PhD & Doctoral Dissertation Examples

The best golden thread in the collection, and you can see the whole of it in about five minutes. Section 1.1 puts the whole claim in a single sentence: hallucination of large language models in software engineering can be understood and mitigated by external guidance, self-reflection, and implicit signals. Sections 1.2, 1.3 and 1.4 are then those three clauses in order, one per study, and each becomes a chapter.

And section 6.1 is called "Revisiting the Thesis Statement", which closes the loop by returning to that exact sentence and saying what the three studies did and did not establish about it. State the claim, structure the work by its parts, come back and answer it. If your chapters feel like separate projects that happen to share a topic, read chapter 1 and chapter 6 here back to back and you will see what is missing. The related work chapter is also unusually well organised, sorted into four themed sections with subsections rather than run as a chronological parade of studies, which is what makes a literature review feel like an argument.

A paper-based engineering thesis whose study chapters all run the same internal skeleton, and the skeleton is the lesson. Chapter 3 carries its fullest form: Summary, Introduction, Background, Related Work, Knowledge Gaps and Research Questions, then results reported against those questions by number (RQ1-3, then RQ4), Discussion, and Limitations and Threats to Validity. Two of those headings are worth stealing whatever your field.

Deriving the research questions in a section explicitly about the knowledge gaps means the reader never has to take on trust that the questions follow from the literature; they watch it happen. And "threats to validity" as a standing section per study, rather than one limitations paragraph at the end of the thesis, is a software engineering convention that reads as confidence rather than weakness. Chapter 1 also splits section 1.1 "Context and Problem Statement" from section 1.2 "Thesis", which keeps the situation and the claim from blurring into each other the way they usually do.

A specialist read, and the one document here that shows you fieldwork going wrong. Chapter 3 carries Fúnez-Flores's research journal, dated. On 7 January 2019 he messaged two student organizations in Honduras and heard nothing, and the entry is him sitting with that: "I still do not have participants. It is frustrating." A reply came days before the semester started on 28 January, and that group then never wrote to him again.

Every methodology chapter you have ever read presents access as a settled fact. This one shows what it cost, across twelve months in the field. The data analysis section is equally concrete: fieldnotes split into descriptive and reflective halves, research memos written from those, NVivo, and member checking done in public: he put his interpretations in front of student activists at livestreamed forums to be argued with. The caveat is real, though. It is 373 pages of dense decolonial theory, and a student who copies the structure without the grounding underneath it will struggle to defend it. Read chapter 3, not the thesis.

Read this one if your project involves a community, an employer, or anyone whose time you are asking for. Carlson worked with 11 of the 12 fellows in one Appalachian fellowship cohort, using photovoice, focus groups and site visits to three of their communities, and chapter 3 is unusually honest about the machinery. Section 3.3.1 works through a genuine bind: a properly participatory study lets participants write the research questions, but the ethics board wants approved questions before you are allowed to talk to anybody.

Her answer, two sets of questions with Table 1 on page 62 setting hers beside the ones the participants generated, is a real solution to a real problem rather than a paragraph about rigor. She also states what participants were paid, per activity, in dollars. Section 3.3.2 records a recruitment plan that half failed (host community members were invited, exactly one said yes) and says so. And the coding section treats narratives, images and relationships as three different problems instead of pretending photographs code like transcripts.

If anyone has told you your research questions, your methods and your results chapter do not line up, read Table 1 on page 80 before you read anything else in this collection. Huang sets out all five research questions in one table, each against its variables, its scale of measurement, and the exact analysis that answers it.

Everything downstream obeys that table: chapter 4 heads each results section with the question written out in full, and chapter 5 is five numbered conclusions, one per question. That is the golden thread on a single page, and it is the cheapest thing here to copy into your own draft tonight. The study behind it is a survey of 1,061 undergraduates at a predominantly white institution and an HBCU, run through exploratory factor analysis, confirmatory factor analysis and structural equation modeling. Section 4.4 is worth a second look for a different reason: she accounts for every one of the 286 discarded responses by reason, which is how you report data cleaning without sounding like you are hiding something.

Go to the limitations passage on page 33. Most limitations sections are an apology in list form: small sample, one country, self-reported data, and no indication of what any of it costs the argument. This one names the consequence every time.

