Elsevier

Biological Psychiatry

Volume 49, Issue 3, 1 February 2001, Pages 194-199
Biological Psychiatry

Nicotine mechanisms in Alzheimer’s disease
The association between smoking and Alzheimer’s disease: effects of study design and bias

https://doi.org/10.1016/S0006-3223(00)01077-5Get rights and content

Abstract

In epidemiologic studies, unrecognized bias can contribute to observed results, causing them to be inaccurate. Analytic study designs, such as the case-control and cohort designs, each carry potential for specific forms of bias. The cohort design is not susceptible to many forms of bias that are experienced by case-control studies. A consistent “protective” effect of smoking on Alzheimer’s disease was documented by many case-control studies. However, the potential effect of biases cannot be separated from the results. Cohort studies now show that smoking may either be unrelated to Alzheimer’s disease onset or possibly generate a modest increased risk. In this review the results and comparisons of various studies and potential biases are discussed.

Introduction

Scientific evidence based on randomized, controlled trials (RCTs) is generally regarded as a high standard of proof of association between a factor and an outcome. In RCTs the investigator randomly assigns comparable patients to one of several treatments. Because valid randomization equalizes the distribution of known and unknown factors related to outcome, differences between the groups can be attributed to the treatment that the investigator applies. However, there are many situations where it is either unfeasible or unethical to apply this research design.

Observational studies to assess the effect of a particular exposure on disease onset are alternatives to RCTs. Careful attention to detail in patient selection, information gathering, and exposure assessment are critical to the validity of observational studies. Studies that assume the temporal relationship between exposure and disease add to our ability to determine whether associations may be causal. Two common observational, analytic study designs used in epidemiology Gordis 1996, MacMahon and Trichopoulos 1996, Rothman and Greenland 1998 are the “case-control” design (or case-referent, case-comparison, etc.) and the “cohort” study design (prospective, concurrent, or historical). In the case-control studies, persons with a specific disease (cases) are assembled along with a comparison group of persons without the disease of interest. History of “exposure” before disease onset is then determined. In cohort studies, exposures are determined for a group without the disease of interest and the incidence of disease is then followed over time. The structure is similar to that of an RCT except that the investigator has no control over “exposure group” assignment.

Between the 1980s and mid-1990s a number of case-control studies were conducted to study the effect of smoking on the onset of Alzheimer’s disease (AD). Lee (1994) produced a review and overall analysis of the potential effect of smoking on AD. He carefully evaluated potential study differences that could have had an effect on results, and then calculated a summary measure of effect based on the reported odds ratios, study size, and confidence intervals (CIs). By definition, the odds ratio is the odds of smoking among the cases divided by the odds of smoking among the control subjects. The odds ratio is interpreted as an estimate of the relative risk; therefore an odds ratio of 2.0 would indicate that smoking was twice as common among cases or that smoking was associated with a twofold increased risk of AD; an odds ratio of 0.5 would imply that smokers enjoyed half the risk of AD that nonsmokers had. The odds ratios seen here are relatively consistent; they reflect a decreased risk for AD in smokers (Lee 1994). Many of the 95% CIs include 1.0, an indication that “no association” cannot be excluded as an explanation. Wide CIs can be caused by small sample sizes. However, combining studies to increase overall sample size increases the power to detect an association when it exists; it also leads to narrower CIs. Because of the consistent odds ratios and evaluation of other evidence available, Lee was justified in conducting a meta-analysis of the data. The net effect observed by meta-analysis was a significant 40% reduction in risk of AD among smokers (Lee 1994). Taken as the sum of many studies, this evidence seemed convincing. Table 1shows the principal case-control studies reviewed by Lee and the estimated relative risk for AD associated with smoking.

Summary measures of effect based on a meta-analysis alone may be misleading. The biases inherent in each of the individual studies can influence the summary estimate. Therefore, though it takes advantage of the large numbers of patients to increase power and narrow CIs, combined analysis does not adjust or eliminate inaccurate results (i.e., due to bias).

