From Yes/No Outcome to Odds Ratios: A Logistic Regression Case Study

A researcher wants to understand why some participants complete a professional development program while others do not. The outcome is not a score or an approximately continuous measurement. Each participant either completes the program or does not complete it.

That distinction determines the modeling strategy. Logistic regression is designed for a binary response and models the probability of the event through the log odds. Ordinary linear regression is not the appropriate default for this outcome because a linear model can produce fitted probabilities outside the possible 0-to-1 range, while modeling a binary response with normal linear regression also conflicts with its distributional and variance assumptions (Lovric, 2011; Moore et al., 2021).

This synthetic case study follows the analysis from outcome coding to odds ratios, confidence intervals, diagnostics, predicted probabilities, and defensible interpretation.

Research question

The study asks:

The emphasis is deliberately on association, not causation. A regression coefficient can describe an adjusted relationship between a predictor and an outcome, but a fitted observational regression does not by itself establish that changing the predictor would cause the outcome to change (Tabachnick & Fidell, 2013).

The synthetic sample contains 900 participants:

• 424 completed the program.
• 476 did not complete the program.

Define the binary outcome before fitting the model

The dependent variable is coded:

• completed = 1: participant completed the program;
• completed = 0: participant did not complete the program.

This coding is consequential. Logistic regression commonly represents the outcome of interest as 1 and the alternative as 0, and the resulting model describes the probability associated with the modeled event (Lovric, 2011; Moore et al., 2014).

Before interpreting any output, therefore, the researcher should confirm which category the software is actually modeling. Software conventions can differ, so the event category should never be inferred merely from the signs of coefficients (Moore et al., 2014; Tabachnick & Fidell, 2013).

Why ordinary linear regression is not the appropriate default

Suppose the researcher coded completion as 1 and non-completion as 0 and fitted an ordinary regression of the form:

Linear probability form:
p = β₀ + β₁X₁ + … + β_kX_k.

The quantity being modeled is now a probability. But probabilities must remain between 0 and 1. With a nonzero linear slope, sufficiently extreme predictor values can make a linear prediction fall below 0 or above 1 (Moore et al., 2021).

There is also a distributional problem. A binary response follows a Bernoulli structure rather than the normal-response structure underlying ordinary Gaussian linear regression; for a binary response, the normality and constant-variance conditions of that model are not satisfied (Lovric, 2011).

Logistic regression solves the probability-bound problem by modeling the log odds rather than probability directly:

Logistic model:
log(p / (1 − p)) = β₀ + β₁X₁ + … + β_kX_k.

The inverse logistic transformation then maps the linear predictor back to a fitted probability between 0 and 1 (Lovric, 2011; Moore et al., 2021).

Why the binary outcome changes the modeling strategy

Feature	Ordinary linear regression	Logistic regression
Quantity modeled	Probability directly when a 0/1 outcome is inserted into a linear model	Log odds of the modeled event
Probability bounds	Predictions can fall below 0 or above 1	Inverse logistic transformation produces fitted probabilities between 0 and 1
Response structure	Ordinary Gaussian-response assumptions do not match a binary response	Designed for a binary response
Typical estimation	Ordinary least squares	Maximum likelihood

Predictor coding and reference groups

Four predictors are used in the synthetic analysis.

Predictor coding in the synthetic model

Predictor	Coding	Interpretation
Coaching	0 = no coaching; 1 = coaching	Comparison of coaching with no coaching
Readiness	Continuous standardized score	Effect of a one-unit increase in readiness
Prior course	0 = no; 1 = yes	Comparison of previous course experience with none
Enrollment route	Standard, referral, outreach	Categorical predictor represented by indicator variables

Categorical predictors with more than two categories can be represented by indicator variables, with one category serving as the reference. Tabachnick and Fidell (2013) describe coding a categorical predictor with one fewer dummy variables than categories; the omitted category defines the comparison baseline. Moore et al. (2014) likewise demonstrate the use of indicator variables for categorical predictors.

Here, standard enrollment is the reference category. The route coefficients therefore answer:

• referral versus standard enrollment;
• outreach versus standard enrollment.

Changing the reference category would change those coefficient comparisons, but it would not change the underlying fitted probabilities produced by the same model.

Model specification

The fitted model is:

log(p_i / (1 − p_i)) = β₀ + β₁(coaching) + β₂(readiness) + β₃(prior course) + β₄(outreach) + β₅(referral).

Here, p_i is the probability that participant i completes the program.

In multiple logistic regression, each coefficient describes its predictor while the other predictors in the model are held constant (Moore et al., 2021; Tabachnick & Fidell, 2013). This point matters: the coaching coefficient is not simply the raw completion-rate difference between coached and uncoached participants. It is the coaching association after adjustment for readiness, prior-course experience, and enrollment route.

