How to write a research paper

Author

Tyler Grimes

Published

October 5, 2026

This note outlines how to write an applied statistics research paper. At this stage, you have already performed a literature review for a research topic, you drafted an analysis plan, you carried out that analysis (possibly adding to/modifying what was planned), and you have your results prepared in a set of tables and figures. Now, it’s time to write.¹

Before you start: what you need in hand

Start writing your paper once you have the following:

Your results, as a set of tables and figures that came from your analysis.
Your notes on what you did: the data source, how the data were cleaned, and which statistical methods you ran. Your analysis code or notebook counts.
Your literature review notes: the papers you read, and what each one found.
A way to manage references, such as Zotero, Mendeley, or EndNote (see Section 10).
A target audience. For a class project this is often your instructor and classmates; for a journal it is that journal’s readers.

1 Terms

Here are some terms used in this guide.

Term	Meaning
Analysis plan	A short document written before the analysis that states the research question, the data, and the statistical methods you intend to use.
Target audience	The specific group of readers you are writing for, and what you can assume they already know.
Tension	The gap, contradiction, or unsolved problem in the existing literature that your paper addresses. It is what makes the reader want to read on. See Section 8.
Reproducible	Someone else could repeat your analysis from your description (and ideally your code) and obtain the same results.
Reference manager	Software that stores your sources and automatically formats the in-text citations and the References section. See Section 10.
Effect size	How big a difference or association is, in the units of the measurement (for example, “0.8 points higher”). A p-value alone does not give this.
Confidence interval (CI)	A range of values that is compatible with the data. A 95% CI is reported alongside an estimate to show its uncertainty.
Association vs. causation	An association means two things vary together. Causation means one produces the change in the other, which usually needs a randomized or carefully designed study.
Limitation	A weakness of your study (for example, a small sample) that affects how far the conclusions can be trusted or generalized.

2 Structure of a research paper

Research papers have a standard structure, which is outlined below.² This structure helps readers quickly navigate a new paper to find what they are looking for. It alleviates some of the burden of reading a paper, because the structure gives the reader a predictable flow of information. If you don’t follow this structure, then the reader will become confused and upset, and they will stop reading.

Abstract: An overview that tells the reader what problem the paper addresses, what was done, the main result, and the conclusion. The abstract should be short, so most details—especially about the methods—will be omitted. (Note: you’ll find published papers with longer abstracts. The expected length and structure of the abstract varies across different research journals.)
Introduction: The reader’s attention is either won or lost here. Within a few sentences you should be building tension that hooks the reader’s interest. You’ll be describing the problem that your paper addresses and convincing the reader that this problem is important to them and they should want to fix it; summarizing the existing literature and background (i.e. the current conversation around this topic); and outlining the solution your paper provides.
Methods: Describe exactly what was done to produce the paper’s results. Things like study design, data sources, data processing, statistical methods, software and version numbers, etc. The Methods should have sufficient information for someone to be able to read it and reproduce all of the results from the paper. Importantly, no actual results should be mentioned in the Methods; it should only state what was done (and be written in past tense).
Results: The analysis should generate tables and figures that record the results, and the Results section walks the reader through those outputs. The results should be stated objectively without explaining why things may have occurred or what implications they have. It should instead focus on what happened and what we observed. Providing interpretations of statistical estimates is OK (for example, giving an interpretation for an estimated regression coefficient), but no conclusions should be drawn from those estimates. We might say “there was a statistically significant difference between group A and group B”, but we should not say “therefore the treatment worked”.
Discussion: The Introduction poses a problem, the Methods state what we did about it, and the Results says what we found. The Discussion closes the loop and ties everything together. Some readers will read the abstract and then jump down to the Discussion before reading the other sections (they’ll only read those if they decide it’s worth their time); you’ll want to keep that reader in mind when writing this section. Start with a brief recap of the problem (summarizing the main points from the introduction), and restate the key findings from the Results. You’ll then explain how those findings fit in with the current literature, what implications they have, and what the reader should take away from this. The strengths and limitations of the study should be discussed (this is essentially a critique on your own methods), which leads into final thoughts on future directions (how this work should be expanded on, or what new challenges it raises).
Conclusion: The Conclusion section is an executive summary: in a few sentences it states what the study accomplished, the key findings, its contribution to the literature, and why that contribution is important.
References: The References section should be automatically populated using a reference manager.

2.1 Templates

These skeletons give you something to fill in. Replace every bracket with specific content from your own paper.

Abstract (four sentences):

[Problem: what is unknown or unresolved, and why it matters]. [What we did: design, data, and main method]. [Main result, with at least one number]. [Conclusion: what the result means for the problem].

Introduction (one paragraph per step; typically three to five paragraphs):

Establish the topic and why it matters. “[Topic] affects [who or what], [with supporting fact] (cite).”
Summarize what is known. “Previous studies found [finding] (cite) and [finding] (cite).”
Build tension. “However, [gap, inconsistency, or overlooked issue]. As a result, [consequence for the reader].”
Resolve it. “In this paper, we [what you did] to [address the gap]. We found [headline result].”

Methods (subsection headings you can reuse): Data source, Data processing, Statistical analysis, Software.

Results (for each table or figure): “[Table/Figure reference] shows [what it displays]. [Key numbers, with units and uncertainty].”

Discussion (in order): Recap of the problem, Key findings, Comparison with the literature, Implications, Limitations, Future directions.

3 Preparation

Before writing, you need to get everything together. I suggest doing this by creating an outline of the paper and dumping contents onto the page. (That way you never have to stare at a blank page.) This step will involve some writing, but do not spend time editing or revising. The goal is to get the bulk of the content organized, so that all the key external information needed to write the paper is contained in the outline. That way, when you start writing, you don’t need to go looking for things.

Here’s what you should prepare for each section:

Abstract: Leave blank.
Introduction: Review and revise the notes from your initial literature review. Put the references in a logical order to build an argument for the existence of the problem(s) that your paper addresses.
Methods: Review your analysis plan and revise it with additional details or changes that were made during the analysis. If you didn’t make an analysis plan, then you should at least have a well documented notebook containing your analysis (e.g. an R markdown or Python Notebook) that you can refer to and retroactively draft an analysis plan. Copy those notes into this section. This should include everything from how the data were obtained and processed to the analyses that were performed (including any simulation studies).
Results: Consolidate the tables and figures that were generated from your analysis, and draft the caption text for each. Organize at a high-level how results will be presented. Insert tables and figures in the approximate order that they will appear. Draft summary text referencing these tables and figures to highlight key findings.
Discussion: Leave blank.
Conclusion: Leave blank.
References: Automatically populate this using a reference manager.

4 Writing the paper

Before you write, you must first determine who your target audience is. What assumptions will you make about your reader? What background knowledge and interests are they expected to have? Use that information to determine the details you need to focus on in order to keep the content understandable and to convey the paper’s value.

When you begin writing, you won’t start at the beginning of the paper. The abstract is impossible to write when you haven’t written the introduction or discussion. The introduction is also difficult when you don’t have the results or discussion written. The most natural starting point is the methods, because that’s where you write about what you’ve done.

