Smoking Is Associated with an Increased Risk of Dementia: A Meta-Analysis of Prospective Cohort Studies with Investigation of Potential Effect Modifiers

Background Previous studies showed inconsistent results on the association of smoking with all-cause dementia and vascular dementia (VaD), and are limited by inclusion of a small number of studies and unexplained heterogeneity. Our review aimed to assess the risk of all-cause dementia, Alzheimer’s disease (AD) and VaD associated with smoking, and to identify potential effect modifiers. Methods and Findings The PubMed, Embase, Cochrane Library and Psychinfo databases were searched to identify studies that provided risk estimates on smoking and incidence of dementia. A random-effects model was used to yield pooled results. Thirty-seven studies were included. Compared with never smokers, current smokers showed an increased risk of all-cause dementia (risk ratio (RR) 1.30, 95% confidence interval (CI) 1.18–1.45), AD (RR 1.40, 95% CI 1.13–1.73) and VaD (RR 1.38, 95% CI 1.15–1.66). For all-cause dementia, the risk increased by 34% for every 20 cigarettes per day (RR 1.34, 95% CI 1.25–1.43). Former smokers did not show an increased risk of all-cause dementia (RR 1.01, 95% CI 0.96–1.06), AD (RR 1.04, 95% CI 0.96–1.13) and VaD (RR 0.97, 95% CI 0.83–1.13). Subgroup analyses indicated that (1) the significantly increased risk of AD from current smoking was seen only in apolipoprotein E ε4 noncarriers; (2) current smokers aged 65 to 75 years at baseline showed increased risk of all-cause dementia and AD compared to those aged over 75 or under 65 years; and (3) sex, race, study location and diagnostic criteria difference in risk of dementia was not found. Conclusions Smokers show an increased risk of dementia, and smoking cessation decreases the risk to that of never smokers. The increased risk of AD from smoking is more pronounced in apolipoprotein E ε4 noncarriers. Survival bias and competing risk reduce the risk of dementia from smoking at extreme age.


Introduction
Dementia is a clinical state characterized by progressive deterioration in cognitive, functional and behavioral abilities [1]. It was estimated that the number of cases of dementia would reach around 81.1 million by 2040 [2]. Alzheimer's disease (AD) and vascular dementia (VaD) could account for approximately 70% and 20% of dementia cases, respectively [1].
The global prevalence of smoking in the population aged more than 15 years is 31.1% and 6.2% for men and women in 2012, respectively [3]. The smoking-dementia relationship has been investigated by many studies [4][5][6][7][8][9][10]. A pooled analysis [11] found that mortality associated dementia was higher in smokers than that in never smokers. Nevertheless, researchers seem to be more interested in the association of smoking with risk of developing dementia. Up to date, two metaanalyses [12,13] on smoking and risk of all-cause dementia, AD and VaD and two meta-analyses [14,15] focusing on smoking and risk of AD have been published. These previous reviews reported that cigarette smoking might be a risk factor for developing dementia. However, several limitations as described below indicate that a more comprehensive meta-analysis is needed.
First, inclusion of a small number of studies was a limitation in previous reviews [12][13][14][15]. A 2014 meta-analysis [15] on smoking and risk of AD and a 2008 meta-analysis [12] on smoking and risk of all-cause dementia, AD and VaD involved 9 and 13 studies, respectively. Yet, more than 20 additional studies were published in the past several years. Second, previous meta-analyses [12][13][14][15] detected significant heterogeneity across studies, but they all did not identify the potential sources of heterogeneity. The unexplained heterogeneity raised questions regarding the reliability of pooled results. Third, a 2007 meta-analysis of 11 studies [13] and a 2008 meta-analysis of 13 studies [12] presented inconsistent results on the association of smoking with all-cause dementia and VaD. Thus, the association of smoking with all-cause dementia and VaD needs to be further investigated. Fourth, the potential dose-response pattern on smoking-dementia association and the modification effect of apolipoprotein E (APOE) ε4 allele on this association remain unclear. Finally, the results of a recent meta-analysis [14] by Cataldo et al on smoking and risk of AD may be biased by not classifying smokers to current and former smokers and the possibility of misclassification caused by inclusion of two large cohort studies [16,17] using death certificates for AD diagnosis.
With those considerations above, our objectives for this study were: (1) to calculate the risk of all-cause dementia, AD and VaD for current versus never smokers, former versus never smokers and ever versus never smokers; (2) to identify potential effect modifiers of association between smoking and all-cause dementia, AD and VaD; (3) to explore the dose-response pattern of the association of smoking and dementia.
to March 25, 2014, with limitation to human subjects imposed. We used the following search terms: "Alzheimer's disease", "dementia", "vascular dementia", "cognitive impairment", "cognitive decline", "cognition", "smoking", "cigarette", "cigarettes", "nicotine" "tobacco" and "smoke". We also reviewed the reference lists of identified studies and pertinent reviews for additional citations. We did not contact original authors through e-mails for extra data parameters.

