ACADEMIC WRITING GUIDE
How to Calculate and Interpret Confidence Intervals in a Dissertation (With Worked Examples)
A dissertation can report a precise-looking average, percentage or regression result while leaving readers unsure how much uncertainty surrounds it. Suppose a survey finds that 60% of responding students consider feedback useful. Is that percentage an exact description of everyone at the university? No: it describes the observed respondents, and any inference beyond them depends on the sampling design and assumptions.
A confidence interval (CI) helps quantify sampling uncertainty around an estimate under a specified statistical model. This guide shows how to calculate, interpret and write up intervals for a mean, a survey proportion and a difference between two group means. Each dataset is invented for teaching; none represents real participants, findings or completed research.
What is a confidence interval?
A point estimate is a single numerical summary, such as a sample mean of 72 points. A confidence interval adds lower and upper limits calculated by a stated method. The interval provides an indication of which values of the population parameter are compatible with the estimate under that method's assumptions.
In the usual frequentist framework, a 95% confidence procedure is designed so that, over many hypothetical repetitions of the same sampling process, approximately 95% of the resulting intervals would contain the fixed population parameter. It does not mean there is a 95% probability that the parameter lies inside this one already-computed interval.
An interval that is relatively narrow tends to indicate more precision than a wide one, when estimates and methods are comparable. But narrowness alone does not prove that sampling was representative, questions were valid, or confounding was controlled. The House of Commons Library's statistical literacy guide provides accessible additional context.
| Reported result | What it communicates | What it does not establish |
|---|---|---|
| Mean = 72 | Observed average among analysed cases | Exact population mean |
| 60% selected an answer | Observed proportion in the specified denominator | Representative population percentage without sampling assumptions |
| 95% CI = 68.80 to 75.20 | Model-based uncertainty surrounding an estimated mean | Probability that the fixed mean is inside this particular realised interval |
| Difference = 6, 95% CI = 1.05 to 10.95 | Estimated contrast and its precision under the stated procedure | A causal treatment effect merely because the interval excludes zero |
When should you report confidence intervals?
Intervals are especially useful for population means, proportions, group differences, correlations, regression coefficients and effect sizes. The type of interval must match the estimand—the quantity your research question asks you to estimate—and the design of the data.
A one-sample mean generally calls for a different calculation from a survey proportion. A paired before-and-after comparison should be analysed as paired data, not as two unrelated groups. Clustered sampling, repeated observations, sampling weights, missingness and model-based estimates can also require specialised interval procedures.
If your question is not yet aligned with its data, first review how to write a dissertation data analysis plan. For sample-size decisions and their impact on precision, see our sampling method and sample size guide.
The basic calculation idea
Many introductory intervals follow this pattern:
Confidence interval = estimate ± critical value × standard error.
The standard error estimates how variable the estimator would be across comparable repeated samples. The critical value comes from an appropriate reference distribution (such as a t or normal distribution) and the chosen confidence level. For some statistics—including small-sample proportions—better-performing interval methods do not follow the simple symmetric pattern.
A useful sequence is to identify the estimand, confirm the design, select the justified interval method, compute the components without rounding early, then interpret the endpoints in the units and direction of the original quantity.
| Research question | Parameter | Illustrative method | Important consideration |
|---|---|---|---|
| What is the average study score? | Population mean | One-sample t interval | Independence and distribution/sampling assumptions |
| What proportion agrees with a statement? | Population proportion | Wilson score interval | Binomial-style assumptions; sampling and nonresponse |
| How far apart are two independent group means? | Mean difference | Welch t interval | Independent groups; differing variances allowed |
| Did the same participants change after a programme? | Mean paired difference | Paired t interval, if appropriate | Differences within people, not independent samples |
| How precise is an adjusted model coefficient? | Conditional model parameter | Model-based coefficient interval | Design, estimation method and model diagnostics |
Worked example 1: A 95% confidence interval for a mean
Hypothetical research question: What is the average assessment score among first-year students represented by the study's sampling design?
Assume a study analyses 40 independent student scores, with a sample mean of 72 points and sample standard deviation of 10 points. This illustrative calculation assumes that a one-sample t procedure is appropriate for the data and recruitment design.
Step 1: Identify the inputs
Analysed sample size, n = 40.
Sample mean = 72 points.
Sample standard deviation, s = 10 points.
Degrees of freedom, df = n − 1 = 39.
Two-sided confidence level = 95%.