Here’s the suggested order for writing, and what you should do in each section:

Methods: Start here. The goal of this section is to provide sufficient information for the reader to reproduce everything you did in your analysis. Describe how the data were collected (if they are from a different study, then briefly summarize how they were collected in that study). Detail any pre-processing steps such as filtering, missing value imputation, variable mutation or derivation, etc. Describe the statistical methods that were used for each analysis performed. References should be included for non-standard statistical methods, as well as for the software or software packages used for computation. No results from the analysis should be mentioned in this section. The methods should be written in past tense—it’s detailing what was done in this study.
Results: What are the main findings of your analysis? You don’t go over every single detail of your results, but instead highlight the key findings from each analysis (with the interests of the target audience in mind). All tables and figures included in the paper should be referred to somewhere in this section. Interpretations of the results are provided, but no conclusions should be drawn or discussed here (that is done in the Discussion section). Every sentence in the Results should be a statement of fact; there should be no conjectures or subjective comments. There also shouldn’t be any external references to the literature here, because we are only writing about what was found in this paper’s analysis. Results should be written in past tense—this is what was found in the analysis.
Discussion: What are the implications of the results? Tie the findings back to the original research question and to other results in the existing literature (you’ll often include additional references here that weren’t used in the introduction). Are the findings consistent with the literature, or are there contradictions? This is where you will provide conjectures or other thoughts to help explain the findings, but those should still be supported by the results or existing literature. Be mindful of the target audience’s interests when deciding what points to emphasize. State any limitations of the analysis, and provide some ideas for future studies.
Introduction: This is the most difficult section to write. The goal is to succinctly summarize the existing literature (i.e. the current “conversation” that your reader is a part of) while building tension—using words like “however”, “although”, “despite”, “yet”, “overlooked”, “inconsistent”, “reported”—to expose the problem that your paper is addressing (see Section 8). The value of your paper is created by the tension that it reveals and resolves. You must (1) build tension in the reader’s mind, and then (2) convince them that this paper provides a resolution. If the reader is convinced of the paper’s value, then they will give it their time and attention. But that attention is quickly lost if the paper is overly verbose, difficult to follow, contains unnecessary information, or is perceived as unimportant. Almost every sentence you write should be a statement of fact supported by one or more references. The remaining sentences provide the narrative that weaves in the value of this paper. The sentences should be organized logically so that it’s easy to follow. The language should be clear and direct while also being persuasive about the existence and importance of the problem being addressed.
Conclusion: In a few sentences, summarize the take-home message of the paper. This focuses on the contribution, significance, and broader impact of the paper.
Abstract: Keep the abstract short—four sentences should be enough. Summarize the problem, what was done, the main result, and the conclusion.
References: Automatically populate this using a reference manager.

5 Example

In this section, we look at a toy example of a hypothetical study. We go through each section of a paper and show a “before” and “after” example text to illustrate what things to look for and think about as you revise.

The hypothetical study: A class project randomly assigned 60 students to either take a 10-minute walk or sit quietly for 10 minutes after lunch. Each student rated their alertness from 1 (very drowsy) to 10 (very alert) at 2:00 pm. The walking group (n = 30) had a mean rating of 6.9 (SD 1.4) and the sitting group (n = 30) had a mean of 6.1 (SD 1.5). Rating differences were analyzed with a two-sample t-test (difference 0.8 points; 95% CI 0.05 to 1.55; p = 0.037).

5.1 Abstract

Before: Alertness is important for students. We did a study on walking and alertness. Walking was found to be somewhat helpful. This could be useful for many people.

After: Afternoon drowsiness is common among students, but few randomized studies have tested whether a short walk reduces it. We randomly assigned 60 students to a 10-minute walk or 10 minutes of quiet sitting after lunch and compared self-rated alertness (1–10 scale) at 2:00 pm. The walking group’s mean alertness was 0.8 points higher (95% CI 0.05 to 1.55; p = 0.037). A short post-lunch walk was associated with higher self-rated alertness in this sample, which supports testing it in a larger and more varied population.

Why the “after” is better: it states the problem, the design, a number with its uncertainty, and a conclusion that doesn’t go further than the data.

5.2 Introduction

Before: Everyone knows that exercise is good for you, and many studies have shown it is important for the brain. Students are often tired in the afternoon, which can be a problem. It would be interesting to see whether walking could help. In this paper we investigate this.

After: Afternoon drowsiness reduces attention during classes held after lunch (Author A, Year). Observational studies report that people who walk more also report higher alertness (Author B, Year; Author C, Year). However, these studies could not separate walking from other differences between active and inactive people, and the few randomized trials used walks of 30 minutes or longer (Author D, Year), which do not fit in a typical lunch break. It is therefore unknown whether a 10-minute walk is enough to raise alertness. In this paper, we randomly assigned students to a 10-minute walk or to quiet sitting after lunch, and measured self-rated alertness at 2:00 pm.

What changed: “Everyone knows” and “important” are replaced by cited facts. The words “however” and “unknown” build tension, and the last sentence resolves it by stating what the paper did.

5.3 Methods

Before: We looked at students and compared walking to not walking. The walking group did better, so we used a t-test.

After: We recruited 60 students from an introductory statistics course in Fall 2026. Students were randomly assigned in a 1:1 ratio to a walking group or a sitting group using a random number generator. After lunch, the walking group walked at a self-selected pace for 10 minutes on a marked campus loop, and the sitting group sat quietly for 10 minutes. At 2:00 pm, each student rated their alertness from 1 (very drowsy) to 10 (very alert) on a paper form. Mean ratings were compared between groups using a two-sample t-test with equal variances, with a two-sided significance level of 0.05. All analyses were conducted in R (version 4.5.0) (R Core Team, Year).

What changed: the “before” contains a result (“did better”), which does not belong in Methods. The “after” is in past tense and gives enough detail that another student could repeat the study.

5.4 Results

Before: Walking clearly worked and made students much more alert, which is probably because exercise wakes up the brain.

After: The walking group (n = 30) had a mean alertness rating of 6.9 (SD 1.4) and the sitting group (n = 30) had a mean rating of 6.1 (SD 1.5) (Table 1). The difference in means was 0.8 points (95% CI 0.05 to 1.55; p = 0.037).

Table 1: Self-rated alertness at 2:00 pm (1–10 scale) by group. These data are hypothetical.

Group	n	Mean alertness	SD
Walking	30	6.9	1.4
Sitting	30	6.1	1.5

What changed: “clearly worked”, “much more”, and “probably because” are removed. The Results state what was observed with numbers and uncertainty, reference an output table/figure, and leave the explanation for the Discussion.

5.5 Discussion

Before: Walking is great and everybody should do it. This proves exercise makes students smarter.

After: Afternoon drowsiness is common among students, and no randomized trial had tested a walk as short as 10 minutes. We found that students assigned to a 10-minute walk rated their alertness 0.8 points higher on a 10-point scale than students assigned to sit (95% CI 0.05 to 1.55). This direction agrees with the observational findings in the Introduction (Author B, Year), and it extends the longer-walk trials (Author D, Year) to a shorter duration. The lower end of the interval is close to zero, so the data are also compatible with a very small effect. This study has limitations: the sample came from a single class, alertness was self-rated and not measured objectively, and students knew which group they were in. Future studies should use an objective measure of alertness, include students from several courses, and test different walk durations.

What changed: the claim “proves … smarter” is replaced by a conclusion that matches the evidence (alertness, not intelligence; association supported by randomization in this sample). Limitations are stated, and the discussion ends with a concrete next step.

