health Survey (CCHS). Available at: http://www23.statcan.gc.ca/imdb/p2SV.pl?Function=getSurvey&SDDS=3226 (Accessed August 15, 2016).
(4.) Statistics Canada. Canadian Tobacco Alcohol and Drugs Survey (CTADS). Available at: http://www23.statcan.gc.ca/imdb/p2SV.pl?Function=getSurvey &SDDS=4440 (Accessed August 15, 2016).
(5.) Gilmore J. Report on Smoking in Canada 1985 to 2001. Statistics Canada, 2002. Available at: http://publications.gc.ca/collections/Collection/Statcan/82- 0077X/82F0077XIE2001001.pdf (Accessed August 15, 2016).
(6.) Statistics Canada. Table 105-0503. Health Indicator Profile, Age- Standardized Rate, Annual Estimates, by Sex, Canada, Provinces and Territories. Statistics Canada, 2015. Available at: http://www5.statcan.gc.ca/cansim/a26?lang= eng&id=1050503 (Accessed January 31, 2017).
(7.) Hill A, Roberts J, Ewings P, Gunnell D. Non-response bias in a lifestyle survey. J Public Health Med 1997; 19(2):203-7. PMID: 9243437.
(8.) Lahaut V, Jansen H, van de Mheen D, Garretsen H. Non-response bias in a sample survey on alcohol consumption. Alcohol Alcohol 2002; 37(3):256-60. PMID: 12003914.
(9.) Kirst M, Mecredy G, Chaiton M. The prevalence of tobacco use co-morbidities in Canada. Can J Public Health 2013; 104(3):210-15. PMID: 23823884.
(10.) Zhao J, Stockwell T, Macdonald S. Non-response bias in alcohol and drug population surveys. Drug Alcohol Rev 2009; 28(6):648-57. PMID: 19930019. doi: 10.1111/j.1465-3362.2009.00077.x.
(11.) Livingstone M, Dietze P, Ferris J, Pennay D, Hayes L, Lenton S. Surveying alcohol and other drug use through telephone sampling: A comparison of landline and mobile phone samples. BMC Med Res Methodol 2013; 13:41. PMID: 23497161. doi: 10.1186/1471-2288-13-41.
(12.) Statistics Canada. Residential Telephone Service Survey (RTSS), 2013. 2014. Available at: http://www.statcan.gc.ca/daily-quotidien/140623/dq140623aeng.htm (Accessed August 15, 2016).
(13.) Smith P, Frank J, Mustard C. Trends in educational inequalities in smoking and physical activity in Canada: 1974-2005. J Epidemiol Community Health 2009; 63(4):317-23. PMID: 19147632. doi: 10.1136/jech.2008.078204.
(14.) Stephens T, Brown KS, Ip D, Ferrence R. The Future of Household Telephone Surveys on Tobacco: Methodological and Contextual Issues. Special Report Series. Toronto, ON: The Ontario Tobacco Research Unit, 2009. Available at: http://otru.org/wp-content/uploads/2012/06/special_future_phone_surveys.pdf (Accessed August 9, 2016).
(15.) Land TG, Landau AS, Manning SE, Purtill JK, Pickett K, Wakschlag L, et al. Who underreports smoking on birth records: A Monte Carlo predictive model with validation. PLoS ONE 2012; 7(4):e34853. PMID: 22545091. doi: 10.1371/ journal.pone.0034853.
(16.) Wong SL, Shield M, Leatherdale S, Malaison R, Hammond D. Assessment of validity of self-reported smoking status. Health Rep 2012; 23(1):47-53. PMID: 22590805.
Received: September 27, 2016
Accepted: February 19, 2017
Thierry Gagne, MSc [1,2]
[1.] Institut de recherche en sante publique de l'Universite de Montreal (IRSPUM), Montreal, QC
[2.] Ecole de sante publique de l'Universite de Montreal (ESPUM), Montreal, QC Correspondence: Thierry Gagne, IRSPUM, 7101 av. du Parc, office 3139, Montreal, QC H3N 1X9, Tel: 514-343-6111, ext. 4565, E-mail: email@example.com Acknowledgements: The author thanks the Interdisciplinary Study of Inequalities in Smoking (ISIS) research group, Jennifer O'Loughlin, Benoit Lasnier and Gerry Veenstra for their help, and acknowledges funding support from a PhD scholarship from the Fonds de Recherche du Quebec - Sante (FrQS).
Conflict of Interest: None to declare.
Caption: Figure 1. CCHS and CTUMS/CTADS estimates of current smoking prevalence. Prevalence estimates include both daily and occasional smoking.
