Modeling smoking in systemic sclerosis: A comparison of different statistical approaches




To determine the effect of different methods of modeling smoking on vascular outcomes in rheumatic diseases.


Data from the Canadian Scleroderma Research Group Registry were used. Patients self-reported their smoking history. Vascular outcomes were severity of Raynaud's phenomenon, presence of finger ulcers, and severity of finger ulcers. Several models were developed to capture the experience of smoking: 1) ever compared to never smoking; 2) current and past smoking compared to never smoking; 3) never, past, and current smoking compared using polynomial contrasts; 4) smoking intensity, duration, and time since cessation assessed separately; and 5) smoking modeled using the Comprehensive Smoking Index (CSI), which integrates intensity, duration, and time since cessation into a single covariate.


This study included 606 patients, of which 16% were current, 42% were past, and 42% were never smokers. Current and past smokers smoked a mean ± SD of 25 ± 17 and 17 ± 18 pack-years, respectively. Smoking duration was shorter in past compared to current smokers (18.3 versus 31.7 years). Past smokers reported having stopped smoking approximately mean ± SD 16 ± 12 years prior, although this ranged from 1 to 50 years. Smoking had no effect on vascular outcomes in the simplest model comparing ever to never smokers. Models that isolated past smokers revealed the presence of a healthy smoker bias in that group. The model using the CSI demonstrated a strong negative effect of smoking on vascular outcomes.


Proper modeling of the effect of smoking is essential in studies of vascular outcomes of rheumatic diseases.


Systemic sclerosis (SSc; scleroderma) is a chronic inflammatory disease characterized by thickening and fibrosis of the skin and internal organs, most commonly the gastrointestinal tract, lungs, and heart. Vascular disease is ubiquitous in SSc and is manifested by Raynaud's phenomenon, digital ulcers, telangiectasias, scleroderma renal crisis, and pulmonary hypertension.

Cigarette smoking is well known to contribute to vascular disease in the general population (1). However, there is no standard way of measuring the effect of smoking in epidemiologic studies (2). Some use categorical variables relating to smoking status (e.g., never, past, current). Others have attempted to quantify the effect of smoking by using variables relating to intensity (e.g., packs per day), duration, and/or time since cessation of smoking. However, even in these simplest cases, problems arise in interpretation. For example, many studies have dichotomized never versus ever smoking, therefore implying that past smoking is equivalent to current smoking; others have defined past smokers as those having stopped from 1 day to 5 years; and others yet have used pack-years, a simple and intuitively appealing cross-product of intensity and duration that, although widely used, remains unproven in terms of accuracy in describing the impact of smoking (3–5).

Studies focused on the association between smoking and disease outcomes are subject to inherent biases (6, 7). A healthy smoker bias hypothesizes that an individual who smokes, now or in the past, may have lungs that were relatively more resistant to the effects of smoking. There is evidence of that bias in the literature (7). A causality bias may occur when patients initiate or discontinue smoking due to symptoms related to the underlying disease (6). Failure to account for these biases may lead to under- or overestimates of the association between smoking and disease outcomes.

The Comprehensive Smoking Index (CSI) is a sophisticated method developed to model smoking in epidemiologic studies (2, 8, 9). The CSI integrates smoking intensity, duration, and time since cessation into a single covariate of smoking effect. This index models the physiologic effect of smoking in a manner that is biologically plausible, flexible and, as a single parameter, easy to interpret. From a statistical point of view, it is parsimonious, avoids issues of multicolinearity between multiple smoking variables, and reduces the possibility of residual confounding resulting from model misspecification (8, 10). Finally, proper implementation of the CSI accounts, at least in part, for some of the potential bias affecting studies on the effect of smoking (8, 11).

