Smoking, alcohol, and coffee consumption preceding Parkinson’s disease

Cases.

We used the medical records linkage system of the Rochester Epidemiology Project to identify all subjects residing in Olmsted County, MN, who developed PD in the period 1976 to 1995.18 Medical care for the population of Olmsted County is provided primarily by the Mayo Clinic and its affiliated hospitals. Additional care providers include a smaller independent clinic and affiliated hospital (the Olmsted Medical Center), a state psychiatric hospital (closed in 1981), three independent practitioners, several nursing homes, and several out-of-county hospitals. All these providers participate in the Rochester Epidemiology Project, which provides the infrastructure for indexing and linking essentially all medical information of the county population.19,20⇓ The indexing of medical diagnoses and surgical interventions is based on the International Classification of Diseases, Adapted Code for Hospitals (H-ICDA).21 We ascertained potential cases of parkinsonism by searching the indexes for 53 H-ICDA diagnostic codes.21 Further details about the study population and the identification of incident cases were reported elsewhere.18 Note that we previously reported incidence data from this population for the 15 years from 1976 to 1990,18 whereas the present study extends for 5 more years through 1995 (a total of 20 years of incidence).

Our diagnostic criteria included two steps: the definition of parkinsonism as a syndrome and the definition of PD within the syndrome. Parkinsonism was defined as the presence of two of four cardinal signs: rest tremor, bradykinesia, rigidity, and impaired postural reflexes. PD was defined as the presence of parkinsonism with all three of the following: 1) no secondary cause (e.g., repeated stroke with step-wise progression, repeated head injury, history of encephalitis, neuroleptic treatment within 6 months before onset, hydrocephalus, brain tumor); 2) no documentation of unresponsiveness to levodopa at doses of at least 1 g/day in combination with carbidopa (applicable only to patients who were treated); 3) no prominent or early (within 1 year of onset) signs of more extensive nervous system involvement (e.g., dementia or dysautonomia) not explained otherwise.18 Onset of PD was defined as the year in which one of the four cardinal signs of PD was first noted by the patient, by family members, or by a care provider (as recorded in the medical record).

Control subjects.

Each case was individually matched by age (±1 year) and sex to a general population control subject residing in Olmsted County, MN, and free of PD, other parkinsonism, or tremor of any type in the index year (year of onset of PD in the matched case). The list of all county residents from which potential control subjects were drawn was provided by the records linkage system and was based on the enumeration of all individuals in contact with the system at least once in the 3 years following the index year.20 This enumeration of the total Olmsted County population at any given point in time has been shown to be complete by comparison with a random digit dialing telephone sample22 and by comparison with the census.23 Therefore, our control subjects were not selected through their diseases or health conditions, but rather on the basis of their residency status. Potential control subjects were selected randomly among all residents fulfilling the matching criteria. Records of potential control subjects were reviewed by a neurologist to exclude the presence of PD, other type of parkinsonism, or tremor of any type before or at the index year. Whenever a potential control was excluded because of parkinsonism or tremor or because county residency in the index year was not confirmed at record review, the next random choice was considered. The presence of dementia or other neurologic diseases was not an exclusion criterion. Because cases and matched control subjects were recruited over a 20-year period, it was possible that a person selected as a control subject at one point in time (index year) would later develop PD and be eligible for inclusion as a case in the study.24 Although we allowed for that possibility, it did not occur.

Reliability of case or control classification.