Logistic regression parameters are commonly estimated using maximum-likelihood methods rather than ordinary least squares (Lovric, 2011; Tabachnick & Fidell, 2013).

What should be checked before interpreting the coefficients?

Logistic regression avoids some assumptions associated with ordinary linear regression, but it is not assumption-free. Tabachnick and Fidell (2013) emphasize several practical conditions that should be examined.

Independence of observations

The usual binary logistic model assumes that responses from different cases are independent. Repeated measurements, matched observations, or other dependence structures require methods that account for that dependence rather than treating every row as unrelated (Tabachnick & Fidell, 2013).

The synthetic case assumes one independent record per participant.

Enough information relative to model complexity

Sparse data, too few cases relative to the number of predictors, and complete separation can lead to very large parameter estimates or standard errors and can prevent maximum-likelihood estimation from converging satisfactorily (Tabachnick & Fidell, 2013).

In this example, there are 900 observations, both outcome categories are well represented, and the fitted model converged without extreme coefficient estimates.

Linearity in the logit for continuous predictors

Logistic regression does not require the binary outcome itself to have a linear relationship with a continuous predictor. Instead, a continuous predictor is assumed to have a linear relationship with the logit of the outcome probability when entered as a simple linear term (Tabachnick & Fidell, 2013).

That distinction is easy to miss. A curved probability relationship can be entirely compatible with a linear relationship on the log-odds scale.

For the synthetic readiness predictor, graphical assessment and an additional nonlinear-term check did not indicate a material departure from the specified linear-logit relationship.

Multicollinearity

Very strong relationships among predictors can destabilize logistic regression coefficients and inflate their standard errors. Tabachnick and Fidell (2013) recommend checking for multicollinearity among predictors rather than assuming that logistic regression is immune to it.

The synthetic design produced very low variance-inflation values:

Variance-inflation values for the synthetic predictors

Predictor	VIF
Outreach route	1.10
Referral route	1.10
Coaching	1.01
Readiness	1.00
Prior course	1.00

There is therefore no indication in this synthetic dataset that predictor redundancy is materially destabilizing the fitted coefficients.

Outlying or poorly predicted observations

Logistic regression can contain cases that the fitted model predicts poorly. Standardized residuals provide one way to identify observations that deserve investigation rather than automatic deletion (Tabachnick & Fidell, 2013).

One synthetic observation had a comparatively large standardized residual of approximately 3.48. It was a participant with low readiness and no coaching or prior-course experience who nevertheless completed the program.

The observation was checked rather than simply removed. Excluding it in a sensitivity refit changed the coaching odds ratio from approximately 2.18 to 2.21, leaving the substantive interpretation essentially unchanged.

Synthetic model results

The fitted coefficients and transformed odds ratios were:

Logistic regression results for the synthetic case study

Predictor	B	Odds ratio	95% CI for OR	p
Coaching: yes vs no	0.781	2.18	1.64–2.92	< .001
Readiness: +1 unit	0.751	2.12	1.81–2.49	< .001
Prior course: yes vs no	0.598	1.82	1.35–2.46	< .001
Outreach vs standard route	-0.220	0.80	0.55–1.17	.250
Referral vs standard route	0.539	1.71	1.21–2.43	.002

The intercept was B = -0.799. Because the intercept corresponds to a specific baseline combination of predictor values, it is usually less substantively interesting than the predictor effects unless that baseline itself is meaningful.

Coefficients and odds ratios are two scales for the same effect

A logistic coefficient (B) operates on the log-odds scale. Exponentiating it gives an odds ratio:

OR = e^B.

Thus, for coaching:

e^0.781 ≈ 2.18.

Logistic coefficients are therefore not probability differences. They describe changes in log odds; their exponentiated versions describe multiplicative changes in odds (Lovric, 2011; Moore et al., 2021; Tabachnick & Fidell, 2013).

Correct interpretation of the coaching result

How the same odds ratio can be overstated

An odds ratio is a ratio of odds, not a ratio of probabilities. Odds are p / (1 − p), so a doubling of odds does not generally mean a doubling of probability (Moore et al., 2014; Moore et al., 2021).

Likewise, “made participants” is causal language. The synthetic regression describes an adjusted association; the model specification alone does not establish a causal intervention effect (Tabachnick & Fidell, 2013).

The distinction becomes clearer when predicted probabilities are examined.

From odds ratios back to predicted probabilities

Odds ratios are compact summaries of model coefficients, but predicted probabilities are often easier to communicate.

Consider participants on the standard enrollment route with no prior course, while varying coaching and readiness. The fitted model gives:

Predicted completion probabilities for selected predictor profiles

Readiness	No coaching	Coaching
-1	17.5%	31.7%
0	31.0%	49.6%
+1	48.8%	67.6%

The coaching odds ratio remains approximately 2.18 throughout these comparisons because the fitted model contains no coaching interaction. Yet the probability difference is not constant.

