Review Reports - Statistical Facilitation in Environmental Science: Integrating Results from Complementary Statistical Analyses Can Improve Ecological Interpretations

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors:

Your manuscript in overall is novelty, but in the present form is weak. Details are as follows:

1- As an overall evaluation the manuscript is very good because it applies topics of data sciences applied to ecology, this is a current topic that has many applications and few ecologists use it.

2- Although the redaction of introduction is very good, ordered and easy to understand there are paragraphs without references as support and this situation implies that these paragraphs are weak because there is no bibliographic support.

3- Material and Methods: this section is complete, easy to understand, it is my personal opinion on the basis of my experience in these topics.

4- Results: these are easy to understand and complete.

5- Discussion is practically nonexistent, because it is a kind of continuation of the explanation of the results section without references as support, really it is the main weak point of the manuscript.

6- References: these are adequate.

Many success

Author Response

REVIEWER 1 - COMMENTS AND SUGGESTIONS FOR AUTHORS

REVIEWER 1- COMMENT 1

1- As an overall evaluation the manuscript is very good because it applies topics of data sciences applied to ecology, this is a current topic that has many applications and few ecologists use it.

COAUTHOR RESPONSE.

Thank you. No change needed or made.

REVIEWER 1- COMMENT 2

COAUTHOR RESPONSE

Thank you for calling our attention to this issue so we can remedy the problem. Upon checking, we note that we have quite a few references in most introduction paragraphs. However, we agree that we need to add references to two introduction paragraphs (7, 16 references added to Overview and Quantitative Challenges of Large-scale Monitoring Data Sets paragraphs; Revised Manuscript Lines 45-64, 86-105).
Please see the changes we made that are described in the table below re references in the introduction (8-24 references per introduction paragraph).



Section	Paragraph		References
	No.	Title	Original	Added	Total

Introduction	1	Overview	3	7	10
"	2	Species Distribuion Models	23	0	24
"	3	Quantitative Challenges	0	16	16
"	4	Statistical Analyses	14	0	14
"	5	Multiple Logistic Regression	12	0	12
"	6	Random Forest	12	0	12
Introduction	7	Objectives	8	0	8

REVIEWER 1- COMMENT 3

3- Material and Methods: this section is complete, easy to understand, it is my personal opinion on the basis of my experience in these topics.

COAUTHOR RESPONSE.

Thank you. No change needed or made.

REVIEWER 1- COMMENT 4

4- Results: these are easy to understand and complete.

COAUTHOR RESPONSE.

Thank you. No change needed or made.

REVIEWER 1- COMMENT 5

COAUTHOR RESPONSE.

Format of the Original Discussion. We have given this reviewer comment much thought. We appreciate that this reviewer may have different views of how to write a discussion than we do. However, we respectfully disagree with this comment as we have addressed many of the issues that this journal’s Guide to Authors says should be included in the discussion (see below):

“Authors should discuss the results and how they can be interpreted in perspective of previous studies and of the working hypotheses.
The findings and their implications should be discussed in the broadest context possible and limitations of the work highlighted.
Future research directions may also be mentioned.”

https://www.mdpi.com/journal/environments/instructions

Below we review how (1) we have included these journal-recommended discussion points in our manuscript, and (2) what we have added in response to reviewer comments (including adding and integrating references).

How We Have Included the Journal-Recommended Discussion Points in Our Manuscript Discussion Section. Relative to discussing our results and interpretation (#1 above) in the broadest context possible (#2 above), the sections listed below interpret our results related to regressors and statistical techniques.

Take Home Message 1 – Strengths of Each Analysis (Objective 1) (Revised Manuscript Lines 437-476)
Take Home Message 2 – Weaknesses of Each Approach (Objective 1) (Revised Manuscript Lines 477-506)

The above discussion sections create the scaffolding for our next two synthetic discussion sections (listed below) that integrate and interpret results on specific regressors by statistical techniques.

Take Home Message 3 - Redundancies are Useful (Objective 2) (Revised Manuscript Lines 507-518)
Take Home Message 4 - Ambiguities Across Analyses Can Guide Future Directions (Objective 2) (Revised Manuscript Lines 519-555)

After stitching together the above four take home messages, in the final two discussion sections, we develop and amplify our most important conclusion (i.e., Neither analysis alone provides the best information. Instead integrating similarities and differences across multiple analyses is the pathway that advances ecological understanding).

What We Learned from Comparing Then Integrating Multiple Analyses (Objective 4) (Revised Manuscript Lines 594-612)
Our Primary Take-Away and Recommendation (Revised Manuscript Lines 637-646)
What is Needed (2 sections; Revised Manuscript Lines 647-673)

Our Approach to Statistical References in the Discussion.

Our study is a hybrid of two foci. Our main focus is an evaluation of what can be learned from comparing and integrating two well-described and frequently-used statistical analyses that are applied to the same ecological dataset. Our secondary focus is ecological interpretation of these results for the specific taxa under examination (i.e., Plains Minnow).
A very detailed description of how our two statistical analyses work would duplicate existing information in statistical textbooks and miss the novelty of the integration we apply here.
Nevertheless, in the original manuscript introduction and methods, we provide a basic description of each statistical approach and provide references on how they work (> 36 statistical references in introduction).
We have citations in the Strengths and Weakness of Each Analysis Discussion sections (7+4= 11 citations per section), but we don’t add references to the Redundancies and Ambiguities and What We Learned sections because these sections analyze and interpret what our data means not what others have found (i.e., others have not addressed this issue in a comparable way).
In response to this reviewer’s comment, we have added references to the Assumptions section of the discussion (25 total statistical references added; Revised Manuscript Lines 556-593).

Our Approach to Ecological References in the Discussion.

In the original manuscript, we also described the life history and ecology of our example taxon with citations in the Study Taxon section (Revised Manuscript Lines 172-190).
Others have not undertaken the detailed comparison of insights on biota as we did here so we can’t integrate our results with literature on like approaches.
Although our contribution is not just about Plains Minnow, we now recognize that readers might be interested in more details on how these statistical results apply to ecology of the Plains Minnow even if they don’t work on this fish.
As such we have added a section entitled Ecological Insights Gained About the Plains Minnow (Revised Manuscript Lines 613-636) and integrated Great-Plains Plains-Minnow literature (10 references added).
Although not related to references, we have added select insights on ecological interpretation and management utility of our study in the discussion (Revised Manuscript Lines 451-458, 514-517, 530-533, 608-612, 660-673).