5.6 Sentence-level fixes

Vague	Precise
The new model performs better.	The new model had a lower mean absolute error than the baseline (2.1 vs. 3.4 units).
Many participants dropped out.	Eighteen of 120 participants (15%) dropped out before the final visit.
The effect was significant.	The difference was statistically significant (p = 0.037). (Or give the size of the effect.)
The results are important.	The results suggest that [specific finding] for [specific audience].
The sample was fairly large.	The sample included 1,204 participants.

6 Revising your draft

Once you have a first draft completed, congratulations! That’s a massive first step. As you review and revise this draft, remember that your writing is not meant to be permanent: it’s meant to move a conversation forward. Don’t try to perfect every sentence, but do address issues that would make a target reader want to stop reading.

6.2 Getting feedback

Ask someone to read your draft before you submit it. A reader who has not seen your analysis is the best test of whether the paper stands on its own.

Choose the reader. A classmate shows you where the text is unclear. An instructor or advisor shows you whether the content is correct.
Ask specific questions. “Can you tell what problem this paper addresses after reading the Introduction?” and “Where did you stop following?” produce more useful answers than “Is it good?”
Give them time and a clean draft. Run the checklist above first, so the reader spends time on content rather than typos.
Respond to comments with a list. For each comment, either make the change or note why you didn’t. If two readers are confused by the same sentence, rewrite it.

6.3 Find and replace

There are words we use in every-day conversations that don’t belong in a research paper. During the revision process, use the “find and replace” feature of your text editor to systematically check for these words. Table 2 lists words that should be deleted on sight; these are words that are subjective or imprecise, and the sentence is often improved by simply deleting them.³ Table 3 lists words that can also be problematic; rather than delete, these should often⁴ be replaced with more precise language.

Table 2: Delete on sight. These words don’t provide scientific meaning.

Category	Problem	Words/Phrases	Instead
Empty intensifiers	Adds subjective emphasis or strength without measurable information.	considerably, dramatically, enormously, exceptionally, extremely, highly, quite, remarkably, substantially, very	Before: “a very large difference.” After: “a difference of 12 units.”
Conversational filler	Reflects informal speech patterns rather than precise scientific communication.	actually, basically, clearly, essentially, in fact, indeed, just, kind of, like, of course, really, simply, sort of	Before: “We just removed the outliers.” After: “We removed observations beyond 3 SD.”
Unsupported certainty markers	Implies that a conclusion is obvious or indisputable without providing justification.	certainly, clearly, obviously, of course, undoubtedly	Before: “Obviously, the model fits.” After: “The model fit well (R^2 = 0.82).”
Unnecessary emphasis	Expresses importance or interest without defining the basis for that judgment.	important, interesting, notable, remarkable, significant†	Before: “An important finding.” After: “A finding that changes how [X] is estimated.”
Vague emphasis	Emphasizes size, importance, or magnitude without specifying a measurable criterion.	considerable, major, minor, substantial	Before: “A major drop.” After: “A drop of 40%.”

† About “significant”

The word significant is allowed when you mean statistical significance, ideally with the test and p-value or interval, as in “the difference was statistically significant (p = 0.037).” Delete it when it is used as a synonym for “important” or “large”.

Table 3: Stop and justify — These words are not automatically problematic, but try to replace them with more precise language.

Category	Problem	Words/Phrases	Instead
Uncertainty and possibility	Indicates uncertainty or potential explanations, but may be unnecessary or insufficiently qualified.	apparently, arguably, can, conceivably, could, maybe, may, might, perhaps, potentially, possibly, presumably, seemingly, would	Before: “This may be because of age.” After: “Age differed between groups (mean 24 vs. 31 years), which could explain the difference.”
Frequency	Describes how often something occurs but lacks a defined frequency or population.	commonly, frequently, generally, largely, mostly, often, rarely, typically, usually	Before: “Missing data were rare.” After: “Missing data affected 3% of rows.”
Magnitude and approximation	Describes size, degree, or closeness but may be replaceable with a specific value or range.	about, almost, approximately, around, fairly, nearly, relatively, roughly, slightly, somewhat	Before: “Nearly half the sample.” After: “47% of the sample.”
Comparisons	Makes comparative claims without specifying the reference point or metric.	better, decreased, improved, increased, higher, larger, lower, smaller, stronger, weaker, worse	Before: “Scores improved.” After: “Scores increased by 4 points from baseline.”
Quality judgments	Evaluates performance, suitability, or importance without defining the criteria used.	adequate, appropriate, effective, efficient, meaningful, optimal, reasonable, robust, suitable, sufficient, useful	Before: “An effective treatment.” After: “A treatment that reduced symptoms by 30%.”
Abstract nouns	Refers to broad concepts that obscures the specific mechanism, variable, or process being discussed.	aspect, behavior, characteristic, concept, context, dynamics, effect, environment, factor, framework, impact, influence, issue, mechanism, phenomenon, process, property, relationship, scenario, situation, system, thing	Before: “The relationship between the variables.” After: “The correlation between age and income.”

7 Common mistakes

These are frequent in first papers. Each has a fix.

Putting results in the Methods, or conclusions in the Results. Instead: Methods say what was done, Results say what was found, and the Discussion says what it means.
Reporting a p-value alone. Instead: Give the effect size and its confidence interval, and then the p-value.
Saying “significant” without saying what kind. Instead: Write “statistically significant” and give the test and p-value, or choose a more specific word.
Claiming causation from observational data. Instead: Use “associated with”, and discuss what else could explain the pattern.
Overstating the conclusion. Instead: Conclude only what the results support, and state the limitations.
Citing papers you have not read. Instead: Read at least the abstract and relevant results of every source, and cite the original source of a finding rather than a paper that mentions it.
Using the wrong tense. Instead: Methods and Results are in past tense (“we assigned”, “the difference was”). The Introduction and Discussion mix tenses: present for established facts, past for specific studies.
Leaving figures and tables unreferenced. Instead: Every figure and table is cited in the Results by its number.
Pasting in text from an AI tool without checking it. AI tools can produce fluent text that contains wrong statements and references that do not exist. Check your course or journal policy on AI use, and verify every claim and every citation yourself, because you are responsible for them.
Writing the Introduction first and getting stuck. Instead: Write the Methods first (see Section 4).

8 The craft of writing effectively

One of the most impactful videos on the way I think about writing is Larry McEnerny’s lecture on the craft of writing effectively. These are the key ideas that have stuck with me:

Writing has two functions:
1. Writing for yourself, to help you think about the world.
2. Writing for others, to change the way they think about the world.
Writing for others is not about conveying your ideas to the reader; it’s about getting the reader to adopt your ideas. This requires a different writing style compared to when you are writing for yourself.
Your writing is not meant to be permanent, because knowledge isn’t cumulative. Knowledge is a conversation moving through time, and your writing is meant to move the conversation forward.
Newness or originality is not important. Value is important. Value drives the conversation forward.
Build tension to create value. Tell the reader what is wrong or missing in the current conversation: draw attention to their problem and the importance of fixing it, then offer your solution. The value of your paper is created by the tension that it reveals and resolves.
If the reader doesn’t think your paper is valuable, then it doesn’t matter how persuasive, organized, or clear it is. They have no motivation to read it, and they won’t.

Tension is built through sentence structure and emphasis, a topic that is masterfully demonstrated in Judy Swan’s workshop lecture Scientific Writing: Beyond Tips and Tricks.