Table 1. Survey characteristics of CCHS and CTUMS/CTADS Survey elements How can it introduce How does this manifest error? in the CCHS and CTUMS/ CTADS? Coverage Coverage issues, either CCHS uses this Target population due to over-or under- operational definition Sampling frame coverage, may result in for non-eligible estimates inferring to individuals: "Excluded a different population from these surveys' than the one targeted. coverage are: persons Over-coverage is living on reserves and defined as the other Aboriginal systematic sampling of settlements in the individuals outside the provinces; full-time target population. members of the Canadian Under-coverage is Forces; the defined as the institutionalized systematic inability to population and persons sample certain living in the Quebec individuals within the health regions of target population. Region du Nunavik and Region des Terres- Cries-de-la-Baie- James. Altogether, these exclusions represent less than 3% of the Canadian population aged 12 and over." CCHS changed its sampling frame in 2013 to cover a higher proportion of Canadians in Nunavut (from 71% to 92%). Notably, CTUMS/CTADS targets Canadians ages 15 and up, CTUMS/CTADS also does not sample residents of the Yukon, Northwest Territories and Nunavut. CCHS was developed to produce reliable estimates at the subprovincial (health region) level; CTUMS/CTADS was developed to produce reliable estimates at the provincial level. This should not introduce error unless sampling at the provincial level provides significantly different sampling probabilities to certain subregions and these probabilities are associated with smoking prevalence. Sampling Sample sizes have a CCHS surveyed Sample size higher chance of approximately 130 000 providing unreliable individuals (i.e., 120 prevalence estimates 000 adults and 10 000 due to random sampling children between 12 and error as they become 17 years of age) in the smaller, especially if 2001, 2003 and 2005 the outcome is rare. cycles, and 65 000 Statistics Canada often individuals in annual reports coefficients of cycles since 2007; variation to help CTUMS/CTADS survey interpret the approximately 20 000 variability of individuals in annual estimates. cycles. Both sample sizes were chosen by Statistics Canada based on their ability to produce reliable estimates at the mandated level (provincial or subprovincial). Non-response Non-response refers to CCHS and CTUMS/CTADS Response rate eligible respondents use a two-step approach who cannot be reached to measure response or who, upon contact, rates because they refused to participate. initially contact The influence of non- potential participants response on estimation by reaching households depends on the randomly, and then distribution of non- participants randomly response. Non-response within that household. that is completely at This produces two random (MCAR) should response rates (RR): not bias prevalence household and person. estimates; non- Multiplying these, response that is CCHS' overall RRs associated with decreased from 84.0% in correlates of smoking 2001 to 66.8% in 2013; but not smoking (MAR) CTUMS/CTADS' overall introduces bias that RRs decreased from can be mitigated if 77.5% in 2001 to 63.1% researchers have in 2013. Results in information on these Zhao et al. (2009) correlates' true suggest that both distribution; non- surveys are liable to response directly non- response bias as associated with the participants in the outcome (e.g., smoking) 2002 CCHS and 2004 CAS is non-random (MNAR) were more often and introduces bias married, educated and that cannot be financially better off corrected. than the general population. (10) Complete information on RRs is available online (see Supplementary Table 2, in the ARTICLE TOOLS section on the journal website). Measurement Measurement error Both surveys use the Questionnaire refers to systematic same standardized Respondent bias due to instrument question and ordering Mode of design (e.g., survey within the full administration questionnaire), the questionnaire. A recent Mode of contact administrator (e.g., study examined phone interviewer, data participants of another entry software) and Statistics Canada other elements of the study, the Canadian contact experience Health Measures Survey (e.g., respondent's (CHMS), and found that capacity to complete self-reported smoking the survey, mode of status was a very administration, mode of strong surrogate of contact, etc.). smoking status when compared with urinary cotinine, a biomarker associated with smoking (sensitivity = 92%, specificity = 98%). (16) This study found that sensitivity could be lower among certain participants, especially young men (76%), but differences were not statistically significant. Gilmore (2002) argued that mode of administration (including face- to-face interviews) and contact (providing advance contacts) in the CCHS could have lowered social desirability bias in contrast to the CTUMS/CTADS. (5)
|Printer friendly Cite/link Email Feedback|
|Publication:||Canadian Journal of Public Health|
|Date:||May 1, 2017|
|Previous Article:||Recommendations: will the tobacco and vaping products act go far enough?|
|Next Article:||Shining new light on newborn screening of cystic fibrosis in the province of Quebec.|