Because the population standard deviation is unknown, use the sample standard deviation and a t critical value. OpenStax's t-interval explanation describes this method.
Step 2: Calculate the standard error
SE = s / √n.
SE = 10 / √40 = 1.5811.
This is the estimated standard error of the sample mean, not the standard deviation of individual students' scores.
Step 3: Obtain the critical value
For 95% confidence with 39 degrees of freedom, the two-sided t critical value is approximately 2.0227. Use a statistical table or software rather than assuming that 1.96 is always appropriate.
Step 4: Calculate the limits
Margin of error = 2.0227 × 1.5811 ≈ 3.20 points.
Lower limit = 72 − 3.1982 = 68.8018.
Upper limit = 72 + 3.1982 = 75.1982.
The 95% confidence interval is approximately 68.80 to 75.20 points.
| Component | Value | Explanation |
|---|---|---|
| Analysed observations | 40 | Independent student scores |
| Sample mean | 72.00 | Point estimate, in score points |
| Sample SD | 10.00 | Between-student spread of scores |
| Standard error | 1.5811 | 10 divided by square root of 40 |
| 95% t critical value, df = 39 | 2.0227 | Two-sided reference value |
| Margin of error | 3.1982 | Critical value × SE |
| 95% interval | 68.80–75.20 | Rounded only for reporting |
Dissertation-ready reporting paragraph
“In this hypothetical sample of 40 students, the mean assessment score was 72.00 points (SD = 10.00). A two-sided, one-sample t-based 95% confidence interval for the mean extended from 68.80 to 75.20 points. Under the stated sampling and modelling assumptions, this interval summarises uncertainty in the estimated population mean; it does not address possible selection bias.”
Why this example works: It identifies the statistic and units, chooses the t distribution for an unknown population SD, distinguishes SD from SE, and avoids claiming certainty about every student in a larger population.
Worked example 2: A confidence interval for a survey percentage
Hypothetical question: What proportion of eligible students considers written assignment feedback useful?
Suppose 120 eligible respondents answer the item and 72 select “useful.” We will illustrate the Wilson score interval, which often has more dependable coverage than the naive normal (Wald) interval, particularly when sample sizes are small or proportions are near zero or one. This does not repair a nonprobability sample or nonresponse bias.
Step 1: Calculate the sample proportion
The sample proportion is 72 / 120 = 0.60 (60%).
The denominator is the 120 valid answers to this item, not necessarily everyone invited to take the questionnaire. Keep these counts separate. See our survey response-rate guide for that distinction.
Step 2: Apply the Wilson method
For a 95% interval, use z ≈ 1.96, p̂ = 0.60 and n = 120. The Wilson limits are:
[p̂ + z²/(2n) ± z × √(p̂(1 − p̂)/n + z²/(4n²))] / [1 + z²/n].
Calculate the denominator first:
Wilson denominator = 1 + 1.96² / 120 = 1.032013.
The resulting Wilson lower and upper limits are approximately 0.51056 and 0.68324.
Using the Wilson method, the 95% interval is approximately 51.1% to 68.3%.
This is an interval for the underlying proportion under the method's assumptions, not a range stating that 51.1–68.3% of the 120 respondents endorsed the item. Among those respondents, the observed fraction remains exactly 72/120 = 60%.
| Quantity | Value | Interpretation |
|---|---|---|
| Valid item responses | 120 | Denominator for this item |
| “Useful” responses | 72 | Numerator |
| Observed proportion | 0.6000 | 60% of the analysed respondents |
| Wilson 95% lower limit | 0.5106 | Approximately 51.1% |
| Wilson 95% upper limit | 0.6832 | Approximately 68.3% |
Dissertation-ready reporting paragraph
“Among 120 respondents with a valid answer to the feedback-usefulness item, 72 (60.0%) selected ‘useful.’ The Wilson 95% confidence interval was approximately 51.1%–68.3%. The interval expresses sampling uncertainty under the chosen method. As participation and nonresponse may not have been random, it should not be treated as eliminating selection bias.”
Why this example works: It separates a response percentage from a survey response rate, specifies Wilson rather than concealing the procedure, and recognises that a confident numerical range does not establish a representative sample. The NIST proportion-interval documentation compares the Wilson approach with simpler alternatives.
Worked example 3: A confidence interval for the difference between two independent means
Hypothetical question: How different are assessment scores between two separately sampled student groups?
