Data Analysis
Academic Research Analysis: From Research Questions to Defensible Results
GOSPELTRADER Research Desk · 12 September 2026 · 12 min read
Quick answer
Academic research analysis is defensible when every hypothesis maps to one pre-specified test, every test's assumptions are checked and reported, every result carries an effect size and confidence interval alongside the p-value, and the write-up presents findings without interpreting them — interpretation belongs in the discussion chapter.
1. Build the analysis matrix before touching the data
The single most useful artefact in a thesis is a table linking each research question to its hypothesis, its variables and their measurement levels, and the test that will answer it. Examiners look for exactly this alignment, and building it first prevents the common failure of collecting data that no available test can address.
| Research question | Variables | Measurement | Test |
|---|---|---|---|
| Does training affect performance? | Training (2 groups), score | Nominal, scale | Independent t-test |
| Do three teaching methods differ? | Method (3), score | Nominal, scale | One-way ANOVA + Tukey |
| What predicts job satisfaction? | 4 predictors, satisfaction | Scale, scale | Multiple regression |
| Is adoption related to sector? | Adoption (Y/N), sector | Nominal, nominal | Chi-square, Cramér's V |
| Does intention predict purchase? | Intention, purchase (Y/N) | Scale, nominal | Binary logistic regression |
2. Power and sample size, computed in advance
A power analysis run after the fact is not evidence; observed power is a deterministic function of the p-value and adds nothing. Compute the required N before collection, state the assumed effect size and its source, and report the calculation in the methods chapter.
Conventional effect sizes (Cohen)
d: 0.20 small 0.50 medium 0.80 large
f: 0.10 small 0.25 medium 0.40 large
r: 0.10 small 0.30 medium 0.50 large
f²: 0.02 small 0.15 medium 0.35 large
Required N at 80% power, α = .05 (two-tailed)
Independent t-test, d = 0.50 → 64 per group (128 total)
One-way ANOVA, 3 groups, f = 0.25 → 52 per group (156 total)
Correlation, r = 0.30 → 84
Multiple regression, 5 predictors, f² = 0.15 → 923. Assumption checking is part of the results, not a private step
State the assumption, the test or diagnostic used, the outcome, and the action taken. When an assumption fails, the remedy is reported rather than hidden — switching to a robust or non-parametric alternative is a legitimate, documented decision.
- • Normality — Shapiro-Wilk for N below 50, plus skewness and kurtosis within ±2 and a Q-Q plot. Remedy: transform, bootstrap, or use a non-parametric test.
- • Homogeneity of variance — Levene's test. Remedy: Welch's correction, reported as such.
- • Independence — a design property; Durbin-Watson near 2 for regression residuals. Remedy: a multilevel or panel model.
- • Linearity and homoscedasticity — residual-versus-fitted plot. Remedy: transformation or a polynomial term.
- • Multicollinearity — VIF below 5, tolerance above 0.20. Remedy: drop, combine or centre predictors.
- • Outliers — standardised residuals beyond ±3.29, Cook's distance above 1. Report the analysis with and without them.
4. Report the result, not the software output
Every inferential statement needs the test statistic, degrees of freedom, exact p-value (to three decimals; p < .001 below that), and an effect size with a confidence interval. Pasted SPSS or R output is not a results chapter.
t-test t(118) = 2.84, p = .005, d = 0.52, 95% CI [0.16, 0.88]
ANOVA F(2, 147) = 6.71, p = .002, η²ₚ = .08
Correlation r(148) = .42, p < .001, 95% CI [.28, .55]
Regression β = .34, t(145) = 4.21, p < .001; model F(4, 145) = 12.60,
p < .001, R² = .26, adjusted R² = .24
Chi-square χ²(2, N = 200) = 11.42, p = .003, Cramér's V = .24
Logistic OR = 1.87, 95% CI [1.21, 2.89], p = .0055. What examiners actually flag
Most corrections at viva concern presentation and discipline rather than arithmetic. These are the recurring ones.
- • Hypotheses in the introduction that no test in Chapter Four addresses.
- • Interpretation and citation inside the results chapter instead of the discussion.
- • p-values reported without effect sizes, or 'p = .000' copied from output.
- • 'Accepting' the null hypothesis rather than failing to reject it.
- • Assumptions asserted as met with no diagnostic reported.
- • Tables duplicating text verbatim, or tables that are unlabelled screenshots.
- • Causal language ('training improved performance') from a correlational design.
- • A sample size with no power justification.
6. Structure of the analysis chapter
A structure that reads well and survives scrutiny: response rate and data screening; demographic profile; reliability and validity of instruments; descriptive statistics for all study variables; assumption testing; hypothesis testing in the order the hypotheses were stated; and a short summary table of supported and unsupported hypotheses.
Each table is numbered, titled above in APA style, referenced in the text, and interpretable without the surrounding prose.
Frequently asked questions
What is the difference between Chapter Four and Chapter Five?
Chapter Four presents what the data show — tests, statistics, tables — with no citation and no explanation of why. Chapter Five interprets those findings against the literature, states implications and limitations, and makes recommendations. Mixing them is one of the most common correction requests.
Should I report exact p-values or thresholds?
Report exact values to three decimal places (p = .043), using p < .001 below that. Never write p = .000, which the software prints when the value is simply smaller than its display precision, and avoid reporting only 'p < .05'.
Do I need to report effect sizes if the result is not significant?
Yes. A non-significant result with a moderate effect size and a wide confidence interval indicates an underpowered study, not the absence of an effect. Reporting both allows the finding to be used in later meta-analysis.
Can I change my analysis plan after seeing the data?
You can, but the change must be disclosed and justified — for example switching to a non-parametric test after a normality violation. Undisclosed changes driven by which result looks better are p-hacking, and inconsistency between the methods and results chapters is what exposes it.
Want this analysis done for you?
We deliver cleaning, assumption testing, modelling and a reporting-ready write-up.