Summary. We believe our original discussion satisfied requirements of the journal for this component of the manuscript. However, we also acknowledge that responding to this and other reviewer comments as described above has improved the discussion substantially.

REVIEWER 1- COMMENT 16

6- References: these are adequate.

Many success

COAUTHOR RESPONSE.

Thank you. No change needed or made.

Submission Date

03 October 2025

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

unfortunately the original version of the paper did not have line numbers. I intended on moving my comments in the pdf bubbles into a line-referenced list, but that's too much work. So, my comments are in the attached pdf.

Half of Table 1 is also unreadable

In general, I support the work, but I'm not convinced it's really as valuable as it could be. For example, a lot of the discussion about advantages and disadvantages of each approach are well-established among machine learning practitioners.

My biggest suggestion is to consider if using a simulated dataset might be a better way to examine how these different techniques respond to different problems in your data.

Comments for author File: Comments.pdf

Author Response

REVIEWER 2- Comments and Suggestions for Authors

Unfortunately the original version of the paper did not have line numbers. I intended on moving my comments in the pdf bubbles into a line-referenced list, but that's too much work. So, my comments are in the attached pdf.

COAUTHOR RESPONSE:

We copied all reviewer comments from the PDF, have pasted them below, and have provided a response to these comments including where changes were made in the revised manuscript where appropriate.

REVIEWER 2- COMMENT 1

Half of Table 1 is also unreadable

COAUTHOR RESPONSE:

The journal pasted the figure this way (only ½ of the table was on the page). However, we have made Table 1 into a portrait table so it should fit the page easily.
We have amended the table for readability as suggested.

REVIEWER 2- COMMENT 2

COAUTHOR RESPONSE:

We agree that these statistical analyses are common in the technical statistical literature and in statistics textbooks, but we do not think a comparative analysis, such as what we present here, is common for an ecological-conservation audience who are not experts in machine learning.
Our decision to use real world applied data to demonstrate the applicability of these types of “advanced” analyses is useful to professionals charged with making management decisions. A comparison of the advantages and disadvantages of each approach is likely to help them understand the application of these techniques to their own conservation problems.

REVIEWER 2- COMMENT 3

My biggest suggestion is to consider if using a simulated dataset might be a better way to examine how these different techniques respond to different problems in your data.

COAUTHOR RESPONSE:

This is another thoughtful idea. However, upon consideration we don’t think it is the right approach for this paper for the reasons below.
We chose to demonstrate the applicability of our analyses to real world “messy” monitoring data. This choice allowed us to demonstrate the applicability of integrating these two techniques to gain a single unified ecological story.
Using simulated data would likely be ignored by many conservation and fisheries managers due to its lack of real-world application.
Finally, because the paper is already complex, we believe that additional information would complicate the paper’s message and reduce understanding for the average reader, who are our target audience.

Reviewer 2 Line-By-Line:

REVIEWER 2- COMMENT 4

Line 20 – 22: this transition is a little awkward and incomplete - these two sentences also fall into the 'two topic sentences' trap which is so hard to avoid in abstracts. the easy fix is something like 'biological conservation seeks to understand, manage, and restore populations of native organisms using many techniques. A main approach is using long-term data collections to inform decision-making.'

COAUTHOR RESPONSE:

Based on the reviewer’s suggestion, this was changed to “Professionals in biological conservation seek to understand, manage, and restore populations of native organisms using many techniques. A common approach is using long-term data collections to inform decision-making” (Revised Manuscript Lines 21-23)

REVIEWER 2- COMMENT 5

Line 22 – 24: such as?

COAUTHOR RESPONSE

We provide more details and examples of all of these issues in the introduction. No changes were made in the abstract in response to this reviewer comment because we wanted to maintain conciseness in the face of abstract word limitations.

REVIEWER 2- COMMENT 6

Line 24 – 26: this is a little jargony - and I think would benefit from some more antecedent exposition talking about things related to the problems that come along with long-term datasets, like they cant be specifically added (e.g., we cant go back in time and sample more locations), but we might be able to add more covariates....but only if those covariates are available, or can be reasonably estimated with models. Another issue is, and I think this is a big part of your work, that very few data are ever perfectly suited for a given statistical test.

COAUTHOR RESPONSE

We have clarified these lines and modified the text as follows “Integrating results from multiple analyses applied to the same dataset (i.e., approaching the same biological problem using different techniques) is one way to address concerns related to field data that violate statistical assumptions and assemble insights based on the weight of evidence” (Revised Manuscript Lines 25-27).

REVIEWER 2- COMMENT 7

Line 30 – 31, 32, 33: such as?

COAUTHOR RESPONSE

We don’t add any text to the abstract to address this review query for two reasons.
- First, the number of words are limited in the abstract.
- Second, throughout the paper, we show a technique for comparing analyses using the same dataset. We think it is clear that the approach we illustrate can be widely used with any organism for a similar dataset.

REVIEWER 2- COMMENT 8

Line 43 – 46: you probably know the term, but i suggest working 'data-driven discovery' in here somewhere.

COAUTHOR RESPONSE

The text was added as suggested “ Biological conservation, a component of environmental science, is a data-driven discipline that seeks to understand, manage, and restore distributions of native organisms across state and regional spatial scales often through data-driven discovery” (Revised Manuscript Line 48-51).

REVIEWER 2 - COMMENT 9

Line 46 – 48: the real issue is with data driven discovery is that we have hypotheses at the end, not conclusions - we identify things which might be true, but need specific hypothesis testing.

COAUTHOR RESPONSE

In field studies, the true state is never known. Even using a weight-of-evidence approach as do here, the only outcome is the most probable justification and potential hypothesis that can “falsify a conclusion.” This is an issue with any scientific discipline beyond physics or mathematics.
We agree it is an interesting philosophical issue. However, no change was made as this concept is beyond the scope of this manuscript and is not essential to address the questions that we asked.
Although the revised manuscript was not edited to specifically address the issues noted in this reviewer comment, we did add two discussion sections issues that integrate our results with some philosophical issues related to statistics:
- What is Needed: Realistic Expectations (Revised Manuscript Lines 647-659)
- What Is Needed: Successful Disciplinary Interpretation of Statistical Patterns (Revised Manuscript Lines 660-673).

