Neurology -- Correspondence for Li et al., 63 (9) 1624-1628

Correspondence published:

Statin therapy and risk of dementia in the elderly: A community-based prospective cohort study: Mahyar Etminan, Mark Fitzgerald, Ali Samii (2 December 2004)

Reply to Etminan et al: Gail Li, Roger Higdon, Walter A. Kukull, ELaine Peskind, Debby Tsuang, Eric B. Larson (2 December 2004)

Statin therapy and risk of dementia in the elderly: A community-based prospective cohort study 2 December 2004

Mahyar Etminan,
Divisions of Clinical Epidemiology, Royal Victoria and Vancouver Hospital, Canada
4.29 687 Pine Ave West H3A 1A1,
Mark Fitzgerald, Ali Samii
Send Correspondence to journal:
Re: Statin therapy and risk of dementia in the elderly: A community-based prospective cohort study

mahyar.etminan{at}mail.mcgill.ca Mahyar Etminan, et al.

We commend Li et al for using robust methodology, especially a time-dependent Cox model. [1] They argue that one reason for the potential biased results of previous case-control studies may have been in the method of selection of the controls. Controls may have been selected in such a way that a control may have had more opportunity to have been prescribed a statin. This may have spuriously shifted the odds ratio towards a protective effect. The authors further elaborate that a nested-case-control study (case control study within a well-defined cohort) may have prevented this bias.
Here we present an alternative explanation for this possible bias referred to as immortal time bias. This bias may have been present in the first large epidemiologic study that showed a protective effect with the use of statins. [2] The study initially followed a cohort of subjects who used at least one prescription of statins (at any time) or non-statin anti -cholesterol drugs from 1992-1998 and later conducted a case-control study within the cohort (nested-case-control study).
Immortal time bias refers to a phenomenon where failure to account for the time-dependency of exposure in a cohort study can affect the rates in the exposed and unexposed groups and therefore bias the rate ratio (RR=Rate Exposed/Rate Unexposed). [3,4] It is possible that a person who entered the cohort in that study did not receive a statin until many months after cohort entry.
In a time-independent analysis, this person would have been considered as exposed when in reality the true person time contribution in the cohort for this individual should have been counted as unexposed person time. Subsequently, the rate in the unexposed group could artificially be made larger potentially biasing the rate ratio towards a protective effect. Because the odds ratio may approximate a rate ratio when the disease of interest is rare, the odds ratio obtained from the nested-case-control study may also be shifted towards a protective effect. Immortal time bias can therefore occur in both a classical cohort study or a nested-case- control study. This bias may be avoided when a time-dependent Cox model is used in the analysis of cohort studies as shown by Li et al.
References
1. Li G, Higdon R, Kukull WA, Peskind E, Van Valen Moore K et al. Statin therapy and risk of dementia in the elderly: a community-based prospective cohort study. Neurology 2004;63:1624-8.
2. Jick H, Zornberg GL, Jick SS, Seshadri S, Drachman DA. Statins and the risk of dementia. Lancet 2000;356:1627-31.
3. Suissa S Effectiveness of inhaled corticosteroids in chronic obstructive pulmonary disease: immortal time bias in observational studies. Am J Respir Crit Care Med 2003;168:49-53.
4. Etminan M. Pharmacoepidemiology II: the nested case-control study. A novel approach in pharmacoepidemiologic research. Pharmacotherapy 2004;9:1105-9.

Reply to Etminan et al 2 December 2004

Gail Li,
Departments of Psychiatry and Behavioral Science, University of Washington
VA Puget Sound Health Care System, 1660 S Columbian Way, Mailstop S-116 MIRECC, Seattle, WA 98108,
Roger Higdon, Walter A. Kukull, ELaine Peskind, Debby Tsuang, Eric B. Larson
Send Correspondence to journal:
Re: Reply to Etminan et al

gli{at}u.washington.edu Gail Li, et al.

We thank Etminan et al for their thoughtful comments. We believe that the immortal bias suggested by them is consistent with our explanation, rather than being an alternative explanation, regarding how the failure to account for timing of exposure in a case-control study could result in a biased estimation of the odds ratio towards detecting a �protective� effect of statins. In a conventional case-control study, the failure to account for time of exposure is mainly due to the failure to establish a comparable time of exposure in the controls.
For example, a case entered the follow-up study at age 65, developed dementia at age 70, started taking a statin at age 71, and exited the study at age 72. In this case, the use of a statin would not be counted as an exposure because statin use occurred after onset of dementia. When selecting a control for this case, the exposure time should be also limited to the same time period by using a �reference� or �index� time, i.e., from age 65 to 70, rather than any time beyond the �reference� time point (age 70). If the control starts a statin at age 71, he (she) should be correctly classified as non statin-exposed, because he (she) did not actually use a statin in that exposure time period. Thus, both cases and controls have equal exposure time with respect to both length of exposure and calendar time.
The study by Jick et al [2] did use the �index date� as a criterion in the selection of matched controls to determine timing of exposure, although there was no explanation of how the index date was determined. The difference in our findings compared to those of Jick et al is not likely due to this type of bias.
However, the immortal time bias discussed by Etminan et al in classic cohort studies justifies the rationale for choosing a time-dependent covariate over a fixed covariate in modeling statin exposure in the Cox regression model in our study. We observed a biased estimate of hazard ratio (HR) when modeling statin use as a fixed covariate (HR = 0.53, 95% confidence interval [CI] 0.32 - 0.87), in contrast to modeling the statin exposure as a time-dependent covariate (HR = 0.90, 95% CI 0.54 - 1.51) in the Cox regression model. We believe the points raised by Dr. Etminan are valid and important, particularly for the design and analysis of future studies.