Start with the purpose of the model
Regression is used for two broad purposes, and the reporting differs between them. An explanatory model estimates the association between an exposure or intervention and an outcome, adjusting for other variables; the coefficient of interest is the result. A predictive model combines several predictors to estimate an individual's outcome or risk; the model's performance is the result. Stating which aim the analysis serves tells readers how to judge it, and determines which guideline applies.
The SAMPL guidelines for basic statistical reporting begin their regression section by asking authors to describe the purpose of the analysis (Lang & Altman, 2015). APA's Journal Article Reporting Standards for quantitative research (JARS-Quant) make the same point for psychology and the social sciences, and ask that the analytic strategy be linked to each hypothesis (Appelbaum et al., 2018). Where the model is meant to predict outcomes for individuals, TRIPOD+AI is the current reporting guideline; it replaced TRIPOD 2015 and covers both regression and machine learning methods (Collins et al., 2024).
Define the variables and the model
Identify the outcome and every predictor, with its units, how it was measured and how it was coded. For categorical predictors, name the reference category. For continuous predictors, say whether they were centred, scaled or transformed, and how non-linear relationships were handled, for example with splines or polynomial terms rather than by cutting the variable into groups (Harrell, 2015). Summarise each variable with appropriate descriptive statistics (mean and standard deviation, or median and interquartile range for skewed data, and counts with percentages for categories) (Lang & Altman, 2015).
Then state the model itself: the type of regression (linear, logistic, Poisson, Cox and so on), the predictors included, any interaction terms, and how clustering or repeated measures were handled. SAMPL suggests reporting the regression equation; for a multiple linear model this takes the form Y = b0 + b1X1 + b2X2 + ... + e. A table of coefficients with the intercept gives the same information. Boe et al. (2024) and Lee (2022) give practical guidance on matching the regression type to the outcome and on what to report for each.
Explain how the model was built
Readers need to know why each predictor is in the model. The most defensible approach in explanatory research is to pre-specify covariates from subject-matter knowledge, for example confounders identified in a causal diagram, and to say so. If variables were selected using the data, describe the procedure exactly: the method, the criterion (for example p < 0.157 or the AIC) and the candidate variables.
Forward, backward and stepwise selection are still common but have well-documented problems. Because variables are kept when their estimated effect happens to be large, the retained coefficients are biased away from zero, and the standard errors, confidence intervals and p-values reported afterwards ignore the selection step and are too optimistic (Heinze et al., 2018; Harrell, 2015). If selection was used, report it honestly, and consider showing the full pre-specified model alongside, or assessing the stability of the selection by bootstrapping (Heinze et al., 2018).
Sample size should also be justified. For prediction models, Riley et al. (2020) provide formulae for the minimum sample size based on the number of candidate parameters, the expected R2 and the outcome proportion, as an alternative to the older rule of thumb of ten events per variable.
Assumptions, outliers, collinearity and missing data
For linear regression, the assumptions concern the model's errors, not the raw variables: a correctly specified (linear) relationship, independent observations, constant residual variance and, for small samples, approximately normal residuals. Ernst and Albers (2017) found that many published studies tested the wrong thing, typically the normality of the outcome or predictors, and that few reported checking assumptions at all. Report how each assumption was assessed (residual-versus-fitted plots, normal Q-Q plots of residuals, plots against each predictor) and what was done if it was not met, such as transformation, robust standard errors or a different model.
Influential observations should be identified with diagnostics such as Cook's distance and handled by a stated rule. Removing a point only because it weakens a result is not acceptable; a sensitivity analysis with and without influential points is more informative. Collinearity among predictors can be assessed with variance inflation factors (Lee, 2022); thresholds of 5 or 10 are conventions rather than strict cut-offs, and the main consequence of collinearity is imprecise, not biased, coefficients.
Report how many participants had missing data on each variable and how this was handled. Complete-case analysis can be biased and wasteful; multiple imputation is often preferable when data are plausibly missing at random, but the imputation model, the number of imputations and the variables used need to be reported (Sterne et al., 2009).
Report coefficients, precision and fit
The core of the results is a table giving, for each predictor, the coefficient, its 95% confidence interval and its p-value (Lang & Altman, 2015; Appelbaum et al., 2018). Unstandardised coefficients, in the units of the outcome per unit of the predictor, are usually the most interpretable; if standardised coefficients are given as well, say how they were standardised. State the number of observations used in the model. For logistic regression the exponentiated coefficients are odds ratios, which approximate risk ratios only when the outcome is uncommon. Give exact p-values rather than "NS", and interpret them alongside the estimates, since a p-value alone does not measure the size or importance of an effect (Wasserstein & Lazar, 2016).