REVIEWER 2- COMMENT 10

Line 48 – 52: this is likely out of place.

COAUTHOR RESPONSE

We believe this sentence follows in logical succession to it’s predecessor. Thus, no change was made.

REVIEWER 2- COMMENT 11

Line 60: shouldn’t this be 'observed' and 'not observed'?

COAUTHOR RESPONSE

We believe this is an acceptable wording choice even though “ present and absent” is the most common use (i.e., presence and absence data). If the reviewer is approaching this from an “occupancy like” standpoint, these analyses do not estimate detection and assume, by default, that sampling was done appropriately and that field crews were detecting a fish if they are present.
For the reason stated above, we maintain the use of “presence” or “absence” for data and “probability of presence” for the models.

REVIEWER 2- COMMENT 12

Line 73 – 75: always? SDMs always address information needs of managers? I certainly agree that they can, but definitely not always. I suggest softening this statement and some of the previous content - at minimum I think it's important to point out the differences between the theory and practice of SDMs. In theory they, like so much scientific work, are incredibly useful tools. In practice, however, they usually are not so good.

COAUTHOR RESPONSE

We have softened this statement: “As a substantial strength for biological conservation, SDMs often address information needs of managers (e.g., distribution, presence, absence of taxa of interest) at the relevant spatial scales needed to make local, state, and regional management decisions, but usefulness may be impacted by the quantitative limitations for any data analysis (e.g., sampling bias, knowledge gaps, and unsampled covariates)” (Revised Manuscript Lines 81-85).

REVIEWER 2- COMMENT 13

Line 77 – 78: I don't think you need to say this right here...it might be a little too specific for the topic sentence of the paragraph. But you do bring up an important question about what counts as an assumption of the stats, like homogeneity of variance vs. something like a 'requirement' stipulated by a power analysis.....and the consequences each can have on the robustness of the results....rare phenomena are necessarily hard to study...so maybe another point you are making here is that there is such a thing as a 'data tipping point' where a species becomes so rare that we cant learn enough to help....so maybe that’s another type of extinction....I’m not sure what that would be called...

COAUTHOR RESPONSE

This topic sentence is a bridge between the previous section and this section.
This reviewer brings up another interesting philosophical thought but we believe addressing this point is beyond the scope of this manuscript and not relevant to the practical audience we address.
Nevertheless, we have altered the text slightly to create a more general topic sentence: “The utility of applying SDMs to biodiversity monitoring data depends on incorporating appropriate caveats that acknowledge the inevitable violations of statistical assumptions” (Revised Manuscript Lines 86-88).

REVIEWER 2- COMMENT 14

Line 80 – 82: I'm not exactly sure what they are, but I think the examples you are using in this paragraph each call to larger phenomena - I think I'm suggesting, as an example, to leave out the call to 'amount of missing data' and just say something about measuring at large scales being more challenging than small scales. One reason, for example, is that as the scale of a study increases (in both time and space) the more signals you'll start integrating into the dataset.

COAUTHOR RESPONSE

The second bullet is specifically meant to reference time and personal limitations of sampling large areas.
We altered the text for clarity as follows “Second, spatial scale can affect physical and biological patterns of variability (Lin et al 2005; Chapman et al 2010; Nakagawa 2014; Bisson et al. 2024) such that logistical limitations of personnel and budget can adversely affect large-scale, mixed-scale and hierarchical sampling (i.e., we can sample often or we can sample across a broad scale, but doing both is increasingly difficult (Revised Manuscript Lines 92-96).

REVIEWER 2- COMMENT 15

Line 83 – 86: the first part repeats the point about statistical assumptions from above. The second half of this is an excellent point - and one I've made repeatedly - even the most perfectly designed programs will not be executed perfectly.

COAUTHOR RESPONSE

This text was altered to make the second point and is now stated as “Fourth, regardless of the statistical or study design and rigor of the sampling regime some assumptions of statistical analysis (e.g., correlations among variables; irregular distributions) often remain unaddressed (Olsen et al. 1999; Biber 2013; Katsis et al 2025; Midway & Daugherty 2025) (Revised Manuscript Lines 99-102).

REVIEWER 2- COMMENT 16

Line 86: not just these.

COAUTHOR RESPONSE

This was not intended to be an exhaustive list.
We altered the revised manuscript text to soften language: “As a result of these limitations among others, statistical results based on monitoring data can be ecologically ambiguous, inconsistent across studies, or uninterpretable, which reduces their utility for conservation of environmental resources (Lindenmayer and Likens 2009)” (Revised Manuscript Lines 102-105).

REVIEWER 2- COMMENT 17

Line 98 – 100: I think you can be more specific here. do you mean how each technique identifies important predictors? Or maybe something more specific about how good they are at predicting occurrence?

COAUTHOR RESPONSE

Yes, we mean identifying important predictors.
We develop how each approach identifies important predictors throughout the manuscript so we don’t add more text here.

REVIEWER 2- COMMENT 18

Line 116 – 119: beware the lure of what’s popular within a discipline.

COAUTHOR RESPONSE

Certainly true. However, this reference can be of benefit when introducing novel concepts, as we do here, to use at least one analysis that most readers will be familiar with. We believe this will also reduce the barriers to entry in understanding the core concepts of this paper because it is one less new concept to learn.

REVIEWER 2- COMMENT 19

Line 119 – 120: interpretation vs. prediction accuracy trade off.

COAUTHOR RESPONSE

This appears to be a general philosophical comment. Thus, no change was suggested or made.

REVIEWER 2- COMMENT 20

Line 127 – 128: they also regularize....I don't think logistic regression does that natively....although I suppose you could do that my doing some model dredging....although that can be dangerous.....but in 'what’s in the data' land, where we limit ourselves to testable hypotheses rather than firm conclusions, we can break a lot of rules.

COAUTHOR RESPONSE

We modified the revised manuscript text to add a caveat related to regularization in response to this reviewer’s comment.
“The resulting ensemble of trees in conjunction with tuning of hyperparameters (regularization) like number of trees, max depth of trees and variables to try per node can reduce the likelihood of overfitting or misspecification compared to a single tree (Faraway 2016; Bonaccorso 2018)” (Revised Manuscript Lines 144-147).

