A master's student sends us her results chapter for review. She has 14 hypotheses, 9 SPSS tables, and a p-value on every page. On the surface it looks complete. But within the first two pages we find ordinal Likert data in an independent-samples t-test, no effect sizes anywhere, and the phrase "highly significant" used six times. Her supervisor will send it back. We know because we have reviewed over 300 thesis statistics chapters, and the same ten mistakes appear in roughly 7 out of 10 of them.
Choosing the Wrong Analysis
1. Using the Wrong Test for the Data Type
A student measures satisfaction on a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree) and runs an independent-samples t-test to compare two groups. The t-test assumes a continuous, interval-level dependent variable. Likert items are ordinal. The distance between "disagree" and "neutral" is not guaranteed to equal the distance between "neutral" and "agree." With a small sample of 28 per group, the t-test produces t(54) = 2.14, p = .037. The Mann-Whitney U test on the same data gives U = 289, p = .061. One says significant, the other does not. The wrong test gave a misleading answer.
The fix: match the test to the measurement level. Nominal data needs chi-square. Ordinal data needs nonparametric tests like Mann-Whitney or Kruskal-Wallis. Continuous data with met assumptions can go into t-tests and ANOVA. Our guide to choosing the right statistical test walks through this decision step by step.
2. Not Checking Assumptions Before Running Tests
We once received a chapter with 11 independent-samples t-tests. Not one had a Shapiro-Wilk test or a Levene's test reported anywhere. The student had clicked Analyze > Compare Means > Independent-Samples T Test in SPSS and trusted whatever came out. Two of those 11 variables had skewness values above 2.0 and kurtosis above 7.0. For those variables, the t-test's normality assumption was clearly violated, and the results were unreliable.
Every parametric test carries assumptions. The t-test assumes normality of the dependent variable within each group and homogeneity of variance across groups. ANOVA adds the assumption of independence of observations. Pearson's correlation assumes a linear relationship and bivariate normality. Before running any parametric test, check normality with Shapiro-Wilk (for samples under 50) or inspect skewness and kurtosis values (for larger samples). Check homogeneity of variance with Levene's test. If assumptions are violated, switch to the nonparametric alternative or apply a transformation.
3. Running Multiple T-Tests Instead of ANOVA
A student has three groups (first-year, second-year, third-year students) and wants to compare their exam scores. Instead of running a one-way ANOVA, she runs three separate independent-samples t-tests: first vs. second, first vs. third, second vs. third. Each test uses α = .05. The probability of at least one Type I error across three tests is not .05. It is 1 − (1 − .05)³ = .143, nearly three times the intended rate. With five groups and ten pairwise comparisons, that probability climbs to .40.
ANOVA exists precisely for this situation. It tests all groups simultaneously with a single F-test, holding the family-wise error rate at .05. If the omnibus F is significant, post-hoc tests (Tukey's HSD, Bonferroni, or Games-Howell depending on equal variances) identify which specific pairs differ. The t-test vs. ANOVA comparison explains when each is appropriate.
Reporting Problems
4. Reporting p < .05 as "Highly Significant"
Statistical significance is binary. The result either crosses your pre-set alpha threshold or it does not. There is no "highly significant" at p = .001 versus "barely significant" at p = .048. Both reject the null hypothesis at α = .05. A p-value of .001 does not mean the effect is stronger than one with p = .04. It means the evidence against the null hypothesis is stronger, which is a different claim entirely. Effect size tells you about strength.
The APA 7th edition (2020, Section 6.36) does not use the term "highly significant." Neither should your thesis. Write "statistically significant" or "not statistically significant." Nothing else. For a deeper look at what p-values actually tell you, see our p-value interpretation guide.
5. Not Reporting Exact P-Values
Writing "p < .05" when SPSS gave you p = .023 throws away information. Your reader cannot tell whether the result just cleared the threshold or cleared it by a wide margin. APA 7th edition requires exact p-values reported to two or three decimal places: p = .023, not p < .05. The only exception is when p is very small. For values below .001, write p < .001 because SPSS often rounds them to .000, and reporting p = .000 implies certainty that does not exist.
In SPSS output, the column labeled "Sig." or "Sig. (2-tailed)" contains the exact p-value. Copy that number directly. If it reads .000, report p < .001. If it reads .047, report p = .047, not p < .05.
6. Missing Effect Sizes
We reviewed a 40-page results chapter last year that reported 23 inferential tests. Zero effect sizes. Every result read like: "t(86) = 2.41, p = .018. The difference was statistically significant." That tells the reader something happened but not how much. APA 7th edition (2020, Section 6.44) states: "For each primary outcome, provide an estimate of the effect size and, where possible, a confidence interval."
