|Home | About | Journals | Submit | Contact Us | Français|
To examine the relationship between smoking and AD after controlling for study design, quality, secular trend, and tobacco industry affiliation of the authors, electronic databases were searched, 43 individual studies met the inclusion criteria. For evidence of tobacco industry affiliation http://legacy.library.ucsf.edu was searched. One fourth (11/43) of individual studies had tobacco affiliated authors. Using random effects meta-analysis, 18 case control studies without tobacco industry affiliation yielded a non-significant pooled odds ratio of 0.91 (95% CI, 0.75–1.10), while 8 case control studies with tobacco industry affiliation yielded a significant pooled odds ratio of 0.86 (95% CI, 0.75–0.98) suggesting that smoking protects against AD. In contrast, 14 cohort studies without tobacco industry affiliation yielded a significantly increased relative risk of AD of 1.45 (95%CI, 1.16–1.80) associated with smoking and the three cohort studies with tobacco industry affiliation yielded a non-significant pooled relative risk of 0.60 (95% CI 0.27–1.32). A multiple regression analysis showed that case-control studies tended to yield lower average risk estimates than cohort studies (by −0.27±0.15, P=.075), lower risk estimates for studies done by authors affiliated with the tobacco industry (by −0.37±0.13, P=.008), no effect of the quality of the journal in which the study was published (measured by impact factor, P=0.828), and increasing secular trend in risk estimates (0.031/year ±.013, P=.02). The average risk of AD for cohort studies without tobacco industry affiliation of average quality published in 2007 was estimated to be 1.72±0.19 (P<.0005). The available data indicate that smoking is a significant risk factor for Alzheimer’s disease.
In 2006, there were 26.6 million people with Alzheimer’s disease (AD). AD prevalence will quadruple by 2050, increasing the burden of care and the associated health costs. Cardiovascular disease is a major risk factor for AD and smoking cigarettes causes cardiovascular disease.[3, 4] Despite a growing body of evidence linking smoking with AD,[5–8] beliefs prevail that smoking protects against AD in both scholarly journals and lay publications.[9–13] For example, a 2008 study of nicotine dependence in Human Genetics begins with the statement, “epidemiological studies reveal that cigarette smoking is inversely associated with AD” the “Alzheimer’s Disease Facts and Figures,” published by the Alzheimer’s Association, does not mention smoking and the May 8, 2008 Oprah Magazine, a trusted source of health information for millions of women, reported that nicotine might protect people from AD. As of October 2008, over 100 web sites stated that smoking was protective for AD. Confusion regarding the role of smoking in AD may discourage cessation attempts among older smokers and contribute to the reluctance of health care providers to treat tobacco dependence in older smokers. A delay in AD onset or progression by just one year would result in nearly 9.2 million fewer cases in 2050. There is a pressing need to understand the risk factors for AD, particularly clarification of smoking’s effect.
Previous reviews[5–13, 18] of the association between smoking and AD have not controlled for both study design and author affiliation with the tobacco industry. In 2002, Almeida et al. identified study design (cohort or case-control) as an important methodological issue because of survival and recall bias that may be responsible for the unclear direction of the association between smoking and AD. But, at the time, only four cohort studies were available, and the direction of the association remained unclear. In 2007, Anstey et al. published a meta-analysis of 19 cohort studies and found that, compared with people who had never smoked, current smokers had an increased risk of AD. Hernan et al. identified 12 prospective studies and found a wide range of risks for AD associated with smoking, which they attributed to bias due to censoring by death. Purnell’s 2008 systematic review of cardiovascular risk factors and incident AD reported on only four cohort studies that used incident cases of AD, three of which reported that current smoking increased the risk of AD and one found no significant relationship.
As early as 1976, the tobacco industry began to invest in AD research,[19, 20] with the goal of developing nicotine-related diagnostics and therapeutics. These research efforts were conducted through the industry’s Council for Tobacco Research (CTR) and the individual tobacco companies (RJ Reynolds [RJR] Biomedical Research Program[19, 20] and Philip Morris Research). CTR’s interest in AD started in 1981 following published anecdotal observations of low rates of tobacco use among patients with AD.[23, 24] In 1987, RJR’s “Smoking and Health's contract program” identified several key research areas, including the “possible relationships between smoking and Alzheimer's Disease.” In 1993, Philip Morris’ interests in AD research were to “not only understand the role of smoking in this disorder but also provide early diagnostics … as well as potential therapeutics (e.g., nicotine analogs).”
Past research has shown that the tobacco industry funding of research on the health risks of smoking is associated with findings that support their position. For example, research showing that factors like genetics, personality type or aging were responsible for diseases such as lung cancer and cardiovascular disease.[26–28] Industry sponsored research is more likely to reach conclusions that are favorable to the sponsor and experience publication bias, publication delays and data withholding. This has been true for sponsorship not only from the tobacco industry[30, 31] but also from the pharmaceutical[32, 33] and chemical industries. The present systematic review estimates the association between cigarette smoking and AD after controlling for study design, quality, secular trend and tobacco industry affiliation of the authors.