REVIEWER 2- COMMENT 21

Line 135: really only two - but one is done twice yielding 3 results

COAUTHOR RESPONSE

We consider the use of standardized data a separate analysis since the output and use of coefficients differs from that of non-standardized logistic regression.

REVIEWER 2- COMMENT 22

Line 136 – 139: I'm not convinced this comparison should be the main thrust of your work - even though it was probably hard.

COAUTHOR RESPONSE

This was a primary goal and adds novelty to the paper.

REVIEWER 2- COMMENT 23

Line 145 – 146: I feel like this problem has already been solved and can be better addressed with a lot of feature engineering and using a technique which has some regularization built in, like xgboost, RF, and elastic net.

COAUTHOR RESPONSE

This is another methodological approach that is much less approachable for practicing field professionals given that the interpretation of many machine learning and black box techniques is more difficult to interpret in relation to real world ecological problems.
A comparison and integration of random forest and logistic regression blends the power of machine learning with interpretability of logistic regression, essentially providing the best of both worlds.

REVIEWER 2- COMMENT 24

Line 179 – 188: what about collinearity? I don’t actually care that you didn’t seem to do anything about it, since in fairness, the native treatment of collinearity in something like RF is a major advantage of using the technique....and how conflicts between variables competing for variability is an easy part to add to your discussion

COAUTHOR RESPONSE

See next comment by this reviewer.
We discuss collinearity and other assumptions in detail in Table 1 and in the discussion section entitled “Assumptions” (Revised Manuscript Lines 556-593).

REVIEWER 2 - COMMENT 25

Line 190 – 191: ok, looks like you did something about collinearity here. variable inflation factor is often used - but again, I think all the pre-work that goes into a 'traditional' statistical technique like a logistic regression is one of the biggest reasons to just use something like RF or XGBoost - but I'm lazy.

COAUTHOR RESPONSE

VIF is another technique for checking for collinearity. We chose Pearson correlation here, but in other analyses we also checked VIF.
No change required or made.

REVIEWER 2 - COMMENT 26

Line 197: which function in which package?

COAUTHOR RESPONSE

Base r (glm) was used. This is cited appropriately.

REVIEWER 2- COMMENT 27

Line 214: using SHapley Additive exPlanations (SHAP) might be a really good way to compare the results of the logistic regressions and the RF....although in fairness, so is something like comparing things like MSE for a test set...but you may not care so much about prediction here

COAUTHOR RESPONSE

SHAP is another approach. We considered using this. However, with the comparative approaches we were already using, we didn’t feel that adding another methodology provided additional benefits.

REVIEWER 2- COMMENT 28

Line 221 -223: sure, but there is also such a thing as being too focused....especially if that focus is based on 'professional judgment'. I'm not saying you are wrong, I'm saying that all approaches come with risks. What risks you are willing to accept depend partly on preference, but also on the objectives of the work. Looking at as many predictors as possible in an exploratory analysis brings a lot of advantages, but also places some very clear constraints on the results....it is very clear that spurious correlations become a real problem....which is another type of problem which comes with expanding the scale of your study....If you throw enough things into the model, you will eventually find something that explains the response variable. Sometimes that's ok. Sometimes it's not.

COAUTHOR RESPONSE

We purposely wanted to avoid throwing a lot of things at the wall given the messiness of monitoring data and the potential for false signals. Thus, we attempted to be broad in the variables evaluated, but within the bounds of what we felt was realistic.

REVIEWER 2- COMMENT 29

Line 227 – 229: you can also tune the number of trees. I feel like there is another tunable parameter in RF....

COAUTHOR RESPONSE

We provide text describing tuning the n_tree hyperparameter (Revised Manuscript Lines 144-146, 249-251).

REVIEWER 2- COMMENT 30

Line 250 – 251: ok so you started playing with feature engineering here. i hope later on you discuss the advantages and disadvantages of this.

COAUTHOR RESPONSE

This appears to be a general philosophical comment. Thus, no change was suggested or made.

REVIEWER 2- COMMENT 31

Line 266 – 272: spatial stuff in rivers introduces 2 challenging constraints: flow direction and channelization. I’m on the fence about the problem of flow direction, especially with motile fishes, but channelization is probably really important. It's also hard to deal with in most off the shelf stats packages. I'm not saying you need to deal with it, but at minimum you need to talk about it in your discussion.

COAUTHOR RESPONSE

We do not have measurements related to channelization of the rivers in our monitoring data set, so we do not have the ability to discuss the potential influence or lack thereof in an informed manner. Additionally, an in-depth discussion of channelization is beyond the scope of the paper. Interesting point, but no change required or made.

REVIEWER 2- COMMENT 32

Line 274 – 280: you can probably drop the unstandardized logistic regression from the paper. I'm not sure how including it improves the work. I think its...umm...standard...to...standardize.

COAUTHOR RESPONSE

We fundamentally disagree. The “unstandardized” regression allows for interpretation on the “real-world” scale and is often what managers and conservation planners are most interested in seeing and using.
As an aside, many professionals prefer to use and report “unstandardized” regressions just as many professionals prefer “standardized.” We do not think there is consensus on this point which is why we present both approaches.

REVIEWER 2- COMMENT 33

Line 359 – 361: you spoke earlier about including quadratic terms though....

COAUTHOR RESPONSE

It can be explained via linear relationship, albeit not well. Thus, we switched to a quadratic term. We were careful to show our step by step approach in order to demonstrate that we are not jumping to a priori

REVIEWER 2 - COMMENT 34

Line 377 – 378: ....the comparison looks pretty good to me. but, this is potentially highlighting a deficiency in MLR vs RF. a cubic term, or some other higher order polynomial might’ve resolved the discrepancy, but that’s just saying that MLR needs more feature engineering than RF.

COAUTHOR RESPONSE

This appears to be a general philosophical comment. Thus no change was suggested or made.

REVIEWER 2- COMMENT 36

Line 387 – 295: ...which is why RF has brown in popularity. Even small improvements, especially given the ease of using RF, is enough to encourage people to use it over the traditional statistical techniques.

COAUTHOR RESPONSE

Potentially, but as noted by the reviewer in earlier comments, biologists and researchers often “fall into the trap” of using what they know.
Additionally, interpretation of logistic regression is much more straightforward as it lacks many of the “black box” issues of concern for random forest.
Using both is the approach we advocate.