Consider the case-control design as shown in Figure 1. The investigator selects cases (persons with AD) and control subjects (persons without AD) and then attempts to determine the smoking history before disease onset in cases and, at a similar time, for the persons called control subjects. If the way in which cases or control subjects are identified and enrolled is related to smoking history, then some bias is likely to exist. If exposure information is obtained differently for cases relative to control subjects, then another bias is likely to exist. Suppose that an investigator carefully included cases who were considered as exposed only if they had a smoking history of at least 2 years before any AD symptoms but, for control subjects, captured their current smoking habits rather than at a comparable point in the past. This could systematically include more control subjects relative to smokers and therefore bias the association.

Recall bias is a well-known contaminant of case-control studies. It stems from differential recall or reporting of exposure status by cases and control subjects. For AD this is a well-recognized problem because AD patients all have memory deficits. Reports given by proxy (spouse, child, other) are often used to obtain smoking history. However, if proxy reports are not used for the control patient as well, then bias could result. Objective exposure measurement (e.g., through specific records or biological measurements) is ideal but difficult to attain.

A more subtle form of bias can result if a cross-sectional sample of AD cases is selected for inclusion in the study. For example, at the start of the case-control study the investigator may attempt to enroll all the living and available AD patients seen in an Alzheimer’s clinic during the last 5 years. This is often described as enrolling prevalent rather than “incident” or new cases. When survival among persons with the disease is related to the exposure of interest, then bias can result. Figure 2 shows a hypothetical representation of the life span of AD patients following diagnosis. Each line represents an individual case from diagnosis to death. The short lines represent short time spans and the long lines represent a greater time between diagnosis and death. Notice that the vertical line representing the cross-sectional sample of existing cases intersects with more longer duration cases than shorter duration ones. Persons who survive only a short time with the disease may differ in many ways from those who survive longer. For example, they may have greater disease severity or more comorbidity. Assessing risk factors primarily from longer surviving cases would not allow the investigator to distinguish between factors related to disease onset and factors related to survival.

One way to evaluate whether biases may have been present in the case-control studies already discussed is to compare their results to those obtained though a study design that is less susceptible to those biases. (Keep in mind that small p values/significance levels are only a probabilistic statement of chance and can result in either biased or unbiased results.) Figure 3 shows a simplified diagram of a cohort study design. Persons without AD are first classified in relation to their exposure to smoking and then reassessed for development of AD, at specified intervals thereafter. This allows observation of incident cases and avoids the prevalent case bias. Because exposure is assessed before onset of disease there is also less chance of reference year bias or recall bias. A major threat to the validity of cohort studies is loss of subjects during the follow-up period. The results of several current cohort studies are shown in Table 2.

Wang et al (1999) conducted an important and informative population-based study in 1987 in Stockholm: the Kungsholmen project, in which persons over age 74 were enrolled. A case-control and a cohort analysis were conducted. Screening of the initial patients revealed 198 prevalent dementia cases, which were eliminated from the incidence portion of the study. These became subjects in a case-control analyses. After exclusion of patients with cognitive impairment and dementia, 343 cognitively intact patients entered the cohort study. From the case-control analyses using the 198 prevalent cases, as compared with nondemented patients of similar age, Wang et al (1999) found that smoking appeared to be a protective factor for AD (odds ratio = 0.6; 95% CI, 0.4–1.1). This result is quite similar to the case-control studies cited previously and to Lee’s summary table (Table 1). However, results obtained from the cohort study showed no decreased risk of AD among smokers (hazard ratio = 1.1; 95% CI, 0.5–2.4). Based on the cohort, Wang then studied mortality among the demented and nondemented, and compared smokers with nonsmokers within each group. It was found that, among the demented, smokers were 3.4 times more likely to die during a 5-year follow-up than demented nonsmokers. In contrast, among the nondemented, smokers’ risk of death differed much less from that of nonsmokers of similar age.