We searched PubMed, PsycINFO, Google Scholar, and Cochrane CENTRAL for reviews and individual studies published as full articles or brief reports using combinations of keywords for smoking (smoking, tobacco, cigarettes and nicotine) and AD (Alzheimer’s disease, dementia, cognition, cognitive impairment, vascular dementia). We also examined reference lists of prior reviews and individual studies and contacted experts in the field. For inclusion, studies needed to be published, use AD as the outcome (not dementia or cognitive decline), human subjects (brain tissue excluded), a measurement of smoking (i.e., ever smoker, current smoker, never smoker), and have a clearly stated study design (case-control or cohort). One prevalence study was included and coded as a case-control study. If both raw and adjusted risk ratios were available, ratios adjusted for the most covariates were used.
The initial search returned 336 articles; 47 met the requirement for detailed review, of which 43 original research studies met our inclusion criteria (Table 1 and Table 2 and Figure 1). One publication was excluded because it was a letter to the editor and the criteria used for diagnosis of AD were unclear. One other publication was excluded because results from the same study were published in two papers, and we included the more completely reported results. For multiple publications from the same sample with the same smoking measures, we used the publication with the longest follow-up; in particular, for the Rotterdam Study, the latest 2007 report was chosen over the 1998 report and for the Canadian Study of Health and Aging the 2002 report was chosen over the three previous reports.[42–44] When the studies were stratified by race,[45, 46] results for the separate populations were included as separate studies. A paper by Wang (1999) provided both case-control and cohort results, which also were treated as separate studies.
For each study, we recorded the smoking exposure measure, method of diagnosing AD and covariates. For each case-control study, we recorded the number of cases and controls, the number of cases and controls who smoked, the source of cases and controls, and multivariate odds ratio (OR) and 95% confidence interval and P value (Table 1). For each cohort study, we recorded the size and source of the cohort, years of follow-up, covariates, multivariate relative risk (RR), and confidence interval (Table 2).
Ten of the 43 studies were reviewed by two individuals, one reviewer blinded to author, journal, and publication date; there was 100% agreement.
We also identified 10 systematic reviews (Table 3).
We searched the Legacy Tobacco Documents Library for evidence of tobacco industry affiliation of the studies’ authors (http://legacy.library.ucsf.edu), a collection of over 51 million pages of previously secret internal documents from the major tobacco companies and organizations that were made available as a result of litigation against the tobacco industry.[48, 49] Tobacco industry affiliation was defined as current or past funding, employment, paid consultation, collaboration, or co-authorship on the included study with someone with then-current or previous tobacco industry funding (within ten years of publication).
The search for possible tobacco industry affiliations was conducted from July 2006 to July 2008 using keywords “Alzheimer’s disease,” “dementia,” “cognition,” and the names of authors and university affiliations of the AD studies in this meta-analysis. Expanded searches were conducted using information in initially identified documents, including named individuals, specific programs, and dates and reference (Bates) numbers. Initial searches produced 10,798 documents. After screening documents based on index entries, we identified 3,324 documents to review. After eliminating duplicates and irrelevant documents, we analyzed 877 documents.
In our meta-analysis we analyzed cohort studies (which yield direct estimates of relative risk) and case control studies (which yield odds ratios) separately. We tested for heterogeneity with the Q test statistic. Random effects meta-analysis was used to estimate pooled risk ratios and 95% confidence limits using the DerSimonian and Laird method because of observed heterogeneity among results among the cohort studies (Stata 10.1 metan). We used Begg’s funnel plot to test for publication bias (Stata metabias).
We tested for the effect of study design, quality, secular trend and tobacco industry affiliation in a weighted (using weights from the meta-analysis using Stata regress option pweight) multiple regression analysis in which the dependent variable was the point estimate of the risk of AD and the independent variables were study design (0=case control, 1=cohort), study quality indexed as the impact factor of the journal (in 2009), where the study was published, secular trend (measured as year of publication, setting 2007 as 0) and tobacco industry affiliation (0=no, 1=yes). We selected 2007 as the base year because that was the year of the most recent individual study. Doing so gives the constant in the regression equation the following interpretation: The average risk estimate one would predict for a cohort study published in 2007 in an average quality journal by a group without tobacco industry affiliation. The choice of base year has no effect on the values, standard errors, or statistical significance of the predictor variables in the model.
Because the type of study design and likelihood of tobacco industry affiliation may have changed systematically over time, we computed variance inflation factors to test for collinearly in the independent variables in the regression analysis. We tested whether the model assumptions were met (linearity of the dependent variable and normality of error terms) by examining a normal probability plot of the residuals.
