Exploring causality of the association between smoking and Parkinson’s disease

Abstract Background The aim of this paper is to investigate the causality of the inverse association between cigarette smoking and Parkinson’s disease (PD). The main suggested alternatives include a delaying effect of smoking, reverse causality or an unmeasured confounding related to a low-risk-taking personality trait. Methods A total of 715 incident PD cases were ascertained in a cohort of 220 494 individuals from NeuroEPIC4PD, a prospective European population-based cohort study including 13 centres in eight countries. Smoking habits were recorded at recruitment. We analysed smoking status, duration, and intensity and exposure to passive smoking in relation to PD onset. Results Former smokers had a 20% decreased risk and current smokers a halved risk of developing PD compared with never smokers. Strong dose–response relationships with smoking intensity and duration were found. Hazard ratios (HRs) for smoking <20 years were 0.84 [95% confidence interval (CI) 0.67–1.07], 20–29 years 0.73 (95% CI 0.56–0.96) and >30 years 0.54 (95% CI 0.43–0.36) compared with never smokers. The proportional hazard assumption was verified, showing no change of risk over time, arguing against a delaying effect. Reverse causality was disproved by the consistency of dose–response relationships among former and current smokers. The inverse association between passive smoking and PD, HR 0.70 (95% CI 0.49–0.99) ruled out the effect of unmeasured confounding. Conclusions These results are highly suggestive of a true causal link between smoking and PD, although it is not clear which is the chemical compound in cigarette smoking responsible for the biological effect.


Background
An overwhelming amount of evidence exists on the inverse association between cigarette smoking and Parkinson's disease (PD). The inverse association is strong and consistent across studies, 1 stronger for current smokers than for former smokers when compared with non-smokers. 1,2 Some studies suggest that smoking duration is more strongly associated with a reduced risk of PD compared with smoking intensity. 3 The overall association appears consistent in men and women 1 and not confounded or modified by educational level. A comparable inverse association was also observed for pipe and cigar smoking in men 4 and for smokeless tobacco. 5,6 An attempt to demonstrate causality of the association has been made using parental smoking as an instrumental variable: it was shown that children of smokers-who are more likely to smoke themselves-are at decreased risk of PD even if they do not smoke. 7 Nonetheless, there is still considerable caution in interpreting this association as protective. Few theories have been postulated to explain the current evidence in a noncausal way and these are summarized with Direct Acyclic Graphs (DAGs) in Figure 1. Some studies failed to replicate the association in cases with an older age of onset 3,8 leading to the hypothesis that smoking might delay, not prevent, PD onset ( Figure 1B). The most intriguing, and more difficult to prove, is a possible confounding effect by a low-risk-taking personality trait that would be regarded as an unmeasured confounder if it is genetically determined or as reverse causation if it is triggered by dopamine shortage 9,10 ( Figure 1C and D). According to this, and coherently with the involvement of dopamine in the brainrewarding circuits, 11 people who will subsequently develop PD tend to have a low-risk-taking personality, which makes them less likely to smoke or more likely to quit. Coherently, before disease onset, people with PD might find it easier to quit smoking compared with those without PD 12 ( Figure 1D). Nonetheless, the inverse association between smoking intensity and PD observed among monozygotic twins argues against a major role of genetics and/or personality. 13 Given that personality trait would have a lesser role in influencing the exposure to passive smoking, demonstrating a decreased risk of PD among those exposed to passive smoking would overcome this effect; however, a previous study failed to find it. 14  Clarifying the causal nature of the association between smoking and PD would contribute to understanding the mechanisms underlying the disease, informing potential targets for preventive or early treatments. Moreover, no data are currently available on the consistency of the inverse association between smoking and PD across clinical subtypes.
The aim of this study is to assess the association between smoking patterns (duration, amount and time since quitting smoking) and PD risk. Specifically, the potential delaying effect; the consistency of smoking patterns among current and former smokers to interrogate any reverse causality; the association with passive smoking; and the consistency of the association across clinical subtypes will be investigated.

Population
The NeuroEPIC4PD study involved 220 494 subjects recruited in Sweden, the UK, the Netherlands, Germany, Spain, Italy and Greece from the general population residing in defined geographical areas between 1992 and 2002 and aged 37-70 years, within the European Presepctive Investigation into Cancer and Nutrition (EPIC) study. 15 Exception was the Utrecht cohort, which was based on breast-cancer-screening participants. 15 The Naples and Utrecht cohorts were restricted to women, whereas all other cohorts involved both sexes. To date, follow-up is 98.5% complete and the median follow-up time of this sample is 12.8 years [inter-quartile range (IQR) 11.5-14.2].