Another general tip for scientific writing is to always be precise and objective. Avoid vague statements that make the reader ask, “what does that mean?” (this is the most frequent feedback I give when reviewing papers). Don’t hedge your statements with modifiers like somewhat, fairly, nearly, etc. Furthermore, don’t use vague words that don’t have a precise meaning. It is better to be clear and possibly wrong, than to be vague and “not even wrong”. That is, if you make a statement that can’t be disproved, then you should clarify it until it can be.

Here are some examples of statements that will make the reader ask “what does that mean?” (in bold is what needs to be fixed):

“The new algorithm is somewhat better than previous methods.” (Better by what metric? What does “somewhat” mean?)
“The treatment produces nearly optimal outcomes.” (nearly optimal means… what? Sub-optimal?)
“The intervention is mostly effective.” (So it’s ineffective in some way? Does “mostly” mean 51% effective? … 99% effective? Effective according to what metric?)
“The estimates are more realistic.” (How is “realism” measured? How much is “more”? More compared to what?)
“The variables exhibit a dynamic relationship.” (… what in the world does that even mean?)
“This method yields meaningful results.” (Meaningful in what way?)
“The analysis provides important insights.” (Why is it important?)

For the last two, you never tell the reader what is meaningful or important. Different readers have different interests. The paper will emphasize key findings that the target audience should be interested in, but it’s left to the individual reader to decide whether those findings are meaningful or important to them.

8.1 Other resources

Here’s a collection of lectures/talks about scientific writing that I’ve learned from:

Judy Swan - Scientific Writing: Beyond Tips and Tricks. The best place to get started for scientific writing. Covers the fundamentals of sentence structure and emphasis.
Larry McEnerny - The Craft of Writing Effectively. Drives home the importance of creating value with your writing.
Simon Peyton Jones - How to write a great research paper. This one is targeted toward computer science research, but it is full of great ideas and gives motivation for why you should write.
Pete Carr - How to Write a Paper in a Weekend. A concise guide to getting a paper ready for publication.
Anthony Newman - How to Write a Great Research Paper, and Get it Accepted by a Good Journal. A longer guide to getting a paper ready for publication.

9 Examples of published papers

Published papers are the best models of this structure, and every scientific journal has examples. After you have read this guide and worked through Section 5, read a few published papers related to your research topic and look for this structure. You’ll find some variation across journals, but the underlying structure and flow of the paper should be recognizable.

How to read a paper as a writer. The papers below are written for researchers and some are technical. You don’t need to follow all of the statistics. Read them in this order, and write one or two sentences of notes for each step:

Abstract. Does it state the problem, what was done, the main result, and the conclusion?
Introduction. Find the sentence where the authors turn from what is known to what is missing (often marked by “however”). That is the tension.
Methods. Is it in past tense? How much detail is given? Could you repeat the analysis?
Results. Find a table or figure, and then find the sentence in the text that summarizes it. Are there any explanations of why, or are they stated as facts?
Discussion. Find the recap of the problem, the comparison with other studies, and the limitations.

If you find these papers too advanced, ask your instructor or librarian for a paper in your own field from a recent issue of a journal that your course already uses. The structure is the same.

In case you don’t have a research topic in mind, here are a few papers to look at:

10 Tools

You can write a good paper with many different tools. Pick the option that matches your experience and your course requirements.

10.1 Option A: a reproducible workflow with Quarto

This option is useful if you already use R or Python, or if you want to learn. It makes every number in the paper trace back to code.

For running the analysis, use a reproducible workflow from start to finish (i.e. from data loading and pre-processing to table and figure generation). All results of the analysis must be captured in a table, figure, or listing, and these should be automatically generated by the code used to run the analysis.
In the analysis code, save figures as .pdf or .png with the desired height, width, and DPI (if using a raster format). Use the gt, flextable, or similar packages to create tables, then store those table objects as .rds files.
Use Quarto to compose the report in a .qmd file. This allows the use of markdown and LaTeX to write and format the report. Directly embed generated figures with caption text. Load table objects within an R code chunk to render the formatted tables with caption text.
Use a reference manager to create a .bib file for references, and use a .csl file for reference formatting.

10.2 Option B: word processor and reference manager

This option is simple and works for most class papers.

Write in Word or Google Docs. Use the built-in heading styles for section titles, and use the caption feature for tables and figures so they are numbered automatically.
Use a reference manager such as Zotero (free) or Mendeley, with its word processor plugin. It inserts citations and builds the References section in the style your course requires, so you never format references by hand.
Save figures as high-resolution .png files, or .pdf when the program allows it, and paste them in with a caption. Keep the code or software steps that produced them in a separate file, so you can regenerate them.
Keep a dated copy of each draft, so you can go back after a bad edit.

Footnotes

In reality, you should be writing all along. It’s useful to begin drafting the paper during the research planning and analysis stage, especially the Methods section. At a minimum, you should be keeping a detailed record of the analyses you’re running and some notes about the findings—an R markdown file or Python Notebook is good for this. However, you don’t want to spend too much time revising every sentence of these notes, because this writing is only for you and not for the final paper.↩︎
Papers in mathematics or theoretical statistics usually won’t have a Methods and Results section, and instead those are replaced by sections that organize the theory being presented. They will, however, have the Introduction and Discussion/Conclusions, and the importance of writing in a way that conveys value to the reader still remains.↩︎
An initial draft of these two tables was created using generative AI.↩︎
I say “often” here because sometimes needless words can be useful for constructing a desired sentence structure. See Judy Swan’s lecture (41:08) for more on using extra words to create emphasis.↩︎