REVIEWER 2- COMMENT 37

Line 418 – 420: that interpretation of these things is based on widely-accepted rules-of-thumb does not make them objective. it makes them pseudo-objective.

COAUTHOR RESPONSE

By default criteria for the decisions frequentist or Bayesian are “pseudo-objective.”
We did not use the word “objective” in the results and remove the adjective “objective” from the discussion.
However, we still emphasize that taking a standardized approach with “agreed upon rules” has advantages in that everyone is doing the same thing provided those rules are chosen prior to analysis (as we did) and not altered to achieve a desired outcome.

Murtaugh, P. A. (2014). In defense of P values. Ecology, 95(3), 611-617.

Verhulst, B. (2016). In defense of P values. AANA journal, 84(5), 305.

REVIEWER 2- COMMENT 38

Line 423 – 424: meaning is not an inherent quality of statistical significance

COAUTHOR RESPONSE

See above response related to pseudo-objectivity.
This is a philosophical comment that is beyond the scope of this manuscript, so no change is needed or made.

REVIEWER 2- COMMENT 39

Line 424 – 427: again, be careful with this. just because its agreed upon doesn't mean it's correct.

COAUTHOR RESPONSE

See above response related to pseudo-objectivity.
This is a philosophical comment that is beyond the scope of this manuscript, so no change is needed or made.

REVIEWER 2- COMMENT 40

Line 428 – 429: I think the first strength is how easy/automated random forest is.

COAUTHOR RESPONSE

We tend to disagree as the automated throw the data at the wall approach can lead to interpretation of false signals very quickly.
Again, for emphasis, we advocate using both approaches and comparing the outputs because each analysis does a different thing well.

REVIEWER 2 - COMMENT 41

Line 442 – 443: ...by default. no special treatments are required.

COAUTHOR RESPONSE

We modified the text to illustrate the point raised by this reviewer: “In our results, we observed that non-linear relationships for percent sand and stream depth could only be addressed in multiple logistic regression by transformation or parametrization prior to model fitting, which is inherently handled by default in random forest” (Revised Manuscript Lines 485-488).

REVIEWER 2- COMMENT 42

Line 453 – 455: the major risk here is that this essentially leads to p-hacking. your work is not p-hacking because you talk about how you adjusted the analysis, but many others are not so forthcoming. In fairness, hacking an analysis can also be done with a RF so it doesn’t protect against bad-faith actors, but still.

COAUTHOR RESPONSE

This appears to be a general philosophical comment. Thus, no change was suggested or made.

REVIEWER 2- COMMENT 43

Line 460 – 461: i actually view this as a distinct advantage. it forces researchers to think about the results, not just rely on things like p<0.05 as if that statistical threshold has magical decision making properties. It's a major fallacy that p = 0.049 matters and p = 0.051 does not.

COAUTHOR RESPONSE

We agree there are many issues related to interpreting these statistical analyses.
See our previous comments and citations related to pseudo-objectivity above.

REVIEWER 2- COMMENT 44

Line 461 – 462: why is this bad? same point as previous lines.

COAUTHOR RESPONSE

See previous comments and citations related to pseudo-objectivity.
This is a philosophical comment that is beyond the scope of this manuscript, so no change is needed or made.

REVIEWER 2- COMMENT 45

Line 462 – 465: you are allowed to decide the answer to both of these questions. The key is just to be upfront about why you are making certain decisions - but also to discuss the good and the bad about what you've done. And be very careful about making decisions like this just because they make you discussion, and by consequence, your life more convenient. 'The first rule is you must not fool yourself. and you are the easiest person to fool..' Richard Feynman said that.

COAUTHOR RESPONSE

This appears to be a general philosophical comment. Thus, no change was suggested or made.

REVIEWER 2- COMMENT 46

Line 467 – 468: i don't understand this statement. you could use training/testing/validation to examine this. you could also go collect more data to test the predictions. as i said earlier, you don’t have conclusions, you have hypotheses.

COAUTHOR RESPONSE

Training/testing/validation does not solve the problem of ground truthing because the true state in these techniques is assumed via the incomplete data.
We could go collect more data, but that assumes the system is steady state and backwards looking inference of new data to ground truth historic data is relevant. This action assumes that time, money, and personal limitations do not exist.
This is a general problem for the analysis of field data samples so it is beyond the scope of our manuscript.

REVIEWER 2- COMMENT 47

Line 468 – 469: i don’t agree that a MLR natively addresses this.

COAUTHOR RESPONSE

We disagree. MLR is inherently not a black box and relations with coefficients in the model are directly interpretable on the real world scale unless standardized.

REVIEWER 2- COMMENT 48

Line 469 – 473: this is true of all analyses and not specific to RF...and is why one would use either cross-validation and/or training testing splits. Complex fits are likely also over-fits. Stated another way, the RF is likely not very generalizable.

COAUTHOR RESPONSE

True, but this is far more likely in RF relative to MLR because the lack of linearization and inherent generalization of maximum likelihood estimation relative to the random forest algorithm.

REVIEWER 2- COMMENT 49

Line 474: I'm not sure they are true redundancies. they are consistencies, sure. but redundant? I'm not sure about that.

COAUTHOR RESPONSE

The description “redundancy” is quite popular in the ecological assembly literature.
The trends are redundant because we are asking the same question via multiple analysis and getting the same answers, which are “redundant” relative to “just pick one analysis and be done with it”.
Nevertheless, you are correct we could have chosen a different word if we so desired.

REVIEWER 2- COMMENT 50

Line 480 – 482: i think one of the main problems you are dealing with here is that you don’t know what the true relationship is. It's possible that all three are wrong, but because they fail in the same way, it looks good, something like a small simulation study with known and prescribed relationships would be better for comparing how these techniques compare to each other and to the real relationships....including how the different techniques handle ugly data, like outliers and non-linear patterns. Then you could use whatever performs the best to look at the relationships for these fish.

COAUTHOR RESPONSE

A simulation study may be useful, but is beyond the scope of this paper.
Also, a simulation study is not practical for practicing environmental biologists, who have limited time and quantitative skills.