Little has been published on the effect of smoking on vascular symptoms in SSc. Studies using categorical variables of never, past, and current smoking status have had contradictory results, with some showing that smoking was associated with worse vascular outcomes and others not (12, 13). One study also modeled smoking using pack-years and showed that only those in the highest tertile of smoking intensity (>25 pack-years) had worse vascular outcomes (13). We recently showed, using the CSI, that smoking has a strong negative impact on vascular symptoms in SSc (14). However, in the course of our analysis, we realized that different approaches to modeling smoking history could have a major impact on the interpretation of the results. The purpose of the present study was to compare different methods of modeling smoking exposure and to demonstrate the importance of proper statistical modeling of smoking on vascular outcomes in rheumatic diseases.


Study design and subjects.

The study design was a cross-sectional study of a national cohort of SSc patients. The study subjects consisted of those enrolled in the Canadian Scleroderma Research Group Registry. Patients in this registry are recruited from 15 centers across Canada. They must have a diagnosis of SSc made by the referring rheumatologist, be age ≥18 years, and be fluent in either English or French. Patients are seen yearly. The patients included in this study were those whose baseline visit was between August 2004 and November 2009.

Smoking status.

Patients recruited into the registry undergo an extensive medical evaluation with standardized reporting of history, physical examination, and laboratory investigations. In their self-reported case report forms, they were asked about their smoking history. They self-reported on current, past, or never cigarette smoking status. For those who reported current or past smoking, they were further asked the average number of packs of cigarettes smoked per day (smoking “intensity”) and the number of years they smoked (smoking “duration”). Past smokers were asked about years since quitting (“time since cessation”).

SSc clinical outcome variables.

We chose 3 important vascular outcome variables in SSc, specifically severity of Raynaud's phenomenon, presence of finger ulcers, and severity of finger ulcers, to demonstrate the effects of various approaches to modeling smoking history. Severity of Raynaud's phenomenon and severity of finger ulcers were reported by the patients using numerical rating scales ranging from 0 (symptom does not limit activities) to 10 (severe limitation). The recruiting physician reported the presence or absence of vascular finger ulcers, defined as follows. In the vascular section of the history, physicians recorded, yes or no, if the patient had digital ulcers in the past that had healed. In the physical examination section, physicians recorded, yes or no, the presence of digital tip ulcerations on the volar aspect of the fingers distal to the proximal interphalangeal joints. Ulcers included active ulcers (denuded area with defined border and loss of epithelialization and loss of epidermis and dermis; excludes fissures, paronychia, extrusion of calcium), healed ulcers (epithelialization of an ischemic ulcer, regardless of residual pain), digital necrosis/gangrene, loss of digital pulp, and/or auto- or surgical amputation.

Statistical analysis.

Descriptive statistics were used to summarize the baseline demographic characteristics, smoking characteristics, and clinical outcomes of the patients, grouped according to smoking status (never, past, or current smokers).

Five definitions of smoking exposure were analyzed in the study: model 1 modeled smoking as a single factor, comparing ever to never smoking; model 2 used two indicator variables comparing current to never smokers and past to never smokers; model 3 modeled smoking using polynomial contrasts, treating never, past, and current smoking as ordered values of a single variable; model 4 used a more complex interaction model of smoking identified from the literature (8) with smoking represented by 4 covariates, specifically never smoked, pack-years smoked, time since cessation, and an interaction term between pack-years and time since cessation; and model 5 represented smoking using the CSI (8, 9).

