Title: Structural equation modeling of parasympathetic and sympathetic response to traffic air pollution in a repeated measures study

Transcription

1 Author's response to reviews Title: Structural equation modeling of parasympathetic and sympathetic response to traffic air pollution in a repeated measures study Authors: Emmanuel S Baja (ebaja@hsph.harvard.edu) Joel D Schwartz (joel@hsph.harvard.edu) Brent A Coull (bcoull@hsph.harvard.edu) Gregory A Wellenius (gregory_wellenius@brown.edu) Pantel S Vokonas (pantel.vokonas@va.gov) Helen H Suh (h.suh@neu.edu) Version: 2 Date: 14 July 2013 Author's response to reviews: see over

2 D E P A R T M E N T O F E N V I R O N M E N T A L H E A L T H H A R V A R D S C H O O L O F P U B L I C H E A L T H July 15, 2013 Prof. Philippe Grandjean Prof. David Ozonoff Editors- in- Chief Environmental Health Dear Dr. Grandjean and Dr. Ozonoff, We are pleased to submit to Environmental Health our revised and edited original article Structural equation modeling of parasympathetic and sympathetic response to traffic air pollution in a repeated measures study, which has been revised in response to the comments of the reviewer. In addition, attached to this cover letter is our response to reviewer comments. All the authors have carefully read the reviewer s comments and have approved and contributed to the editing and revision of the manuscript. In addition, we would like to reiterate that no author has an existing relationship with the industry. Furthermore, this manuscript is not under consideration elsewhere and none of the paper s contents have been previously published. If you have any questions regarding our response to the reviewer s comments and the revised submitted article please do not hesitate to me at ebaja@hsph.harvard.edu. We will be waiting and eagerly anticipating for the decision of the journal. Sincerely, Emmanuel S. Baja, Sc.D. Exposure, Epidemiology, and Risk Program Department of Environmental Health Harvard School of Public Health Suite 426, Landmark Center West 401 Park Drive Boston, MA Phone: (617) Fax: (617)

3 Response to Reviewer Comments We thank the reviewers for their careful review of our manuscript and their thoughtful comments. We provide a response to each comment below. Reviewer 1 The manuscript as presented is a mixture of numerous apparently conflicting results that the authors need to sort out better for the reader. The authors also need to note past work using a more common form of structural equation modeling (factor analysis), which has been applied in the past to relate multiple exposure variables with health variables. For example, the work of the EPA Workshop on the Source Apportionment of the Health Effects of PM (see Environ Health Perspect Dec; 113(12): ) which summarized the results of time series analyses of health effects with source factors. To help the reader understand the differences between the approach taken and conventional factor analyses others have applied in the past, the paper should apply a conventional Varimax Factor analysis to the exposure data and apply that in the analysis, and the similarities/difference between that and the SEM used here noted. Reply: We appreciate the comments and suggestions of the reviewer. We have incorporated the suggestion of the reviewer and have included in the discussion section past works (factor analysis) that relate multiple exposure variables with health variables. We, however, did not conduct additional analyses using factor analytic approaches, as (1) these approaches (as include in our revised discussion) have been used extensively in previous studies and (2) our analyses includes analysis of single pollutants that are often used as tracers of source-related impacts. We believe that doing so would add unnecessary complexity to an already complex manuscript and is beyond the scope of this project. As to the interpretation of the seemingly mixed results between SEM Traffic factor vs. BC, it seems likely that they can be interpreted as the difference between traffic in general (including NOx from gas combustion vehicles, with little BC emitted) and diesel PM (most closely associated with the BC variable). For example, past published analyses of urban elemental carbon samples, though not from Boston, has found that over 90% was due to diesel emissions in particular (J Expo Sci Environ Epidemiol Jul;20(5): ). Thus, the authors should consider the possible explanation for the seeming inconsistency between the SEM General Traffic health relationship with cardiac metrics, vs. for BC in particular, as indicating the higher toxicity of diesel traffic emissions vs. gasoline powered vehicle emissions. Reply: We have revised our paper accordingly.

