Rockwood K, Kirkland S, Hogan DB, MacKnight C, Merry H, Verreault R, Wolfson C, McDowell I. Use of Lipid-Lowering Agents, Indication Bias, and the Risk of Dementia in Community-Dwelling Elderly People. Arch Neurol. 2002;59(2):223-227. doi:10.1001/archneur.59.2.223
Recent reports suggest a possibly protective effect for statins in patients with Alzheimer disease. This association could be due to indication bias, ie, people who elect to take lipid-lowering agents (LLAs) may be healthier than those who do not, so that it may be these other health factors that explain their lower risk of dementia.
To examine the association between the use of LLAs and dementia, adjusting for other markers of health, and to investigate factors associated with LLA use.
A cohort study of LLA use and a case-control study of dementia in relation to LLA use, in a secondary analysis of the Canadian Study of Health and Aging.
A nationally representative population-based survey of Canadians 65 years and older.
To examine features associated with statin use, we evaluated data on 2305 people for whom health information, drug use, and cognitive status were known. To examine the relationship between LLA use and dementia, we selected incident cases of dementia (n = 492, of whom 326 had Alzheimer disease) that occurred between the first and second waves of the study. Control subjects were 823 persons examined during the first and second phases of the Canadian Study of Health and Aging who had no cognitive impairment.
Use of LLAs was significantly (P<.001) more common in younger (65-79 years) than in older (≥80 years) people. It was not associated with other factors indicating a healthy lifestyle, but was associated with a history of smoking and hypertension. Use of statins and other LLAs reduced the risk of Alzheimer disease in subjects younger than 80 years, an effect that persisted after adjustment for sex, educational level, and self-rated health (odds ratio, 0.26; 95% confidence interval, 0.08-0.88). There was no significant effect in subjects 80 years and older.
While the possibility of indication bias in the original observations cannot be excluded, it was not demonstrated in LLA use in this study. Lipid-lowering agent use was associated with a lower risk of dementia, and specifically of Alzheimer disease, in those younger than 80 years. Further research is warranted.
RECENT STUDIES have raised the possibility that the use of statins, but not other lipid-lowering agents (LLAs), may prevent dementia. In a cross-sectional analysis1 of more than 57 000 patients, the prevalence of probable Alzheimer disease (AD) in those 60 years and older taking lovastatin or pravastatin sodium, but not simvastatin, was 3.8 per 1000. This was significantly different from the overall prevalence of 8.1 per 1000 in the total patient population, and also significantly different from the prevalence of AD among patients using other cardiovascular and/or antihypertensive drugs. Similarly, Jick et al2 reported, in a nested case-control study derived from a computerized database of 368 general practitioners, that the relative risk of dementia for those prescribed statins was 0.29 (95% confidence interval [CI], 0.13-0.63).
While these results suggest a protective effect of statins on dementia, they are also subject to important limitations. The representativeness of patients in these databases is unclear, and information on important potential confounding factors, such as educational level, was not available. Cross-sectional and case-control designs are susceptible to bias, particularly indication bias (or "confounding by indication"), which occurs when a drug is prescribed for a reason that itself is associated with the outcome of interest. As argued in the editorial that accompanied the article by Wolozin et al,1 while the data were gathered (1996-1998), physicians were more likely to have prescribed statins to patients who were "more highly educated, attentive, inquisitive and concerned about their future health."3(p1411)
The Canadian Study of Health and Aging (CSHA) is a national, population-based, representative cohort study4,5 of dementia in elderly people that can provide insight into the extent to which indication bias is responsible for the association between statins and dementia. We reviewed the data therein to investigate whether statin use is associated with demographic and lifestyle factors, such as educational level and self-reported health, and to quantify the relationship between the use of LLAs and dementia.