Study selection
Studies were included if they (1) had a prospective cohort study design and were published in English or Chinese; (2) examined smoking status and identified that participants were free of dementia at baseline; and (3) reported minimum information necessary to obtain risk ratio (RR) on smoking and incident dementia (all-cause dementia, AD or VaD).
The process of study screening was independently conducted by two reviewers (GCZ and YW). An initial screening by scanning titles and abstracts was conducted to exclude irrelevant studies. We conducted a second screening by reading the full text to exclude unrelated articles. Any disagreement about eligibility of studies was resolved by consensus.

Data extraction
Two reviewers (GCZ and YW) independently extracted information. Discrepancies between two reviewers were settled by discussion. Collected information was presented as follows: last name of the first author, publication date, study location, mean age of participants at baseline, sex, race, number of cases and participants, number of current smokers, source of cohort, female proportion in study population, maximum length of follow-up, loss to follow-up rates, smoking category, outcome of interest, diagnostic criteria, the most fully adjusted risk estimates with corresponding 95% confidence interval (CI), and adjustment factors.

Statistical analysis and quality assessment
A random-effects model was used to yield summary effect size. We used I 2 statistic to quantitatively describe heterogeneity across studies [19]. High heterogeneity existed when I 2 is more than 75%, moderate heterogeneity when I 2 ranged from 50% to 75% and low heterogeneity when I 2 was less than 50%. To reflect the stability of our results and to identify potential sources of between-study variability, we conducted sensitivity analyses through three methods, namely ignoring a single study in turn, repeating our analyses through a fixed-effects model, and using various exclusion criteria. Where possible, subgroup analyses were also performed to explore underlying sources of heterogeneity. A p-value for heterogeneity between subgroups was calculated through meta-regression.
Given that all types of dementia are relatively uncommon events, the hazard ratio (HR) and odds ratio (OR) were roughly equal to RR [20,21]. The HR and OR were therefore directly regarded as RR when RR was unavailable. We derived unadjusted RR from the corresponding exposure distribution when HR, OR and RR were unavailable. We used data from the report with the longest follow-up duration when multiple reports originated from the same population. If stratified results by smoking status (i.e., current and former) were separately reported, we summarized these stratum data through a random-effects model to produce an average estimate for ever smoking. Similarly, for two studies [4,22] presenting measures by exposure level, we combined these risk estimates through a random-effects model to obtain the overall value for our meta-analysis. One study [5], whose authors only provided measures for men and women separately, was treated as two separate reports.
We conducted a dose-response analysis based on the method previously described by Orsini et al [23]. Given that the original researchers did not report person-year by exposure level, we approximately derived such data from mean duration of follow up and number of participants at each exposure level. We designated the midpoint of lower and upper boundaries as the assigned dose because all available data parameters of tobacco consumption were reported as range. Furthermore, if the highest range was open-ended, we considered that it shared the same width as the adjacent range. We fitted log-linear dose-response model to regress the log RR on the exposure level.
To reflect the impact of current smoking on the incidence of dementia at a population level, population attributable fraction (PAF) was calculated by the following formula: PAF = {P×(RR-1)/[P×(RR-1)+1]}. Here, P and RR denote the prevalence of current smoking and the summary RR, respectively. On the basis of included studies that reported prevalence of current smoking at baseline, we chose the median prevalence because the prevalence distribution was skewed.
We judged the methodological quality of included studies through the Newcastle-Ottawa quality assessment scale [24]. This tool could award a maximum of nine stars for each cohort study: four stars for the selection of study cohorts, two stars for the comparability of study groups and three stars for the ascertainment of outcome. If a study obtained six or more stars, it was considered to be of high quality.
Publication bias was tested with Begg's test [25] and Egger's test [26]. Data synthesis and analysis were performed via STATA software (version12.0, StataCorp, College Station, TX). Statistical significance level was set at p<0.05 under two-sided test unless otherwise specified.