Cohen's d belongs with t-tests. Partial eta-squared (ηp²) belongs with ANOVA. Cramér's V goes with chi-square. R² goes with regression. Academic Stats Agent calculates and reports these automatically, but if you are working in SPSS, check the "Estimates of effect size" box under Options in the relevant dialog. For t-tests, you will need to calculate Cohen's d by hand or use the formula d = 2t / √df as an approximation. See our effect size guide for the full reference table.
Data Handling Errors
7. Ignoring Missing Data
SPSS handles missing data silently. If your dataset has 200 respondents but 35 left question 12 blank, SPSS will run the analysis on 165 cases without a warning. The output table will show N = 165 in small print, but most students never notice. The problem compounds when different variables have different patterns of missingness. One test runs on 182 cases, another on 164, another on 191. Your sample keeps shrinking, and the results may be biased if the data is not missing completely at random.
Before running any analysis, check the extent and pattern of missing data. In SPSS, go to Analyze > Descriptive Statistics > Frequencies and look at the "Missing" row. If more than 5% of values are missing on a variable, you need a strategy: listwise deletion, pairwise deletion, mean imputation, or multiple imputation, depending on the mechanism and proportion. Our missing data guide covers how to decide.
8. Using the Wrong Descriptive Statistics
A student reports "the mean satisfaction score was 3.7 (SD = 0.9) on a 5-point Likert scale." The mean of ordinal data is technically meaningless because the intervals are not equal. For a single Likert item, the correct central tendency measure is the median, and the correct dispersion measure is the interquartile range (IQR). The median might be 4.0 (IQR = 3.0–4.0), which paints a different picture than the mean of 3.7.
There is a common exception. When you sum or average multiple Likert items into a composite scale (e.g., a 10-item anxiety questionnaire scored 10–50), many methodologists treat the composite as approximately interval. In that case, the mean and standard deviation become acceptable. The distinction matters: single items get medians, composite scales can get means. Reporting the wrong descriptive for your measurement level is a clear sign to reviewers that the student did not think carefully about data types.
| Data Type | Central Tendency | Dispersion | Example Variable |
|---|---|---|---|
| Nominal | Mode | Frequency / Percentage | Gender, Faculty |
| Ordinal (single item) | Median | IQR, Range | Single Likert item |
| Ordinal (composite scale) | Mean (if treated as interval) | SD | 10-item satisfaction scale |
| Interval / Ratio | Mean | SD | Age, Test score, GPA |
Interpretation and Presentation
9. Confusing Correlation with Causation in the Discussion
A Pearson correlation of r = .47 (p < .001) between hours of social media use and anxiety scores does not mean social media causes anxiety. It means they move together. Anxious students might scroll more as a coping mechanism. A third variable, like loneliness, might drive both. Without random assignment and experimental control, the design is cross-sectional and correlational. The data cannot distinguish cause from consequence.
The mistake shows up in the discussion chapter. The student writes: "Social media use increases anxiety among students, as shown by the significant positive correlation." That sentence makes a causal claim from a correlational finding. The correct framing is: "Higher social media use was associated with higher anxiety scores." Save causal language for experimental designs with manipulated independent variables and proper control groups. Our correlation vs. regression article explains the distinction in more detail.
10. Copy-Pasting SPSS Tables Instead of Formatting Results in Text
SPSS output tables are designed for analysts, not for thesis readers. They contain far more information than your results section needs, and they look nothing like APA format. We regularly see results chapters where the student has pasted 15 raw SPSS tables with labels like "Group Statistics," "Independent Samples Test," and "Correlations" directly into the Word document. Some are cropped oddly. Some still have the grey SPSS background. None contain effect sizes because the default SPSS output for t-tests does not include them.
APA format requires results reported in text, not in software output. A proper result looks like this: "Students who received the intervention (M = 76.4, SD = 8.2) scored significantly higher than the control group (M = 71.1, SD = 9.0), t(88) = 2.93, p = .004, d = 0.61." That single sentence replaces two SPSS tables (Group Statistics and Independent Samples Test) and adds the effect size that SPSS omitted. If you need to include a table, build it yourself in Word with only the columns your reader needs. Our SPSS output reporting guide shows exactly how to translate each type of output into APA text.
Key takeaway: Most thesis revision requests come down to the same ten errors: wrong test for the data type, unchecked assumptions, misused significance language, missing effect sizes, ignored missing data, inflated Type I error from multiple t-tests, causal claims from correlational data, vague p-value reporting, mismatched descriptives, and raw SPSS output instead of APA-formatted text. Fix these before submission and you remove the most common reasons supervisors send chapters back. Academic Stats Agent handles most of them automatically, but knowing what to look for makes you a better researcher regardless of the tool you use.