We are writing this letter in response to the recently published article by Barry et al. (2018). Our review of their publication has identified shortcomings associated with their reanalysis of the data from Health Canada's Community Noise and Health Study (CNHS), which challenge their overall conclusion that proximity to wind turbines was associated with quality of life in the CNHS.

The primary objective of the CNHS was to assess self-reported and objectively measured outcomes associated with exposure to wind turbine noise (WTN). In March 2016, the results from Health Canada's CNHS were published in the Journal of the Acoustical Society of America as a special section on WTN. These six papers followed two previous CNHS publications related to quality of life (Feder et al., 2015) and sleep (Michaud et al., 2016a). In the CNHS outdoor A-weighted wind turbine sound pressure levels at dwellings and distances between dwelling and the nearest wind turbine were highly correlated (Spearman ρ = −0.90, p < 0.0001). Therefore, no additional information was gained by presenting both exposure metrics in each paper insofar as there were no differences in the results whether analysed by either distance or WTN.

In support of open science, data originating from the CNHS are available to interested parties through the Statistics Canada Research Data Centres (Statistics Canada, 2018). Recently, Barry et al. (2018) published a reanalysis of the response to wind turbine exposure using the CNHS dataset. Their reanalysis focused on distance to the nearest wind turbine as the exposure metric of interest based on their suggestion that WTN may serve as a surrogate for other factors that are more closely associated with annoyance. This approach has merit and is consistent with our observation that in the CNHS the prevalence of annoyance toward WTN ranked below annoyance toward other wind turbine features, even in areas where WTN levels were highest (Michaud et al., 2016b). Although we noted limitations evaluating the response to wind turbines based upon setback distance, we also used distance to wind turbines as an exposure metric in the derivation of an exposure–response relationship for aggregate annoyance toward multiple wind turbine features (Michaud et al., 2018). Our concern with the recent publication by Barry et al. relates not to their use of distance, but to their analytical methods.

Barry et al. noted in the methods section of their paper that distance to the nearest wind turbine was the primary predictor to build each model. They reported that distance was used as a continuous variable and logarithmically transformed to normalize the distribution. However, the assumption for any linear regression model is not that the continuous explanatory variables need to follow a normal distribution, but that the residuals of the linear regression model need to be normally distributed. Therefore, if a transformation is required to fulfill this assumption, it is the response variable and not the distance variable that should be transformed. Indeed, linear regression is robust to violation of the normality assumption (Schmidt and Finan, 2018). For logistic regression, this assumption does not apply and there is no reason to logarithmically transform the independent variables in any logistic regression model. As discussed below, when the independent variable is transformed there is no longer a straightforward way to interpret the observed association with the dependent variable.

Quality of life was assessed in the CNHS with the World Health Organization Quality of Life-Bref questionnaire. This questionnaire is used to derive scores for environment, social, psychological and physical quality of life domains, whereby higher scores suggest better quality of life on any particular domain. Barry et al. stated in their results section that proximity to wind turbines was inversely associated with scores on the environment domain (β = −1.23, SE = 0.145, p = 0.046). The authors interpreted this to mean that for every kilometre a person lives from a wind turbine there was an associated 1.23 increase in the score on the environment domain. As distance was on a logarithmic scale in their regression model, the increase in the score would depend on the value of distance. In other words, the relationship between distance and score would not be linear, which means that the increase in score for a one unit change in distance from 1 to 2 km would be different from the increase in score for a one unit change in distance from 2 to 3 km, or from 3 to 4 km, etc. There is no simple interpretation of a regression model where the independent variable is logarithmically transformed other than in terms of percent change: a one percent change in the independent variable (distance in this case) was associated with a β*ln(1.01) change in the dependent variable (quality of life domain score), where β is the coefficient estimate for the log-distance variable. Moreover, the beta coefficient values reported by Barry et al. for the environment domain (1.23) and physical domain (1.26) are unattainable given the observed variability in the scores on either domain. According to Table I in Barry et al., the scores in the environment domain varied by at most 1 unit, even for variables that were much more significant than log-distance (e.g., income or age group). When the independent variable is logarithmically transformed, taking the exponential of the coefficient for this variable does not back transform the independent variable to its original scale.

TABLE I.

Simple linear regression models for the World Health Organization Quality of Life-Bref Domains.

ModelDomainanbbetaSEp-value
distance Physical 1184 −0.070 0.052 0.174 
Psychological 1184 −0.011 0.042 0.798 
Social 1181 −0.055 0.048 0.254 
Environment 1184 0.038 0.038 0.311 
ln(distance) Physical 1184 −0.295 0.139 0.035 
Psychological 1184 −0.116 0.112 0.300 
Social 1181 −0.212 0.131 0.105 
Environment 1184 0.110 0.102 0.280 
distance province benefit Physical 1184 −0.045 0.052 0.379 
Psychological 1184 0.001 0.042 0.982 
Social 1181 −0.047 0.049 0.338 
 Environment 1184 0.049 0.038 0.197 
ln(distance) Physical 1184 −0.179 0.140 0.203 
province Psychological 1184 −0.078 0.113 0.489 
benefit Social 1181 −0.180 0.133 0.177 
Environment 1184 0.150 0.103 0.144 
ModelDomainanbbetaSEp-value
distance Physical 1184 −0.070 0.052 0.174 
Psychological 1184 −0.011 0.042 0.798 
Social 1181 −0.055 0.048 0.254 
Environment 1184 0.038 0.038 0.311 
ln(distance) Physical 1184 −0.295 0.139 0.035 
Psychological 1184 −0.116 0.112 0.300 
Social 1181 −0.212 0.131 0.105 
Environment 1184 0.110 0.102 0.280 
distance province benefit Physical 1184 −0.045 0.052 0.379 
Psychological 1184 0.001 0.042 0.982 
Social 1181 −0.047 0.049 0.338 
 Environment 1184 0.049 0.038 0.197 
ln(distance) Physical 1184 −0.179 0.140 0.203 
province Psychological 1184 −0.078 0.113 0.489 
benefit Social 1181 −0.180 0.133 0.177 
Environment 1184 0.150 0.103 0.144 
a

Barry et al. provided insufficient information on their final regression models for the environment and physical domains to enable replication.

b

Sample size for the four World Health Organization Quality of Life-Bref domains differs from the study sample size of 1238 due to missing values: personal benefit (n = 53), social (n = 4), physical, psychological, and environment (n = 1).