Suppose Group A has 64 students, mean score 76 and sample SD 12; Group B independently has 64 students, mean score 70 and sample SD 16. The research question asks for Group A minus Group B. We use a Welch t interval, which does not require equal population variances, assuming independence and appropriate sampling/distribution conditions.
Step 1: Estimate the difference
The observed mean difference is 76 − 70 = 6 points.
Keeping the subtraction direction explicit prevents an apparently reversed interpretation later.
Step 2: Estimate its standard error
For independent samples, the estimated SE is:
SE = √(sA²/nA + sB²/nB).
SE of the difference = √(12²/64 + 16²/64) = 2.50 points.
A Welch–Satterthwaite calculation gives approximately 116.84 degrees of freedom; the two-sided 95% t critical value is approximately 1.9805.
Step 3: Calculate and interpret the interval
Using Welch's t critical value of approximately 1.9805, the margin is about 4.95 points. Unrounded margin ≈ 4.95119 points.
Unrounded lower limit ≈ 1.04881 points. The approximate 95% interval for the difference is 1.05 to 10.95 points.
| Component | Group A | Group B |
|---|---|---|
| Analysed n | 64 | 64 |
| Mean score | 76 | 70 |
| Sample SD | 12 | 16 |
| Definition of contrast | Group A minus Group B | — |
| Estimated difference | 6 points | — |
| Welch estimated SE | 2.50 | — |
| Welch approximate df | 116.84 | — |
| 95% CI for A − B | 1.05 to 10.95 points | — |
Dissertation-ready reporting paragraph
“In this hypothetical comparison, Group A (n = 64, M = 76, SD = 12) scored an average of 6 points higher than Group B (n = 64, M = 70, SD = 16). The estimated difference (A − B) had a Welch 95% confidence interval of approximately 1.05 to 10.95 points. The positive interval is consistent with higher scores in Group A under this model, but this observational comparison does not establish why the groups differed.”
Why this example works: It defines the comparison, uses the proper independent-group uncertainty formula, avoids forcing equal-variance assumptions and distinguishes a difference in observed means from a causal effect.
How to interpret intervals that include or exclude zero
For many two-sided effect measures expressed as differences, zero means no difference. If a conventional 95% confidence interval excludes zero, it often corresponds to a two-sided p-value below .05 for the same test, assumptions and estimation procedure. If it includes zero, the result can be compatible with no difference—but also with other values inside the interval.
Do not turn that observation into a universal rule. For odds ratios or risk ratios, a ratio of 1, not zero, commonly represents no association. And an interval excluding a null value does not automatically imply a large, important, unbiased or causal effect.
A confidence interval should be read for direction, magnitude, precision and meaningful thresholds. For example, if improving scores by at least five points is considered practically worthwhile, the Group A minus Group B interval (1.05–10.95) includes both values below and above that threshold. The estimate may favour Group A, but the interval does not establish that the practically important threshold has been met with high precision.
If your dissertation uses effect-size measures, our effect sizes guide explains how effect estimates and their substantive meaning differ from statistical significance.
Five decisions that make a confidence interval trustworthy
1. State what population or process the estimate concerns
A confidence interval for a convenience sample does not magically justify inference to all students in a country. Define the target population, recruitment and selection mechanism, and any constraints on generalisation.
2. Explain the statistical method
Name a one-sample t interval, Wilson score interval, Welch interval, or another method as appropriate. Report whether it is two-sided or one-sided when relevant; do not silently mix methods.
3. Use the correct effective sample and unit
Missing answers can change the denominator. Students nested within classes are not necessarily independent observations. Repeated measures from the same participant require procedures that respect the dependence. Weighting and complex survey designs may affect standard errors.
4. Check assumptions and influential observations
For means, assess whether the data, sample size and dependence structure make t-based inference reasonable. For models, attend to specification and residual or influence diagnostics where relevant. Neither a large sample nor a narrow interval fixes systematic measurement or selection problems.
5. Keep precision distinct from certainty
Report intervals to sensible decimal places in the original units, and reflect the uncertainty in the conclusion. Higher confidence levels typically create wider intervals from the same data under the same procedure. Larger informative samples often improve precision, but only within the appropriate statistical design.
How to present a confidence interval in your dissertation
Place the estimate and interval close together, define any abbreviations and keep the numbers consistent between text and tables. A short results paragraph often works better than repeating every step of the calculation.
A reusable pattern is:
“Among [analysed n and population description], the estimated [parameter] was [point estimate, units] (95% CI [lower, upper]), obtained using [method]. Under [study design and relevant assumptions], this suggests [careful interpretation]. [Important limitation] restricts [generalisation or causal inference].”