Model fit is summarised by R2 (or adjusted R2, which penalises the number of predictors) for linear models, with the overall F test if relevant. These describe fit in the data used to build the model and are optimistic about performance elsewhere. For predictive models, TRIPOD+AI asks for internal validation by bootstrapping or cross-validation, and for discrimination and calibration measures with confidence intervals (Collins et al., 2024). Reporting in this area is still weak: Andaur Navarro et al. (2022) found a median adherence to TRIPOD items of 38.7% across 152 machine learning prediction studies.
A worked example
The figures below are hypothetical and purely illustrative. Suppose a cross-sectional study of 120 adults models systolic blood pressure (mmHg) on age, weekly hours of physical activity and sex, using ordinary least squares. With three predictors the residual degrees of freedom are 120 - 3 - 1 = 116.
| Predictor | b (95% CI) | SE | t(116) | p |
|---|---|---|---|---|
| Age (per year) | 0.42 (0.20 to 0.64) | 0.11 | 3.82 | < .001 |
| Physical activity (per hour/week) | -0.15 (-0.33 to 0.03) | 0.09 | -1.67 | .098 |
| Sex: male vs female (reference) | 2.10 (0.22 to 3.98) | 0.95 | 2.21 | .029 |
A results sentence might read: "Each additional year of age was associated with 0.42 mmHg higher systolic blood pressure (95% CI 0.20 to 0.64), adjusting for physical activity and sex." Three points are worth noting. First, each confidence interval is b ± 1.98 × SE, using the t distribution with 116 degrees of freedom. Second, the interval for physical activity runs from a reduction of 0.33 mmHg to an increase of 0.03 mmHg per hour; it does not show that activity has no effect, only that the data cannot distinguish a small benefit from none. Third, adjusted R2 (1 - (1 - 0.31) × 119/116 = 0.29) is slightly lower than R2 because it allows for the number of predictors.
What to report
| Report | Why it matters |
|---|---|
| Aim (explanatory or predictive) and guideline followed | Determines how the model should be judged |
| Outcome and predictors with units, coding and reference categories | Allows coefficients to be interpreted |
| Model type, functional form, interactions and handling of clustering | Allows the model to be rebuilt |
| How predictors were chosen (pre-specified or selected, with criteria) | Selection affects estimates and p-values |
| Assumption checks, influential points and collinearity diagnostics | Shows the model is adequate for the data |
| Missing data per variable and method used | Complete-case analysis may be biased |
| Coefficients with units, 95% CIs and exact p-values; n analysed | Shows magnitude and precision |
| R2 or adjusted R2; validation and calibration for prediction models | Describes fit and expected performance |
| Software, version and packages | Supports reproducibility |
Figures help when chosen well. A scatter plot with the fitted line and confidence band suits simple regression (Lang & Altman, 2015); for multiple regression, residual plots and plots of predicted values across the range of a predictor (holding others fixed) are more informative than raw scatter plots.
Common mistakes
- Testing the outcome or predictors for normality instead of examining the residuals.
- Using stepwise selection and then reporting the selected model's p-values as if it had been pre-specified.
- Reporting standardised "beta" coefficients without units or saying how they were standardised.
- Omitting the reference category for categorical predictors.
- Dichotomising continuous predictors at arbitrary or data-driven cut-points.
- Presenting R2 from the development data as evidence that a prediction model is valid.
- Not stating how many observations were excluded for missing data.
- Interpreting a non-significant coefficient as proof of no association.
A useful final check is to ask whether a colleague with your data and your methods section could reproduce every number in the coefficient table. If they would have to guess any modelling decision, the reporting needs more work.
Key takeaways
- Say whether the model is explanatory or predictive, and use TRIPOD+AI for prediction models (Collins et al., 2024).
- Define every variable, its coding and functional form, and how covariates were chosen (Lang & Altman, 2015).
- Data-driven variable selection biases estimates and p-values; pre-specify where possible (Heinze et al., 2018).
- Check assumptions on the residuals, not on the raw variables (Ernst & Albers, 2017).
- Report coefficients with units and 95% confidence intervals, not p-values alone (Wasserstein & Lazar, 2016).
- Treat in-sample R2 as description, not validation.