As part of an ongoing genetic and epidemiologic study, we are inviting all surviving PD cases and control subjects involved in the present study to come to the Department of Neurology at the Mayo Clinic for a standardized evaluation by one of three movement disorders specialists. Subjects willing to be examined but unable to travel are examined at their house (or institution). Presence or absence of PD or of parkinsonism is assessed without access to the medical record maintained in the records linkage system. As of March 1, 2000, we contacted 94 of the 196 cases; among those not contacted, 85 were deceased at the time of the study, and 17 are pending contact. Of these 94 cases, 53 underwent the examination. Fifty-two cases of PD were found to fulfill our diagnostic criteria for PD (98%); however, one case fulfilled the criteria for parkinsonism, but the diagnosis by type remained uncertain (2%). We also contacted 109 of the 196 control subjects; among those not contacted, 71 were deceased at the time of the study, and 16 are pending contact. Of these 109 control subjects, 56 underwent the examination. None of the control subjects examined were found affected by PD or parkinsonism, except for one who developed PD ∼7 years after the index year (and after 1995).

Exposure ascertainment.

The complete medical dossiers of cases and control subjects were reviewed by a neurologist to abstract information about smoking, alcohol, and coffee consumption preceding the onset of PD motor symptoms (or the index year for control subjects). Information about tobacco consumption included cigarette smoking, pipe smoking, cigar smoking, and tobacco chewing or snuff use. For cigarette smoking, we obtained information on number of cigarettes per day and number of years smoked. Whenever the number of cigarettes per day varied over time, we computed an average number of cigarettes per day (weighted by duration). If, for example, a man smoked 20 cigarettes/day between age 18 and 25, 30 cigarettes/day between age 26 and 40, and 10 cigarettes/day between age 41 and 72, the weighted average was ∼17 cigarettes/day for 55 years. Pack–years were directly abstracted from the record when available or computed by multiplying the average number of packs per day (cigarettes per day divided by 20) times the number of years smoked.

Information about alcohol consumption was collected with an attempt to define the amount of intake. However, because quantitative information was not homogeneously available in medical records, we used approximate doses such as “heavy,” “moderate,” and “light” or approximate frequencies of consumption such as “regular,” “occasional,” and “rare.” We also abstracted diagnoses of alcoholism.

For coffee consumption, we collected information about number of cups per day. Whenever the number of cups per day varied over time, we computed an average dose (weighted by duration) as described above for smoking.

Reliability of exposure information.

We assessed the reliability of information about cigarette smoking and other tobacco consumption obtained at medical record review (in the records linkage system) by interviewing a subsample of cases and control subjects. As part of an ongoing study, we conducted an independent telephone interview with a total of 199 individuals (88 direct interviews and 111 interviews with a proxy informant). Two proxy informants and one subject interviewed directly were unable to provide information about tobacco use; of the remaining 196 medical record–interview pairs, 91% (178/196) were in agreement about cigarette smoking ever versus never (91 negative and 87 positive agreements). However, the record documented cigarette smoking but the telephone interview missed it for 14 individuals (7%), and the interview yielded a positive response but the record did not contain the information for 4 subjects (2%). Agreement was also good for pipe smoking (89%), cigar smoking (90%), and tobacco chewing or snuff use (98%). Unfortunately, a similar reliability study was not possible for alcohol or coffee consumption because these two habits were not included in the telephone interview.

Data analysis.

Consistent with the matched design, matched pairs analyses were performed, and the OR was used to estimate the relative risk. For each study variable, we calculated an OR, a 95% CI, and a p value (two-tailed test, α = 0.05) using conditional logistic regression.24,25⇓ For primary analyses, we considered habits as ever versus never. For cigarette smoking, we also analyzed separately past and current smoking. Dose–effect contrasts were investigated for number of cigarettes per day, number of years smoked, number of pack–years (of cigarettes), and number of cups of coffee per day. Individuals missing information about a given exposure or dose were excluded from the corresponding analysis. Frequencies of missing information in cases and control subjects were compared using Fisher’s exact tests.