The study design and data collection protocol of the CSHA are described in detail elsewhere.4,5 Briefly, in the first phase (CSHA-1: February 1991 to May 1992), subjects were recruited from the community and institutions based on age-stratified (65-74, 75-84, and ≥85 years) random samples in 36 urban and surrounding rural areas in all 10 Canadian provinces. Subjects were 65 years and older as of October 1, 1990, and were fluent in English or French. In all, 9008 community residents and 1255 residents of institutions were surveyed. Participants living in the community were first screened for cognitive impairment by the Modified Mini-Mental State Examination.6 Those screening positive for cognitive impairment (Modified Mini-Mental State Examination score of <78) were invited to a clinical examination (n = 1106) and were assessed according to a detailed protocol7 that used standard criteria to diagnose dementia8 and its causes, chiefly AD9 and vascular dementia.10 This protocol was also followed for those unable to complete the screening test (eg, because of deafness [n = 59]) and for a random sample of subjects who were cognitively normal, according to the results of the screening test (n = 494). All institutionalized participants underwent the clinical assessment without first being screened.
This clinical examination included the determination of functional status, health history, and medication use and an informant interview by a nurse; a medical history and physical examination by a physician; and, for most subjects (those scoring ≥50 on the Modified Mini-Mental State Examination [n = 1879]), neuropsychological testing. In addition, subjects without dementia completed the Self-assessed Risk Factor (SARF) questionnaire. The SARF questionnaire contains information on several lifestyle factors, including alcohol and tobacco use, exercise level, and vaccination status.11
Of the 2914 subjects participating in the clinical examination, 921 had no cognitive impairment, 861 had cognitive impairment but not dementia, and the remainder were diagnosed as having dementia. Those who screened negative and were not invited to the clinical examination were presumed to not have dementia.
In 1996, a follow-up data collection (CSHA-2) was undertaken. Of the original participants, 1846 had died and 1135 refused to participate or could not participate. Community subjects presumed to not have dementia at CSHA-1 were screened and examined as previously described. The screening interview also included information on self-reported health and level of education. All those who underwent a clinical examination during CSHA-1 were also invited to the CSHA-2 clinical assessment, which was similar to the one conducted during CSHA-1. A total of 2305 subjects participated in the clinical examination. The clinical interview included an inventory of all prescribed drugs, including LLAs.
To assess the possibility of indication bias, we used a cohort design, with exposures (demographic and lifestyle factors) from CSHA-1 and the outcome (LLA use) at CSHA-2. The 2305 subjects participating in the clinical examination were included in this portion of the analysis. Self-assessed Risk Factor data were only available for 1354 of these subjects, as neither the 350 with dementia at CSHA-1 nor the 210 originally institutionalized were asked to complete the SARF questionnaire. The remaining 391 subjects did not return the questionnaire.
To assess whether LLAs were associated with dementia, we used a case-control design. Control subjects were those without cognitive impairment at the CSHA-2 clinical interview who were also (by screening interview or clinical examination) not impaired at CSHA-1. Cases were defined as those with incident dementia, ie, those who at baseline did not have dementia, based on either the clinical examination results or a screening Modified Mini-Mental State Examination score of 78 or higher, and at CSHA-2 were diagnosed as having dementia. The decision to use only recent-onset dementia cases was made to minimize confounding by indication, given the observation that those already diagnosed as having dementia in CSHA-1 were less likely to have undergone aggressive treatment of vascular risk factors.12
Of the 2305 individuals who underwent a clinical examination at CSHA-2, 350 were excluded because they had dementia at CSHA-1 and 640 were excluded because they had cognitive impairment, but no dementia, at CSHA-2.13 Of the remainder, 823 were controls and 492 met the case definition of incident dementia. Of these 492 subjects, 326 were diagnosed as having AD.
Comparisons regarding demographic and lifestyle characteristics for LLA users and nonusers were conducted using χ2 tests for categorical data and 2-sided t tests for continuous data. Similarly, baseline characteristics of cases and controls were compared using t tests or χ2 tests as appropriate. Bivariate analyses to assess the impact of potential confounders were first conducted using the Mantel-Haenszel approach. The association between statin use and dementia was examined by calculating crude and adjusted odds ratios (ORs) using logistic regression analysis (SAS statistical software, version 8.01; SAS Institute Inc, Cary, NC). Odds ratios and 95% CIs for them were calculated. Because LLA use was likely to be less controversial in younger patients, and thus less susceptible to confounding by indication, the analysis was stratified by age, dichotomized as younger than 80 years vs 80 years and older.