REVIEWER 2- COMMENT 51

Line 495 – 504: there are other things to consider here as well, such as if other features would help - like blocking the data by tributary or maybe if including interactions would change anything. Seems pretty clear that these fish really like south central Kansas (at least relative to other areas).

COAUTHOR RESPONSE

We agree that there are other statistical analyses and refinements available. We tried to choose common statistical approaches that were widely used.
We also tried to keep the analysis as simple as possible.
The points the reviewer suggested are thoughtful but including all possible additions is beyond the scope of the present manuscript.

REVIEWER 2- COMMENT 52

Line 506: to capture?

COAUTHOR RESPONSE

Changed as suggested (Revised Manuscript Line 547).

REVIEWER 2- COMMENT 53

Line 506 – 509: Some people are very uncomfortable with this sort of model dredging - I'm not, but some are. It's worth considering the advantages and disadvantages of an approach like this - although in fairness, adjusting a model iteratively is a main point of residual diagnostics.

COAUTHOR RESPONSE

We agree with the last point and no other reviewer was uncomfortable with this. No change made.

REVIEWER 2- COMMENT 54

Line 509 – 511: that these relationships are likely complex should not be surprising

COAUTHOR RESPONSE

This appears to be a general philosophical comment. Thus, no change was suggested or made.

REVIEWER 2 - COMMENT 55

Line 514: there are also issues with what data are even available to use as predictors. And if they are measured at a scale the fish care about. These are not specific to the stats used, but maybe certain types of analyses are more likely to throw false positives than others. I'm guessing RF is more likely to find spurious correlations, but MLR is more likely to make them look really important (...because of rules like p<0.05). maybe. I don’t know.

COAUTHOR RESPONSE

This would require a very in depth discussion relative to the value gained and is not specific to what was done in this manuscript. No change made.

REVIEWER 2 - COMMENT 56

Line 516 – 517: how many is too many? Are some violations worse than others?

COAUTHOR RESPONSE

Again, this is an intriguing philosophical comment. We agree that it would be valuable to discuss these issues in statistics and research design classes.
As the reviewer knows, there are no clear answers to these questions. Related to the first question, ecologists always have fewer samples than statisticians recommend. Related to the second question there is no clear guidance on “ranking violations.” Hence this discussion is beyond the scope of this paper.
Having said this, we propose here that comparing the results of different approaches can help diagnose whether conclusions are believable or questionable.
As this is a philosophical comment, which is beyond the scope of this manuscript, no change is suggested or made.

REVIEWER 2 - COMMENT 57

Line 525 – 526: Natively? I'm not sure it does this - collinearity can break a MLR cant it?

COAUTHOR RESPONSE

Not natively but based on methodological approach, as noted in the statement. No change made.

REVIEWER 2- COMMENT 58

Line 532 – 533: but they aren't that clear. Another option is to look at some other metric like root mean squared error or mean absolute error.

COAUTHOR RESPONSE

We disagree. See comments and references above related to p-value and pseudo-objectivity.

REVIEWER 2 - COMMENT 59

Line 535 – 537: can doesn't mean does.

COAUTHOR RESPONSE

This statement does not contain absolutes as implied, so no change was needed or made.

REVIEWER 2- COMMENT 60

Line 547 – 550: it's funny because what you are describing here is the bias against how science is actually supposed to be performed. it is exactly the ambiguities which drive investigation. sadly that is not part of the 'publish or perish' paradigm.

COAUTHOR RESPONSE

The reviewer makes many intriguing philosophical comments throughout their detailed review. Although addressing the explicit philosophical issues identified by this reviewer is beyond the scope of this paper, we add two sections to the end of the revised manuscript discussion that addresses some philosophical issues as they relate to our data.
- What is Needed: Realistic Expectations (Revised Manuscript Lines 647-659).
- What Is Needed: Successful Disciplinary Interpretation of Statistical Patterns (Revised Manuscript Line 660-673).

REVIEWER 2- COMMENT 61

Line 550 – 556: IMO your paper is saying to use RF. even its errors tend to be useful....it does have a tendency to overfit, but if you remember that you are deriving hypotheses here, and not conclusions, casting a wide net is a huge advantage. ....keeping in mind the earlier comments about not fooling yourself. sadly another part of science which is obscured by professional science is the need for skepticism especially of our own work and especially when it tells us what our livelihoods depend on.

COAUTHOR RESPONSE

No, our paper is not saying to use RF.
We are saying use both approaches as they both have different strengths, then integrate the results of both analyses in the final ecological interpretation (Revised Manuscript Lines 637-646).

REVIEWER 2- COMMENT 62

Line 561: again, not really. people, for example, will ignore a high p-value when it's convenient to do so, and will also ignore a low pvalue when it is also convenient. Statistical thresholds are supposed to provide some guardrails and shortcuts, but in reality they do neither.

COAUTHOR RESPONSE

This reviewer is bringing up many interesting points about whether we are trained adequately and restrained ethically. This is very important and interesting but goes beyond the points we address here.
See previous comments and references above related to p-values and pseudo-objectivity.

REVIEWER 2- COMMENT 63

Line 567 – 571: again, I think you are being too gracious with not discounting or dismissing MLR. like I suggested already, a small simulation study could help resolve that - you could easily manipulate things like sample sizes and see how sensitive/robust each technique is to the various problems. but undoubtedly what you will find is that statistics are least able to help us when we need them most: when sample sizes are small and/or when data are noisy.

COAUTHOR RESPONSE

See previous comments above on simulation studies.
We think MLR has merit for ecological interpretation. Given that applied statistics are tools to help technical experts understand data patterns, increased interpretability is important.

REVIEWER 2- COMMENT 64

Table 1: ...some unknown fraction of this table is missing.

COAUTHOR RESPONSE

Yes, the journal pasted this table so that half of it was off the page.
Yes, the format of the table is changed to portrait so the journal editors will have no trouble pasting Table 1 in a readable format.
Edits in amount of information and font size were also made to increase the readability of this table as suggested by the reviewer.

REVIEWER 2- COMMENT 65

Figure 1: I suggest making the datapoints bigger and turning down the alpha.

COAUTHOR RESPONSE

The figure was edited to address this reviewer’s comment.

REVIEWER 2- COMMENT 66

Figure 3: I think you can overlay Figures 1 and 3 using colors to discriminate the techniques....might make a nice way to compare and contrast RF and MLR...certainly makes the next paragraph easier to understand.