The CSI integrates the component variables of smoking intensity, duration, and time since cessation into a single covariate of smoking effect. Therefore, the large variation in the smoking exposure in both current and past smokers is taken into account. The CSI has the following functional form (as described by Leffondre et al [8] and Hoffmann and Bergmann [11]): CSI = (1 − 0.5dur*/τ)(0.5tsc*/τ)ln(int + 1), with tsc* = max(tsc − δ, 0) and dur* = max(dur + tsc − δ, 0) − tsc*, where dur = duration, tsc = time since cessation, int = intensity, tsc* = lag-adjusted cessation time, and dur* = lag-adjusted duration. Implementation of the CSI involves the estimation of two parameters, a half-life (tau [τ]) and lag time (delta [δ]), via maximization of model fit. Tau can be interpreted as the rate at which the health impact of smoking decays over time (i.e., rate = ∼1/τ), while delta can be interpreted as the length of delay between the occurrence of smoking and the manifestation of its impact on health. In particular, the lag time parameter functions to model the short-term causality bias where changes in smoking behavior (i.e., cessation or initiation) can occur in direct response to changes in disease symptoms (8). Tau and delta values are estimated for each health outcome; therefore, each health outcome has a unique CSI with a particular functional form, which reflects the mechanism by which smoking affects that outcome. Regression analysis then proceeds using the CSI as a single covariate, whose regression coefficient (b − CSI) reflects the association of smoking with the response variable tested. In-depth discussion of the estimation of tau and delta and of the mathematical form of the CSI (where it corresponds to the variable X2) has been published elsewhere (8, 9, 14).

All regression models were adjusted for age, sex, ethnicity, disease duration (from the onset of the first non–Raynaud's phenomenon disease manifestation), and limited or diffuse skin involvement, based on the definition by LeRoy et al (15). Logistic regression was used to model the binary outcome of the presence/absence of finger ulcers, while the patient assessment of severity ranging from 0–10 was power transformed to ensure that residuals were homogeneous and normally distributed. Box-Cox power transform exponents were 0.0 for the severity of Raynaud's phenomenon and −0.9 for the severity of finger ulcers; a default shift value of 1 was used. In the CSI regression models, an indicator variable for never smoked was also included to ensure a smooth transition in model fit between low exposure smokers and never smokers (2).

The models were compared based on the following criteria: first, model specificity (i.e., the ability of a model to detect an effect of smoking) measured by the magnitude of the regression coefficient P value(s) and type 3 P value (i.e., the P value representing the significance of the aggregate set of smoking covariates in the model). Second, model goodness-of-fit, measured by the Akaike's information criterion (AIC). A lower AIC value represents a better-fitting model, with a difference of >3 points showing some evidence of better fit and a difference of >10 points indicating a significantly better fit (16). Third, the degree of multicolinearity, assessed using the variance inflation factors (VIFs) of the smoking covariates. Fourth, the ability to reduce the potential for model misspecification. This was assessed by way of a small simulation study designed to examine the performance of the models under two different scenarios and measured by the average mean square error (MSE) (17, 18). Using the smoking characteristics of the patients in this study, we generated simulated data first under the assumption that model 5 was the true model (scenario 1) and then assuming that model 4 was the true model (scenario 2). We then compared the average MSE of the models in estimating the true response means across the subjects for 100 different simulated data sets under each scenario. For purposes of brevity, only the results for models 2, 4, and 5 are shown. Details of the simulation study are provided in Supplementary Appendix A (available in the online version of this article at The fifth and final criterion is overall interpretability, represented by how easily and coherently the regression results could be interpreted in a clinical and biologic sense.

Of the patients with some data in the database at the time of this study, 606 had complete data and were included in this analysis. Comparison of the study population and patients excluded due to missing data (n = 434) showed no meaningful differences. All of the statistical analyses for real and simulated data were performed with R, version 2.10.0 (19).

Ethical considerations.

Ethics committee approval for the data collection protocol was obtained at each site and each patient provided informed written consent to participate in this study.


This study included 606 SSc patients, of which 87% were women, 90% were white, the mean ± SD age was 55 ± 12 years, the mean ± SD disease duration was 11 ± 9 years, and 36% had diffuse disease (Table 1). Of these, 16% were current, 42% were past, and 42% were never smokers.