The first: her design is cross-sectional, and she says why that is not a stylistic choice — there is no time-series dimension because consistent firm-level biodiversity data separate from broader ESG scores does not yet exist. So the analysis "cannot rule out omitted variable bias or reverse causality", and the fix is named too: panel data, instrumental variables, or a quasi-experimental design.

The second is better still. Her regional results hold up statistically, but the mechanism behind them has not been tested, and she says her own explanation for it is "theoretically plausible but requires further decomposition". Writing that sentence about your own favourite idea is harder than it looks.

It is also a useful model for writing about something the data barely covers yet. She is measuring biodiversity dependence through ENCORE, a dataset she notes "remains under development", across more than 10,000 listed firms — and the caution shows in the claims rather than only in the caveats.

Read this one for the problem it has rather than the answer it reaches: two datasets that do not line up, and a decision made in the open.

Open it at "Methodological Innovations in Measuring Financial Literacy" (page 12). He is joining the National Financial Capability Study to COMPUSTAT. The NFCS is run "every three years", so it exists for 2009, 2012, 2015, 2018 and 2021; COMPUSTAT is annual and his sample runs 2009 to 2024. Merge them and two years in three have a hole where the literacy measure should be.

He names the problem — "due to variations in survey participation, missing data is a recurring challenge" — then uses Fully Conditional Specification to impute the gaps, and, the part worth copying, reports the results both ways: Panel A on the observed survey data, Panel B on the imputed series, side by side in the tables. The reader can see what the imputation did rather than taking his word for it.

At 103 pages it is also the longest of the professional doctorates here, and the closest in feel to a traditional thesis if that is the shape you have been asked for.

Read this for a single artefact: the literature summary table, Table B.1, running from page 188. Its columns are Paper, Setting, Dataset and Source, Method, Explanatory variables, Outcome variables, and the column most students omit, Sign of effect, which records that droughts raise intimate partner violence where floods show a null effect, and notes for each study what moderated it.

A table like that does something prose cannot: it makes studies comparable, and it makes a gap visible rather than asserted. When you can see that every prior paper used one dataset and one method, the case for your own design writes itself. Build one early whatever your field; it also stops you rereading the same twelve papers because you cannot remember what they found. The topic pairing is instructive too: development economics, weather shocks and gender-based violence in one dissertation shows how an essays-format thesis holds together on a question rather than a dataset.

Open it at "Literature Review, Regulatory Rules, and Hypotheses" (page 131), the best answer here to "synthesize the literature, don't summarize it." Xiong is writing about trading halts. He opens: "The effect of trading halts is debatable." Then the first school and its authors: halts ease market tensions, reduce volatility, protect liquidity providers. Then the opposing school: halts merely delay volatility and can create a magnet effect dragging prices toward the suspension threshold.

Then: "There can be a middle ground between the above two schools": Harris, who splits volatility into informed and uninformed trading and shows the halt helps for one and harms the other. Position, counter-position, resolution, with the hypotheses built on the unresolved part. Copy the three-move structure directly; it works in any field with a live disagreement. Also the cleanest example here of the three-essay format.

The theory entry. If your dissertation is a formal model rather than an empirical study, this is the shape: a baseline model, its results, and then "Appendix A. Extensions and Robustness" (page 59), where the model is extended rather than the regressions re-run.

The extension is instructive in itself: the main model covers the takedown half of a notice-and-takedown copyright regime, and the appendix adds the counter-notice procedure as a settlement game: the owner sues, offers to drop the suit if the platform closes, the creator accepts or refuses, and the creator responds only if the threat to litigate is credible, meaning the owner's expected gain must exceed litigation costs. Decision nodes are color-coded by player, with nature's moves marked separately. Two things to take: robustness in a theory paper means showing your result survives a richer model, not a different standard error; and a labelled game tree is worth several pages of prose.

The model here for explaining how you built your sample, which is the paragraph most students write as one sentence and lose marks for.

Open it at "Data and Methodology" (page 12). She starts from the Public Plans Data, over 3,700 observations covering 2001 to 2018. Then she cuts it, and says why each time. The years 2001 and 2002 go, and the sentence doing it is worth stealing: "To obtain a balanced data set, where each plan is observed during the same time-period, 2001 and 2002 data were excluded." That leaves 2,800 observations across 2003 to 2018. The District of Columbia goes "because 16 years of plan data were not reported". Federal employee plans are out because the source does not track them.