Conditional logistic regression models were used to investigate the independent effect of a risk or protective factor after adjustment for one or several other factors or to adjust for confounding variables (e.g., education). Whenever variables were dichotomized or polychotomized, the cutoffs were derived from the pooled distribution of cases and control subjects (e.g., using the median, tertiles, or quartiles). We also used a conditional logistic regression model to adjust for two additional putative risk factors for PD found in this same data base and reported elsewhere: anxiety disorders (ever versus never) and depressive disorders (ever versus never).26

Because we found consistently higher incidence rates of PD in men than women of all ages in this same Olmsted County population,18 we reanalyzed our data for men and women separately. In addition, because of the possible role of genetic factors in early-onset PD27 and the possible age-dependent effect of smoking on PD,3 we reanalyzed our data stratifying the cases by age at onset of PD in four quartiles. The effect of coffee on age at onset of PD among cases was explored using direct displays of data and box plots.28

Consideration of methods.

Our study has a number of strengths compared with previous case-control studies of risk habits that almost invariably were not population based, used prevalent PD cases, and collected data through an interview.1 First, we included all incident cases of PD from a defined population over a defined time window and control subjects derived from the same population over the same time. Referral bias is expected to be minimal in this population-based case-control study.39 Although cases of PD that did not come to medical attention may have been missed, we showed that our case ascertainment through the records linkage system was complete and representative of the population.18 We emphasize that our control subjects were general population control subjects and they should have been affected by smoking-related diseases (or coffee- or alcohol-related diseases) with frequencies representative of the general population with the same age and sex. Second, our study design reduced the risk of prevalence–incidence bias.39 If smoking has a negative prognostic effect on survival after onset of PD (reduced survival because of smoking–PD interaction), prevalent series of PD cases may underrepresent individuals who smoked. This negative effect on survival would tend to create a spurious inverse association in a case-control comparison based on prevalent cases.40,41⇓

Third, we collected information about exposures through a retrospective review of medical records in the records linkage system; this approach reduced the risk of recall bias.39 The study did not involve any interview or recall of past events. Documentation of smoking, alcohol, and coffee consumption was generated historically as part of routine medical care and before the onset of PD in cases. Therefore, the information was independent of knowledge about subsequent disease status. Fourth, strict matching of cases and control subjects for age (and calendar year) prevented differences across groups due to secular trends in habits and behaviors.

Our study has, however, a number of limitations. First, because record abstracting of exposures for cases and control subjects was performed by a neurologist, we cannot exclude some unmasking in the data collection. This potential unmasking would increase the risk of an exposure suspicion bias.39 The concern is that the abstractor would search the records of cases more intensively than records of control subjects. Because the information abstracted for this study is clearly present in the record and does not require interpretation or clinical judgment, we do not think this was a problem. In addition, the information abstracted from the record agreed with that independently obtained from a telephone interview in a subsample of cases and control subjects (reliability study for smoking information).

Second, information about smoking, alcohol, and coffee was not available for all study subjects; however, differences in frequency of missing data between cases and control subjects were not significant. Individuals with missing data were excluded from analysis. In addition, some of our cases or control subjects may have been misclassified concerning smoking, alcohol, and coffee consumption or about the dose of exposure. Misclassification (primarily undercounting of exposures) may have resulted from inaccurate reporting (denial of use) or from inaccurate documentation in the records. However, because data on exposure were generated prospectively (historically), we may reasonably assume that the misclassifications were independent of the subsequent disease status and should have biased our analyses toward conservative findings (driving the OR toward 1.0). In addition, our reliability study comparing record information with telephone interview information about the same individuals suggests that the record documents smoking more often than the interview, thus arguing against an undercounting of exposures. On the other hand, some of our negative findings for dose–effect trends may have been due to the combined effect of missing data and possible misclassification. For example, this may explain our failure to detect a significant dose–effect trend for pack–years of smoking.

Third, our findings based on a retrospective review of records in a records linkage system could be biased if cases had a longer time of enrollment in the system than control subjects. Because the likelihood of generating habits information over life is related to the number of years in which a person has been in contact with the system, this is a potential source of bias. We investigated the duration of enrollment in the records linkage system for cases and control subjects and found no difference.