Case ascertainment and sample size
A total of 881 PD cases was ascertained in the participating EPIC centres. 16 The present analysis has been conducted on a total sample of 214 533 subjects (including 715 incident PD cases) after removing 147 prevalent PD cases, 5359 subjects (including 19 PD cases) with missing information on smoking status at recruitment. Moreover, 221 subjects with PD-like conditions [Multi-System Atrophy (MSA) N ¼ 24; Progressive Sopra-nuclear Palsy (PSP), N ¼ 21; vascular parkinsonism, N ¼ 34; Lewy Body Dementia (LBD), N ¼ 34; essential tremor, N ¼ 27; PD with essential tremor, N ¼ 9; and unclassified parkinsonism, N ¼ 72] were also removed from the analysis. The sample resulted in a total of 2 666 206 person/years. Procedures for PD case ascertainment in the EPIC cohort have been described elsewhere. 16 In brief, in each centre, potential cases were identified through record linkage and validated through clinical record review by a neurologist expert in movement disorder who collected additional clinical data, including age of onset (defined as age when the first motor symptom was noticed) and clinical subtype at onset (tremor-dominant, postural instability/gait disturbance, akinetic-rigid forms). 16

Smoking characteristics
Answers to a number of questions on present and past smoking habits were collected at recruitment in the EPIC study. These included smoking status at recruitment (never, former and current smoker), age when they started smoking and quit, and number of cigarettes/day smoked at different ages. This latter information was not collected in Sweden, which was therefore excluded from all analyses on smoking intensity (n ¼ 53 291). Starting from this core information, a number of variables were derived: duration of smoking (never smokers, smokers for <20, 20-29, 30þ years) missing for 4620 individuals; smoking intensity as mean lifetime cigarettes/day (never smokers, <12, 12þ cigarettes/day) missing for 10 876 individuals; time since quitting smoking, namely number of years elapsed from quitting smoking and recruitment to the cohort (never smoker, 19þ, 9-18, <9 years) missing for 2221 individuals; age when quit smoking (never smoker, <33, 34-43, 44þ years) missing for 2221 individuals; and age when started smoking (never smoker, 20þ, 17-19, <16 years) missing for 3011 individuals. Information on second-hand smoke (SHS) exposure was available only in a few centres: participants were asked whether any of their parents smoked when they were children in Italy, the Netherlands and Sweden (N ¼ 59 329), whereas information on current SHS exposure at home or work was available only for participants recruited in Italy and Sweden (N ¼ 40 816).
Additional information collected at baseline and relevant for this analysis is the highest educational level attained (none/primary, technical, secondary, university).

Statistical analysis
Cox-regression models using age as the underlying time variable, adjusted for level of education and sex, and stratified for centre and age at recruitment, were run in order to investigate the effects of the main smoking variables in relation to PD onset. Models investigating smoking status, duration and amount of smoking, time and age since quitting smoking for former smokers and age when started smoking were investigated and p-values for trend across categories calculated where appropriate. Analyses were repeated using never smokers as the reference category where appropriate, in men and women separately, and restricted to tremor-dominant and akinetic-rigid forms of PD at onset. Heterogeneity across country was tested using the approach proposed by Smith et al. 17 Heterogeneity was assessed by the likelihood ratio of two stratified models: one with country-specific estimates and one with overall estimates. Under the null hypothesis of no heterogeneity, this statistic follows approximately a chi-square distribution on (k -1)*(j -1) degrees of freedom (where k is the number of categories of smoking variable and j is the total number of countries).
In order to investigate a potential delaying effect of smoking on PD onset, possible non-proportionality was assessed using the Schoenfeld residuals. 18 Also, the analysis on the main three smoking variables was repeated on the mid-age of PD onset after excluding subjects with an onset at 70þ years (<70 years, N ¼ 385) or on late PD onset, after excluding those with an age of onset younger than 70 years (70þ years, N ¼ 330). Studying separately subjects with a young age at onset (50 years) was not possible, as there were only 12 such cases.
For indirectly exploring reverse causality, the Cox regression exploring the dose-response relationships between smoking intensity and duration were repeated among current and former smokers at recruitment separately.
Both variables on SHS (in infancy and at recruitment) where studied in relation to PD onset in Cox-regression models repeated in never smokers only in an attempt to overcome unmeasured and residual confounding of the main association.
Finally, for exploring the possible competing risk of mortality in the smoker group, a competing-risk survival analysis was carried out using death as a competing event and the Fine and Gray regression model. 19 A sensitivity analysis was conducted repeating the main Cox models using definite and very likely PD diagnosis only (389 PD cases). For further detail on how cases were labelled, please refer to the methodological paper. 16 All analyses were done using STATA 12 IC and R version 3.3.2 (R Foundation for Statistical Computing, Vienna, Austria).
No direct patient involvement was needed to run this study, which was based on data previously collected.