Then she does the thing that makes a sample real to a reader: she describes its shape. About 3.5 retirement plans per state; one each for Hawaii, Idaho, Mississippi, Oregon and Wyoming; fifteen for California and ten for Texas. You finish the page able to picture the dataset, which is the whole job of that section.

It is also a good example of a practitioner question — a working professional asking which lever actually closes a funding gap — answered with public data rather than proprietary access.

The model for structural estimation, and more generally for interpreting a number instead of reporting it. Zahur estimates a conduct parameter for LNG sellers and gets 0.72 to 0.90. He does not stop there.

He tells you that 1 means firms behave exactly as Cournot competition predicts and 0 means fully competitive, and therefore that 0.72-0.90 means sellers are close to Cournot but somewhat more competitive, and that fully competitive conduct is implausible because the estimate is nowhere near zero. Three sentences turning a coefficient into a finding. Most results chapters report significance and move on, leaving the examiner to work out what the magnitude means. Note also how he handles model simplicity: he runs a deliberately parsimonious specification, says so, and reports that adding covariates does not measurably change the fit. Defending a simple model on evidence beats apologizing for it.

Thirty-three pages, seven numbered sections, no chapters. If you have been told a doctorate must be four hundred pages, read this one first: it is built like a journal article, and it was examined as a doctorate.

Open it at "III. Description of Data and Empirical Testing Methodology" (page 7) for the thing it does best, which is justifying every variable from the research question rather than from availability. She wants to know whether market risk drives macroeconomic stability, so she picks the two variables the Federal Reserve is actually mandated to manage — "the macro variables that stem directly from the Federal Reserve's 'dual mandate' to ensure low unemployment and price stability" — and then splits inflation in two, survey-based CPI against the breakeven rate that reflects bond market expectations.

The detail worth copying is on the same page. Her VIX series runs from January 1990, but the breakeven inflation data only start in January 2003, and instead of quietly using the shorter window she says why: earlier estimates "suffered from serious liquidity constraints of TIPS", with three citations. A reader never has to wonder whether she noticed.

The shortest document in the collection, and it earns its place on one section. "Project Site and Population" (pages 12-13) is the most concrete setting description in the collection: a 37-bed adult medical unit in an urban community teaching hospital, 57 nurses, roughly 2,500 discharges a year, patients 65% white and 21% black, 60% aged over 65, average stay five days, medical residents rotating in two-week blocks.

It even says who chose the unit and on what basis. Students are told to "describe the setting" and produce two sentences; this shows what the instruction means, and why it matters: every one of those facts constrains what the intervention could be and what the results generalize to. The measurement section states the outcome as a formula rather than a sentence. Do not use this as a model for length; use it for the setting section.

Open this at "Limitations" (page 32) if you have been told your limitations section is too defensive or too vague. Bird does three things most students do not. She names the limitation: a small sample, despite an extended deadline and a lunch laid on to recruit.

She then identifies which specific claim in her own project is under-supported and why: nitrous oxide was the change providers most wanted, and the only evidence for it in IUD placement comes from an adolescent study, so it cannot carry an adult recommendation. And she says the whole area is under-researched and names that as a finding rather than an excuse. That is limitations written as analysis: it tells a reader exactly how far to trust the conclusion, which is the entire job of the section.

Read this if your supervisor has written "reflexivity?" in your margin and you do not know what a satisfactory answer looks like. Herron's methodology chapter contains a subsection headed "Bracketed Experience of Researcher" (page 36) in which she states, in the first person, that she was raised Christian and working class, converted to Catholicism, spent eight years in a Benedictine religious community, worked two years in hospice, and believes the spirit continues after death, all while studying nurses' experiences of dying patients' visions. She is naming every prior belief that could shape what she hears, where a reader can check it against her findings.

Almost every methods textbook tells students to bracket their assumptions; almost no thesis shows what that looks like, and students end up writing a paragraph of throat-clearing. This is the real thing, and it takes about a page. Chapter 3 also argues its way to Husserl's descriptive phenomenology rather than announcing it. Nothing here for you if your work is quantitative.

The clearest model here of what a DNP quality-improvement project actually is, which matters because it is a different genre from a research dissertation and almost nobody explains the difference. It is not trying to discover something; it is trying to change something in one place and measure whether the change held.