Similarly, the interpretation of the odds ratio (OR) from the logistic regression models of Table II in their results section was misleading. An OR of 0.19 was interpreted as follows: “The odds of reporting being annoyed by a turbine are reduced by about 20% for every kilometer a person lives further away from a wind turbine.” As mentioned above for linear regression, logarithmically transforming distance in logistic regression means the relationship between distance and annoyance is non-linear and it is misleading to interpret the OR per unit change. It would be more precise to describe this association as follows: for each doubling in distance, the odds of reporting to be highly annoyed by WTN changed by a factor of 0.32. Again, there is no reason to logarithmically transform distance in any logistic regression model.

Barry et al. set out to reanalyse the CNHS data using distance to the nearest wind turbine as an exposure metric, where distance was not grouped into five exposure categories because the variability within each category is unaccounted for by grouping. This alternative approach was reasonable. Our primary concern is not with their rationale for the reanalysis, but with the transformation of the exposure parameter and the authors' interpretation of their results.

We have analysed each of the four domains from the World Health Organization Quality of Life-Bref questionnaire without grouping distance into categories. It was not possible to replicate the final model for the environment domain presented by Barry et al., which included the list of variables in the final model, but did not identify the covariates considered in the stepwise process. For the physical domain the authors only presented the estimate for the association with log-distance (β = 1.26, SE = 0.20, p = 0.043) without identifying the variables in the final model, or the covariates considered in the stepwise analysis. In Table I we provide distance and log-distance (only for comparative purposes) in a simple linear regression model with and without personal benefit and province forced into each model. The results in Table I are consistent with our previous conclusion that there was no observed association regarding exposure to WTN and the four quality of life domains (Feder et al., 2015). It is worth highlighting that regardless of the model used, the slope associated with the physical domain was negative, meaning that scores on this domain actually appeared to decrease as distance between dwelling and wind turbine increased, which is in contrast to a positive slope of 1.26 reported by Barry et al. Furthermore, when the analysis was conducted without logarithmically transforming the distance variable, there was no apparent statistical association between distance to wind turbines and any of the four quality of life domains. Based on these results, we maintain that there was insufficient evidence in the data collected in the CNHS to suggest that proximity to wind turbines was associated with health-related quality of life measures, as suggested by Barry et al.

1.
Barry
,
R.
,
Sulsky
,
S. I.
, and
Kreiger
,
N.
(
2018
). “
Using residential proximity to wind turbines as an alternative exposure measure to investigate the association between wind turbines and human health
,”
J. Acoust. Soc. Am.
143
(
6
),
3278
3282
.
2.
Feder
,
K.
,
Michaud
,
D. S.
,
Keith
,
S. E.
,
Voicescu
,
S. A.
,
Marro
,
L.
,
Than
,
J.
,
Guay
,
M.
,
Denning
,
A.
,
Bower
,
T. J.
,
Lavigne
,
E.
,
Whelan
,
C.
, and
van den Berg
,
F.
(
2015
). “
An assessment of quality of life using the WHOQOL-BREF among participants living in the vicinity of wind turbines
,”
Environ. Res.
142
,
227
238
.
3.
Michaud
,
D. S.
,
Feder
,
K.
,
Keith
,
S. E.
,
Voicescu
,
S. A.
,
Marro
,
L.
,
Than
,
J.
,
Guay
,
M.
,
Denning
,
A.
,
McGuire
,
D.
,
Bower
,
T.
,
Lavigne
,
E.
,
Murray
,
B. J.
,
Weiss
,
S. K.
, and
van den Berg
,
F.
(
2016b
). “
Exposure to wind turbine noise: Perceptual responses and reported health effects
,”
J. Acoust. Soc. Am.
139
(
3
),
1443
1454
.
4.
Michaud
,
D. S.
,
Feder
,
K.
,
Keith
,
S. E.
,
Voicescu
,
S. A.
,
Marro
,
L.
,
Than
,
J.
,
Guay
,
M.
,
Denning
,
A.
,
Murray
,
B. J.
,
Weiss
,
S. K.
,
Villeneuve
,
P. J.
,
van den Berg
,
F.
, and
Bower
,
T.
(
2016a
). “
Effects of wind turbine noise on self-reported and objective measures of sleep
,”
SLEEP
39
(
1
),
97
109
.
5.
Michaud
,
D. S.
,
Marro
,
L.
, and
McNamee
,
J.
(
2018
). “
Derivation and application of a composite annoyance reaction construct based on multiple wind turbine features
,”
Can. J. Public Health
109
(
2
),
242
251
.
6.
Schmidt
,
A. F.
, and
Finan
,
C.
(
2018
). “
Linear regression and the normality assumption
,”
J. Clin. Epidemiol.
98
,
146
151
.
7.
Statistics Canada Research Data Centre reference
(
2018
). “
The research data centres (RDC) program
,” http://www.statcan.gc.ca/eng/rdc/index (Last viewed 11 June 2018).