Table 1. Baseline characteristics of the entire cohort and according to smoking status
 Entire cohort (n = 606)Never smoker (n = 255)Past smoker (n = 252)Current smoker (n = 99)
Demographics and disease characteristics    
 Age, mean ± SD years55.25 ± 12.3456.18 ± 13.2255.84 ± 12.2151.36 ± 9.27
 Women, no. (%)527 (87.0)240 (94.1)209 (82.9)78 (78.8)
 White, no. (%)543 (89.6)222 (87.1)229 (90.9)92 (92.9)
 Disease duration, mean ± SD years10.94 ± 9.4511.68 ± 10.3111.21 ± 9.328.35 ± 6.74
 Diffuse disease, no. (%)218 (36.0)88 (34.5)97 (38.5)33 (33.3)
Vascular symptoms    
 Severity of Raynaud's phenomenon (range 0–10), mean ± SD2.83 ± 2.932.68 ± 2.792.62 ± 2.883.76 ± 3.26
 Presence of finger ulcers, necrosis, or amputation, no. (%)387 (64.1)161 (63.6)154 (61.1)72 (72.7)
 Severity of finger ulcers (range 0–10), mean ± SD1.93 ± 2.951.79 ± 2.751.75 ± 2.852.77 ± 3.54

There was great variability in terms of smoking exposure within the subgroups of current and past smokers (Figure 1). The mean ± SD amount of cigarettes smoked was 25 ± 17 pack-years for current smokers and 17 ± 18 pack-years for past smokers, and more past smokers were heavy smokers (≥2 packs/day) than current smokers. Smoking duration varied enormously, from 1 to 60 years, and was less for past smokers than current smokers (18.3 versus 31.7 years). Most of the current smokers were long-time smokers, while many past smokers smoked only for a short period. Time since cessation is an attribute of past smokers only and is shown in Figure 1E. On average, past smokers stopped smoking approximately mean ± SD 16 ± 12 years prior to their baseline registry visit, although this varied widely, from 1 to 50 years. The wide ranges in smoking intensity, duration, and time since cessation underscore the fact that categorizing smokers as never, past, or current is an extremely crude approximation of smoking status.

Figure 1.

Distribution of smoking component variables (intensity, duration [dur], and time since cessation [tsc]) for past and current smokers. A, distribution of smoking intensity for past smokers, B, distribution of smoking intensity for current smokers, C, distribution of smoking duration for past smokers, D, distribution of smoking duration for current smokers, and E, distribution of smoking cessation time for past smokers.

Results and comparison of the regression models.

Overall, the results of the regression models examining the effect of various ways of modeling the effect of smoking on vascular outcomes in SSc were consistent with each other (Table 2). However, it can also be seen that the more refined models captured the effect of smoking more strongly and provided greater insight into the effect of smoking.

Table 2. Summary of the regression results of the 5 different models of smoking*
 BPVIFType 3 Pτ, yearsδ, yearsAIC
  • *

    VIF = variance inflation factor; AIC = Akaike's information criterion; PY = pack-years; TSC = time since cessation; CSI = Comprehensive Smoking Index.

  • Estimated regression coefficient for ever vs. never smokers.

  • Estimated regression coefficient for current vs. never smokers.

  • §

    Estimated regression coefficient for past vs. never smokers.

  • Estimated regression coefficient for the linear contrast covariate.

  • #

    Estimated regression coefficient for the quadratic contrast covariate.

  • **

    Estimated regression coefficient for the PY covariate.

  • ††

    Estimated regression coefficient for the TSC covariate.

  • ‡‡

    Estimated regression coefficient for the PY × TSC interaction term.

  • §§

    Estimated regression coefficient for the CSI.

Model 1: ever vs. never smoking       
 Severity of Raynaud's phenomenon (range 0–10)0.0070.921.046   1,503.2
 Presence of finger ulcers−0.020.921.041   741.5
 Severity of finger ulcers (range 0–10)−0.010.691.046   612.8
Model 2: current vs. never, past vs. never       
 Severity of Raynaud's phenomenon (range 0–10)      1,498.5
 Presence of finger ulcers      738.9
 Severity of finger ulcers (range 0–10)      612.0
Model 3: polynomial contrasts       
 Severity of Raynaud's phenomenon (range 0–10)      1,498.5
 Presence of finger ulcers      738.9
 Severity of finger ulcers (range 0–10)      612.0
Model 4: never + PY + TSC + PY × TSC       
 Severity of Raynaud's phenomenon (range 0–10)      1,503.2
  PY × TSC−0.0006‡‡0.041.5610.07   
 Presence of finger ulcers      736.9
  PY × TSC−0.00050.471.4720.03   
 Severity of finger ulcers (range 0–10)      612.4
  PY × TSC0.00010.321.5610.30   
Model 5: CSI       
 Severity of Raynaud's phenomenon (range 0–10)0.52§§0.0021.032 10.21,498.3
 Presence of finger ulcers1.200.0021.045 19733.1
 Severity of finger ulcers (range 0–10) 117615.0