4 Reviewer 2 Abstract, results: For the third sentence I'd suggest "The strongest association was observed for the 4-hour moving average" or "The strongest association observed was for the 4-hour moving average." Reply: We have incorporated the suggestion. Also change "linear mixed models" to "Bayesian linear mixed models" in the second sentence. Reply: We have changed linear mixed models to Bayesian linear mixed models. The last sentence in this section is pretty awkward. Reply: We have revised the last sentence to The effect of traffic on parasympathetic tone was stronger among diabetic as compared to non-diabetic participants. Specifically, an IQR increase in traffic pollution in the 48-hr prior to the clinic visit was associated with a 44.3% (95% PI: -67.7%, -4.2%) lower mean parasympathetic tone among diabetics and a 7.7% (95% PI: -18.0%, 41.4%) higher mean parasympathetic tone among non-diabetics. Abstract, conclusion: Try "BC is adversely associated with LF/HF" or other wording to avoid implying causation. Reply: We have revised the first sentence to BC was associated with adverse changes LF/HF in the elderly. The last sentence should be deleted or rewritten to say something meaningful. Are the SEMs more useful here than typical LMMs here? Why? Reply: We have deleted the last sentence. Background, I would remove the mention of LMMs from the last sentence. The comparison of SEMs with LMMs is not really part of your analysis so its mention here distracts the reader from the key idea of your paper. Or if this really is your purpose, your tables should be revised to include LMM results and directly address these comparisons. Reply: We have modified this sentence as suggested.

5 Background, "Notably, in models with 3 or more surrogates of the latent exposure, SEMs reduce the attenuation due to measurement error which is an ongoing concern in air pollution studies [9]." Please provide a page number for this reference. There are likely some additional model conditions required for reduced attenuation for example, are the measurement error distributions assumed to be independent across surrogates and across time? If they are 100% correlated across surrogates, for example, why would SEMs provide any benefit? It would be helpful to provide more detail on this claim, and to assess whether or not any conditions necessary for reduction of attenuation are likely to hold for air pollution studies such as yours. Reply: We have modified the last sentence of the paragraph to provide additional details about measurement error. The sentence now read This is a typical phenomenon in measurement error corrections, it tends to de-attenuate point estimates coupled with wider confidence intervals due to reflecting additional uncertainty as a result of measurement error correction. The hope is that the total mean squared error (bias squared + variance) is lower than uncorrected estimate. Study Population, could fasting or recent abstention from smoking have effects on HRV? If so, is cardiac rhythm measured at the same time of day for each visit? Reply: Fasting or recent abstention from smoking has been shown to have effects on HRV. For example, eating increases parasympathetic tone, especially after big meals. Smoking has similar effects. As a result, participants were required to fast and not smoke prior as part of the study protocol of the Normative Aging Study. Cardiac rhythm is measured at the same time of day for each visit. Outcome Data, for how long was cardiac rhythm measured at each visit? Reply: A sentence was added in the Outcome Data section the methods section of the paper, The cardiac rhythm was measured for approximately 7 minutes. Exposure and Weather Data, you indicate multiple air pollutant monitoring stations but it's unclear how these are used to assign individual observed moving average pollutant exposures X_ijt. Exposure assignment is critical for study reliability and should be described clearly. Are participant's home/work addresses used at all, or is the pollutant exposure at any given time just a simple average across all monitoring sites? Reply: The pollutant exposure was averaged across all the monitoring sites. We did not use the participant s home and work addresses in the determination of exposure. This fact has been clarified in the Methods section of the revised manuscript. Statistical Analysis, 2nd paragraph, you probably want "SEMS in a Bayesian framework" instead of "SEMs in Bayesian framework." Reply: We have revised the sentence accordingly.

6 Equation 4, why would human discomfort (ATemp) have a causal effect on traffic exposure? Do both AT and DPT have causal effects on traffic exposure? If so, is the constructed ATemp variable the best way to model those effects? Reply: Ambient and dew point temperature have causal effects on traffic exposure. Since ATemp is a linear combination of AT and DPT, we believe that the constructed ATemp variable is an appropriate way to model these effects. ATemp has been used in other studies for similar reasons (Ren C et al. Am J Epidemiol May 1;173(9): Luttmann-Gibson H et al. Occup Environ Med Sep;67(9):625-30). Statistical Analysis, "The relationships between latent outcome variables, and their observed markers of measures were also modeled." What are "observed markers of measures?" Is a marker different than a measure? Reply: We have revised the sentence to The relationships between latent outcome variables, and their observed markers were also modeled. Eqns. 1, 3, 5, and 8, is tau the variance or its reciprocal ("precision")? Reply: The tau in Eqs. 1, 3, 5, 7, and 8 is the precision. We have corrected the error in the involved equations. Equation 9 and 10, using ATemp (instead of AT and DPT) is plausible here. Statistical Analysis, "We also included a subject level random intercept, bi, to represent subject-specific permanent effects for each participant to accommodate the repeated measures design of the study." Why is b_i assumed to be the same for LF/HF (Eq. 10) and parasympathetic tone (Eq. 9) for each participant? This seems overly restrictive. Is it needed for identifiability? Reply: We do not assume that b i is the same for LF/HF (Eq. 10) and parasympathetic tone (Eq. 9) for each participant. We have revised the Statistical Analysis section to clarify this fact. Statistical Analysis, "Moreover, Eqs. (9) and (10) implicitly depend on participant i because the scheduled time visits t depend on participant i." You might elaborate on this point if it's important. Are you saying that this is another source of within-subject correlation because multiple pollutants are measured at the same visit day t on a particular visit? Reply: Yes, this could be another source of within-subject correlation.