--- title: "How to write a research paper" author: "Tyler Grimes" date: "2026-10-05" format: html: number-sections: true image: how-to-write-a-paper.png --- This note outlines how to write an applied statistics research paper. At this stage, you have already performed a literature review for a research topic, you drafted an analysis plan, you carried out that analysis (possibly adding to/modifying what was planned), and you have your results prepared in a set of tables and figures. Now, it's time to write.^[In reality, you should be writing all along. It's useful to begin drafting the paper during the research planning and analysis stage, especially the Methods section. At a minimum, you should be keeping a detailed record of the analyses you're running and some notes about the findings—an R markdown file or Python Notebook is good for this. However, you don't want to spend too much time revising every sentence of these notes, because this writing is only for you and not for the final paper.] ::: {.callout-note} ## Before you start: what you need in hand Start writing your paper once you have the following: - **Your results**, as a set of tables and figures that came from your analysis. - **Your notes on what you did**: the data source, how the data were cleaned, and which statistical methods you ran. Your analysis code or notebook counts. - **Your literature review notes**: the papers you read, and what each one found. - **A way to manage references**, such as Zotero, Mendeley, or EndNote (see @sec-tools). - **A target audience.** For a class project this is often your instructor and classmates; for a journal it is that journal's readers. ::: # Terms {#sec-glossary} Here are some terms used in this guide. | Term | Meaning | |:--------------|:-----------------------------------------------------------| | Analysis plan | A short document written *before* the analysis that states the research question, the data, and the statistical methods you intend to use. | | Target audience | The specific group of readers you are writing for, and what you can assume they already know. | | Tension | The gap, contradiction, or unsolved problem in the existing literature that your paper addresses. It is what makes the reader want to read on. See @sec-effective. | | Reproducible | Someone else could repeat your analysis from your description (and ideally your code) and obtain the same results. | | Reference manager | Software that stores your sources and automatically formats the in-text citations and the References section. See @sec-tools. | | Effect size | How big a difference or association is, in the units of the measurement (for example, "0.8 points higher"). A p-value alone does not give this. | | Confidence interval (CI) | A range of values that is compatible with the data. A 95% CI is reported alongside an estimate to show its uncertainty. | | Association vs. causation | An association means two things vary together. Causation means one produces the change in the other, which usually needs a randomized or carefully designed study. | | Limitation | A weakness of your study (for example, a small sample) that affects how far the conclusions can be trusted or generalized. | # Structure of a research paper {#sec-structure} Research papers have a standard structure, which is outlined below.^[Papers in mathematics or theoretical statistics usually won't have a Methods and Results section, and instead those are replaced by sections that organize the theory being presented. They will, however, have the Introduction and Discussion/Conclusions, and the importance of writing in a way that conveys value to the reader still remains.] This structure helps readers quickly navigate a new paper to find what they are looking for. It alleviates some of the burden of reading a paper, because the structure gives the reader a predictable flow of information. If you don't follow this structure, then the reader will become confused and upset, and they will stop reading. - **Abstract**: An overview that tells the reader what problem the paper addresses, what was done, the main result, and the conclusion. The abstract should be short, so most details—especially about the methods—will be omitted. (Note: you'll find published papers with longer abstracts. The expected length and structure of the abstract varies across different research journals.) - **Introduction**: The reader's attention is either won or lost here. Within a few sentences you should be building tension that hooks the reader's interest. You'll be describing the problem that your paper addresses and convincing the reader that this problem is important to them and they should want to fix it; summarizing the existing literature and background (i.e. the current conversation around this topic); and outlining the solution your paper provides. - **Methods**: Describe exactly what was done to produce the paper's results. Things like study design, data sources, data processing, statistical methods, software and version numbers, etc. The Methods should have sufficient information for someone to be able to read it and reproduce all of the results from the paper. Importantly, no actual results should be mentioned in the Methods; it should only state what was done (and be written in past tense). - **Results**: The analysis should generate tables and figures that record the results, and the Results section walks the reader through those outputs. The results should be stated objectively without explaining *why* things may have occurred or what implications they have. It should instead focus on *what happened and what we observed.* Providing interpretations of statistical estimates is OK (for example, giving an interpretation for an estimated regression coefficient), but no conclusions should be drawn from those estimates. We might say "there was a statistically significant difference between group A and group B", but we should *not* say "therefore the treatment worked". - **Discussion**: The Introduction poses a problem, the Methods state what we did about it, and the Results says what we found. The Discussion closes the loop and ties everything together. Some readers will read the abstract and then jump down to the Discussion before reading the other sections (they'll only read those if they decide it's worth their time); you'll want to keep that reader in mind when writing this section. Start with a brief recap of the problem (summarizing the main points from the introduction), and restate the key findings from the Results. You'll then explain how those findings fit in with the current literature, what implications they have, and what the reader should take away from this. The strengths and limitations of the study should be discussed (this is essentially a critique on your own methods), which leads into final thoughts on future directions (how this work should be expanded on, or what new challenges it raises). - **Conclusion**: The Conclusion section is an executive summary: in a few sentences it states what the study accomplished, the key findings, its contribution to the literature, and why that contribution is important. - **References**: The References section should be automatically populated using a reference manager. ## Templates These skeletons give you something to fill in. Replace every bracket with specific content from your own paper. **Abstract** (four sentences): > [Problem: what is unknown or unresolved, and why it matters]. [What we did: design, data, and main method]. [Main result, with at least one number]. [Conclusion: what the result means for the problem]. **Introduction** (one paragraph per step; typically three to five paragraphs): 1. *Establish the topic and why it matters.* "[Topic] affects [who or what], [with supporting fact] (cite)." 2. *Summarize what is known.* "Previous studies found [finding] (cite) and [finding] (cite)." 3. *Build tension.* "However, [gap, inconsistency, or overlooked issue]. As a result, [consequence for the reader]." 4. *Resolve it.* "In this paper, we [what you did] to [address the gap]. We found [headline result]." **Methods** (subsection headings you can reuse): Data source, Data processing, Statistical analysis, Software. **Results** (for each table or figure): "[Table/Figure reference] shows [what it displays]. [Key numbers, with units and uncertainty]." **Discussion** (in order): Recap of the problem, Key findings, Comparison with the literature, Implications, Limitations, Future directions. # Preparation {#sec-prep} Before writing, you need to get everything together. I suggest doing this by creating an outline of the paper and dumping contents onto the page. (That way you never have to stare at a blank page.) This step will involve some writing, but do not spend time editing or revising. The goal is to get the bulk of the content organized, so that all the key external information needed to write the paper is contained in the outline. That way, when you start writing, you don't need to go looking for things. Here's what you should prepare for each section: - **Abstract**: Leave blank. - **Introduction**: Review and revise the notes from your initial literature review. Put the references in a logical order to build an argument for the existence of the problem(s) that your paper addresses. - **Methods**: Review your analysis plan and revise it with additional details or changes that were made during the analysis. If you didn't make an analysis plan, then you should at least have a well documented notebook containing your analysis (e.g. an R markdown or Python Notebook) that you can refer to and retroactively draft an analysis plan. Copy those notes into this section. This should include everything from how the data were obtained and processed to the analyses that were performed (including any simulation studies). - **Results**: Consolidate the tables and figures that were generated from your analysis, and draft the caption text for each. Organize at a high-level