Two things to copy. The intervention is driven by a named theoretical framework, Prochaska and DiClemente's Transtheoretical Model, and he walks through all six of its stages before saying how the model is applied on his units. That is the step most QI projects skip, and skipping it leaves an intervention that is a PowerPoint with no theory of why it should work. And the goals and expected-outcome sections state the target numerically before the results are in (90% of staff on two named units, within four weeks), so the results chapter has something to be judged against. He is also honest that the evidence base for de-escalation training is thin, and runs the project anyway.

The one to read if no validated instrument exists for what you want to measure and you are wondering whether you are allowed to build your own. She wanted to measure provider knowledge before and after an education session, found no reliable existing measure, built a 17-item pretest and posttest from her own literature review, and then stated plainly that "reliability and precision were not confirmed in this study" (page 12) rather than implying the instrument was validated.

That admission is why this is here, and also why to read it critically: the sample is six providers at one practice, and the defence of the Likert scale leans on a general reliability figure from an unrelated study. Read the Measurement Instruments section for how to construct and describe an instrument, and read the results knowing six responses cannot support much. A weak study honestly reported teaches more than a strong one written up as if nothing went wrong.

The only example here of the professional doctorate portfolio, and if you are on a DClinPsy, DEdPsych or DHealthPsy this is the format nobody explains properly. It is not one long thesis. It is three pieces: a systematic literature review written as a standalone publishable article, an empirical study formatted as a journal paper, and a set of structured reflexive sections behind them: a plain English summary, a critical appraisal of the study's design, and a statement of impact and dissemination. A portfolio abstract covering the whole thing and a statement of what the candidate personally contributed open the document.

Read the contents page first and you will understand the shape faster than from any handbook. The review half is done properly: PRISMA-guided, with study selection, a quality appraisal stage and study characteristics reported separately. The empirical half is a measure-validation study, which is worth reading even if you never touch a questionnaire, because of the order it reports things in: removal of invalid data, normality testing, then univariate and multivariate outliers, and only after all that the dimensionality, reliability and validity results. The housekeeping comes before the findings and is shown, not summarised. Note too that section 4.2 is called "Epistemology" and sits inside the method rather than in a separate philosophy chapter, which is how a journal-format paper handles it.

The clean experimental-psychology shape, and short enough to read in an evening. There is no separate methodology chapter, because in this format each experiment carries its own: Experiment 1 has Method, Participants, Design and Materials, Procedure, then Results, and Experiment 2 follows the same pattern.

If your supervisor has said "your methods belong with each study, not in a chapter three", this is what they mean. Two details worth copying. The tasks are named and described individually under Procedure (Generation Task, Reading Span, Change Detection) rather than being bundled into a paragraph, so a reader can see exactly what a participant did and in what order. And the introduction is organised by construct rather than chronologically, splitting long-term memory and working memory each into encoding and retrieval before the two are brought together in "The Current Study", which is what makes a two-variable research question feel inevitable instead of arbitrary.

Read this if you are frightened of finding nothing. Röseler set out to work out which personality traits make people more susceptible to anchoring, ran meta-analyses across the eight most studied traits, and found no evidence for any of them.

That is the thesis. It passed, and it is more useful than most positive results, because he does not stop at the null: he argues the reason is that susceptibility to anchoring cannot currently be measured reliably, and then discusses what would have to change for the question to be answerable at all. A null result plus a diagnosis of why is a contribution; a null result plus a shrug is not, and the difference between those two things is visible on the page here. It is also worth opening if you are writing a publication-based thesis: his five research articles sit in an appendix, and the 45 pages of frame in front of them are what turn a stack of papers into a single argument. That frame is a clean model of organizing a literature before you meta-analyse it, too: he sorts the eight traits into three paths according to where the prediction came from, exploratory, general judgment and decision-making theory, or anchoring theory specifically.

A model of the multi-experiment thesis, which is the standard shape in experimental psychology and one that students coming from a single-study Master's find hard to picture. Four eye-tracking experiments, each testing prediction from a different source of information: the meaning of the verb, number marking, case marking, and implicit causality. They are not four separate papers stapled together, and chapter 9 proves it with an overview table setting all four side by side against the source of information and the level of prediction each one probed.

That table is the thing to look at first: it is what a coherent programme of studies looks like from above. Two practical details worth copying. The experimental items for every experiment are printed in full in the appendices, so another lab could rerun it. And the license line on the copyright page carves out quoted material from other authors, which is the correct and often-forgotten way to release your own work openly without over-claiming rights you do not hold.