At CSHA-2, 57 people were taking statins and 15 were taking other LLAs. The characteristics associated with LLA use at CSHA-2 are presented in Table 1. Lipid-lowering agent users were significantly younger than nonusers and were more likely to live in the community, to have hypertension, and to be a current or a former smoker. Educational level, self-reported health, and sex were not significantly associated with use of an LLA.
The characteristics of cases and controls used in the analysis of the association of dementia with LLA use are outlined in Table 2. Cases were slightly older and had a higher prevalence of "cognitive impairment, no dementia" than controls at baseline. The prevalence of stroke was much higher in those with other dementias, reflecting the preponderance of vascular dementia in this group.
The unadjusted analyses showed that use of statins was associated with a lower odds of AD (OR, 0.26; 95% CI, 0.09-0.72) and all types of dementia (OR, 0.21; 95% CI, 0.08-0.54). Similar ORs were estimated for the use of any LLA and AD or dementia (OR, 0.25; 95% CI, 0.10-0.62).
When stratified by age, the protective effect of taking either a statin or another LLA was observed only for all types of dementia in those younger than 80 years (OR, 0.24; 95% CI, 0.07-0.80); for those 80 years and older, the OR was 0.43 (95% CI, 0.11-1.58). After further adjustment for sex, educational level, and self-rated health, the OR for those younger than 80 years was 0.26 (95% CI, 0.08-0.88); for those 80 years and older, the OR was 0.50 (95% CI, 0.13-1.88). While the effect remains protective in those 80 years and older, the CI includes 1.0. (Similar trends are observed when the analysis is restricted to AD only, statin use only, or other LLA use. The inclusion of institutionalized subjects, who had a high prevalence of dementia but who were infrequently taking statins, narrows the CIs to less than 1.0.)
These data suggest that LLAs may have a role in protecting against the expression of dementia in elderly people. The data go against such an association being explained simply by confounding by indication: not only was LLA use not associated with other markers of a healthy lifestyle but it was associated with other vascular risk factors, which increasingly are implicated in all types of dementia.14- 16
Confounding by indication is a consideration in observational studies, and parallel scenarios in which protective effects are documented with other agents and outcomes suggest that it must be addressed carefully. For example, in a meta-analysis17 of 10 observational studies, postmenopausal estrogen use decreased the risk of dementia by 29%, but randomized controlled trials18,19 in women with dementia have failed to replicate this finding. Similarly, estrogen use was long held to be protective for ischemic heart disease in women based on many, but not all, observational studies20,21 that reported such an effect. In contrast, randomized treatment trials,22,23 again, have not shown a protective effect. Moreover, the ongoing Women's Health Initiative,24,25 a randomized clinical trial of 27 500 healthy women, recently issued a press release to communicate an early finding of a small increase in heart attacks, strokes, and blood clots in women taking hormones compared with nonusers.
Our data must be interpreted with caution. Although confounding by indication was addressed by controlling for known risk factors and markers of healthy behaviors, we cannot account for unknown and/or unmeasured factors. Because this is a secondary analysis, and because it covers a period in which the use of LLAs in elderly people was controversial,3 the numbers of LLA users are smaller than would be needed to draw more robust conclusions. Similarly, we were unable to make full use of the cohort design of the CSHA, because, at baseline, of the 2914 subjects for whom drug information was available, only 17 were taking statins and only 14 were taking other LLAs. (The unadjusted analyses suggest a protective effect: the relative risk was 0.74 [95% CI, 0.54-1.01]). On the other hand, the data come from a representative population-based sample, cover a similar period as the original reports, and allow confounding by indication at least to be addressed. Restricting the adjusted analysis to community-dwelling respondents is a conservative strategy, and the persistence of the protective effect enhances the possibility that the observed association is not simply due to confounding.