For review papers, we tested for a significant association between author affiliation with the tobacco industry and the conclusion of their review using a two tailed Fisher Exact Test, with reviews dichotomized into “protective” or “not protective or ambigous.”
There was no evidence of publication bias (Figure 2).
Of the 43 individual studies of smoking and AD (Figure 1), 26 were case-control studies (Table 1) and 17 were cohort studies (Table 2). Eleven of the 43 studies (26%) were conducted by tobacco industry affiliated investigators. Out of 11 tobacco affiliated studies only 3 disclosed an affiliation (27%). Of the 8 case control studies with tobacco affiliation, just 3 disclosed the affiliation (37%); of the 3 cohort studies with tobacco affiliation, none disclosed the affiliation.
The 18 case-control studies without tobacco industry affiliation yielded a nonsignificant pooled odds ratio of 0.91 (95% CI 0.75–1.10; Table 1 and Figure 1). In contrast, the 8 case-control studies with tobacco industry affiliation yielded a significant pooled odds ratio of 0.86 (95% CI 0.75–0.98; Table 1 and Figure 1), suggesting that smoking protects against AD. The 14 cohort studies without tobacco industry affiliation indicated a significantly increased relative risk of AD of 1.45 (95% CI 1.16–1.80; Table 2 and Figure 1) associated with smoking. In contrast, the three cohort studies with tobacco industry affiliation yielded a non-significant pooled relative risk of 0.60 (95% CI 0.27–1.32; Table 2 and Figure 1).
Table 3 presents the results of the multiple regression analysis that estimates the simultaneous effects of study design, study quality, secular trend and tobacco industry affiliation. Controlling for the other variables, on the average case-control studies tended to yield lower risk estimates than cohort studies (by −.27, P=.075). Study quality, measured by the impact factor of the journal in which the paper was published, was not associated with the magnitude of the risk estimate (P=.828) . There was a secular trend in risk estimates, with the newer studies showing higher risks (increasing at 0.031/year, P=.020). Controlling for these other factors, affiliation with the tobacco industry was associated with significantly lower risk estimates, by −.37 (P=.008). Controlling for these other factors, this analysis indicates that the average risk of AD based on a non-industry affiliated cohort study in 2007 is 1.72 (P<.0005, the constant in the regression equation). All the variance inflation factors were 1.39 or less, indicating that these different factors were independent of each other.
Ten systematic reviews of the association between smoking and AD were published between 1992 and 2008 (Table 4); 4 had tobacco industry affiliated authors. The association between author affiliation with the tobacco industry and the conclusion of their review was significant (P=0.005). Among the 6 reviews without tobacco industry affiliation, 1 found “no clear effect,”  2 found “no protective effect” of smoking for AD,[5, 18] and 3 found smoking to be a significant risk factor for AD.[6–8] The 4 reviews with known tobacco industry affiliation concluded that smoking “protected against AD.”[10–13]
Controlling for study design, quality, secular trend and tobacco industry affiliation, we found a significant increase in AD risk associated with smoking. Controlling for all these factors, current or ever tobacco smoking was a significant risk factor for AD (RR= 1.72; 95% CI 1.33–2.12 for cohort studies of average quality in 2007). Case-control studies tended to yield lower risks than cohort studies and most of the tobacco affiliated studies used case-control design. Even after controlling for study design, tobacco industry affiliation was independently associated with lower risk estimates. In contrast to the significant increase in risk of AD that our analysis demonstrated, the combination of case-control design and tobacco industry affiliation yielded a protective odds ratio of 0.86 (95% CI 0.75–0.98).
The need to account for these factors in evaluating the evidence is demonstrated by the fact that if one simply combined all 43 studies in a single random effects meta-analysis, one would obtain an inaccurate null result (risk ratio = 1.05; 95% CI 0.91–1.20).
In 1992, Hirayama et al. (no tobacco affiliation), published the first large cohort study and found smoking to be a risk factor for AD (RR = 1.61; 95% CI 1.10–2.38). In reference to Hiriyama’s soon to be published paper, PN Lee a long-time, paid statistical consultant for the tobacco industry,[52–54] informed Philip Morris Tobacco Company in a secret report, that “prospective studies are often more scientifically valid than case-control studies.” However, Philip Morris continued to fund case-control studies and in 1994, RJR Biological Research Division initiated a project to investigate the “Utility and Feasibility of Funding Epidemiology Studies on smoking and Alzheimer’s Disease.” This project included a review of current literature and several interviews with tobacco industry affiliated scientists. The recommendation from this project was that, it was no longer “feasible… to fund either a prospective or retrospective epidemiology study on cigarette smoking and AD.”