how results will be presented. Insert tables and figures in the approximate order that they will appear. Draft summary text referencing these tables and figures to highlight key findings. - **Discussion**: Leave blank. - **Conclusion**: Leave blank. - **References**: Automatically populate this using a reference manager. # Writing the paper {#sec-writing} Before you write, you must first determine who your target audience is. What assumptions will you make about your reader? What background knowledge and interests are they expected to have? Use that information to determine the details you need to focus on in order to keep the content understandable and to convey the paper's value. When you begin writing, you won't start at the beginning of the paper. The abstract is impossible to write when you haven't written the introduction or discussion. The introduction is also difficult when you don't have the results or discussion written. The most natural starting point is the methods, because that's where you write about what you've done. Here's the suggested order for writing, and what you should do in each section: - **Methods**: Start here. The goal of this section is to provide sufficient information for the reader to reproduce everything you did in your analysis. Describe how the data were collected (if they are from a different study, then briefly summarize how they were collected in that study). Detail any pre-processing steps such as filtering, missing value imputation, variable mutation or derivation, etc. Describe the statistical methods that were used for each analysis performed. References should be included for non-standard statistical methods, as well as for the software or software packages used for computation. No results from the analysis should be mentioned in this section. The methods should be written in past tense—it's detailing what *was* done in this study. - **Results**: What are the main findings of your analysis? You don't go over every single detail of your results, but instead highlight the key findings from each analysis (with the interests of the target audience in mind). All tables and figures included in the paper should be referred to somewhere in this section. Interpretations of the results are provided, but no conclusions should be drawn or discussed here (that is done in the Discussion section). Every sentence in the Results should be a statement of fact; there should be no conjectures or subjective comments. There also shouldn't be any external references to the literature here, because we are only writing about what was found in *this* paper's analysis. Results should be written in past tense—this is what *was* found in the analysis. - **Discussion**: What are the implications of the results? Tie the findings back to the original research question and to other results in the existing literature (you'll often include additional references here that weren't used in the introduction). Are the findings consistent with the literature, or are there contradictions? This is where you will provide conjectures or other thoughts to help explain the findings, but those should still be supported by the results or existing literature. Be mindful of the target audience's interests when deciding what points to emphasize. State any limitations of the analysis, and provide some ideas for future studies. - **Introduction**: This is the most difficult section to write. The goal is to succinctly summarize the existing literature (i.e. the current "conversation" that your reader is a part of) while building tension—using words like "however", "although", "despite", "yet", "overlooked", "inconsistent", "reported"—to expose the problem that your paper is addressing (see @sec-effective). *The value of your paper is created by the tension that it reveals and resolves.* You must (1) build tension in the reader's mind, and then (2) convince them that this paper provides a resolution. If the reader is convinced of the paper's value, then they will give it their time and attention. But that attention is quickly lost if the paper is overly verbose, difficult to follow, contains unnecessary information, or is perceived as unimportant. Almost every sentence you write should be a statement of fact supported by one or more references. The remaining sentences provide the narrative that weaves in the value of this paper. The sentences should be organized logically so that it's easy to follow. The language should be clear and direct while also being persuasive about the existence and importance of the problem being addressed. - **Conclusion**: In a few sentences, summarize the take-home message of the paper. This focuses on the contribution, significance, and broader impact of the paper. - **Abstract**: Keep the abstract short—four sentences should be enough. Summarize the problem, what was done, the main result, and the conclusion. - **References**: Automatically populate this using a reference manager. # Example {#sec-example} In this section, we look at a toy example of a hypothetical study. We go through each section of a paper and show a "before" and "after" example text to illustrate what things to look for and think about as you revise. **The hypothetical study:** A class project randomly assigned 60 students to either take a 10-minute walk or sit quietly for 10 minutes after lunch. Each student rated their alertness from 1 (very drowsy) to 10 (very alert) at 2:00 pm. The walking group (n = 30) had a mean rating of 6.9 (SD 1.4) and the sitting group (n = 30) had a mean of 6.1 (SD 1.5). Rating differences were analyzed with a two-sample t-test (difference 0.8 points; 95% CI 0.05 to 1.55; p = 0.037). ## Abstract > **Before:** Alertness is important for students. We did a study on walking and alertness. Walking was found to be somewhat helpful. This could be useful for many people. > > **After:** Afternoon drowsiness is common among students, but few randomized studies have tested whether a short walk reduces it. We randomly assigned 60 students to a 10-minute walk or 10 minutes of quiet sitting after lunch and compared self-rated alertness (1–10 scale) at 2:00 pm. The walking group's mean alertness was 0.8 points higher (95% CI 0.05 to 1.55; p = 0.037). A short post-lunch walk was associated with higher self-rated alertness in this sample, which supports testing it in a larger and more varied population. *Why the "after" is better:* it states the problem, the design, a number with its uncertainty, and a conclusion that doesn't go further than the data. ## Introduction > **Before:** Everyone knows that exercise is good for you, and many studies have shown it is important for the brain. Students are often tired in the afternoon, which can be a problem. It would be interesting to see whether walking could help. In this paper we investigate this. > > **After:** Afternoon drowsiness reduces attention during classes held after lunch (Author A, Year). Observational studies report that people who walk more also report higher alertness (Author B, Year; Author C, Year). However, these studies could not separate walking from other differences between active and inactive people, and the few randomized trials used walks of 30 minutes or longer (Author D, Year), which do not fit in a typical lunch break. It is therefore unknown whether a 10-minute walk is enough to raise alertness. In this paper, we randomly assigned students to a 10-minute walk or to quiet sitting after lunch, and measured self-rated alertness at 2:00 pm. *What changed:* "Everyone knows" and "important" are replaced by cited facts. The words "however" and "unknown" build tension, and the last sentence resolves it by stating what the paper did. ## Methods > **Before:** We looked at students and compared walking to not walking. The walking group did better, so we used a t-test. > > **After:** We recruited 60 students from an introductory statistics course in Fall 2026. Students were randomly assigned in a 1:1 ratio to a walking group or a sitting group using a random number generator. After lunch, the walking group walked at a self-selected pace for 10 minutes on a marked campus loop, and the sitting group sat quietly for 10 minutes. At 2:00 pm, each student rated their alertness from 1 (very drowsy) to 10 (very alert) on a paper form. Mean ratings were compared between groups using a two-sample t-test with equal variances, with a two-sided significance level of 0.05. All analyses were conducted in R (version 4.5.0) (R Core Team, Year). *What changed:* the "before" contains a result ("did better"), which does not belong in Methods. The "after" is in past tense and gives enough detail that another student could repeat the study. ## Results > **Before:** Walking clearly worked and made students much more alert, which is probably because exercise wakes up the brain. > > **After:** The walking group (n = 30) had a mean alertness rating of 6.9 (SD 1.4) and the sitting group (n = 30) had a mean rating of 6.1 (SD 1.5) (@tbl-example). The difference in means was 0.8 points (95% CI 0.05 to 1.55; p = 0.037). | Group | n | Mean alertness | SD | |:----------|:--:|:--:|:--:| | Walking | 30 | 6.9 | 1.4 | | Sitting | 30 | 6.1 | 1.5 | : Self-rated alertness at 2:00 pm (1–10 scale) by group. These data are hypothetical. {#tbl-example} *What changed:* "clearly worked", "much more", and "probably because" are removed. The Results state what was observed with numbers and uncertainty, reference an output table/figure, and leave the explanation for the Discussion. ## Discussion > **Before:** Walking is great and everybody should do it. This proves exercise makes students smarter. > > **After:** Afternoon drowsiness is common among students, and no randomized trial had tested a walk as short as 10 minutes. We found that students assigned to a 10-minute walk rated their alertness 0.8 points higher on a 10-point scale than students assigned to sit (95% CI 0.05 to 1.55). This direction agrees with the observational findings in the Introduction (Author B, Year), and it extends the longer-walk trials (Author D, Year) to a shorter duration. The lower end of the interval is close to zero, so the data are also compatible with a very small effect. This study has limitations: the sample came from a single class, alertness was self-rated and not measured objectively, and students knew which group they were in. Future studies should use an objective measure of alertness, include students from several courses, and test different walk durations. *What changed:* the claim "proves ... smarter" is replaced by a conclusion that matches the evidence (alertness, not intelligence; association supported by randomization in this sample). Limitations are stated, and the discussion ends with a concrete next step. ## Sentence-level fixes | Vague | Precise | |:--------------------------------------|:-------------------------------------------------------| | The new model performs better. | The new model had a lower mean absolute error than the baseline (2.1 vs. 3.4 units). | | Many participants dropped out. | Eighteen of 120 participants (15%) dropped out before the final visit. | | The effect was significant. | The difference was statistically significant (p = 0.037). *(Or give the size of the effect.)