Four numbered experiments again (nine, counting the sub-studies), but included for a different reason: experiment 1 does not answer the research question at all. It is a test of whether the method itself is reliable, the Headturn Preference Procedure used with infants, checked across three sessions before it is trusted to carry anything.

Almost nobody does this, and almost everybody should at least think about it. If your whole thesis rests on one instrument, showing that the instrument behaves is not a detour, it is the foundation. The other thing to take from it is how experiments 2, 3 and 4 are motivated: adults first, then nine-month-olds, then six-month-olds, with each study's design justified by what the previous one found rather than by a plan drawn up in advance. The write-up says so in as many words, which is how you show a reader that a programme of work developed rather than simply happened.

The clearest three-article dissertation in the collection, and the one to open if you have been told to write one and have never seen the shape: three self-contained studies, each with its own introduction, methods, results and discussion, bound by a common question. The first article is a systematic review conducted to PRISMA (section 1.2, page 10), with the search strategy, the search results and the coding scheme written out: study structure variables (year, authors, sample size, methods, geography) separated from review data variables (comparators, program components, barriers, intentions).

Students are routinely asked for a "systematic" review and produce a narrative one; this shows the difference, and shows a review used as a first paper that establishes the gap the next two fill rather than as background. The second and third articles are both municipal surveys, the third combining its responses with community characteristics data, so the three together also show a review used to set up two rounds of primary data collection.

Read the Methods section (page 47) if your data does not exist in a form anybody has released, and you are trying to work out whether assembling it yourself counts as method. Yan needed the reference lists of forty years of Chinese sociology. CNKI, unlike Web of Science, does not supply reference text, so she downloaded and OCR'd the articles, then scraped download and citation counts off the result pages because those are displayed but not downloadable, then merged that with the official metadata.

What makes it a model rather than a war story is what follows: she states the sample of 9,250 articles, then every inclusion and exclusion with its reason. Conference summaries out, because they are not substantive. Book reviews in, because in Chinese sociology a book review is a long essay comparable to a research article. Zero-reference articles out, because citation counts jump discontinuously at one reference. Final analytic sample: 3,746. That is what "justify your sample" means.

The longest document in the collection and not one to read end to end. Read "Analysis and Positionality" (page 48) and stop. Galeana Rodriguez had previously worked with the World Bank and expected that to make Miskitu communities distrust him.

It did not. Instead people assumed he worked for an NGO and kept steering him into normative arguments about the property regime, so his actual difficulty was the opposite of the one he had prepared for. That alone is worth reading, because positionality is usually taught as a fixed thing you declare rather than something that behaves unpredictably in the field. Then he does the part almost nobody does: he says his positionality set limits on the study, and names two specific lines of inquiry he decided not to pursue because of it. Positionality connected to scope, rather than self-disclosure that changes nothing about the research.

Two conventions here that confuse students, both worth seeing done. First, the methods are in an appendix, "Appendix I: Plan Bee (Methods)" on page 157, not in chapter three. That is normal in a theory-led ethnography and it is not disorganization; the argument runs on theory and the methods are supporting evidence.

Second, the positionality. Andrews is not observing beekeepers, she is one: bee club officer, mentor, educator, and the writer of the curriculum for the state's Master Beekeeping program. She sets out what that access bought: beekeepers, research scientists and extension agents; a real sense of how discouraging the work is; research questions that came from the community rather than the literature. Then, without softening it, what it cost: "it has been difficult. I have had to repeatedly confront and articulate my own stance(s)" while mentoring the people she was studying. Insider research is common, and this is the honest version of writing it up.

The best model here if your work is statistical. Go to section 3.6, "Limitations" (page 61).

Berry opens by naming the assumption the whole chapter rests on, that the classifier's error rate is known, then explains why it cannot be fully resolved (cross-validation estimates it but uncertainty remains; a calibrated model just moves the problem to estimating calibration accuracy), then adds the threats specific to his data: bots and inactive accounts that look like real members of a demographic group and bias the estimate. Two short paragraphs, and a reader knows exactly how far to trust the result. No apology, no "further research is needed" filler. It is also a clean example of the three-paper dissertation with per-chapter appendices, where mathematical derivations sit at the end of each chapter rather than interrupting the argument. Useful if you have been told to "put that in an appendix" and are unsure what survives in the main text.