ACADEMIC WRITING GUIDE

How to Interpret Logistic Regression Results in a Dissertation (With Worked Examples)

By Acadelyra Editorial TeamPublished

A dissertation may ask whether students who receive academic feedback are more likely to submit an assignment on time, whether participation in a training programme is associated with course completion, or whether access to a learning resource predicts passing an assessment. When the outcome has two possible categories, binary logistic regression can be an appropriate analysis.

The difficulty often begins after the software finishes. Its output may contain coefficients labelled B, Exp(B), odds ratios, confidence intervals, likelihood statistics and classification measures. None should be copied into the dissertation without understanding what it means.

This guide explains how to choose, read, calculate, check and report binary logistic regression results. The worked examples are hypothetical teaching illustrations. They are not findings from real student participants or output from a fitted research dataset.

What is binary logistic regression?

Binary logistic regression models the relationship between explanatory variables (predictors) and the probability of an outcome coded as one of two categories, commonly 0 and 1. For example, a researcher may code submitted on time = 1 and not submitted on time = 0.

Unlike ordinary least-squares linear regression, logistic regression does not model the outcome probability directly as an unrestricted straight line. Instead, the model uses the log-odds of the event:

log[p / (1 − p)] = β₀ + β₁X₁ + β₂X₂ + …

Here, p is the probability of the outcome coded 1, β₀ is the intercept, and each β describes a change in log-odds associated with a predictor, holding the other included predictors fixed.

Check the event coding first. If a programme reports the odds of not submitting on time rather than the odds of submitting on time, every interpretation changes direction.

Logistic regression is not the automatic choice for every categorical outcome. Multinomial models may fit nominal outcomes with more than two categories, while ordinal models may be suitable for ordered categories. Your research question, outcome definition and intended interpretation should drive the choice. Our guide to writing a dissertation data analysis plan can help connect model selection to an actual research objective.

Probability, odds and odds ratios: three different quantities

The words sound similar but refer to different things.

  • Probability: the proportion of possible outcomes expected to be events, on a scale from 0 to 1.
  • Odds: probability divided by one minus probability, p/(1 − p).
  • Odds ratio (OR): the ratio of the odds in two conditions, or the multiplicative change in odds for a one-unit change in a predictor under the fitted model.

A 60% probability means p = 0.60; the corresponding odds are 0.60/0.40 = 1.5, or 1.5 to 1. A 30% probability means odds 0.30/0.70 ≈ 0.429. The odds ratio comparing these two cases is 1.5/0.429 ≈ 3.5.

An odds ratio of 3.5 does not mean the probability is 3.5 times as high. In this illustration, the probability ratio is 0.60/0.30 = 2. Odds and risks differ, especially when the outcome is common.

Illustrative probabilities, odds, odds ratios and probability ratios
MeasureGroup AGroup BWhat to say
Probability0.600.3060% versus 30%
Odds1.500.429Event odds are 1.50 versus 0.429
Odds ratio, A versus B3.50Reference: 1.00A has 3.5 times the odds of the event
Probability ratio, A versus B2.00Reference: 1.00A has twice the probability in this illustration

These figures are calculated from illustrative probabilities, not taken from a research sample.

How to read the output step by step

1. Confirm the outcome and reference group

Find how the dependent variable was coded and which value the model considers the event. Identify the reference category of any categorical predictor. An odds ratio for feedback received = yes is interpreted relative to the omitted or designated no category, not in isolation.

2. Check the analytic sample

Report the number of observations that entered the final model, not simply the number recruited. Determine whether unanswered survey items, invalid entries or exclusions changed the sample. Compare the analytic denominator with your missing-data plan. Our guides to cleaning research data and handling missing survey responses explain why exclusions must be transparent.

3. Read the coefficient B and Exp(B)

A coefficient B is in log-odds units. Software often reports Exp(B) = eᴮ, which is the odds ratio.