* | | The results are important. | The results suggest that [specific finding] for [specific audience]. | | The sample was fairly large. | The sample included 1,204 participants. | # Revising your draft {#sec-revisions} Once you have a first draft completed, congratulations! That's a massive first step. As you review and revise this draft, remember that your writing is not meant to be permanent: it's meant to move a conversation forward. Don't try to perfect every sentence, but do address issues that would make a target reader want to stop reading. ## Revision checklist {#sec-checklist} Work through this list on each draft. Each item is a yes/no question; if the answer is not what the item expects, fix it. **Audience and story** - [ ] Does the writing align with the interests and background knowledge of the target audience? - [ ] Are you building tension in the introduction as you outline the current literature and motivate the problem that your paper addresses? - [ ] Are the sentences and paragraphs organized in a logical structure? - [ ] Are there any unnecessary sentences? (Delete them.) - [ ] Are there any vague or hedged statements? (See @sec-find-replace.) **Methods and Results** - [ ] Are there any results in the Methods? (There should not be.) - [ ] Is there enough information in the Methods for someone to be able to independently re-run all of the paper's analyses? - [ ] Is the Methods section written in past tense? - [ ] Is there any discussion in the Results? (There should not be.) - [ ] Are all the tables and figures included in the paper referenced in the Results? - [ ] Does every number in the text match the table or figure it came from? **Discussion, Conclusion, and References** - [ ] If the reader skips down to the Discussion, will they find a cohesive story of this paper's findings and contributions? - [ ] Are the conclusions of the paper supported by the paper's results? (Make sure the conclusions aren't overstated.) - [ ] Does the conclusion of the paper address the problem that the introduction said the paper would address? - [ ] Is every citation a source you have read, and does it appear in the References? ## Getting feedback Ask someone to read your draft before you submit it. A reader who has not seen your analysis is the best test of whether the paper stands on its own. - **Choose the reader.** A classmate shows you where the text is unclear. An instructor or advisor shows you whether the content is correct. - **Ask specific questions.** "Can you tell what problem this paper addresses after reading the Introduction?" and "Where did you stop following?" produce more useful answers than "Is it good?" - **Give them time and a clean draft.** Run the checklist above first, so the reader spends time on content rather than typos. - **Respond to comments with a list.** For each comment, either make the change or note why you didn't. If two readers are confused by the same sentence, rewrite it. ## Find and replace {#sec-find-replace} There are words we use in every-day conversations that don't belong in a research paper. During the revision process, use the "find and replace" feature of your text editor to systematically check for these words. @tbl-delete lists words that should be deleted on sight; these are words that are subjective or imprecise, and the sentence is often improved by simply deleting them.^[An initial draft of these two tables was created using generative AI.] @tbl-justify lists words that can also be problematic; rather than delete, these should often^[I say "often" here because sometimes needless words can be useful for constructing a desired sentence structure. See Judy Swan's [lecture (41:08)](https://youtu.be/jLPCdDp_LE0?si=YPUkbQqH6NdLNGwM&t=2468) for more on using extra words to create emphasis.] be replaced with more precise language. | **Category** | **Problem** | **Words/Phrases** | **Instead** | |:-----------|:--------------------------------|:----------------------------------------------------|:------------------------------| | Empty intensifiers | Adds subjective emphasis or strength without measurable information. | considerably, dramatically, enormously, exceptionally, extremely, highly, quite, remarkably, substantially, very | *Before:* "a very large difference." *After:* "a difference of 12 units." | | Conversational filler | Reflects informal speech patterns rather than precise scientific communication. | actually, basically, clearly, essentially, in fact, indeed, just, kind of, like, of course, really, simply, sort of | *Before:* "We just removed the outliers." *After:* "We removed observations beyond 3 SD." | | Unsupported certainty markers | Implies that a conclusion is obvious or indisputable without providing justification. | certainly, clearly, obviously, of course, undoubtedly | *Before:* "Obviously, the model fits." *After:* "The model fit well ($R^2 = 0.82$)." | | Unnecessary emphasis | Expresses importance or interest without defining the basis for that judgment. | important, interesting, notable, remarkable, significant† | *Before:* "An important finding." *After:* "A finding that changes how [X] is estimated." | | Vague emphasis | Emphasizes size, importance, or magnitude without specifying a measurable criterion. | considerable, major, minor, substantial | *Before:* "A major drop." *After:* "A drop of 40%." | : Delete on sight. These words don't provide scientific meaning. {#tbl-delete} ::: {.callout-note} ## † About "significant" The word *significant* is allowed when you mean **statistical significance**, ideally with the test and p-value or interval, as in "the difference was statistically significant (p = 0.037)." Delete it when it is used as a synonym for "important" or "large". ::: | **Category** | **Problem** | **Words/Phrases** | **Instead** | |:-----------|:--------------------------------|:----------------------------------------------------|:------------------------------| | Uncertainty and possibility | Indicates uncertainty or potential explanations, but may be unnecessary or insufficiently qualified. | apparently, arguably, can, conceivably, could, maybe, may, might, perhaps, potentially, possibly, presumably, seemingly, would | *Before:* "This may be because of age." *After:* "Age differed between groups (mean 24 vs. 31 years), which could explain the difference." | | Frequency | Describes how often something occurs but lacks a defined frequency or population. | commonly, frequently, generally, largely, mostly, often, rarely, typically, usually | *Before:* "Missing data were rare." *After:* "Missing data affected 3% of rows." | | Magnitude and approximation | Describes size, degree, or closeness but may be replaceable with a specific value or range. | about, almost, approximately, around, fairly, nearly, relatively, roughly, slightly, somewhat | *Before:* "Nearly half the sample." *After:* "47% of the sample." | | Comparisons | Makes comparative claims without specifying the reference point or metric. | better, decreased, improved, increased, higher, larger, lower, smaller, stronger, weaker, worse | *Before:* "Scores improved." *After:* "Scores increased by 4 points from baseline." | | Quality judgments | Evaluates performance, suitability, or importance without defining the criteria used. | adequate, appropriate, effective, efficient, meaningful, optimal, reasonable, robust, suitable, sufficient, useful | *Before:* "An effective treatment." *After:* "A treatment that reduced symptoms by 30%." | | Abstract nouns | Refers to broad concepts that obscures the specific mechanism, variable, or process being discussed. | aspect, behavior, characteristic, concept, context, dynamics, effect, environment, factor, framework, impact, influence, issue, mechanism, phenomenon, process, property, relationship, scenario, situation, system, thing | *Before:* "The relationship between the variables." *After:* "The correlation between age and income." | : Stop and justify — These words are not automatically problematic, but try to replace them with more precise language. {#tbl-justify} # Common mistakes {#sec-mistakes} These are frequent in first papers. Each has a fix. - **Putting results in the Methods, or conclusions in the Results.** *Instead:* Methods say what was done, Results say what was found, and the Discussion says what it means. - **Reporting a p-value alone.** *Instead:* Give the effect size and its confidence interval, and then the p-value. - **Saying "significant" without saying what kind.** *Instead:* Write "statistically significant" and give the test and p-value, or choose a more specific word. - **Claiming causation from observational data.