Model specificity.

The P values for each of the covariates shown in Table 2 represent the specificity of the model to detect an effect of smoking. An overall increase in specificity (or decrease in P values) for the effect of smoking was seen as model refinement increased from models 1 to 5, with the CSI having the smallest P values among all of the models. While some individual smoking covariates were statistically significant in model 4, the type 3 P values indicate that the overall effect of smoking in that model was not necessarily so. The unexpected protective effect of past smoking in models 2 and 3 and the strong effect of time since cessation in reducing the impact of smoking in model 4 should be noted. Note also that none of the smoking models detected a statistically significant effect of smoking on the severity of finger ulcers outcome.

Model goodness-of-fit.

AIC values comparing the goodness-of-fit of the smoking models are shown in Table 2. The CSI model showed consistently good model fit with the lowest AIC values for both the severity of Raynaud's phenomenon and presence of finger ulcers outcomes, although for the former outcome there were also 2 other models within 3 AIC units of the CSI. For the severity of finger ulcers outcome, the CSI model had the highest AIC. However, all AIC values from all models were within 3 units of each other, and therefore were approximately equivalent.


VIFs were computed for all smoking model covariates and are shown in Table 2. VIFs for the CSI model were exceedingly low, ranging from ∼1.01–1.04, and comparable to those of model 1 (ever versus never) and model 3 (polynomial contrasts). VIFs for the linear combination of pack-years and interaction model (model 4) were markedly higher for all terms, ranging from ∼1.48–1.73, indicating that multicolinearity was acting to reduce statistical power.

Accuracy in estimating the effect of smoking.

The MSEs shown in Table 3 measured the accuracy with which a model was able to predict the effect of smoking on simulated data. Scenario 1 shows the results when models 2, 4, and 5 were used to predict data simulated under the assumption that the CSI model was the truth. The MSE when using model 4 was 3.1 times higher than when the correct model 5 was used (1.15 divided by 0.37 = 3.1). Scenario 2 shows the results when models 2, 4, and 5 were fit to simulated data generated by the interaction model (model 4). In this case, the CSI performed reasonably, as its MSE was only 1.9 times higher than the correct model. Therefore, the relative increase in error in incorrectly assuming the CSI model as the truth was substantially lower than when incorrectly assuming the interaction model as the truth. Further details of the simulation study are provided in Supplementary Appendix A (available in the online version of this article at

Table 3. Mean square error measures for models 2, 4, and 5 using simulated data generated by the CSI model (scenario 1) and the interaction model (scenario 2)*
Smoking modelNo.Scenario 1, mean square error (× 100)Scenario 2, mean square error (× 100)
  • *

    CSI = Comprehensive Smoking Index; PY = pack-years; TSC = time since cessation.

  • CSI (τ = 2, δ = 1).

  • Model 4.

Current, past, never21.381.96
PY + TSC + PY × TSC41.150.30


Although models 1 to 3 appeared simple, the nonsignificant effect of ever smoking and the apparent protective effect of past smoking were difficult to interpret. Presumably, the heterogeneity within the smoking categories was not properly captured by these models. Model 4 showed a significant effect in some of the individual smoking covariates, most notably time since cessation. However, the type 3 P values that assessed the overall significance of smoking remained nonsignificant for 2 of the 3 outcomes. Moreover, the protective effect of pack-years for severity of Raynaud's phenomenon and the presence of finger ulcers in model 4 was also difficult to interpret. The CSI model was the most easily interpretable and coherent model insofar as identifying the negative effect of smoking on the various vascular outcomes of interest. In addition, the estimated half-life and lag time parameters provided considerable insight into the effect of smoking on these outcomes, as explained below.