Results
Demographic characteristics and smoking habits for men and women in the EPIC cohort and PD cases are described in Table 1. Former smokers at recruitment had a 20% reduced risk of developing PD during follow-up compared with never smokers; current smokers had a halved risk compared with never smokers (Table 2). These results were highly consistent in men and women (Table 3) and no heterogeneity was detected across countries ( Table 4). The difference in incidence rates across countries is more likely due to local differences in case-ascertainment procedures rather than true difference in incidence, as discussed in. 16 Studied individually, all smoking variables were found to be inversely associated with the risk of PD with clearcut dose-response relationships. For age when started and quit smoking, a monotonic trend across categories was not evident ( Table 2). The analysis of residuals of Schonefeld showed no evidence of non-proportionality over the follow-up period. The smoothed curves for former smokers ( Figure 2A) and for current smokers ( Figure 2B) were flat, showing that beta-coefficient (log hazard ratio) estimates did not vary during follow-up (time) (Figure 2). Smoking variables were associated with inverse risk of both mid-age and late-onset PD; however, all the estimates are stronger in the latter. All the risk estimates, conversely, remain highly consistent for the akinetic-rigid and tremordominant forms at onset ( Table 5). The Postural Instability/Gait Disturbance (PIGD) form could not be studied individually, as it included only 42 subjects. 16 The competing-risk analysis using mortality as a competing factor yielded much stronger point estimates but largely overlapping 95% confidence intervals (CIs) for all the active smoking variables: smoking for 30þ years or 12þ cigarettes/day is associated with a 55% reduced risk of PD compared with never smokers ( Table 2).
Hazard ratios (HRs) of smoking intensity and duration from Cox models stratified for smoking status at recruitment are shown in Figure 3. Point estimates in current smokers are consistently lower compared with those in former smokers, although the pattern of risk reduction is highly comparable across the two groups, all trends had p 0.001 and no interaction was detected between smoking duration and intensity and smoking status (p-value for interaction 0.823 and 0.537, respectively).
Analysis of passive smoking, although hampered by limited power, showed no association between exposure to passive smoking in infancy and risk of PD. However, an inverse association was found between passive-smoking exposure at home or at work and risk of PD (HR 0.70, 95% CI 0.49-0.99), which was replicated among never smokers only (HR 0.71, 95% CI 0.46-1.10).
The sensitivity analysis including definite and very likely PD only yielded strikingly similar results (Table 3). All associations were, if anything, strengthened despite the widening of CIs due to the smaller sample size. An inverse association between age when quitting smoking and risk of PD was also suggested by the sensitivity analysis.

Discussion
This study provides unique data on the inverse association between cigarette smoking and risk of PD in a large, well-established cohort study, supporting previous findings, 3,4,8 and allows testing of explanations other than a direct protective effect. Overall, data coming from the NeuroEPIC4PD study show a robust inverse association between smoking status at recruitment and PD risk, with a dose-response relationship between PD risk and smoking duration and intensity. Of particular interest is the replication of the main findings of the inverse relationship between smoking and PD among different subtypes of the disease. This is a novel finding, as, to our knowledge, clinical subtypes have not been investigated to date in such an epidemiological setting.

Delaying effect of smoking
The fact that proportional assumption hypothesis is verified demonstrates that the risk does not vary over the follow-up period, and this argues against a delaying effect of smoking on PD onset ( Figure 1B). Moreover, at odds with some previous reports, 3,8 our findings of an inverse relationship between smoking variables and risk of PD are not weakened when the analysis is restricted to old-age onset PD (70þ years). Taken together, these results are not supportive of the hypothesis that smoking might delay, rather than prevent, PD onset, as previously suggested. 3,8 However, despite this piece of evidence being important and informative per se, the distinction between delaying and preventing any disease onset is somewhat artificial, as these mechanisms might coincide from both a clinical and a biological point of view.