  • B > 0 corresponds to OR > 1: higher modeled event odds.
  • B < 0 corresponds to OR < 1: lower modeled event odds.
  • B = 0 corresponds to OR = 1: no estimated change in odds for that predictor.

The interpretation is conditional on other included predictors when the model is adjusted. It is not proof of cause and effect.

4. Interpret uncertainty, not just the p-value

A confidence interval around an odds ratio indicates the precision of its estimate under the model and interval procedure. For a conventional two-sided 95% interval, inclusion of 1 corresponds to no clear departure from an odds ratio of 1 by that interval criterion; the exact relationship to a p-value depends on the test and interval methods used.

Do not write that a non-significant coefficient proves there is no association. A wide interval may be consistent with meaningful increases and decreases in odds.

5. Review the model and its diagnostics

A model may have a likelihood-ratio test, log likelihood, AIC, pseudo-R² or a classification table. These do different jobs. AIC is generally for comparing models fitted to the same outcome and comparable data using a consistent likelihood basis; a smaller AIC is not automatically proof of good calibration. Pseudo-R² is not the share of outcome variation explained in the ordinary linear-regression sense.

Check whether the model is appropriate: plausible functional form for continuous predictors on the logit scale, independent observations or a model accounting for dependence, sparse cells and possible separation, influential observations, reasonable sample support, and substantive justification for included predictors. Prediction-focused studies should also examine calibration and discrimination where relevant. An accuracy percentage alone can be misleading when most observations belong to one category.

Worked example 1: Calculate an odds ratio from a 2 × 2 table

Hypothetical question: Is access to structured feedback associated with submitting an assignment on time?

The following counts are invented solely to demonstrate calculation.

Hypothetical feedback access and on-time submission counts
Feedback accessSubmitted on timeNot on timeTotal
Yes483280
No305080
Total7882160

Step 1 — Calculate the event probability. With feedback, 48/80 = 0.60 or 60%. Without feedback, 30/80 = 0.375 or 37.5%.

Step 2 — Convert to odds. With feedback, 48/32 = 1.50. Without feedback, 30/50 = 0.60.

Step 3 — Divide the odds. OR = 1.50/0.60 = 2.50 (equivalently 48 × 50 / (32 × 30) = 2.50).

Interpretation: In these invented counts, students with feedback access have 2.5 times the observed odds of on-time submission compared with those without feedback access. That is not the same as saying they are 2.5 times as likely to submit on time: the observed probability ratio is 0.60/0.375 = 1.60.

This unadjusted table cannot tell us whether access to feedback itself caused timely submission. Students with access may differ on other relevant characteristics. A fitted multivariable model may address specified measured factors but cannot automatically eliminate confounding.

Adaptable results prose: “In the illustrative sample, on-time submission occurred among 48 of 80 students (60.0%) with feedback access and 30 of 80 (37.5%) without it. The unadjusted odds ratio was 2.50, indicating greater observed odds of timely submission in the feedback-access group. This descriptive association does not establish causality.”

Worked example 2: Interpret a hypothetical multiple logistic regression table

Suppose a researcher fitted a model predicting on-time assignment submission using weekly study hours, feedback access and whether the student was in a later year of study. The following coefficient estimates, odds ratios and intervals are fabricated coherent teaching values, not statistical output from the 160-count example above.

Separate hypothetical adjusted logistic regression estimates and intervals
PredictorBExp(B), odds ratioIllustrative 95% OR intervalInterpretation
Weekly study hours0.0951.10[1.02, 1.19]Per additional reported hour, about 10% higher odds, adjusted
Feedback access (yes vs no)0.6932.00[1.20, 3.33]Approximately double the adjusted odds
Later-year student (yes vs no)−0.2230.80[0.48, 1.33]Estimated 20% lower adjusted odds, with substantial uncertainty

The odds ratios follow eᴮ to rounding: e^0.095 ≈ 1.10, e^0.693 ≈ 2.00, and e^(−0.223) ≈ 0.80. The intervals are illustrative and correspond approximately to plausible symmetric intervals on the log-odds scale; they do not arise from a real fitted dataset.