Figure 2 showed the potential that longer duration cases tend to be most included in case-control studies. Wang’s results establish that such a bias may be present in at least some of the case-control studies. Wang et al (1999) concluded that persons with AD who are also smokers often die more quickly and therefore would be unavailable for case-control studies. This would give the appearance that smoking among cases was much less common than it actually is and, in turn, that smoking among control subjects was more common than among cases. Thus, a case-control study could conclude that smoking has a “protective” effect.

Ott et al (1998) reported on another cohort study that was conducted in Rotterdam and included all eligible persons over 55 years of age. Six thousand eight hundred seventy persons who provided information concerning smoking were observed for incidence of dementia and AD. The length of follow-up was approximately 2 years at this report. In comparing former smokers and “never” smokers, Ott et al (1998) observed a relative risk of 1.4 (95% CI, 0.9–2.0). Current smokers enjoyed a relative risk of 2.2 (95% CI, 1.3–3.6) relative to never smokers. This study indicates a twofold increased risk for AD among smokers; men showed higher risk than women. Ott et al then examined the simultaneous effects of smoking and apolipoprotein E (APOE) genotype. Among persons with an ϵ4 allele (a risk factor for AD), in either current or former smokers, the effect of smoking was null. Among persons without an ϵ4 allele the risk of AD due to smoking appeared elevated. The biological reasons for effect modification (interaction) of the smoking–AD association by APOE ϵ4 are not obvious. Apolipoprotein E is involved in cholesterol transport, but exactly how APOE acts in the pathogenesis of AD is not known. One might speculate that both APOE4 and smoking act on cholesterol level, a risk factor for vascular disease, but in the presence of APOE4 the additional effect of smoking on cholesterol level is less. Then, one must also hypothesize a vascular component to AD to complete the mechanism. Much more foundation is needed to establish the true nature of these observed associations.

Merchant et al (1999) conducted a smaller cohort study in New York City; it included 1062 persons, who were observed for approximately 2 years. Follow-up examination and diagnosis were conducted in a manner similar to that of Ott et al (1998). Table 2 shows the risk of AD in former smokers relative to “never” smokers was consistent with a conclusion of “no association.” However, when “current” smokers (smoking at baseline) were compared with never smokers, a modest increased risk was seen for current smokers. The relative risk estimates presented were adjusted for education and ethnicity. Merchant et al (1999) was also able to stratify by APOE ϵ4–containing genotype. The result was consistent with that obtained by Ott et al (1998): increased risk was seen among those with non–ϵ4-containing genotypes, but essentially a null association between smoking and AD among those who carry the ϵ4 allele.

Launer et al (1999) reported an analysis of data compiled from similar cohort studies conducted in Denmark, France, The Netherlands, and the United Kingdom—the Eurodem cohort. The contributing cohorts were designed and conducted under the principle that their data would be combined, and thus noncomparability across study sites was not a major detriment (although adjustment for study site was included in analyses). The analysis included 13,147 subjects, who were observed for approximately 2 years, resulting in 24,731 person-years of observation. As with the Ott et al (1998) and Merchant et al (1999) studies, no increased risk of AD due to smoking was seen in former smokers relative to “never” smokers. Modestly increased risk was observed, however, for those who were smokers at baseline when compared with never smokers. Concerned that the effect of increased risk could be due to vascular disease, Launer also analyzed the data excluding those AD patients with contributing cardiovascular disease; the results did not change materially. Analysis by gender showed somewhat greater relative risk estimates among men than among women. However, there was not a sufficient increase to conclude that an interaction between gender and smoking (effect modification by gender) was a credible conclusion. When the data were stratified by “family history” of dementia, the increased risk in current smokers was confined to the groups designated as having no family history. These results are analogous to the findings of APOE genotype Merchant et al 1999, Ott et al 1998 but may represent a different phenomenon because family history may be a surrogate for many genetic and environmental factors.

