The Week Before Your Viva: What to Revise and What to Let Go

Seven days to go, and you have opened the regression chapter of a textbook again and spent twenty minutes reading about multicollinearity, a topic that takes up one sentence of your thesis. That is not preparation, it is anxiety wearing the costume of revision. What will actually happen in that room is that two or three people will ask you about specific choices in your work (why that sample, why that analysis, how far your conclusions can travel) and you will either answer from memory or fumble through a 200-page PDF. The difference between those two is decided this week, and it is not decided by reading more theory.

The numbers you should be able to say without looking them up

There is a short list of figures examiners expect to hear straight from you, with no pause and no scrolling. Start with your N, but with its history attached: “214 participants” is not enough, the whole route has to come out. “We recruited 240 and analyzed 214: 18 did not complete the post-test and 8 failed the inclusion criteria once the questionnaires were checked.” That level of detail kills the follow-up question before it is asked. If you have subsamples, know their sizes too: the n per group, per wave, and per analysis that did not use the full sample.

Next come the degrees of freedom of your main tests, which are the signature of your design. A t(212) says you compared two independent groups with 214 people; an F(2, 211) says three groups compared in that same sample with no covariate, because the moment you add one the denominator drops to 210. If someone asks how many cases went into that model and your degrees of freedom do not match what you just said, a crack opens that can swallow twenty minutes of your viva. Checking this takes half an hour and it is the best half hour of the week.

Then your effect sizes with their confidence intervals, not your p values. The sentence you want ready sounds like this: “The intervention group improved relative to control, t(212) = 2.87, p = .005, d = 0.39, 95% CI [0.12, 0.66]; by Cohen's conventions that sits between small and medium, and the lower bound does not rule out something close to trivial.” The second half is the part that shows whether you understand your result or merely copied it. If you are unsure how to say a standardized mean difference out loud, the guide to reporting Cohen's d in APA 7 has the wording and the conventions from Cohen (1988), presented as what they are: rough field benchmarks, not thresholds anyone is obliged to obey.

Finish with reliability in your own sample for each instrument (alpha or omega, your value and not the one from the original validation paper: if yours came out at .68, far better that you say it first) and one power figure. If you ran an a priori calculation, know the effect you assumed and where it came from; if you did not, know the smallest effect your N could detect at a reasonable power. Six or seven numbers in total. They fit on one sheet of paper, and that sheet is your real revision material.

The decisions you have to justify, not the theory behind them

Almost no uncomfortable question in a viva is theoretical. They are questions about choices you made: why this analysis and not another, what you did with missing data, why you excluded those eight cases, why you did or did not correct for multiple comparisons, why you transformed that variable. The theory of every test is in any textbook and your examiners know it; what is in no textbook is the reasoning that took you to your decision with your data.

Prepare each one in three beats: what you did, why it was reasonable there, and what would have changed had you done the opposite. “I used multiple imputation because missingness reached 11% and the pattern did not appear to depend on the outcome; with complete cases the coefficients moved in the third decimal and the conclusions were unchanged.” That last clause, the sensitivity check, defuses the question before the “but what if” arrives. You do not need to have run everything: you need to know what would happen.

Assumptions deserve their own pass, because that is where most people improvise. Know what each test you ran required, how you checked it, and what you found, including the parts that came out ugly. If residual normality was doubtful and you kept the parametric model because the sample size could carry it, say exactly that. If you want a checklist to tick off, the walkthrough of statistical assumptions and how to check them covers what to verify for each test.

One detail that changes the temperature of the room: decisions are defended, not apologized for. “It is a limitation, I know, I should have done something else” is an invitation to dig. “I chose this for these two reasons, the alternative would have added this, and I would consider it for the next study” closes the topic. Same information, opposite outcomes.

What is not worth revising

Formulas. Nobody is going to ask you to derive the t distribution or write a maximum likelihood estimator by hand. If someone asks what a p value actually is, they are testing your one-sentence interpretation (the probability of a result at least as extreme as yours if the null hypothesis were true), not your ability to derive anything. Memorizing formulas in the last week eats hours and improves no answer.

Syntax. It does not matter whether you used SPSS, R, Mplus, Stata or jamovi: no examiner will ask for the argument name you used to request bootstrapping. At most you will be asked which software you used and why, and that is a ten-second answer. If your committee is the kind that asks to see code, that scenario is solved by having a tidy, commented script, not by memorizing it.