How to interpret these results: For the study-hours variable, a one-hour increase is associated with an estimated 1.10-fold change in odds, holding the other modeled variables constant. The claim is not “each extra hour raises the probability by 10 percentage points.” For feedback access, the adjusted odds ratio compares students with and without feedback while holding the other predictors fixed. For year of study, the interval includes 1, so the illustration does not provide a precise direction of association under the conventional interval criterion.

Worked reporting paragraph: “In a hypothetical multivariable logistic regression for on-time assignment submission, reported weekly study hours were positively associated with the modeled event (adjusted OR = 1.10 per hour, illustrative 95% CI [1.02, 1.19]). Feedback access was also associated with higher event odds (adjusted OR = 2.00, illustrative 95% CI [1.20, 3.33]). The estimate for later-year status was less precise (adjusted OR = 0.80, illustrative 95% CI [0.48, 1.33]). These figures are teaching examples rather than observed empirical results.”

In a real dissertation, include the actual analytic sample, model specification, estimation procedure, reference categories and relevant diagnostics. Only supply p-values or fit statistics when they were genuinely calculated from the fitted model.

Worked example 3: Convert model predictions back to probabilities

Students often report an odds ratio correctly and then accidentally call it a probability change. Predicted probabilities can make interpretation more concrete.

Suppose a separate hypothetical one-predictor model is specified as:

logit(p) = −1.3863 + 0.6931 × feedback, where feedback = 0 for no access and 1 for access.

The inverse-logit formula is p = 1/(1 + e^(−η)), where η is the linear predictor.

Separate hypothetical one-predictor model probabilities
FeedbackLinear predictor ηPredicted odds e^ηPredicted probability
No (0)−1.38630.2500.200 or 20.0%
Yes (1)−0.69320.5000.333 or 33.3%

Calculate the contrast: The predicted odds ratio is 0.500/0.250 = 2.00. The probability difference is approximately 33.3% − 20.0% = 13.3 percentage points. The probability ratio is approximately 0.333/0.200 = 1.67.

The model therefore doubles the odds across these two settings but does not double the probability. The baseline probability matters. These conditional predictions use a made-up intercept and coefficient; they are not predictions for identifiable students.

For a real multiple-predictor model, specify the values at which other predictors are held and how predicted probabilities or average marginal effects were calculated. Never pretend that a single odds ratio implies one universal percentage-point change across all starting probabilities.

Model diagnostics and responsible interpretation

Before presenting a model as credible, check the issues that matter for your design:

Logistic regression diagnostics and reporting actions
Diagnostic questionWhy it mattersReporting action
Was the modeled event coded correctly?Reversing 0 and 1 reverses the substantive interpretationDeclare event and reference groups
How many observations entered the model?Missingness and exclusions alter the evidence baseReport analytic N and material exclusions
Are continuous predictors appropriately modeled?Logit-linearity assumptions may be inappropriateExamine functional form; justify nonlinear terms if needed
Are cases or outcomes sparse?Separation can destabilize estimatesReport concerns and justify appropriate methods
Is there substantial collinearity?Highly overlapping predictors can make coefficients impreciseCheck predictor relationships and model specification
Is this prediction or explanation?A strong predictive model need not support causal explanationMatch evaluation to the research aim
Are probabilities well calibrated?Predictive accuracy alone can conceal poor probability estimatesReport calibration and discrimination where applicable

There is no single universal diagnostics checklist or cutoff that validates all logistic regressions. Use the conventions suitable for the model and sample, and distinguish what was examined from what was merely planned.