Literature search
We identified 4,417 relevant citations after removing duplicates. A total of 4,351 citations were further excluded after reviewing their title and abstract. The remaining 66 citations were assessed in more detail for eligibility by reading the full text. Of these, 31 were excluded. We added 2 studies [27,28] through the process of scanning the reference lists of pertinent reviews. Finally, 37 studies were used for the final data synthesis (Fig. 1).

Study characteristics
The main characteristics of included studies are shown in Table 1. Our meta-analysis involved 960,280 individuals, and documented 14,935 all-cause dementia cases, 5,816 AD cases, and 1,406 VaD cases during follow-up varying from 2 years [9] to 40 years [29]. The sample size of included studies ranged from 163 [30] to 848,505 [5]. The average score for methodological quality of included studies was 6.22, with 24 studies showing high quality (S1 Table). 29 studies reported information on loss to follow-up. Of these, 14 had low loss to follow-up rate (≦20%). All cohorts were derived from general populations except the two cohorts from hospitals [22,31]. The mean age of participants at baseline ranged from 42.5 years [31] to 84.0 years [32]. Note that three studies [31,33,34] and one study [29] recruited volunteers and twins as their participants, respectively. 34 studies reported information about sex ratios. Of these, the proportion of women ranged from 0% [10,22,[33][34][35][36][37] to 81.34% [32]. The overall proportion of women in these 34 studies was 42.12%.
Adjusted RRs were available for 22 studies and most adjusted for age (21 studies) and sex (15 studies). The diagnostic criteria for dementia varied across studies, but Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) [38] and Diagnostic and Statistical Manual of Mental Disorders, Third Edition Revised (DSM-III-R) [39] were the most   (Fig. 4). However, the increased risk of AD was of  2 Value is expressed as maximum. 3 Dementia was determined by a battery of neuropsychological measures and a standardized neurological examination. 4 The term "other" in the "Adjustment factors" column refers to all the confounders except age, sex, education, APOE ε4, BMI, diabetes, alcohol and hypertension. 5 Value refers to sample size at baseline. 6 The risk estimates were available just for former smoking and the risk of all-cause dementia and AD.

Sensitivity analyses
Sensitivity analyses by exclusion of any single study in turn did not materially change the pooled risk estimates of all-cause dementia, AD and VaD for current and former smokers (Table 2). Furthermore, when we examined the effect of various exclusion criteria (see Table 2 for details of various exclusion criteria) on the summary results and repeated our analysis through the fixed-effects model, the aforementioned initial associations still remained. Similar results were observed for the association between ever smoking and all-cause dementia. However, the associations of ever smoking with AD and VaD were not stable. Specifically, the pooled RR of AD for ever smoking ranged from 1.10 (95% CI 0.97-1.26) to 1.14 (95% CI 1.02-1.29) in the analysis of omitting a single study in turn. Moreover, the analysis through the fixed-effects model produced a statistically significant RR of 1.17 (95% CI 1.09-1.24). For ever smoking and VaD, the result by exclusion of the study by Juan et al [9] was marginal (RR 1.13, 95% CI 0.99-1.29).