Use actual numerical outputs from your completed analysis, not values copied from these teaching examples. For broader results-writing conventions, review how to report statistical results in a dissertation and how to present research data in tables and figures.
Common confidence-interval mistakes
Treating a 95% CI as a 95% probability statement. A conventional frequentist interval has a long-run coverage interpretation; it is not a posterior probability distribution.
Using 1.96 in every situation. A small-sample mean with unknown population SD commonly calls for a t critical value. Complex methods may call for entirely different uncertainty procedures.
Using a simple normal interval for every proportion. For rare outcomes or small samples, the basic Wald interval may perform poorly; consider a justified alternative such as Wilson.
Confusing standard deviation with standard error. SD describes variation among observed values; SE concerns uncertainty in the estimated statistic under the assumed sampling process.
Claiming the interval contains 95% of students' scores. A confidence interval for the population mean is not a prediction interval or reference range for individual observations.
Calling non-significance evidence of no effect. A wide interval containing zero may also contain large, meaningful effects. Equivalence and noninferiority claims require their own prespecified margins and appropriate tests.
Hiding missingness or changing the denominator. Explain how many observations contributed to the estimate, especially when different questionnaire items have different response counts.
Claiming causal effects from observational comparisons. An interval for a group difference still cannot remove uncontrolled confounding or reverse causality.
Frequently asked questions
Is a 95% confidence interval always better than a 90% interval?
Not inherently. A 95% interval is commonly reported and is typically wider than a 90% interval computed by a comparable method on the same data. Choose the confidence level according to the research plan, discipline and inferential purpose rather than selecting it after seeing the results.
What if my 95% confidence interval is very wide?
The estimate may be imprecise. Review the sample size, outcome variation, data structure, model and measurement quality. Report the uncertainty candidly. Do not narrow the interval by choosing an unjustified method or excluding inconvenient data.
Can I report a confidence interval without a p-value?
Often yes, depending on your analysis, requirements and reporting standard. An effect estimate with a well-specified interval can convey direction, magnitude and uncertainty more directly than an isolated p-value. Some research questions still call for a corresponding test or a formal decision rule.
Do I need raw data to calculate a confidence interval?
Sometimes adequate summary statistics suffice, as in the one-sample mean worked example. Other methods need raw observations, contingency counts, model covariance estimates, weights or more complex design information. Never reconstruct an interval from insufficient inputs or assume its assumptions are met without checking.
Can a confidence interval be used in qualitative research?
Ordinary statistical confidence intervals generally apply to quantitative estimands and sampling models. Rich interview accounts are not usually represented by a confidence interval for a theme. If a qualitative study also counts or estimates proportions, the meaning of any interval depends on its design and justified quantitative assumptions.
Does a narrow confidence interval mean my results are valid?
No. A narrow interval can be precise around a biased estimate. Sampling design, measurement validity, data quality, missingness, model assumptions and ethics remain essential.
Final dissertation checklist
Have I clearly identified the quantity my interval estimates?
Have I chosen a method suited to the outcome and study design?
Have I verified the sample count, denominator, units, assumptions and critical value?
Have I carried full precision through intermediate steps and rounded only in the report?
Are the estimate and two interval limits consistent across text and tables?
Have I distinguished sampling uncertainty from selection bias and measurement error?
Have I interpreted the interval relative to meaningful effects, not only the null threshold?
Have I avoided fabricated analyses, unjustified causal claims and probability misinterpretations?
Final thoughts
A confidence interval becomes useful only when the reader knows what is being estimated, how uncertainty was calculated and what the research design can justify. Learn the method, check the arithmetic and explain the limits of the evidence in ordinary language. Your dissertation will be stronger for accurate uncertainty reporting—not for adding another statistic to the page.
If you need help improving the clarity of your analysis plan, results reporting or explanation of statistical findings, explore Acadelyra's academic support services. You remain responsible for your data, research decisions, original analysis and institution's academic-integrity requirements.
Sources and further reading
Penn State STAT 200: Inference for One Sample — sampling uncertainty and mean intervals.
OpenStax: A Single Population Mean Using the Student's t-Distribution — t critical values and confidence intervals for means.
NIST: Proportion Confidence Interval — normal, adjusted and Wilson intervals.
House of Commons Library: Confidence Intervals and Statistical Significance — plain-language statistical literacy guidance.