Reverse causality
If an inverse causal relationship-accounting for subjects with a preclinical dopaminergic change who therefore might find it easier to quit smoking-was responsible for the observed inverse association between smoking and PD, the dose-response relationship between smoking duration and intensity should not hold true among former smokers ( Figure 1C). The fact that the risk of PD was reduced among current and former smokers argues against this possible explanation. Furthermore, the inverse association between time since cessation and PD reinforces the idea that reverse causality is not a likely explanation of the findings: having quit smoking 9-18 years before recruitment into the study (therefore up to 30 years before disease onset) still confers a reduced risk of PD compared with never   smokers. This results are in line with previous observational studies that showed an inverse association between parental smoking and PD in the offspring; 7 also, the use of parental smoking as an instrumental variable overcomes the potential for a reverse-causality effect.

Unmeasured confounding
Whereas it was not possible to account for personality trait, its unmeasured confounding effect can be overcome by using exposure to passive smoking in relation to PD onset. Risk propensity is likely to influence one's attitude towards active smoking, whereas passive smoking is more likely to be related to these personal characteristics in a weaker way (e.g. smokers tend to have smoking partners).
The inverse association between passive smoking and PD onset, whose point estimate has been replicated among never smokers only, argues against considering personality trait as a major confounder. These results are in line with previous reports showing how adjusting for sensationseeking score only slightly attenuated the inverse association between smoking and PD suggesting an independent effect 20 and with observations that personality traits such as neuroticism and introversion do not explain the inverse association between smoking and PD risk. 21

Biological plausibility
A number of substances present in tobacco have been proposed as potentially responsible for the inverse association between smoking and PD. One of these is 2,3,6-trimethyl-1,4-naphthoquinone (TMN), an inhibitor of monoamine oxidase (MAO) A and B activity. 22 TMN partially protects against 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP)-induced neurodegeneration in mice by reducing endogenous dopamine metabolism and consequently decreasing oxidative stress. Synthetic MAO B inhibitors are currently used in the treatment of PD, providing symptomatic relief, but they may also protect against nigrostriatal damage decreasing dopamine metabolism, as suggested by delayed need for antiparkinsonian drugs in a recent clinical trial. 23 Another candidate is nicotine itself, given the close anatomical relationship between the nicotinic cholinergic and dopaminergic neurotransmitter systems in the striatum. Nicotine influences also the dopaminergic activity by acting at nicotinic receptors on dopaminergic terminals and modulating dopamine release. 24,25 The role of nicotine is being investigated in a randomized trial in patients with early PD, but a role of other tobacco components cannot be excluded. Being exposed to passive smoke is associated with a reduced risk of 30% (HR 0.70, 95% CI 0.49-0.99) and being a light smoker with a 20% reduced risk (HR 0.80, 95% CI 0.62-1.02) ( Table 2). Although the difference could be due to limits in the design (data on passive smoking were available for a subset of the sample), it cannot be excluded that passive smoking has a stronger effect than one would expect from a pure equivalence of levels of exposure. Passive smoking has been demonstrated to be as mutagenic as active smoking, 26 although earlier studies suggest that the overall chemical composition of passive smoking might not represent only the diluted composition of side-stream smoking, given the sorbing and desorbing properties of some volatile and semi-volatile organic compounds in passive smoking. 27 The main strengths of this study are the prospective design, the validated clinical outcome, 28 the large sample and the detailed information on smoking patterns. This allowed a powered recall-bias-free analysis of smoking patterns in relation to PD onset. The main limitation of this study, however, is the lack of repeated smoking measurements over time, which might introduce some exposure misclassification, decreasing our ability to study smoking patterns in relation to PD onset. This is particularly true for outcomes ascertained many years after recruitment. However, the smoking pattern analyses repeated separately for PD cases ascertained within and after 8 years since recruitment yield highly consistent results (data not shown).

Conclusions
In conclusion, the present findings are consistent with a protective effect of smoking on the risk of PD. Point estimates of smoking status are strong, with a strong exposure-response relationship of smoking intensity and Table 5. Hazard ratios (HRs) and relative 95% confidence intervals (CIs) for Cox regressions analysing risk of PD at early and older age of onset and in tremor-dominant or akinetic-rigid forms duration. The consistency across different disease subtypes suggests that the putative protective effect might spread to the entire clinical spectrum of the disease. Finally, the inverse association found between passive smoking and PD is supported by a consistent finding among never smokers and points towards a true biological effect not mediated by personality type. Although smoking to prevent PD cannot be recommended given the multiple adverse effects of smoking, our results confirming an inverse association warrants further research on the mechanisms involved. In particular, the use of Mendelian randomization and biomarkers of long-term cigarette-smoke exposure should provide compelling final evidence on the inverse association between smoking and PD.

Funding
No specific funding was available for this study. The researchers are independent from any funding sources with regard to this study.