Subgroup analyses
The results of subgroup analyses on association of smoking with all-cause dementia and AD are summarized in Tables 3 and 4. Current smokers aged 65 to 75 years at baseline showed increased risk of all-cause dementia and AD compared to those aged over 75 or under 65 years (Fig. 5), despite the fact that the difference between these subgroups was not significant (all p for heterogeneity>0.05). Current smokers without APOE ε4 allele showed a significantly increased risk of AD (RR 2.01, 95% CI 1.34-3.03), whereas the association between current smoking and AD in APOE ε4 carriers was not significant (RR 1.51; 95% CI 0.69-3.28). The association between former smoking and all-cause dementia and AD persisted in all subgroups, with no evidence of heterogeneity between subgroups from meta-regression, apart from the subgroup for all-cause dementia stratified by the mean age of participants (p = 0.03). In the analysis stratified by adjusted risk estimates, we found evidence of significant heterogeneity between subgroups for the association between current smoking and all-cause dementia (p = 0.03) and AD (p = 0.01). Sex, study location, race and several key study characteristics, including sample size, mean duration of follow-up and loss to follow-up rate, were not the source of heterogeneity for pooled RR of all-cause dementia and AD for current, former and ever smoking. Excluding studies relying on other diagnostic criteria than NINCDS-ADRDA did not change the association of smoking and AD. The moderate heterogeneity of the pooled RR of AD decreased to 16.6% (I 2 value) when we excluded studies that recruited volunteers. We did not identify any source of heterogeneity for the pooled RR of all-cause dementia and AD through sensitivity analyses.
doi:10.1371/journal.pone.0118333.t002 Additionally, we found no significant difference between the subgroup using DSM-IV criteria and the subgroup using DSM-III-R criteria. The small number of studies of ever smoking and VaD precluded our interpretation to the results of subgroup analyses (S2 Table). We did not perform subgroup analysis with respect to current smoking, former smoking and VaD since limited studies [4,5,7,9,10] were available.

Dose-response analysis
Only two studies [4,22] were included in the dose-response analysis of current smoking and all-cause dementia. Under the fixed-effect dose-response meta-regression model, the risk of dementia significantly increased by 34% for every 20 cigarettes per day (RR 1.34, 95% CI 1.25-1.43; p for nonlinearity = 0.36, goodness-of-fit chi 2 = 5.91, p = 0.43; Fig. 6). For current smoking and incident AD and VaD, only one study was available, which showed a dose-response relationship [4].

Publication bias
No evidence of publication bias was found for any association by Begg's test and Egger's test (all p>0.05).