Doll et al (2000) reported the association between smoking and AD, but based diagnosis of AD on whether dementia was listed as a cause of death on the death certificate. This study was based on the original study of British physicians that began in 1951 and included 34,439 male patients at the outset. Follow-up was performed at 6- to 12-year intervals following study initiation. Patients’ current smoking status was recorded as smoking for 10 years or more before death, and former smokers had stopped smoking roughly 34 years before death. This study relied on death certificates for the diagnosis of dementia and AD. Doll et al were quite aware that death certificates could be inadequate to determine the type of dementia. Therefore, they conducted a substudy of the death certificates from 1996 to 1998. In the substudy, additional information was obtained from certifying physicians or other hospital physicians concerning dementia diagnosis. Results of the substudy were then applied to estimate the proportion of cases with death certificate entries indicating dementia that were likely to be AD. As of December 1998, 24,133 deaths had occurred from the original cohort; this constituted the base of the Doll study. Death certificates identified only 473 subjects as demented. The substudy results allowed those 473 to be further classified as AD or as “other” dementias. As shown in Table 2, there was no increased risk of AD associated with smoking (based on death certificate diagnosis of AD).

Although this study has many distinct advantages, such as long-term follow-up and careful exposure measurement well before disease onset, it also causes some concern about diagnostic misclassification. The substudy of dementia diagnosis based on death certificate entries was an excellent step toward validity. Doll does not propose a method to detect unrecognized dementia and AD prevalence among the approximately 24,000 other deaths that had occurred at time of analysis. The reliance on death certificate entries alone to identify cases may have failed to capture many patients with AD. For example, persons who die of cancer while in the early to moderate stages of dementia may not have dementia entered as a cause of death on their death certificate.

Calendar year of death may have also influenced the likelihood that AD or dementia would be recorded on a death certificate (Doll et al 2000). Death certificates completed between 1951 and 1978 may be considerably less likely to include dementia as a cause, since diagnostic fashion tended to regard some “senility” as a part of aging. This study contrasts others cited previously, in that its outcome is not incidence of the clinical diagnosis of dementia and AD. Nondifferential misclassification of diagnosis (missing persons who died with unrecorded dementia, irrespective of smoking) could be expected to drive an observed association toward the null.

Section snippets

Discussion

We have discussed some of the potential biases associated with case-control studies, and all the results appeared rather consistent. History of smoking was more frequent among control subjects than it was among cases. Stated causally, smoking appeared to be protective for AD when evidence was limited to case-control studies. Biases inherent in selection of patients, such that exposure to smoking was favored in control subjects, through differential survival (Wang et al 1999), appear to be one

Acknowledgements

Aspects of this work were presented at the symposium “Nicotine Mechanisms in Alzheimer’s Disease,” March 16–18, 2000, Fajardo, Puerto Rico. The conference was sponsored by the Society of Biological Psychiatry through an unrestricted educational grant provided by Janssen Pharmaceutica LP.

The author’s work is also supported in part by grants from the National Institute on Aging (AG06781, AG07584, AG16976).

References (24)

  • A. Ott et al.

    Smoking and risk of dementia and Alzheimer’s disease in a population-based cohort studyThe Rotterdam Study

    Lancet

    (1998)
  • L.A. Amaducci et al.

    Risk factors for clinically diagnosed Alzheimer’s diseaseA case-control study of an Italian population

    Neurology

    (1986)
  • L. Barclay et al.

    Tobacco use in Alzheimer’s disease

    Prog Clin Biol Res

    (1989)
  • D.E. Brenner et al.

    Relationship between cigarette smoking and Alzheimer’s disease in a population-based case-control study

    Neurology

    (1993)
  • G.A. Broe et al.

    A case-control study of Alzheimer’s disease in Australia

    Neurology

    (1990)
  • V. Chandra et al.