COAUTHOR RESPONSE

This suggestion complicates results that we are trying to simplify. Thanks for trying to help, but we keep the two plots separate.

REVIEWER 2- COMMENT 67

Figure 4: ....the comparison looks pretty good to me. but, this is potentially highlighting a deficiency in MLR vs RF. a cubic term, or some other higher order polynomial might’ve resolved the discrepancy, but thats just saying that MLR needs more feature engineering than RF.

COAUTHOR RESPONSE

Interesting musing. No change made.

REVIEWER 2- COMMENT 68

Table 2 & 3. I suggest you find a way to convert these tables and figures into a simple heat map which makes these differences obvious.

COAUTHOR RESPONSE

The comparison between estimates and metrics of importance between unstandardized and standardized regression in the tables is complex as is. Any additional information would make the information overly confusing.
We provide specific estimates here to demonstrate the real-world data verses smoothed data, a comparison that we think is important and which can’t be shown in the format suggested above
We also think that translation of these tables into heat maps would both reduce the grain of information available to readers and obfuscate important differences.
Thanks for the suggestion, but we keep our original figure format after making all specific edits suggested by reviewers.

Note in Closing: Reviewer 2 spent a great deal of time commenting on our manuscript (see all 68 of their specific comments!!). We appreciate their efforts and have tried to respond thoughtfully to all comments that applied to the specific points addressed in this manuscript. We believe we have successfully revised the manuscript on all points raised except for those comments that are philosophical and would require a different (but interesting) manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Although the authors clearly explain the methodology, the discussion of the limitations of the random forest method could be more detailed, particularly regarding the interpretation of the ecological meaning of complex models.

The work is methodologically strong but gives less attention to the broader ecological context and potential implications for watershed management and species protection policies.

The example focuses on only one species, so the generality of the conclusions could be better supported by including another case study.

Also, Table 1 must be improved graphically and visually

Author Response

Comments and Suggestions for Authors

REVIEWER 3 – COMMENT 1

REVIEWER 3 – COMMENT 2

The work is methodologically strong but gives less attention to the broader ecological context and potential implications for watershed management and species protection policies.

COAUTHOR RESPONSE COMMENTS 1-2

The reviewer raises some big questions that are largely unresolved issues in the environmental profession [e.g., (1) ecological interpretation of statistics; (2) turning what we know into what we should do). In the revision, we add some text addressing both of the above comments.

We describe the strengths and weaknesses of both statistical approaches in detail (Revised Manuscript Lines 437-506) so we do not add more discussion about the limitations of the random forest model in the revised manuscript
We interpret the integrated statistical results ecologically by showing where the analyses agree (redundancies; Revised Manuscript Lines 507-518), and also clarify where biologists should be skeptical about trends (ambiguities; Revised Manuscript Lines 519-555).
We have added a discussion text section entitled Ecological Insights Gained About the Plains Minnow (Revised Manuscript Lines 613-636) that reviews what our results mean ecologically.
We also add text discussing ecological interpretation and management implications at Revised Manuscript Lines 451-458, 514-517, 530-533, 608-611.
Interpreting what complex analyses mean for the discipline and how they can be applied to planning management actions is a challenge for our profession. We add two sections to our discussion that discuss future needs related to realistic expectations and successful interpretation (Revised Manuscript Lines 647-673).
Related to big picture issues like watershed management and species protection policies, we add some text at the end of the discussion to point out the mismatch between what focused statistical analyses have found, data available for across-site analysis, and what is needed to implement statewide understanding and conservation actions (Revised Manuscript Lines 663-671).
Although no one (including us) can fully answer the questions raised by this reviewer, we have made a substantial effort and have considerable text that addresses this reviewers’ comments.

REVIEWER 3 – COMMENT 3

The example focuses on only one species, so the generality of the conclusions could be better supported by including another case study.

COAUTHOR RESPONSE

Adding other species would make the already long and complex results overwhelming. We appreciate the sentiment expressed relative to generality across taxa, but we intentionally keep the manuscript at an understandable level of information by just reviewing data on a single species.
We give substantial details on our methods for analysis, comparisons, and interpretations so others can repeat and report what we did relative to other species.

REVIEWER 3 – COMMENT 4

Also, Table 1 must be improved graphically and visually

COAUTHOR RESPONSE

We agree that changes in formatting were needed.
The journal pasted this table so that half the table was off the page. We have made changes in orientation, font size, and wording to make this table more readable to address this reviewer concern.

Submission Date; 03 October 2025; Date of this review 19 Oct 2025 10:50:40

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The paper is surely worth publishing, however, I would suggest making it more palatable for readers less versed in modern methods of numerical analysis. This could be achieved by explaining some analytical steps, and devoting more attention to ecological interpretation of the obtained results. Some particular points:

Methods
1. Page 5 last paragraph: “We balanced the number of presence and absence sites by randomly removing absences” – please provide rationale behind this action. One can argue that the ratio absence/presence sites is also important information.
Discussion would be fuller with the following points:
1. On pages 12/13, in the paragraph “Resolving Ambiguity Among Analyses to Better Understand Data Trends …” the Authors, comparing the results obtained by the three studied methods, have singled out the standardised multiple logistic regression. They suggest that it is better than the others because only it ranks some evaluation criteria. How this ranking translates into the plains minnow ecology?
2. In the paragraphs “Resolving Ambiguity Among Analyses – Stream Depth” and “Resolving Ambiguity Among Analyses – Percent Sand”, the Authors supplemented initial analyses by adding squared parameters. The combination parameter + squared parameter panned out analytically, but again – how this mirrors the influence of said parameters upon the fish ecology?
Tables
1. Table 1 does not show in its entirety.
2. Tables 2 and 3 – repeating column names in the table heading is redundant.
Figures
1. Figures 1, 3, 4
  - make axes’ labels black
  - repeating individual graph names in the figure heading is redundant.
2. Figure 2 – repeating individual graph names in the figure heading is redundant.
3. Figure 5

make axes’ labels black.

enlarge font of the axes’ labels.

Language – only one remark – be consistent and stick to the “dam-free” spelling with a hyphen.

