This workflow shows how to analyze three categorical outcomes common in education dissertations: a binary indicator, an ordered proficiency level, and an unordered postsecondary pathway. It emphasizes outcome coding, reference categories, model assumptions, probability-based interpretation, sparse cells, clustering, and calibrated reporting. The synthetic data are designed for practice and do not represent real students or schools.
You will be able to select a model from the outcome’s measurement structure, fit binary logistic, proportional-odds, and multinomial models, translate coefficients into adjusted probabilities, evaluate major assumptions and diagnostics, and produce dynamic APA-style text.
| outcome | category | n | percent |
|---|---|---|---|
| On track | 0 | 1040 | 49.5 |
| On track | 1 | 1059 | 50.5 |
| Proficiency | Advanced | 829 | 39.5 |
| Proficiency | Basic | 345 | 16.4 |
| Proficiency | Below basic | 553 | 26.3 |
| Proficiency | Proficient | 372 | 17.7 |
| Pathway | Four-year college | 770 | 36.7 |
| Pathway | Other | 360 | 17.2 |
| Pathway | Two-year college | 443 | 21.1 |
| Pathway | Workforce | 526 | 25.1 |
The event coded 1 is “on track.” Proficiency levels retain their natural ordering. The pathway outcome has no defensible ordering, so two-year college is an explicit comparison category. Changing a reference category changes coefficient labels, but not fitted probabilities.
| variable | missing_n | percent |
|---|---|---|
| attendance | 89 | 4.2 |
| pathway | multilingual | n | flag |
|---|---|---|---|
| Two-year college | No | 306 | FALSE |
| Two-year college | Yes | 137 | FALSE |
| Four-year college | No | 529 | FALSE |
| Four-year college | Yes | 241 | FALSE |
| Workforce | No | 352 | FALSE |
| Workforce | Yes | 174 | FALSE |
| Other | No | 236 | FALSE |
| Other | Yes | 124 | FALSE |
Complete-case modeling below is used to keep the tutorial focused. A dissertation should justify its missing-data strategy and ordinarily consider multiple imputation when assumptions are plausible.
| term | odds_ratio | conf_low | conf_high | p_value |
|---|---|---|---|---|
| (Intercept) | 0.025 | 0.008 | 0.084 | 0.000 |
| prior_score | 2.836 | 2.529 | 3.192 | 0.000 |
| attendance | 60.642 | 15.748 | 236.940 | 0.000 |
| ses | 1.319 | 1.196 | 1.457 | 0.000 |
| multilingualYes | 0.999 | 0.806 | 1.238 | 0.993 |
| interventionYes | 1.253 | 1.029 | 1.527 | 0.025 |
| school_support | 1.095 | 0.993 | 1.209 | 0.069 |
Odds ratios are multiplicative changes in odds, not percentage-point changes in probability. Continuous predictor units should be meaningful before interpreting them.
| metric | value |
|---|---|
| AUC | 0.766 |
| Brier score | 0.197 |
| Sensitivity | 0.705 |
| Specificity | 0.677 |
| term | odds_ratio | conf_low | conf_high | p.value |
|---|---|---|---|---|
| prior_score | 2.407 | 2.199 | 2.640 | 0.000 |
| attendance | 12.030 | 3.917 | 37.109 | 0.000 |
| ses | 1.231 | 1.134 | 1.337 | 0.000 |
| multilingualYes | 0.999 | 0.835 | 1.196 | 0.994 |
| interventionYes | 1.215 | 1.029 | 1.435 | 0.021 |
| school_support | 1.171 | 1.078 | 1.273 | 0.000 |
A positive coefficient indicates greater odds of being in a higher rather than any lower proficiency category, assuming that cumulative contrast is constant across cut points.
| term | Df | logLik | AIC | LRT | Pr(>Chi) |
|---|---|---|---|---|---|
| NA | -2424.641 | 4867.283 | NA | NA | |
| prior_score | NA | NA | NA | NA | NA |
| attendance | NA | NA | NA | NA | NA |
| ses | NA | NA | NA | NA | NA |
| multilingual | NA | NA | NA | NA | NA |
| intervention | NA | NA | NA | NA | NA |
| school_support | NA | NA | NA | NA | NA |
Treat this test as one source of evidence. Large samples can flag small departures, and small samples can miss consequential ones. Examine category-specific fitted probabilities and consider partial proportional-odds or multinomial models when departures matter substantively.
| outcome | term | relative_risk_ratio | p_value |
|---|---|---|---|
| Four-year college | (Intercept) | 2.356 | 0.264 |
| Four-year college | prior_score | 2.064 | 0.000 |
| Four-year college | attendance | 0.629 | 0.591 |
| Four-year college | ses | 1.200 | 0.004 |
| Four-year college | multilingualYes | 1.060 | 0.674 |
| Four-year college | interventionYes | 1.022 | 0.862 |
| Four-year college | school_support | 1.270 | 0.000 |
| Workforce | (Intercept) | 1.931 | 0.415 |
| Workforce | prior_score | 0.862 | 0.032 |
| Workforce | attendance | 0.552 | 0.513 |
| Workforce | ses | 0.773 | 0.000 |
| Workforce | multilingualYes | 0.973 | 0.850 |
| Workforce | interventionYes | 1.066 | 0.634 |
| Workforce | school_support | 1.127 | 0.072 |
| Other | (Intercept) | 1.999 | 0.433 |
| Other | prior_score | 0.749 | 0.000 |
| Other | attendance | 0.324 | 0.257 |
| Other | ses | 0.915 | 0.222 |
| Other | multilingualYes | 1.077 | 0.641 |
| Other | interventionYes | 1.028 | 0.850 |
| Other | school_support | 1.055 | 0.468 |
Each relative risk ratio compares the named pathway with two-year college. It is not an odds ratio for the named category in isolation because all outcome probabilities compete and sum to one.
Students are nested within schools. The models above include a school-level predictor but do not by themselves account for residual dependence. Depending on the question and number of clusters, use cluster-robust standard errors, generalized estimating equations, or generalized mixed models. Avoid treating a small number of clusters as if they supported asymptotic cluster-robust inference.
| term | Estimate | Std. Error | z value | Pr(>|z|) |
|---|---|---|---|---|
| (Intercept) | -3.676 | 0.658 | -5.585 | 0.000 |
| prior_score | 1.042 | 0.055 | 18.868 | 0.000 |
| attendance | 4.105 | 0.734 | 5.593 | 0.000 |
| ses | 0.277 | 0.052 | 5.331 | 0.000 |
| multilingualYes | -0.001 | 0.104 | -0.009 | 0.993 |
| interventionYes | 0.226 | 0.109 | 2.075 | 0.038 |
| school_support | 0.091 | 0.076 | 1.202 | 0.229 |
In the synthetic binary analysis, intervention participation was associated with 1.25 times the adjusted odds of being on track, 95% CI [1.03, 1.53], p = .025. The model’s apparent-sample AUC was 0.77; this is descriptive and should not be presented as external validation. In the ordinal analysis, a one-SD increase in prior achievement corresponded to 2.41 times the cumulative odds of a higher proficiency category, 95% CI [2.20, 2.64], p < .001.
Use multilevel categorical models for meaningful residual clustering; generalized estimating equations for population-average correlated outcomes; discrete-time survival models for event timing; latent class models when the categories are unobserved; and causal methods when the estimand is an intervention effect rather than an adjusted association. Rare events, quasi-separation, survey weights, and complex samples require specialized planning.