Discussion
The present meta-analysis identified that current smokers consistently showed increased risks of all-cause dementia, AD, and VaD compared with never smokers, while former smokers did not show increased risks. We also found that ever smokers showed increased risks of all-cause dementia and VaD, but the associations were weaker than that for current smokers. The increased risk of AD in ever smokers was marginal.
A diagnosis of dementia can never be valid and definitive without subsequent histopathological confirmation and this is challenging for researchers who rely on clinical diagnoses [43]. Accordingly, participants generally received a clinical diagnosis of dementia on the basis of various diagnostic classification systems. Several diagnostic criteria of dementia including DSM-  IV criteria, DSM-III-R criteria, and NINCDS-ADRDA criteria are widely used. However, some studies observed that the prevalence of dementia varied remarkably when researchers relied on different diagnostic criteria [44,45]. A large number of studies included in the present study allowed us to examine the effects of different diagnostic classification systems on the risk of dementia through subgroup and sensitivity analysis. We found that the diagnostic criteria of dementia did not modify the smoking-dementia association. Previous studies indicated that the incidence of dementia rises with increasing age [46][47][48][49]. However, in subgroup analyses stratified by age at baseline, we observed an unanticipated decline in the risk of all-cause dementia and AD among current and ever smokers aged from 65-75 to over 75 years, most notably for current smokers. A possible interpretation for above phenomenon is that the risk of all-cause dementia and AD should rise with increasing age, but survival bias [50] and competing risk [51] apparently reduce the risk of dementia from smoking at extreme age. In fact, several previous studies [52,53] have observed similar phenomena. Hernan and colleagues [53] observed that the weighted average RR of incident AD gradually decreased with increasing age. Similarly, Taylor and colleagues [52] found a positive association among individuals aged less than 50 years but an inverse association among those aged more than 50 years when investigating the relationship between smoking increasing allele and smoking initiation. Generally, survival bias will be introduced when researchers recruit older smokers into their study cohorts at baseline, and competing risk will occur if a smoker dies of smoking-associated events (e.g., lung cancer) before receiving a diagnosis of dementia during follow-up [51]. Given these facts, future epidemiologic studies on "bad" exposures and age-related outcomes should take into account the influence of survival bias and competing risk when making interpretations to their results.
The modification effect of APOE ε4 allele on the association between smoking and dementia remains controversial. As suggested by our subgroup analyses, the association of current smoking and increased risk of AD remained significant only in APOE ε4 noncarriers. Our findings are consistent with results of previous prospective studies [7,8,54]. The underlying mechanisms for the aforementioned results are unclear. APOE ε4 carriers are at an elevated risk of AD [55][56][57]. A potential explanation for our results is that the increased risk of AD due to APOE ε4 allele makes the increased risk due to other risk factors insubstantial. We could not assess the combined effects of smoking and APOE ε4 allele on all-cause dementia and VaD because of few individual studies exploring it. Further research is warranted to clarify the modification effect of APOE ε4 allele on association of smoking with dementia.
In the present study, smokers were grouped into current, former and ever smokers. Our meta-analyses observed increased risk of dementia in ever but not former smokers. Considering ever smokers consist of current and former smokers, this phenomenon may be attributable to the influence of current smokers. Indeed, as shown by our forest plots, risks of dementia in ever smokers were lower than those in current smokers but higher than those in former smokers. The potential biological mechanisms regarding smoking and increased risk of dementia have been proposed [58,59]. In brief, smoking could disturb the balance between generation and reduction of oxidants and free radical species. The consequent overload of oxidants and free radical species triggers oxidative stress (OS). Increased OS contributes to formation of senile plaque and neurofibrillary tangles, signifying the occurrence of dementia. In addition, increased risk of dementia resulting from smoking may be realized through cardiovascular disease (CVD) because smoking-related OS has been hypothesized as an underlying mechanism for CVD [60], and CVD contribute to an increased risk of dementia [61,62]. The heterogeneity of the studies regarding adjustment of the risk estimates for underlying confounders is a possible limitation of our review. Smoking is associated with other dietary, lifestyle and social factors, such as higher intake of alcohol, lower body weight, higher blood pressure and lower education level. The subgroup analyses stratified by adjustment for body mass index, hypertension, alcohol and education showed consistent results on current smoking and all-cause dementia, as well as former smoking and all-cause dementia and AD, but the results for current smoking and AD only remained significant in subgroups with adjustment for alcohol and education. In addition, the limited number of studies adjusting for body mass index and hypertension could be subject to over-adjustment because both could be mediators in the causal pathway between smoking and dementia. Considering these facts, we could not completely rule out the possibility that the inadequate control for various underlying confounders may bias our findings. A promising and useful approach to reduce confounding bias is Mendelian randomization [63,64], and it can assess the potential causal associations between exposures and end points. Fortunately, genome-wide association studies have identified that the CHRNA3 rs1051730 genotype is a common genetic marker strongly associated with smoking behavior [65,66]. Thus, it raises a possibility of using Mendelian randomization to explore the nature of smoking-dementia association.
Our study also has several strengths. Inclusion of only prospective cohort studies can avoid recall bias and chicken-egg question occurring in case-control studies. We included a large number of studies, so we had enough statistical power to identify smoking-dementia association. Subgroup analyses promote a better understanding of smoking-dementia association. The stability of pooled results as evidenced by sensitivity analyses indicates the robustness of our findings. No statistical evidence of publication bias and inclusion of studies that treated smoking as a covariate further support the robustness of our findings. In addition, we explored the dose-response pattern between smoking and risk of all-cause dementia for the first time, despite the fact that only two studies were available.

Conclusions
Our meta-analysis indicates that smokers show an increased risk of dementia, and quitting smoking decreases the risk to that of never smokers. The increased risk of AD from smoking is more pronounced in APOE ε4 noncarriers. Both survival bias and competing risk reduce the risk of dementia from smoking at extreme age. The smoking-dementia relationship could not be modified by sex, race, study location and the diagnostic criteria of dementia. Future study is warranted to clarify the modification effect of APOE ε4 allele on association of smoking with dementia.