Tests you never ran, and their assumptions. Revising repeated-measures ANOVA from scratch when you fitted a mixed model is wasted time, with one exception: the obvious alternative to your main analysis, which does deserve attention (more on that below). The rest of the postgraduate syllabus is not on the table.

Decimals in secondary tables, exact citations from your whole introduction, the publication year of every instrument. If you are asked about one correlation in Table 3 you can look it up; having the document in front of you is standard in almost every viva or defense. What you cannot look up is “what was your main hypothesis?”. The working rule: from memory, whatever defines your work; look up whatever merely details it.

The consistency pass: half an hour that beats three evenings of revision

Before revising anything, check that your document does not contradict itself. It is the highest-return job of the week, because a numerical inconsistency is one of the few things that can drag a whole examination somewhere you do not get out of gracefully. An examiner who finds impossible degrees of freedom does not let it go: they stay there, and from that moment they read everything else with a magnifying glass.

Four checks, in this order. First, that degrees of freedom match the N you report and the missingness you report, test by test. Second, that every p value is consistent with its statistic and its degrees of freedom; recomputing them is quick, and that is where the transcription error you made eight months ago surfaces. Third, that any figure appearing twice (abstract and results, text and table) is the same figure. Fourth, uniform rounding, with two decimals for most statistics and no leading zero on anything that cannot exceed 1.

Do this pass on paper or on screen with a pencil, not in your head. And if you find something, it is cheap to fix: a transcription error is handled in the room with one sentence (“there is a typo in the degrees of freedom in Table 4, the correct value is 211, and I will correct it in the final version”) and the matter closes there. The expensive version is discovering it live because someone else pointed at it.

Use the same pass to tab the document: bookmarks on the descriptives table, the main analysis table, the model figure and the limitations list. Those are the four places you will want to reach in under three seconds while three people watch you do it.

Three questions that knock people sideways, and how to answer them

“Why didn't you do X?” X is nearly always the analysis that particular examiner would have run: a mixed model instead of your ANOVA, an SEM instead of your regressions, a Bayesian analysis instead of your null hypothesis test. It is not an ambush, it is how they think. The answer that works acknowledges the alternative, says what it would add, and explains why yours fitted your question and your data. What sinks people is pretending it does not exist. If the analysis they are proposing is one you do not command, the territory is mapped in what to do when you are asked for an analysis you do not master: the honest route always scores higher than the bluff.

“Is that effect meaningful?” Repeating that it was significant will not do. Translate the number into the scale your outcome is measured on and into the context of your field: how many scale points that is, whether anyone would notice it, how it compares with what similar interventions report. A d = 0.39 from a brief, cheap intervention may be interesting; the same d from twenty sessions of therapy is a different conversation. Have that comparison ready before they do.

“You cannot conclude that from this design.” This usually follows a causal verb that slipped into a cross-sectional study: predicts, produces, improves. The answer is not to defend the sentence, it is to show you distinguish association from causation and to rephrase out loud: “you are right, the correct wording is that the relationship is associational, and the design can neither rule out third variables nor establish temporal order”. Conceding fast and precisely closes the point; arguing extends it.

There is a fourth that only unsettles people who did not see it coming: what you make of a result that was not significant. If your thesis has one, prepare it as a finding with an interpretation (power, interval width, equivalence) rather than a failure to be steered around.

Two mock runs out loud, and no more

Knowing an answer and saying it are two different skills, and only the second is trained by speaking. A silent rehearsal, re-reading your own pages, gives you the feeling of mastery and teaches you nothing; the first time you say “the indirect effect was significant, with a bootstrap interval of...” out loud you find out where your tongue trips and which sentence you have not actually built.

Two mock runs are enough. In the first, question yourself with your phone recording: twenty minutes on methods and results, answers capped at two minutes. Listening back is uncomfortable and that is where the learning is. In the second, find a person. Someone from another discipline is better, because they force you to explain without jargon, and that is where it shows whether you understand your analysis or only recite its label.