    Case-control study of late onset “probable Alzheimer’s disease”

    Neurology

    (1987)
  • M. Dewey et al.

    Risk factors for dementiaThe Liverpool study of continuing health in the community

    Int J Geriatr Psychiatry

    (1988)
  • R. Doll et al.

    Smoking and dementia in male British doctorsProspective study

    BMJ

    (2000)
  • L. Ferini-Strambi et al.

    Clinical and epidemiological aspects of Alzheimer’s disease with presenile onsetA case control study

    Neuroepidemiology

    (1990)
  • L.R. French et al.

    A case-control study of dementia of the Alzheimer type

    Am J Epidemiol

    (1985)
  • L. Gordis

    Epidemiology

    (1996)
  • A.B. Graves et al.

    Alcohol and tobacco consumption as risk factors for Alzheimer’s diseaseA collaborative re-analysis of case-control studies. EURODEM Risk Factors Research Group

    Int J Epidemiol

    (1991)
  • Cited by (49)

    • The shared genetic architecture of modifiable risk for Alzheimer's disease: a genomic structural equation modelling study

      2022, Neurobiology of Aging
      Citation Excerpt :

      Factor 1 had positive loadings for risk factors that are strongly associated with reduced life expectancy (smoking, deprivation, less education and physical inactivity) (Holford et al., 2014; Lewer et al., 2020; Warburton et al., 2010) together with a negative loading for AD. This factor might reflect only an apparent protective effect of these factors on AD due to premature mortality and survival bias (Hernán et al., 2008; Kukull, 2001; Weuve et al., 2015), or could simply reflect a clustering of these traits independent of AD. Factor 2 had highly stable positive loadings for a majority of the risk factors and a positive loading for AD, suggesting that it might represent a common genetic pathway to AD across metabolic, psychiatric and lifestyle traits.

    • Epidemiology for the clinical neurologist

      2016, Handbook of Clinical Neurology
      Citation Excerpt :

      Prevalence bias can distort associations between risk factors and diseases. In the 1990s, several case-control studies demonstrated a protective association between smoking and Alzheimer's disease (Kukull, 2001). It was later understood that smokers who developed Alzheimer's were dying earlier than nonsmokers with Alzheimer's because of other diseases associated with smoking.

    • Alzheimer's Disease and the Search for Environmental Risk Factors

      2015, Environmental Factors in Neurodevelopmental and Neurodegenerative Disorders
    • History of cigarette smoking in cognitively-normal elders is associated with elevated cerebrospinal fluid biomarkers of oxidative stress

      2014, Drug and Alcohol Dependence
      Citation Excerpt :

      Cigarette smoking in the US promotes at least a 10-year reduction in life expectancy (Jha et al., 2013), which may create a survivor bias due to premature death. In other words, the study of the effects of smoking in elders will be inescapably biased toward the healthiest smokers—those individuals who survived or did not experience significant smoking-related morbidity (Chang et al., 2012; Kukull, 2001). Therefore, the effects of smoking on F2-isoprostane levels in this elder sample may be underestimated due to survivor bias.

    • Smoking and increased Alzheimer's disease risk: A review of potential mechanisms

      2014, Alzheimer's and Dementia
      Citation Excerpt :

      For all the aforementioned case-controlled and cohort studies, there was considerable variability in the sample sizes, the duration that participants were followed (for cohort studies), and the covariates (e.g., APOE genotype, alcohol consumption, sex, biomedical risk factors) measured and/or controlled for in statistical analyses. Survivor bias has been indicated to promote an underestimation of the smoking-related risk for AD in both case-controlled and cohort studies [88–92]. Specifically, elders who die prematurely from smoking-related diseases are a major source of attrition; this reduces the proportion of smokers who may have ultimately developed AD, creates attrition in cohort studies, and those smokers who do survive are biased toward healthier individuals [92].

    View all citing articles on Scopus
    View full text