In an earlier review of AD and smoking studies, cohort designs were found to be less susceptible to bias than case-control designs. A potential problem with case-control studies is recall bias. AD patients all have memory deficits and reports given by proxy are often used to obtain smoking history. However, if proxy reports are not used for the control subjects as well, then bias can result. Differential mortality is a problem when investigating the effects of smoking in Alzheimer’s disease because of the very low incidence rates of AD before age 75. Persons with AD often die more quickly and are unavailable for case-control studies, which may lead to the false interpretation that smoking among cases is less common than it actually is and that smoking among control subjects is more common than among cases. Cohort studies often involve only a short time period of only 3–5 years, whereas, case control studies may cover activities over a period of decades. Debanne et al. conducted a simulation study that showed that cohort studies also understate the risk of AD in smokers because of early deaths. (They did not compare the magnitude of the bias in cohort vs. case-control studies.) When survival among persons with the disease is related to the exposure of interest, downward bias can result.
Our study has several limitations. The primary data collector (JKC) was not blinded. The data extracted, however, were objective (i.e., risk ratios). In a study of blinding in 5 meta-analyses, blinding made no difference in the pooled risk estimates.
In the document search for tobacco industry affiliation, we may not have retrieved every relevant document. Some materials may have been destroyed or concealed by the tobacco companies. After 1998, the tobacco companies were aware that documents might eventually be made public and became more careful about what they wrote down.
The application of meta-analysis to observational studies can be problematic because of biases in both the original studies and publication bias.[57, 63, 64] There is, however, no evidence of publication bias in the studies we considered (Figure 2).
Random effects meta-analysis is considered the appropriate approach for estimating risks when there are heterogeneous results among studies, but does not explain the reasons for this heterogeneity. By dividing the studies according to design type (case-control or cohort) and whether or not there was tobacco industry affiliation, we found both groups of case-control studies to be homogenous. Both groups of cohort studies remained heterogeneous (albeit less so); we did not explore the reasons for heterogeneity within these subsets of studies. Study design type, industry affiliation and year of publication explained a significant portion of the heterogeneity among the 43 studies.
There was a nonsignificant trend toward lower risk estimates being associated with case-control studies (which yield ORs) than cohort studies (which yield RRs). The OR is an estimate of the RR when the RR is near 1.0; to the extent that the ORs from the case-control studies are not good estimates of the RR, this coefficient in the regression model (Table 3) quantifies the net effect of differences in results due to study design and biases due to using the OR as an estimate of RR.
One should interpret this significant association between conclusions of systematic reviews and tobacco industry affiliation cautiously. In the individual studies there is a secular trend in publishing, type of design, and tobacco industry funding, which may be confounded in this analysis of the reviews. There is a clear time trend in the publishing of case-control and cohort studies, with more of the former appearing in earlier years. Specifically, of the 26 case-control studies in Table 1, 23% were published in the 1980s, 58% in the 1990s, and 19% in the 2000s. In contrast, of the 17 cohort studies in Table 2, 6% were published in the 1980s, 41% in the 1990s, and 53% in the 2000s. It could be possible that the year of publication is associated with the conclusion of a reduction in AD risk for smokers, since all the industry-affiliated reviews were published in 1994 or earlier. This finding may be also be due to the fact that the tobacco industry stopped supporting reviews after 1994, as the cohort studies were beginning to appear.
For the last two decades, the tobacco industry has been actively funding research that supports the position that cigarette smoking protects against AD, and for the past two decades, the scientific literature has reported conflicting results as to the direction of the association between smoking and AD. Consequently, older smokers and their health care providers have been unaware that smoking is a modifiable risk factor for AD. There is an association between tobacco industry affiliation and the conclusions of individual studies and, probably, review papers of AD. Controlling for industry affiliation, study design and other factors, smoking is not protective against AD; it is a significant and substantial risk factor for Alzheimer’s disease.
Funding/Support: This research was supported by the California Tobacco Related Disease Research Program grant no. 16RT-0149; National Cancer Institute grant no. CA-87472 and National Institute Drug Addiction K23-DA018691.
Role of the Sponsor: The funding agencies had no role in the design of the study, the conduct of the research, or the preparation of the manuscript.
Author Contributions: Dr Cataldo had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Study concept and design: Cataldo, Prochaska, Glantz.
Acquisition of data: Cataldo.
Analysis and interpretation of data: Cataldo, Prochaska, Glantz.
Drafting of the manuscript: Cataldo, Prochaska, Glantz.
Critical revision of the manuscript for important intellectual content: Cataldo, Prochaska, Glantz.
Statistical analysis: Glantz
Obtained funding: Cataldo, Prochaska, Glantz.
Administrative, technical, or material support: Cataldo, Prochaska, Glantz.
Study supervision: Cataldo, Prochaska, Glantz.
Financial Disclosures: None reported.
Disclaimer: The content is solely the responsibility of the authors and does not represent the official views of NCI, California TRDRP or NIDA
Additional Contributions: None.