What does not work is rehearsing without knowing what against. Before the mock run, have in front of you the likely questions for your specific work rather than a generic list (the questions that come up in almost every defense make a decent starting grid once you map them onto your own thesis), and know your examiners: a UK viva is usually a closed room with two examiners going line by line through the thesis for two or three hours, a US defense often opens with a public presentation before the committee questions, and in Australia the written examiners' reports still carry much of the weight even where an oral has been added. Same statistics, different room.

Time yourself. The two-minute answer is a skill in itself: someone who talks for seven minutes about a methods choice signals insecurity and, on top of that, hands over three new questions for every extra thing mentioned. Finish and stop.

The seven days, and what you carry in

Days 7 and 6. A full re-read of your own thesis with a pencil, the numerical consistency pass, and building the number sheet: N and its route, degrees of freedom for every main test, effect sizes with intervals, reliabilities and your one power figure. One sheet, handwritten if you can: writing it out forces you to decide what earns a place on it.

Days 5 and 4. The decisions. Write a three-line justification for each one (analysis chosen, missing data, exclusions, corrections, transformations) and add its “what would have changed if”. This is the real studying, and it is not studying statistics: it is studying your thesis.

Days 3 and 2. The two mock runs out loud, plus your list of real limitations, yours and not the textbook ones. One well-chosen limitation argued well adds to your case; three generic ones subtract from it.

The day before. No statistics at all. Go over the presentation, check the logistics (room, projector, link, who is examining and which methodological school each of them comes from) and sleep. A rested brain retrieves a number that an exhausted one cannot find even after six passes.

On the day. You carry three things: the thesis tabbed at those four places, the number sheet, and a notebook to write questions down as they are asked (writing buys you two seconds of thought and stops you answering only the last clause of a long question). Nothing else. If it is not in your head by now, it is not going to get there in the corridor.

Frequently asked questions

How much time should I spend revising statistics before my viva?

Far less than the anxiety demands. Two sessions on your own numbers and your own document (the number sheet and the consistency pass) plus two more writing out the justification for each methodological decision will cover it. The rest of the week pays better in mock runs out loud than in re-reading textbooks: general theory is rarely what fails in a viva, what fails is not having built the sentence that justifies a specific choice.

Do I need to memorize the formulas?

No. No serious examiner asks you to derive a distribution or write an estimator by hand. What they may ask for is a one-sentence interpretation of what you report: what a p value actually means, what a 95% confidence interval tells you, why an effect size does not grow with sample size the way significance does. That is understanding, not memory, and you prepare it by explaining it out loud to someone outside your field.

Can I bring notes or a sheet with my numbers into the room?

Yes, and it is standard practice: the thesis on paper or on screen, a sheet with your main figures, and a notebook for writing questions down. Looking up a secondary number in front of examiners costs you nothing. What does read badly is looking things up to answer what defines your work: your N, your design, your main hypothesis, your headline effect. From memory whatever defines it, look up whatever details it.

What do I say if I cannot remember a number during the viva?

You look it up, out loud and without drama: “that is in Table 3, one moment”. Two seconds of page-turning bothers nobody. What is remembered badly is inventing an approximate figure that then fails to match the table, because from there the examiners start checking everything else. If you genuinely do not have it, say so and offer to supply the correct value for the final version.

What if the examiners ask about an analysis I did not run?

Acknowledge the alternative, say what it would add, and explain why your choice was reasonable for your question and your data. If you can also say what you would expect that other analysis to show, better still. What does not work is improvising command of a technique you do not use: any panel spots it by the second follow-up, and the honest route (“I have not worked with it, it strikes me as a sensible line for the next study”) costs far less.

Is a UK viva different from a US defense in what gets asked about statistics?

The format differs more than the content. A viva is usually a closed session with an internal and an external examiner working through the thesis in detail, a US defense often starts with a public presentation before the committee takes over, and in Australia the examiners' written reports carry much of the weight even where an oral has been added. In all three the statistical questions land in the same places: sample size justification, assumptions, the alternative analysis, effect sizes and how far your conclusions can travel.

Walking in with the whole textbook revised impresses nobody; walking in knowing your N, your degrees of freedom, your effect sizes and the reason behind every decision changes how that hour sounds from the first minute. And if what is keeping you awake is one specific analysis you know will not survive a follow-up question, that gets looked at before the viva rather than after: that is what my statistical consulting is for.