Observation of External Wounding on Green Ash (Fraxinus pennsylvanica Marshall) Trees Associated with Tree Injection Systems
Round 1
Reviewer 1 Report
The manuscript is characterized by a high-quality introduction and discussion, both supported by a sufficient amount of literature. The methodology is also (except for the statistical part) very well described. In contrast, statistical analysis suffers from some problems.
Statistical approach:
ANOVA part: I see no point in using One-way ANOVA. 1) For ANOVA, an important assumption is the normal data distribution (or normal distribution of residuals), which will hardly be fulfilled for tree condition and tree canopy thinning (explicitly stated by the authors that these are ratio/percentage), wound number (quite low counts), correct treatment application (0/1) and wound cause (categorical). 2) If the normality of the data is not met, generalized linear models should be used (which the authors know because they use them later in the text) – with beta distribution (for ratios), Poisson/negative binomial distribution (counts) and binomial distribution (0/1). For Wound cause, most likely the chi-square test must be used. 3) If only two groups of data (two treatment systems) are compared, a completely sufficient solution would be to use the t-test (in the case of data normality) or the Wilcoxon test. 4) On the other hand, if the authors wanted the greatest possible accuracy, they may use generalized linear mixed models, in which individual municipal regions would appear as a random term since the differences between them are not very interesting for interpretation.
(G)LM part: 1) If they do not want to work with municipal regions as a random term but as a fixed term – explanatory variable, then it does not make much sense to select Chicago as "Control" and compare the others with it (L136) – it just follows the default setting in R. By setting the contrasts appropriately, you can achieve comparing each region to a common average that makes more sense. 2) It is not specified how the model (logistic and linear) was built - was a stepwise selection based on AIC used? Was it forward or backward selection, or did the authors build the model arbitrarily without using these tools? Did they consider possible interactions between the explanatory variables when building the models?
L155: „Trunk diameter in the open system trees was significantly less (P=0.003) by 5 cm (2 in) compared with the sealed system trees.“ This seems to me to be quite essential. Is trunk diameter response variable of treatment (different treatments cause differences in trunk diameter) or rather an explanatory variable (for trees with a smaller diameter = younger, an open system is chosen, while for larger = older trees, a sealed system is preferred)? If the diameter of the trunk was not controlled at the beginning of the experiment, then it is better to assume that it was already different before the treatment. If we assume that younger (smaller diameter) trees are in better condition than older (bigger), then this is a significant confounding factor, limiting the interpretation of the difference between treatments. It is therefore necessary to include it in the model as a covariate and perform an ANCOVA type GLM model.
Minor comments
L137: I find it unfortunate that the sentence is split in half by table 1.
L157: I don't understand why Table 2 is repeated twice in the text.
L162: The logistic model showed that trunk diameter, treatment system, and the Milwaukee region significantly explained the presence of wounds… Sounds really weird – I would recommend … and the region significantly explained the differences in presence of wounds. Compared to Chicago, trees in Milwaukee … - or even better, relate the comparisons to the average (by setting the contrasts, see above).
L225: „Tanis and McCollough (2016)“ This quote is not converted to Forests format.
Author Response
Thank you for your suggestion for improvement. Below are our response and changes made for improvement.
Statistical approach:
ANOVA part: I see no point in using One-way ANOVA. 1) For ANOVA, an important assumption is the normal data distribution (or normal distribution of residuals), which will hardly be fulfilled for tree condition and tree canopy thinning (explicitly stated by the authors that these are ratio/percentage), wound number (quite low counts), correct treatment application (0/1) and wound cause (categorical). 2) If the normality of the data is not met, generalized linear models should be used (which the authors know because they use them later in the text) – with beta distribution (for ratios), Poisson/negative binomial distribution (counts) and binomial distribution (0/1). For Wound cause, most likely the chi-square test must be used. 3) If only two groups of data (two treatment systems) are compared, a completely sufficient solution would be to use the t-test (in the case of data normality) or the Wilcoxon test. 4) On the other hand, if the authors wanted the greatest possible accuracy, they may use generalized linear mixed models, in which individual municipal regions would appear as a random term since the differences between them are not very interesting for interpretation. >>> Thank you for the suggested improvement for statistical examination. The ANOVA was just used for simplicity to explore if differences exist between the two groups and similarly the comparison of mean values in the two groups (open or sealed). We did test for Homogeneity of Variances and found the Levene Test to be significant, thus no violation of that assumption. While an assumption for normality is important, it is more so with small sample sizes (e.g., 30). Our study was a magnitude higher with well over 300 observations (an exception was wound height with fewer but over 100 and we did not consider this variable in models). The ANOVA is quite robust for normality violations with large samples such as in this study. We did find various normality deviations, in particular with tree condition and tree thinning. We randomly selected treated trees from the population (e.g. ~ 100 trees in Milwaukee from a population of ~ 35,000) thus believe due process was completed for independence. Having said all this, we did change table one to report findings from a Mann-Whitney, which we did run and found similarities to our table one reported significance, except for tree canopy and tree canopy thinning. The Mann-Whitney was used since we had independent samples. Table one as developed as intended to report the descriptive differences between the two groups (open and sealed) which was further explored through further tests (Logistic, Linear Regression, and Chi-square were appropriate) to test for significance when controlling for other variables. However, we did change to a non-parametric approach as suggested. Thus, we do truly appreciate your time with expressing concern, however, after consideration changed as requested.