How to organize logistic regression findings in a dissertation

A clear results section generally follows this sequence:

  1. Research question and outcome: Define the binary event and predictor purpose.
  2. Sample: State N used for the model, eligibility and relevant missing-data decisions.
  3. Model: Name binary logistic regression, specify reference categories and major covariates.
  4. Results table: Present B or OR, suitable uncertainty estimates and actual p-values where applicable.
  5. Interpretation: Explain odds, not probability, and specify the comparison or increment.
  6. Model assessment: Summarize relevant fit, calibration, assumptions or estimation limitations.
  7. Caution: Explain what the observational design and measurement cannot establish.

For related reporting conventions, see our guide on reporting statistical results in a dissertation. Do not bury an essential coefficient inside pages of raw SPSS output. Use a clear table and explain the handful of findings that address the research objectives.

Common mistakes students make

Calling an odds ratio a risk ratio. An odds ratio of 2 is a doubling of odds, not necessarily probability. Show baseline probability or predicted probabilities if the practical implication needs explaining.

Forgetting event coding. A coefficient predicting dropout cannot be interpreted as predicting retention without changing the interpretation.

Ignoring reference categories. For a categorical predictor, the comparison is always relative to a defined category.

Treating p-values as the entire answer. Report magnitude and uncertainty as well as the inferential result; do not hide findings that were not statistically significant.

Calling pseudo-R² ordinary R². These statistics are constructed differently and should not be translated into an automatic “percent explained” claim.

Reporting classification accuracy without context. A model predicting the majority class for everyone can achieve high accuracy and still be unhelpful.

Claiming causation from adjustment. Statistical control for included variables does not rule out all confounding, selection problems or reverse causation.

Mixing examples from unrelated datasets. A 2 × 2 table, multivariable model and predicted-probability calculation must not be presented as if they all came from the same fitted dataset unless they actually did.

Dissertation logistic regression checklist

  • Is the binary outcome and event coding explicit?
  • Are reference categories and predictor units stated?
  • Are sample size, exclusions and missingness documented?
  • Are odds ratios correctly calculated or copied from the verified model?
  • Do intervals and p-values match the model output and test procedures?
  • Are odds kept distinct from probabilities and percentage-point changes?
  • Are model assumptions and relevant diagnostics addressed?
  • Are effect direction, magnitude and uncertainty reported together?
  • Are observational interpretations appropriately limited?
  • Are tables and prose numerically consistent?

Frequently asked questions

Can I use logistic regression with Likert-scale data?

It depends on the outcome, not simply whether a questionnaire uses Likert items. Binary logistic regression requires a binary modeled outcome. A single ordered response with several categories may call for another model. Avoid dichotomising meaningful categories merely to make a method convenient.

What does Exp(B) mean in SPSS?

For standard binary logistic regression, Exp(B) is the exponentiated coefficient, interpreted as an odds ratio for the specified predictor increment or category comparison, conditional on other variables in the model.

Is an odds ratio below 1 bad?

No. It means the estimated event odds decrease as the predictor changes in the specified way. Whether that is desirable depends entirely on what event was coded 1 and the study context.

Can I say an odds ratio of 1.50 means a 50% increase?

You can say 50% higher odds for the specified comparison, not 50% higher probability. The latter generally requires a separate probability calculation.

Should I report a model that was not statistically significant?

Yes, when it was part of the justified analysis. Report estimates, intervals and relevant diagnostics honestly, avoiding the claim that non-significance proves no association.

How many predictors can I include?

There is no universal safe count. Consider the event distribution, available information, design, prespecified scientific questions, overfitting risk and estimation method. A large nominal sample does not guarantee enough information for a complex model if events are rare.

Final thoughts

Understanding logistic regression means understanding what event was modeled, whose odds are compared, how uncertain the estimate is, and what the design permits you to conclude. When these details are clear, the tables become meaningful evidence rather than mysterious software output.

Use your actual research data and verified model output, explain the result in plain academic language, and distinguish statistical association from causal claims. If you need support with interpreting output, organizing your dissertation results or reviewing the clarity of your methods and tables, explore Acadelyra's academic-support services. Your research decisions, data and academic responsibilities remain your own.

Further reading