Interpretation of the CSI model.

Figure 2 shows the distribution of the smoking effect predicted by the CSI model for one of the vascular outcomes, specifically the presence of finger ulcers. For this outcome, the half-life (τ) was 1 year and the lag time (δ) was 9 years. Figures 2A and B show the effect of smoking in current smokers. Increasing smoking intensity increases the odds ratio for the presence of finger ulcers. Duration has little effect because of the low half-life; there is little cumulative effect of smoking over time. Figures 2D, 2E, and 2F show the effect of smoking in past smokers. Figure 2A indicates that there are ∼2 subgroups of past smokers, those with more recent cessation who, like current smokers, experience an effect of smoking that increases with intensity, and those with longer cessation times who exhibit virtually no effect of smoking. Duration also has little effect in past smokers for the same reasons as above. The effect of cessation time is shown more clearly in Figure 2F: here it can be seen that patients with cessation times less than ∼9 years (i.e., the lag time) have increased symptom severity due to smoking, whereas those with higher cessation times have no effect. Finally, note that as smoking exposure diminishes (intensity and duration approaching 0 or time since cessation approaching infinity) the odds ratio drops below 1, indicating the healthy smoker bias that is detected. All of the vascular variables in our study were found to have a low half life (τ = ∼1 year) and thus exhibited similar behavior to the example described.

Figure 2.

Effect of smoking predicted by the Comprehensive Smoking Index model on the presence of finger ulcers, expressed as the change in odds ratio (OR). In these plots, each individual patient is represented by a data point, showing the heterogeneity within current and past smoking subpopulations. The effect of smoking is plotted versus intensity, duration, and time since cessation for both current (A, B, and C) and past (D, E, and F) smokers in order to show trends with these measures of smoking exposure.


This study highlights the importance of proper modeling of smoking in SSc. Of great importance in this study, past smokers were highly different from current smokers and in fact slightly healthier on average than never smokers, presumably due to the range of smoking exposure among past and current smokers and to the presence of a healthy smoker bias. Therefore, model 1, which aggregated past and current smokers, incorrectly failed to detect any effect at all of smoking.

Models 2 and 3, which distinguished between never, past, and current smoking groups, confirmed a negative effect of smoking in current smokers, but failed to detect an effect in past smokers. Again, these models in fact supported the presence of a healthy smoker bias, given the U-shaped trend in the effect of smoking, with past smokers appearing healthier than both never and current smokers. This effect achieved statistical significance in the model using never, past, or current smoking as ordered values of a single variable (model 3), possibly due to the assumption of ordering.

More complex models that accounted for smoking intensity, duration, and cessation times were the most sensitive for detecting the effect of smoking. In particular, the CSI (model 5) estimated a highly nonlinear effect of smoking cessation time on outcomes, with past smokers approximately divided into recent quitters who exhibited an effect of smoking and those with longer cessation times who showed virtually no effect. The CSI model therefore accounted for the heterogeneity in smoking effect within past smokers and clarified the reason why past smokers, when taken together, showed no effect. The findings of model 4, which showed that time since cessation was statistically significant and acted to diminish the effect of smoking, were consistent with this. Although more complex to implement, results obtained using the CSI model were, however, easier to interpret than those for model 4. In addition, the CSI exhibited numerous other advantageous properties, including increased specificity, reduced amounts of multicolinearity, similar or superior goodness-of-fit, and reduced potential for model misspecification.