We did use a chi-square to test for differences for correct treatment application and wound cause (categorical). We had a typo with wound cause listed as one if the ANOVA tested variables which it was not done (and not in table one) and deleted that reference. The regions were used since we did not have data on the number of times trees were treated, but do know that trees in Sioux Falls were treated most likely only once (EAB just arrived), Minneapolis (likely 2 times on average per tree) and the Milwaukee and Chicago regions more so (likely 4 to 6 times). Thus, this is the sole reason for included region to see if a difference existed.
(G)LM part: 1) If they do not want to work with municipal regions as a random term but as a fixed term – explanatory variable, then it does not make much sense to select Chicago as "Control" and compare the others with it (L136) – it just follows the default setting in R. By setting the contrasts appropriately, you can achieve comparing each region to a common average that makes more sense. 2) It is not specified how the model (logistic and linear) was built - was a stepwise selection based on AIC used? Was it forward or backward selection, or did the authors build the model arbitrarily without using these tools? Did they consider possible interactions between the explanatory variables when building the models? >>> (1) Chicago was selected since emerald ash borer was found there sooner (~2006) than the other locations (Milwaukee ~ 2010, Minneapolis ~ 2015) and Sioux Falls ~ 2020; these dates are approximate as EAB was in these metro regions earlier but discovered on these approximate dates). Thus, using region as an explanation. Ideally, we would have had the number of treatments as each treatment produces 1 injection site for every 2 inches of diameter and this ideally repeated every two years based on label rates and methods. Thus, using Chicago as the comparison location. We could easily use any other region (Sioux Falls for example), and then the other three location would possibly be significant since more injections and thus more potential external wounds. And yes, R is funny that way with looking alphabetically and yes we often will code to sort if there is one we want to compare against, however in this case we deliberately picked Chicago since EAB was found there sooner. (2) We did mention in the methods that a logistic model was used but failed to explain the model testing was a full model with no simplification. We think/hope the methods clearly state the tested variables used and these are reported in table 3 with no simplification.
L155: „Trunk diameter in the open system trees was significantly less (P=0.003) by 5 cm (2 in) compared with the sealed system trees.“ This seems to me to be quite essential. Is trunk diameter response variable of treatment (different treatments cause differences in trunk diameter) or rather an explanatory variable (for trees with a smaller diameter = younger, an open system is chosen, while for larger = older trees, a sealed system is preferred)? If the diameter of the trunk was not controlled at the beginning of the experiment, then it is better to assume that it was already different before the treatment. If we assume that younger (smaller diameter) trees are in better condition than older (bigger), then this is a significant confounding factor, limiting the interpretation of the difference between treatments. It is therefore necessary to include it in the model as a covariate and perform an ANCOVA type GLM model. >>> It would be nice if the populations were the same, however, perhaps it was not clear that we used tree populations that were available under treatment programs and randomly selected a sample of the population. There was no preference that the open is used in smaller trees and sealed in larger. Was just with the populations studied and samples pulled for the tow treatments that the sealed sample was larger. What this means for injections on a tree is on average one extra injections sites based on the label rate that specifics one injection for every 2 inches (5 cm) of trunk diameter at 4.5 feet (1.37 m). We added a limitation to the discussion that trees in the sealed system were larger in diameter. However, while the difference is significant, this alone does not account for differences in mean wound number per tree or mean total wound area. Also, tree condition in a model is confounding since stem sound such as from tree injection related cracking will lower tree condition by ~5 to 10% using the standard CTLA tree condition method as cited in the methods. Thus, this is the reason tree condition is not included in the model.
Minor comments
L137: I find it unfortunate that the sentence is split in half by table 1. >>> Corrected to join the methods together
L157: I don't understand why Table 2 is repeated twice in the text. >>> Removed the extra Table 2 reference. Was an artifact of formatting and moving text. Thank you for catching this.
L162: The logistic model showed that trunk diameter, treatment system, and the Milwaukee region significantly explained the presence of wounds… Sounds really weird – I would recommend … and the region significantly explained the differences in presence of wounds. Compared to Chicago, trees in Milwaukee … - or even better, relate the comparisons to the average (by setting the contrasts, see above). <<<< Thanks for the editorial suggestions, change made in paper
L225: „Tanis and McCollough (2016)“ This quote is not converted to Forests format. >>>> Corrected
Reviewer 2 Report
Review comments Forests manuscript 1976353
I reviewed the manuscript entitled: Observation of external wounding on green ash (Fraxinus pennsylvanica Marshall) trees associated with tree injection systems
By R.J. Hauer, J.J. Ball, E. North
The manuscript describes an observational study to evaluate the effects of two methods of treating green ash for protection from the non-native insect emerald ash borer.
