Chapter 4: Results

Chapter 4: Results

This is the chapter that reports on your results. It does not, at most universities, provide discussion of them, it just provides the results of your research. In the case of quantitative research it also provides your statistical analysis supporting. For a qualitative thesis it reports on how you conducted your coding and development of themes (or similar). Note, if you are doing a qualitative based thesis that addressing each question from your interviews for each participant and the need to often include quotes can make for a long chapter 4.

Results

Discussion of what your results actually mean is normally left to the concluding chapter unless you are at a university that uses more chapters than five (this is not atypical in the UK). Before you can get your results and analyze them, however, you need to gather your data. This sounds easy but can be quite challenging. It is challenging for quantitative studies as it is easy for a doctoral candidate to create survey questions that are not in good alignment with their survey instrument and/or their research questions (or theoretical framework!). Qualitative studies pose their own significant challenges especially should be in areas such as, for example, the design – and execution – of interviews. A survey that poses questions that are 'off the mark' in terms of the core research problem being addressed can vastly complicate statistical analysis or even make it impossible.

Quantitative chapter 4

How to structure your thesis.

Generally, a quantitative results chapter is structured by either the research questions or the hypotheses you posed. What will you provide to your reader?

Sample content for a quantitative results chapter:

It will mostly be charts and tables, with the text referring to them rather than repeating what they show — describing your data, your sample, and your analysis. Organization and clarity count and each table and chart must be clearly labeled so readers of the text can easily follow your exposition.

When to get a Statistician

Qualitative chapter 4

Chapter 4 is where AI use becomes actively hazardous to your dissertation — and your graduation

Minor misuse of AI in the early chapters of your dissertation may cause a painful interview with your Chair and a lot of work to rectify matters. Major misuse in those chapters? We leave that to your imagination.

However, now we come to actual research results. Even minor AI misuse here is potentially genuinely threatening to the outcome of your dissertation journey.

And it is easy to walk into. You have your data. You are tired, the analysis is hard, and you ask a model to help you interpret what you found — you describe your results, or paste in a summary, or upload a table. Back comes fluent, confident text. Figures. Significance levels. Interpretation.

Here is the part that catches people. If you did not give it your actual dataset, it is not analyzing anything. It is producing text that resembles statistical output, because producing plausible text is what it does. The numbers will look right. Some of them will be invented.

And if you did upload your data, you have a different problem, not a solved one: you now cannot easily tell which figures came from your data and which came from the model filling gaps. You must be able to reproduce every number in your chapter from your own dataset, using your own software, without the model in the room. If you cannot, it does not go in.

Same failure as ghost citations. Different consequence entirely — and how bad it gets depends almost entirely on when it is caught.

Very lucky? You find it yourself, or we do. A bad afternoon and a rewrite.

Lucky? Your Chair finds it, decides it was a mistake rather than misconduct, and lets you redo the whole mess. Painful, slow, survivable.

Unlucky? The further down the chain it travels, the worse it gets. Your committee has less latitude than your Chair, and less reason to extend it. School Review has less again, and no relationship with you at all.

And if nobody finds it until after you have graduated and published on it? That is the version that ends careers. It stays attached to your name permanently, and there is no version of that conversation you get to be part of.

The whole gradient runs one way: the earlier it is caught, the more of your work survives. Which is the entire argument for checking it yourself before anyone else has to.

Analysis is the one part you cannot hand over — and that now applies to both domains

Coding transcripts is tedious. Running and re-running statistical models is tedious. Both are also where the analysis actually happens, and that combination is precisely what makes them dangerous to outsource.

In qualitative work

This page already notes the criticism of NVivo — that it flattens data and misses the insight human coding produces. Handing coding to an AI is that same objection, considerably amplified.

• Inductive coding means starting with no presuppositions. Codes emerge from your interviewees' own words. A language model is made of presuppositions. It has read millions of documents and it will bring their categories to your transcripts whether you want them or not.
• Immersion is the method. Reading every answer to question one, then every answer to question two, is not preparation for the analysis. It is the analysis. Skip it and you have not saved time — you have removed the step where understanding happens.
• You will not be able to defend it. When a committee member asks why that quote sits under that theme, "that is where the coding put it" is not an answer.

In quantitative work — the same objection, differently shaped

This used to be a smaller problem because the tools could not do it. That has changed. Upload a dataset and you will get descriptive statistics, model output, tables, and an interpretation. The arithmetic may well be correct.

What you lose is everything that happens on the way there:

• You never actually look at your data. You do not see the distribution that is nowhere near normal. You do not spot the outlier that turns out to be a data entry error, or the variable with forty percent missing, or the two participants who answered every item identically.
• You do not discover that an assumption is violated, because you never tested it — and a model run on violated assumptions produces confident output that is simply wrong.
• You cannot say why you chose that test, because you did not choose it. "Why this model rather than the obvious alternative?" is a routine defense question and it has no good answer if something else made the decision.
• You will not recognize a result that should have surprised you. Researchers who know their data notice when a number looks off. That instinct is built by handling the data, and it cannot be borrowed.

In both domains the principle is identical: the tedious step is the step where you come to understand your own research. Removing it does not accelerate the work. It removes the work.

What AI can do with either kind of material: organize quotes you have already selected, check that every theme you claim is supported by quotes you identified, format tables, verify that reported figures match the ones in your output, and check your numbering and cross-references. It cannot decide what your data means, and it should not be the reason you believe anything about your own findings.