It is interesting to note that smoking had a strong negative effect on 2 of the 3 outcomes selected, specifically severity of Raynaud's phenomenon and the presence of vascular ulcers as assessed by physicians, and was weakest for the third patient-reported severity of finger ulcers outcome. The first two outcomes are clearly vascular outcomes and the negative effect of smoking is highly plausible. Poor specificity of the third outcome as a measure of vascular disease may explain why smoking did not have a significant effect on it. Indeed, in addition to vascular abnormalities, finger ulcers in SSc also occur from severe skin tethering, trauma, and calcinosis. Smoking is less likely to have an effect on these additional mechanisms of ulceration. Yet patients, in reporting the severity of their finger ulcers, are unlikely to have distinguished vascular from other mechanisms of ulceration. Therefore, the importance of additional nonvascular factors in this variable may explain why the effect of smoking was weaker on this third patient-reported outcome.

The “healthy smoker effect” has been described to account for the observation that some patients with higher smoking intensity have been shown to have better lung function (7), thereby suggesting that those who smoke self-select in part because they have lungs that are relatively resistant to the effects of smoking. This (self)-selection bias can cause past smokers to appear healthier than never smokers, since past smokers are intrinsically healthier to begin with (due to the healthy smoker effect). Moreover, past smokers in this study population were found to have a shorter smoking duration than current smokers (18.3 versus 31.7 years) and cessation times as long as 50 years. Therefore, it is possible that past smokers were also healthier compared to current smokers.

The implementation of the CSI, which includes a never smoking factor, accounts for the healthy smoker effect. As smoking exposure decreases, the CSI also decreases. Therefore, the effect of smoking in past or current smokers with extremely low smoking duration or intensity or with extremely long cessation times also approaches zero. Since never smokers are assigned a CSI of zero, the never smoking term effectively provides an adjustment for any intrinsic differences between never and ever (i.e., past and current) smokers.

The advantageous characteristics of the CSI outlined above as well as its capability to model inherent biases in studies of smoking provide a strong case for implementing the CSI, when possible. In the absence of sufficient time or statistical expertise, simpler models of smoking could be used with the following caveats: there may be marked heterogeneity in smoking effect within past and current smoking categories; dependence on intensity, duration, and time since cessation may be highly nonlinear such that linear models likely have reduced power due to imprecise fitting of the data; and various biases inherent to smoking status may be present and unaccounted for.

The fact that this study was a cross-sectional analysis of data in one rheumatic disease, specifically SSc, may be viewed as a limitation. It is difficult to know if the advantages found for the CSI in these data would necessarily be found in analyses of other data sets and diseases. This remains to be demonstrated.

The findings of this study need to be viewed in the larger context of the considerable importance of vascular outcomes in rheumatic diseases. Smoking is a well-known contributor of vascular disease and vascular disease is ubiquitous in rheumatic diseases. We believe that the issues that we have identified (heterogeneity, nonlinearity, and bias) must be accounted for in the proper characterization of this important confounder of vascular outcomes in rheumatic diseases.


All authors were involved in drafting the article or revising it critically for important intellectual content, and all authors approved the final version to be published. Dr. Hudson had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Study conception and design. Hudson, Lo, Steele.

Acquisition of data. Hudson, Baron.

Analysis and interpretation of data. Hudson, Lo, Baron, Steele.


Actelion Pharmaceuticals and Pfizer, Inc., had no role in the design of the study, analysis of the data, preparation of the manuscript, approval of the study for publication, and decision to submit for publication.



Investigators of the Canadian Scleroderma Research Group are as follows: M. Baron, J.-P. Mathieu, M. Hudson, S. Ligier, T. Grodzicky: Montreal, Quebec; J. Pope: London, Ontario; J. Markland: Saskatoon, Saskatchewan; D. Robinson, S. Mittoo: Winnipeg, Manitoba; N. Jones: Edmonton, Alberta; N. Khalidi, E. Kaminska: Hamilton, Ontario; P. Docherty: Moncton, New Brunswick; A. Masetto: Sherbrooke, Quebec; D. Smith: Ottawa, Ontario; E. Sutton: Halifax, Nova Scotia; M. Fritzler: Advanced Diagnostics Laboratory, Calgary, Alberta, Canada.