7 Statistical Analysis, "Additionally, to demonstrate that our Bayesian approach to estimation yields trustworthy parameter estimates, we also compared the results from the above BLMMs to Frequentist framework estimates (Frequentist linear mixed models, LMMs)..." This statement doesn't make any sense at all to me. If this is really your intent you should describe what criteria could be used to determine the "trustworthiness" of the BLMM results based on the frequentist estimates. Do you actually view the frequentist estimates as the gold standard? Reply: We have revised and removed to demonstrate that our Bayesian approach to estimation yields trustworthy parameter estimates in the sentence. We do not view the frequentist estimates as the gold standard. Equation 12, your notation for LF/HF_it is confused--in Eq. 8 it was a measured value but here it is an expected value. The definition of each variable should be consistent throughout. I'd suggest using E(HF_it) and E(LF/HF_it) in Eqns Eqns 11-12, using a different set of parameters for each equation would make it clear that these were fit as two separate models. Reply: We have revised the notations for equations 11 and 12 accordingly. Statistical Analysis, why did you choose only HF and LF/HF as outcomes for the frequentist and Bayesian LMMs? Weren't SDNN and rmssd also of interest? Reply: We selected LF/HF as an outcome for the linear mixed models because it was the only marker of sympathetic tone. Since HF was the reference scale for Parasympathetic tone in our structural equation models, we chose HF as an outcome for LMM and BLMM to represent the marker of parasympathetic tone. SDNN and rmssd are important measure of HRV and they are of interest especially in structural equation modeling in part of our study. Results, the 69.6% ever smoking rate is a concern. What proportion of the HRV visits are for current smokers? For how long did participants abstain from smoking prior to the clinic visit? Was abstention confirmed by cotinine measurement? Given that cigarette smoke may be the predominant source of BC, CO, NO, and NO2 for current smokers a crude "ever/never" confounding adjustment may not be sufficient. Reply: The participants were asked to abstain from smoking the day prior to their clinic visit. We never confirmed the abstention by cotinine measurement. We agree with the reviewer that cigarette smoke may be a source of BC, NO, NO 2 and CO for current smokers, but exposure to these pollutants from active smoking is likely uncorrelated with ambient levels of these pollutants measures at stationary monitors and thus cannot be a confounder. Moreover, only 4% of the participants were current smokers at the time of their visit.

8 Table 1, what are the units of measurement for BC, CO, NO, and NO 2? Reply: We have added the units of measurement for BC, CO, NO and NO 2. Traffic Pollution and Parasympathetic Tone Measurement Models, the first sentence is quite awkward. I would try this in 2-3 sentences. Reply: We have revised the first sentence accordingly. Table 2, here it appears that tau is being used for the precision (1/variance) rather than the variance. This is not consistent with your definitions in the text. Reply: We have corrected the definitions in the text to reflect that tau is precision. Effect of Traffic Pollution on Parasympathetic Tone and Sympathetic Tone Marker LF/HF, the posterior probabilities of positive effects are worth highlighting in the text--these are easily interpreted and useful quantities that can only be obtained using Bayesian methods--a clear advantage of your proposed approach. Reply: We have revised the section to highlight the posterior probabilities of positive effects. Table 3, is the BC/HF relationship the strongest among the various traffic pollutants and parasympathetic HRV outcomes? Is that why you chose to report it here, and not the other effect estimates? If you fit the other models as well you should report that. Reply: We chose BC as a tracer of traffic pollution, given its widespread use as a marker of traffic pollution in previous studies. Similarly, we chose HF as a marker of parasympathetic function based on findings from previous studies. Model Comparison, add text such as "(not shown here)" if the LMM results aren't included in the paper. Reply: We have added the text (results not shown here) to reflect the comparison between LMM and BLMM. Also it's somewhat misleading to describe the SEM-BLMM comparison in terms of which shows larger effect estimates for "parasympathetic tone," considering that the BLMM results appear to use only HF whereas the SEM results relied on HF, SDNN, and rmssd. Reply: We agree that direct comparison of the results for the SEM and BLMM is challenging given the differences in exposure and outcome noted by the reviewer. However, we think it is of value to describe and compare the results if one would use the SEM approach versus the BLMM approach.