What it is genuinely good for here

All mechanical, all checkable, and all worth doing — especially in a chapter where, as noted above, committee members do focus on formatting nitpicks:

• Table and figure formatting to APA or whatever your program requires.
• Numbering. Sequential, no gaps, no duplicates.
• Cross-references, both directions. Every table mentioned in your text exists. Every table that exists is mentioned. Both. This is worth doing properly.
• Labels. Every table and chart clearly and consistently labeled, as this chapter demands.
• Terminology consistency with your earlier chapters.
• Explaining a statistic you have half-forgotten — what a test assumes, what a term means, why an assumption matters. That is general knowledge, and you can check it against any textbook.

Under no circumstances use AI on your findings. Use it on your text. That is the whole line, and it does not move.

And if your grasp of the statistics is not solid, the answer is a statistician, not a chatbot. As set out above: low-cost insurance, and worth every penny.

What if my data does not support my hypothesis?

What do I do if my results are not significant?

Frequently the answer is genuinely interesting, and frequently it is more interesting than the original hypothesis — because a result that surprises you is a result that surprises the field.

Practical steps:

• Report what you found, plainly, including the null result. That is what Chapter 4 is for
• Then ask what would explain it. Was the relationship absent, or was your instrument unable to detect it? Was the sample too small? Did something else emerge in the data that you were not looking for?
• Look at what your data does support. Often there is a real finding sitting beside the one you expected, and nobody noticed because nobody was looking for it
• Take it to your Chair early, and take the honest version. A null result explained well is defensible. A null result concealed until the defense is not

Can a dissertation be saved when the data does not fit?

A dissertation can be reframed around what the data actually shows. The introduction is rewritten, the research questions are adjusted to what was really asked, the discussion follows the evidence.

That is real work and it is not a disaster. It is considerably less work than collecting new data, and considerably less risky than forcing the original claim.

Where it becomes hard is when the study was not really yours to begin with.

A case worth knowing about — when nobody helps

A candidate at a well-known online university came to a topic they cared about and understood. The Chair reshaped it — into something related, but not the same question, and not one the candidate recognized as theirs.

The reshaped topic got researched, because that is what candidates do when a Chair takes a view.

The data did not support it. It could not have — the reshaped hypothesis was not the question the design was built to answer.

But the results said something. Something real, and something the field would have wanted to know.

Nobody helped find it. Not the Chair, not the committee. The findings were treated as a failure of the study rather than as findings, and the fact that they pointed somewhere uncomfortable for the institution did not help.

The candidate left the program.

And then the part that settles it

The study is now in final draft, heading for journal submission.

Once somebody actually asked what the results said — rather than why they failed to say the expected thing — there was a genuine finding in the data. It is being written up properly, and it is going out.

That is the test, and it is why this case is worth telling.

The data was never the problem. The analysis was not the problem. The candidate was not the problem. There was a paper in that dataset — in the same dataset the program had written off as a failed study.

What failed was the interpretation — and it failed because nobody was looking for what was there. They were looking for what they expected, did not find it, and stopped.

That candidate could have had a doctorate. The material was there.

What that case actually shows about who finishes

Not that the candidate could not do doctoral work. There is a study in final draft, heading for a journal, built from a dataset an institution had given up on. That is doctoral work — and a return elsewhere is planned, with a sharper question, on a subject thought hard about since.

And the research says the same thing. Among admitted candidates, measured ability does not predict who finishes. What predicts it is momentum, absence of excessive distress, and a project that makes sense to you — and this program removed all three, in order.

• Momentum — gone, once the data did not fit and nobody could say what to do next
• Distress — mounting, with no path forward and nobody in their corner
• Coherence — gone from the start. It was never their project after the Chair rewrote it

And there was a fourth thing, underneath all three: the topic did not engage them.

Not because they were uninterested in research. They had come with a question they cared about. What they were left with was a question somebody else cared about, and they had to carry it for years.

This candidate was not failed by the difficulty of doctoral work. They were failed by a system that generated exactly the conditions competent people leave under.

What if my Chair chose my topic and I do not care about it?

This is common, and almost nothing is written about it.

Many candidates end up researching a topic their Chair chose.

Sometimes with good reason:

• The original was unworkable or too broad
• It was methodologically impossible with the resources available
• It was outside the Chair's ability to supervise

That is supervision doing its job, and a candidate should listen.

Sometimes the reasons are less about you:

• A Chair with a research program who needs the work done
• A dataset already available
• A topic the Chair is more comfortable with

None of that is corrupt, and all of it produces the same result for you.

Why a topic you do not care about is a real risk

Two things to take from this

If your data does not fit, the first question is whether the hypothesis was ever right for your design. Sometimes a null result is a finding. Sometimes it is a proposal-stage problem arriving late — and knowing which changes what you do next entirely.

And if your Chair has reshaped your topic into something you do not recognize, take that seriously now rather than later.

Chairs reshape topics for good reasons frequently — scope, feasibility, methodological fit. That is supervision working, and you should listen.

But a topic you cannot explain in your own words, or cannot say why it matters, is a topic that will fail you at Chapter 4 — because you will not be able to interpret findings for a question you never really asked.

The test is simple and it takes a minute. Can you say, without notes, what your study asks and why it matters? If you cannot, that is worth resolving before you collect a single data point — with your Chair, in writing, while changing it still costs you little.

See our Methodology at Proposal Stage page on making the questions, framework, and instrument fit each other — and you.