We restricted our analysis to incident cases at CSHA-2, on the grounds that prevalent cases would have been less likely to be prescribed a statin in the period under consideration. Earlier analysis12 of the CSHA data set suggested that, at least among those with severe dementia, treatment of vascular risk factors occurs less often than in those with mild or moderate dementia. One possibility, however, is that statin use might also have been less likely among those with cognitive impairment but no dementia at CSHA-1, who are more likely to be incident dementia cases. When we repeated the analysis performed between statin or other LLA use and incident dementia in community-dwelling subjects (stratified by age), but restricted incident cases to those with no cognitive impairment at baseline, the effect was still protective (OR, 0.36) but, likely as a result of fewer subjects, the CI included 1.00 (95% CI, 0.11-1.23).
In addition, even though we used only incident cases, we are unable to account for cessation bias, ie, the possibility that the apparently protective effect of LLAs is overstated because those who were taking LLAs and developed dementia had their statins discontinued. Again, the numbers are too small to allow conclusions to be drawn; of the 2 patients who were taking LLAs at CSHA-1, were alive, and developed dementia by CSHA-2, the LLA was discontinued in 1, compared with 2 of the 10 who were taking LLAs at CSHA-1 and had not developed dementia at CSHA-2.
Small numbers also prevent us from commenting specifically on the merits of any one statin over another. Interestingly, whereas simvastatin was not evidently protective in one of the earlier reports, a recent experimental study26 found that simvastatin strongly reduces the abnormal β-amyloid proteins used in animal models. In contrast to the General Practice Research Database study,2 we found that LLAs other than statins were associated with a protective effect, but again the small numbers prevent us from being more definitive.
It would be preferable to be able to adjust the LLA-dementia relationship for all the lifestyle factors recorded in the SARF questionnaire, but as the questionnaire was completed by only 681 of the 823 cases and 255 of the 452 controls, this was not possible. Analysis of the SARF data for these groups, however, shows an interesting trend: in general, AD cases were more likely to have a less healthy lifestyle than controls, whereas those cases with dementia other than AD were as likely to have a healthy risk factor profile as controls. For example, 69% of controls exercised regularly, compared with 66% of those with other dementias and 56% of those with AD. Similar trends were observed for smoking and vaccination status. People with AD were less likely to have used alcohol (nonusers, 70%) than those with other dementias (nonusers, 42%) and controls (nonusers, 60%). Thus, the protective effect seems to exist in groups with dementia regardless of risk factor profiles.
The CSHA data lend support to the observation of a protective association between statins and dementia and suggest that it might be extended to other LLAs. Although longitudinal studies27- 29 investigating the link between hypercholesterolemia and AD give conflicting results, the tantalizing possibility that LLAs might help to prevent dementia requires further research.
Accepted for publication September 17, 2001.
Author contributions: Study concept and design (Drs Rockwood, Kirkland, MacKnight, Wolfson, and McDowell, and Ms Merry); acquisition of data (Drs Rockwood, Hogan, MacKnight, Verreault, Wolfson, and McDowell, and Ms Merry); analysis and interpretation of data (Drs Rockwood, Kirkland, Hogan, MacKnight, Wolfson, and McDowell, and Ms Merry); drafting of the manuscript (Drs Rockwood and MacKnight); critical revision of the manuscript for important intellectual content (Drs Rockwood, Kirkland, Hogan, MacKnight, Verreault, Wolfson, and McDowell, and Ms Merry); statistical expertise (Drs Kirkland, MacKnight, Verreault, and Wolfson, and Ms Merry); obtained funding (Drs Rockwood, Wolfson, and McDowell); administrative, technical, and material support (Dr McDowell); study supervision (Drs Rockwood and McDowell); data collection (Drs Rockwood, Hogan, MacKnight, and Verreault); and clinical input (Drs Rockwood, Hogan, and MacKnight).
Corresponding author and reprints: Kenneth Rockwood, MD, FRCPC, Geriatric Medicine Research Unit, Queen Elizabeth II Health Sciences Centre, 5955 Veterans' Memorial Ln, Room 1421, Halifax, Nova Scotia, Canada B3H 2E1 (e-mail: email@example.com).