** *Instead:* Use "associated with", and discuss what else could explain the pattern. - **Overstating the conclusion.** *Instead:* Conclude only what the results support, and state the limitations. - **Citing papers you have not read.** *Instead:* Read at least the abstract and relevant results of every source, and cite the original source of a finding rather than a paper that mentions it. - **Using the wrong tense.** *Instead:* Methods and Results are in past tense ("we assigned", "the difference was"). The Introduction and Discussion mix tenses: present for established facts, past for specific studies. - **Leaving figures and tables unreferenced.** *Instead:* Every figure and table is cited in the Results by its number. - **Pasting in text from an AI tool without checking it.** AI tools can produce fluent text that contains wrong statements and references that do not exist. Check your course or journal policy on AI use, and verify every claim and every citation yourself, because you are responsible for them. - **Writing the Introduction first and getting stuck.** *Instead:* Write the Methods first (see @sec-writing). # The craft of writing effectively {#sec-effective} One of the most impactful videos on the way I think about writing is Larry McEnerny's lecture on [the craft of writing effectively](https://www.youtube.com/watch?v=vtIzMaLkCaM). These are the key ideas that have stuck with me: - Writing has two functions: 1. Writing for yourself, to help you think about the world. 2. Writing for others, to change the way they think about the world. - Writing for others is *not* about conveying your ideas to the reader; it's about **getting the reader to adopt your ideas**. This requires a different writing style compared to when you are writing for yourself. - Your writing is not meant to be permanent, because knowledge isn't cumulative. **Knowledge is a conversation moving through time**, and your writing is meant to move the conversation forward. - Newness or originality is not important. Value is important. Value drives the conversation forward. - **Build tension to create value**. Tell the reader what is wrong or missing in the current conversation: draw attention to *their* problem and the importance of fixing it, then offer your solution. The value of your paper is created by the tension that it reveals and resolves. - If the reader doesn't think your paper is valuable, then it doesn't matter how persuasive, organized, or clear it is. They have no motivation to read it, and they won't. Tension is built through sentence structure and emphasis, a topic that is masterfully demonstrated in Judy Swan's workshop lecture [Scientific Writing: Beyond Tips and Tricks](https://www.youtube.com/watch?v=jLPCdDp_LE0). Another general tip for scientific writing is to always be precise and objective. Avoid vague statements that make the reader ask, "what does that mean?" (this is the most frequent feedback I give when reviewing papers). Don't hedge your statements with modifiers like *somewhat*, *fairly*, *nearly*, etc. Furthermore, don't use vague words that don't have a precise meaning. It is better to be clear and possibly wrong, than to be vague and "not even wrong". That is, if you make a statement that can't be disproved, then you should clarify it until it can be. Here are some examples of statements that will make the reader ask "what does that mean?" (in **bold** is what needs to be fixed): - "The new algorithm is **somewhat better** than previous methods." (Better by what metric? What does "somewhat" mean?) - "The treatment produces **nearly optimal** outcomes." (nearly optimal means... what? Sub-optimal?) - "The intervention is **mostly effective**." (So it's *ineffective* in some way? Does "mostly" mean 51% effective? ... 99% effective? Effective according to what metric?) - "The estimates are **more realistic**." (How is "realism" measured? How much is "more"? More compared to what?) - "The variables exhibit a **dynamic relationship**." (... what in the world does that even mean?) - "This method yields **meaningful** results." (Meaningful in what way?) - "The analysis provides **important** insights." (Why is it important?) For the last two, you never tell the reader what is meaningful or important. Different readers have different interests. The paper will emphasize key findings that the target audience should be interested in, but it's left to the individual reader to decide whether those findings are meaningful or important to them. ## Other resources Here's a collection of lectures/talks about scientific writing that I've learned from: - [Judy Swan - Scientific Writing: Beyond Tips and Tricks](https://www.youtube.com/watch?v=jLPCdDp_LE0). The best place to get started for scientific writing. Covers the fundamentals of sentence structure and emphasis. - [Larry McEnerny - The Craft of Writing Effectively](https://www.youtube.com/watch?v=vtIzMaLkCaM). Drives home the importance of creating value with your writing. - [Simon Peyton Jones - How to write a great research paper](https://www.youtube.com/watch?v=1AYxMbYZQ1Y). This one is targeted toward computer science research, but it is full of great ideas and gives motivation for *why* you should write. - [Pete Carr - How to Write a Paper in a Weekend](https://www.youtube.com/watch?v=UY7sVKJPTMA). A concise guide to getting a paper ready for publication. - [Anthony Newman - How to Write a Great Research Paper, and Get it Accepted by a Good Journal](https://www.youtube.com/watch?v=oVF4c6JTAHs). A longer guide to getting a paper ready for publication. # Examples of published papers {#sec-examples} Published papers are the best models of this structure, and every scientific journal has examples. After you have read this guide and worked through @sec-example, read a few published papers related to your research topic and look for this structure. You'll find some variation across journals, but the underlying structure and flow of the paper should be recognizable. **How to read a paper as a writer.** The papers below are written for researchers and some are technical. You don't need to follow all of the statistics. Read them in this order, and write one or two sentences of notes for each step: 1. **Abstract.** Does it state the problem, what was done, the main result, and the conclusion? 2. **Introduction.** Find the sentence where the authors turn from what is known to what is missing (often marked by "however"). That is the tension. 3. **Methods.** Is it in past tense? How much detail is given? Could you repeat the analysis? 4. **Results.** Find a table or figure, and then find the sentence in the text that summarizes it. Are there any explanations of *why*, or are they stated as facts? 5. **Discussion.** Find the recap of the problem, the comparison with other studies, and the limitations. If you find these papers too advanced, ask your instructor or librarian for a paper in your own field from a recent issue of a journal that your course already uses. The structure is the same. In case you don't have a research topic in mind, here are a few papers to look at: - [Modern modelling techniques are data hungry: a simulation study for predicting dichotomous endpoints](https://link.springer.com/article/10.1186/1471-2288-14-137) - [Development and validation of prediction models for health-related quality of life outcomes after breast cancer surgery and reconstruction](https://www.sciencedirect.com/science/article/pii/S0748798326000855) - [Effective Sample Size for the Kaplan-Meier Estimator: A Valuable Measure of Uncertainty?](https://www.tandfonline.com/doi/full/10.1080/00031305.2025.2542390#d1e385) - [Instability of the AUROC of Clinical Prediction Models](https://onlinelibrary.wiley.com/doi/full/10.1002/sim.70011) - [Machine Learning for Early Detection of Cognitive Decline in Parkinson's Disease Using Multimodal Biomarker and Clinical Data](https://www.mdpi.com/2227-9059/12/12/2758) # Tools {#sec-tools} You can write a good paper with many different tools. Pick the option that matches your experience and your course requirements. ## Option A: a reproducible workflow with Quarto This option is useful if you already use `R` or `Python`, or if you want to learn. It makes every number in the paper trace back to code. - For running the analysis, use a reproducible workflow from start to finish (i.e. from data loading and pre-processing to table and figure generation). All results of the analysis must be captured in a table, figure, or listing, and these should be automatically generated by the code used to run the analysis. - In the analysis code, save figures as `.pdf` or `.png` with the desired height, width, and DPI (if using a raster format). Use the `gt`, `flextable`, or similar packages to create tables, then store those table objects as `.rds` files. - Use `Quarto` to compose the report in a `.qmd` file. This allows the use of markdown and LaTeX to write and format the report. Directly embed generated figures with caption text. Load table objects within an `R` code chunk to render the formatted tables with caption text. - Use a reference manager to create a `.bib` file for references, and use a `.csl` file for reference formatting. ## Option B: word processor and reference manager This option is simple and works for most class papers. - Write in **Word** or **Google Docs**. Use the built-in heading styles for section titles, and use the caption feature for tables and figures so they are numbered automatically. - Use a reference manager such as **Zotero** (free) or **Mendeley**, with its word processor plugin. It inserts citations and builds the References section in the style your course requires, so you never format references by hand. - Save figures as high-resolution `.png` files, or `.pdf` when the program allows it, and paste them in with a caption. Keep the code or software steps that produced them in a separate file, so you can regenerate them. - Keep a dated copy of each draft, so you can go back after a bad edit.