9 Discussion, it's a bit strange to compare your results to only 3 of many published studies (and recent review papers) on air pollution and HRV. Do you feel the other studies are flawed? Reply. The Discussion now references a number of additional studies in this area. Discussion, "LMM," "linear mixed model," and "linear mixed modeling" are often used in this section, sometimes in reference to frequentist LMMs and sometimes for BLMMs. This is quite confusing--for example "impacts assessed using the linear mixed models (BLMMs and LMMs) though statistically insignificant when using LMMs were associated with decreased parasympathetic tone marker, HF, and associated with increased sympathetic tone marker, LF/HF, and were consistent across all examined exposure windows, as evidenced by their respective estimated means (Table 3)," where Table 3 only shows SEM and BLMM results. I think this would all be simpler if you drop the frequentist LMMs and consistently refer to the remaining version as BLMMs. Reply: We have dropped the frequentist LMM in the discussion and consistently refer to BLMMs throughout the text. Discussion, delete "we showed that" from "Additionally, we showed that SEMs have the ability to analyze unbalanced longitudinal data or repeated measures, with multiple outcomes and/or exposure markers, and benefit from a reduction in measurement error bias." Reply: We have deleted we showed that in the discussion part of the paper. Discussion, "In addition, our study used a random subject intercept which means that the contrasts are a mixture of within and between subjects; for this reason, confounding factor bias should be small relative to that in a purely cross-sectional study design [21]." Please provide a page number for reference 21. This claim makes sense to me if you have 1) a large number of repeated measurements per subject and 2) time-invariant confounders, but I'm not sure that it holds otherwise. Reply: We have provided a page number for reference 21. We have 214 participants with one, 216 with two, 259 with three, and 11 with four repeated measurements. We agree with the reviewer and have revised the sentence. Discussion, "Because participants lived 22 km on median straight line distance from our ambient monitoring site," do you mean up to 22 km? Reply: Yes.

10 Additional File 1, this description is very helpful, though a bit surprising to see the prior identified in an appendix rather than the main text. It doesn't seem like you've given the priors much thought--there's a large published literature on air pollution and HRV that could be used to constrain some of the model parameters, particularly for the BLMM, which closely follows typical models. Reply: We included the implementation particulars in the appendix to meet manuscript length requirements and to allow more advanced readers of our paper the opportunity to look at the appendix to obtain more details of our analysis. We acknowledge a large published literature on air pollution and HRV however we did not used the information to constrain some of the model parameters. Truncation of the factor loading priors is very reasonable, but I'm not sure the truncated priors are best described as "non-informative." Why shouldn't readers view your truncation as a method of incorporating strong prior information about the sources of BC, CO, NO, and NO2, and about the underlying biological connection among HF, SDNN, and rmssd? Reply: We agree with the reviewer, the truncated priors are informative priors, because we expect all surrogates to be positively correlated. Therefore, the factor loadings/source contributions of traffic-related pollutants and parasympathetic tone markers cannot be negative. The first sentence of the Implementation Particulars section was removed in the Additional File 1. Additional Files, please include your WinBUGS code. Reply: We have included our WinBUGS code as an additional File.

11 Reviewer 3 I found no need for either Major Compulsory Revisions or Minor Essential Revisions. In terms of Discretionary Revisions, I request that the authors provide some commentary on their ability to demonstrate the influence of NO and NO2 to influence their effect markers. Table 1 indicates that the NO and NO2 concentrations were all low, and were reported to only one significant figure. They need to be able to show that proper conclusions be drawn from reliance on such a limited database? Reply: We apologize for the misperception. The concentrations of both NO and NO 2 are in the parts per billion (ppb). We rounded off the figures to 2 decimal points in the ppm concentration to match the ppm concentration of CO.