Author Response

REVIEWER 4, COMMENT 1

COAUTHOR RESPONSE

Our study is a hybrid of two foci. Our main focus is an evaluation of what can be learned from comparing and integrating two well-described and frequently-used statistical analyses, applied to the same ecological dataset. Our secondary focus is ecological interpretation of these results for the specific taxa under examination (i.e., Plains Minnow).
A very detailed description of how our two statistical analyses work would duplicate existing information in statistical textbooks and miss the novelty of the integration we apply here. Most of the details of the analysis (how they work, output) are in the statistical textbooks that we cite. We don’t want to duplicate existing efforts, so we do not add more statistical details abut analytical steps.
We agree that ecological interpretation is useful. To address this, we have added the following section entitled Ecological Insights Gained About the Plains Minnow (Revised Manuscript Lines 613-636).
We also add text discussing ecological interpretation and management implications at Revised Manuscript Lines 451-458, 514-517, 530-533, 608-611.

Philosophically, we address the importance and difficulty of ecological interpretation in a new discussion section entitled What Is Needed: Successful Disciplinary Interpretation of Statistical Patterns (Revised Manuscript Lines 660-673).

REVIEWER 4, COMMENT 2:

Methods

Page 5 last paragraph: “We balanced the number of presence and absence sites by randomly removing absences” – please provide rationale behind this action.

COAUTHOR RESPONSE

All general and generalized linear models (including linear and logistic regression) are adversely affected when the sample sizes in classes or treatments are very different.
For this reason, sampling and experimental designs propose that the number of samples in each class or treatment be “balanced” or of near equal sizes.
Of several appropriate methods of balancing, we elected to randomly remove absences across the dataset to make the fewest assumptions about the environmental and geospatial distributions of data. This balancing approach avoided zero-inflation concerns which bias statistical results.
Thus, in response to this reviewer comment, we add the following modified text: To avoid zero-inflation, we balanced the number of presence and absence sites by randomly removing absence sites to equal the number of presence sites (Revised Manuscript Lines 209-211).
Because we use a random site removal technique, we have identified elsewhere that the means are similar in original and reduced datasets for presence and absence sites suggesting that the reduction is an acceptable way to meet the assumptions of the analysis.
We agree that the ratio of absence/presence sites can be important information and could be used if we were using an analysis where number of sites was the response variable.

REVIEWER 4, COMMENT 3

Discussion would be fuller with the following points:

On pages 12/13, in the paragraph “Resolving Ambiguity Among Analyses to Better Understand Data Trends …” the Authors, comparing the results obtained by the three studied methods, have singled out the standardised multiple logistic regression. They suggest that it is better than the others because only it ranks some evaluation criteria. How this ranking translates into the plains minnow ecology?

COAUTHOR RESPONSE

We want to emphasize that we think all analyses have strengths and weaknesses and we don’t favor any single analysis. In the discussion section entitled What We Learned from Comparing Then Integrating Multiple Analyses (Objective 4), we show what was gained from using all three analyses (Revised Manuscript Lines 594-612).
Our primary recommendation is to integrate the results of multiple statistical tests on the same dataset because that no statistical tool alone provides all of the ecological information that is available. We added a section near the end of the discussion entitled Our Primary Take-Away and Recommendation to clarify this recommendation (Revised Manuscript Lines 637-646).
The ecological insights gained arise from identifying similarities in regressors across analyses (redundancies; Revised Manuscript Lines 507-518) and unresolved areas across analysis (ambiguities; Revised Manuscript Lines 519-555) that require more investigation.
We also added a section entitled Ecological Insights Gained About the Plains Minnows (Revised Manuscript Lines 613-636) that also addresses what this means ecologically to Plains Minnow.
We also add text discussing ecological interpretation and management implications at Revised Manuscript Lines 451-458, 514-517, 530-533, 608-611.

REVIEWERS 4, COMMENT 4

In the paragraphs “Resolving Ambiguity Among Analyses – Stream Depth” and “Resolving Ambiguity Among Analyses – Percent Sand”, the Authors supplemented initial analyses by adding squared parameters. The combination parameter + squared parameter panned out analytically, but again – how this mirrors the influence of said parameters upon the fish ecology?

COAUTHOR RESPONSE

This is a good point. The ecological impact of adding squared parameters is to create a convex or hump-shaped response. This squared metric suggests that whether probability of fish presence increases with the regressor variable or decreases with the regressor variable depends on where regressor values are relative to the maximum (top of the hump). This interpretation was added to the discussion (Revised Manuscript Lines 530-533).

We also suggest in the discussion that seeking to sample regressor values at the maximum (top of the hump) and inflection points (where the slope changes dramatically) should be priorities in future sampling (Revised Manuscript Lines 608-611).

REVIEWERS 4, COMMENT 5

Table 1 does not show in its entirety.

COAUTHOR RESPONSE

The journal pasted this table so only half of it was visible.
The table was reformatted as a portrait to fit on the page.
We have also edited the wording and fonts to increase readability as this reviewer suggested.

REVIEWERS 4, COMMENT 6

Tables 2 and 3 – repeating column names in the table heading is redundant.

REVIEWERS 4, COMMENT 7

Figures 1, 3, 4
1. make axes’ labels black
2. repeating individual graph names in the figure heading is redundant.

REVIEWERS 4, COMMENT 8

Figure 2 – repeating individual graph names in the figure heading is redundant

REVIEWERS 4, COMMENT 9

Figure 5

make axes’ labels black.

enlarge font of the axes’ labels.

COAUTHOR RESPONSE TO ALL FIGURE COMMENTS 6-9

We have edited the figures to respond to the above reviewer suggestions (Comments 6-9).

REVIEWERS 4, COMMENT 10

Language – only one remark – be consistent and stick to the “dam-free” spelling with a hyphen.

COAUTHOR RESPONSE

Changed as suggested. Good suggestion. Thanks.

Submission Date

03 October 2025

Date of this review

13 Nov 2025 14:46:35

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors:

Your manuscript was improved in according to the reviewer comments, I noted it in marked text.

I suggest accept it in the present form.

Reviewer 2 Report

Comments and Suggestions for Authors

One last comment, although it's a biggie.

While I'm sure the journal will accept the paper, I don't fully agree that some of my previous comments labelled as 'philosophical' were philosophical - many were technical. Regardless, many of those comments were not adequately addressed. Consequently, I do not believe the paper is suitable for publication.