• At readiness = -1, the fitted probability changes from about 17.5% to 31.7%, a difference of roughly 14 percentage points.
• At readiness = 0, it changes from 31.0% to 49.6%, a difference of roughly 19 percentage points.
• At readiness = +1, it changes from 48.8% to 67.6%, again a different probability contrast.

The inverse logistic transformation converts the linear predictor into a probability, producing the characteristic nonlinear relationship (Lovric, 2011; Tabachnick & Fidell, 2013).

Predicted probabilities should therefore be tied to clearly specified predictor values rather than reported as though one odds ratio corresponded to a universal probability increase.

Interpreting the other predictors

Readiness

The readiness coefficient is B = 0.751, corresponding to:

OR = e^0.751 = 2.12.

Holding the other predictors constant, a one-unit increase in the standardized readiness score is associated with approximately 2.12 times the odds of completion (95% CI [1.81, 2.49]).

For a continuous predictor, an odds ratio describes the multiplicative change in odds associated with a one-unit increase in that predictor (Lovric, 2011; Tabachnick & Fidell, 2013). Consequently, the definition of “one unit” should always be clear.

Prior course experience

Participants with previous course experience had approximately 1.82 times the adjusted odds of completion compared with participants without previous experience (95% CI [1.35, 2.46]).

Again, this is an odds comparison, not a statement that their probability of completion was 82 percentage points higher.

Enrollment route

Referral participants had approximately 1.71 times the adjusted odds of completion compared with the standard-route reference group (95% CI [1.21, 2.43]).

For outreach participants, the estimated OR was 0.80, but its 95% confidence interval ranged from 0.55 to 1.17. Because the interval includes 1, the data do not provide a precise directional estimate separating outreach from the standard route at the conventional 95% confidence level.

Confidence intervals belong next to odds ratios

A point estimate alone hides uncertainty.

For coaching, the estimated OR is 2.18, while its 95% confidence interval is 1.64 to 2.92. Logistic-regression confidence intervals can be constructed for regression coefficients and then exponentiated to express the interval on the odds-ratio scale (Moore et al., 2021).

That interval conveys much more than “p < .001.” It shows the range of effect sizes compatible with the fitted model and its sampling uncertainty.

The same principle changes the interpretation of outreach. Its point estimate of OR = 0.80 might tempt someone to write that outreach “reduces completion odds by 20%.” But the confidence interval of 0.55 to 1.17 crosses 1. The estimate is not sufficiently precise to support a firm directional conclusion.

A practical interpretation template

For a binary predictor, a useful reporting structure is:

After adjustment for [other modeled predictors], the [exposed/1] group had [OR] times the odds of [event coded 1] compared with the [reference/0] group (95% CI [lower, upper]).

For a continuous predictor:

Holding the other modeled predictors constant, a one-unit increase in [predictor] was associated with a multiplicative change of [OR] in the odds of [event] (95% CI [lower, upper]).

When probability is the quantity readers actually care about, supplement the odds ratio with predicted probabilities for meaningful predictor profiles.

What this analysis does not establish

Conclusion

When the research outcome is genuinely binary, model choice should begin with that fact.

Logistic regression models the log odds of the event while producing fitted probabilities constrained to the meaningful 0-to-1 range. Its predictors can include continuous and categorical variables, but coding decisions, reference groups, independence, sample information, linearity in the logit for continuous predictors, multicollinearity, and unusual observations still require attention (Lovric, 2011; Moore et al., 2021; Tabachnick & Fidell, 2013).

The most important interpretive discipline comes after estimation. A coefficient is a change in log odds. Its exponential is an odds ratio. Neither is automatically a probability difference.

In this synthetic example, saying that coaching was associated with 2.18 times the adjusted odds of completion is accurate. Saying that coaching made participants 2.18 times more likely to complete overstates what the odds ratio—and the research design—actually establishes.

For many audiences, the strongest presentation is therefore a combination: report the odds ratio with its confidence interval, identify the reference group and modeled event, and then translate the fitted model into predicted probabilities for substantively meaningful cases.

References

Lovric, M. (Ed.). (2011). International encyclopedia of statistical science. Springer. https://doi.org/10.1007/978-3-642-04898-2

Moore, D. S., McCabe, G. P., & Craig, B. A. (2014). Introduction to the practice of statistics (8th ed.). W. H. Freeman.

Moore, D. S., McCabe, G. P., & Craig, B. A. (2021). Introduction to the practice of statistics (10th ed.). Macmillan Learning.

Tabachnick, B. G., & Fidell, L. S. (2013). Using multivariate statistics (6th ed.). Pearson.