Overall, I found the manuscript written very concisely, but understandable. I found no unneeded, extra words. I welcomed the brevity in all sections of the manuscript, particularly for the introduction. This was an observational study where treatments were not specified and applied consistently, as the authors explain. I found the study design adequate for the stated objectives. It is unfortunate that a control treatment could not be part of the study to better evaluate the number of wounds per tree – that is, were all wounds observed assumed to result from boring treatment for EAB. However, I understand why that was not possible because lack of treatment would have invited infestation by EAB, and death of the tree. I consider a strong part of the study design was sample collection in 4 cities, but that was also a weakness because of increased variation resulting from EAB treatment by various operators, methods, tools, etc, which the authors stated. Despite many sources of uncontrolled variation, the authors found considerable significant results associated with the open and closed method of EAB injection treatment. The authors clearly identified limitations of the study and stated logical questions that could direct future work. Results from this study should be of interest arborists considering injection treatment of other species of trees, but these results are most important to arborists in cities where green ash are present.
Author Response
Overall, I found the manuscript written very concisely, but understandable. I found no unneeded, extra words. I welcomed the brevity in all sections of the manuscript, particularly for the introduction. This was an observational study where treatments were not specified and applied consistently, as the authors explain. I found the study design adequate for the stated objectives. It is unfortunate that a control treatment could not be part of the study to better evaluate the number of wounds per tree – that is, were all wounds observed assumed to result from boring treatment for EAB. However, I understand why that was not possible because lack of treatment would have invited infestation by EAB, and death of the tree. I consider a strong part of the study design was sample collection in 4 cities, but that was also a weakness because of increased variation resulting from EAB treatment by various operators, methods, tools, etc., which the authors stated. Despite many sources of uncontrolled variation, the authors found considerable significant results associated with the open and closed method of EAB injection treatment. The authors clearly identified limitations of the study and stated logical questions that could direct future work. Results from this study should be of interest arborists considering injection treatment of other species of trees, but these results are most important to arborists in cities where green ash are present. <<<< Thank you for the review and complete explanation of our thoughts which we agree with.
Round 2
Reviewer 1 Report
I appreciate the effort of the authors to solve my comments. It is a pity that the authors did not provide the manuscript in the form of a tracked version to make more obvious what has been changed.
ANOVA/non-parametrical solution: I appreciate the author’s extensive explanation. I cannot agree with their sentence: We did test for Homogeneity of Variances and found the Levene Test to be significant, thus no violation of that assumption. The null hypothesis of the test is, that the population variances are equal. On the other hand, they chose a non-parametric solution in the end, so it doesn't matter. And Mann-Whitney test (equivalent to the Wilcoxon rank sum test) is a good choice. However, the description of the methodology is changed but the results are not.
Regions as explanatory/random term and Chicago as control: Now I understand why the differences among individual regions were of interest to the authors, and thus I also understand why they deliberately picked Chicago as a control - it certainly helped that they now added this explanation (even if only cursorily) to the methodology.
Building the model: It would help here if the authors explicitly mentioned why the model was built in the exact order of factors that the authors chose. I asked about AIC/forward selection etc. because I was curious about how they decided which of the factors would be ranked first, second and third. The reason of my concern was, that the default type of ANOVA, which tests a (generalized) linear model, tests individual explanatory variables sequentially, after taking into account the influence of the previous one. Thus, the p-value of individual factors may be different in different models, depending on their order. The model constructed by the authors test the effect of the treatment system, after counting the effect of trunk diameter and region (which appear as covariates in the model). This is very well done, but the decision that led the authors to choose the appropriate order of factors is not explicitly described (that some of them are covariates).
Author Response
I appreciate the effort of the authors to solve my comments. It is a pity that the authors did not provide the manuscript in the form of a tracked version to make more obvious what has been changed. >>>> Not sure what happened with the upload, did have track changes on and I checked that the uploaded version on the Forests site did show the changes. Hopefully this upload does have tracked changes showing the changes.
Regions as explanatory/random term and Chicago as control: Now I understand why the differences among individual regions were of interest to the authors, and thus I also understand why they deliberately picked Chicago as a control - it certainly helped that they now added this explanation (even if only cursorily) to the methodology.>>> It was in the last version submitted as … (Chicago as comparison region since this was the first region with EAB)
Building the model: It would help here if the authors explicitly mentioned why the model was built in the exact order of factors that the authors chose. I asked about AIC/forward selection etc. because I was curious about how they decided which of the factors would be ranked first, second and third. The reason of my concern was, that the default type of ANOVA, which tests a (generalized) linear model, tests individual explanatory variables sequentially, after taking into account the influence of the previous one. Thus, the p-value of individual factors may be different in different models, depending on their order. The model constructed by the authors test the effect of the treatment system, after counting the effect of trunk diameter and region (which appear as covariates in the model). This is very well done, but the decision that led the authors to choose the appropriate order of factors is not explicitly described (that some of them are covariates). >>>> Not sure if we were clear on the prior response but the model was used to test if treatment system explained differences in tree wounds with tree diameter and region used to control for larger trees have more injection sites and region as a surrogate for number of times injected. The model was not simplified. We were not attempting to make a predictive model, rather using the logistic and linear models as two methods to test if these variables explain wounds. We hope this clarification makes sense.
