Hierarchical models work by representing data-generating structure at multiple levels at once. Individual observations belong to groups, groups belong to larger populations, and parameters for those groups are modelled as related rather than either identical or completely independent. The result is partial pooling: noisy groups borrow information from the wider population and are shrunk toward a common distribution, while groups with abundant distinctive evidence remain closer to their own data. Hierarchical models therefore turn clustering from a nuisance to be corrected into structure that can be estimated.
Students sit inside classes.
Classes sit inside schools.
Repeated measurements sit inside people.
Patients sit inside hospitals.
Neighbourhoods sit inside cities.
Studies sit inside literatures.
The data table may look flat.
The world that produced it is not.
The governing question: which observations belong together, how much do their groups genuinely differ, and how much information should each group borrow from the others?
The current Stan User’s Guide describes the essential mechanism directly: when group cells contain few observations, hierarchical priors stabilise and sharpen parameter estimates, with the data determining the amount of partial pooling. Stan’s partial-pooling case studies show the three canonical possibilities—complete pooling, no pooling and partial pooling—and make clear that hierarchical modelling estimates the population distribution of group effects rather than forcing one of the two extremes.
Quick Read
NESTED / GROUPED DATA → DEFINE LEVELS → GROUP-SPECIFIC PARAMETERS → POPULATION DISTRIBUTION OF PARAMETERS → ESTIMATE BETWEEN-GROUP VARIATION → PARTIAL POOLING → SHRINK NOISY GROUPS MORE → RETAIN INFORMATIVE GROUP DIFFERENCES → PREDICT NEW GROUPS / UNITS → CHECK GROUP STRUCTURE → INTERPRET
1. Hierarchical Models Begin With Dependence
Two students in the same classroom share a teacher, timetable and local environment.
Two measurements from the same patient share biology.
Two products from the same factory share process conditions.
Rows within the same group are therefore often more similar than unrelated rows.
2. Flat IID Models Can Pretend Related Rows Are Independent
Treat 1,000 students from ten schools as 1,000 independent observations.
The standard errors can become too small.
School-level variation can be mistaken for individual variation.
Hierarchical modelling puts the grouping structure inside the probability model.
3. Complete Pooling Says Every Group Shares the Same Parameter
Suppose twenty schools each have a different observed mean score.
A complete-pooling model ignores those school identities and estimates one common school effect.
It assumes genuine between-school variation is zero after modelled predictors.
This can be too rigid when groups truly differ.
4. No Pooling Says Every Group Is Unrelated
Estimate one completely separate school effect for each school.
A school with five students gets no help from a school with five hundred.
Small groups can produce wild estimates because their local data are noisy.
No pooling throws away information that groups may belong to a common population.
5. Partial Pooling Lives Between Those Extremes
Hierarchical models estimate each group separately enough to preserve real differences and jointly enough to share information.
Stan’s case studies describe exactly this middle ground: group-specific parameters differ, but data from all groups inform the population distribution from which those effects arise.
The amount of pooling is learned rather than fixed to all or nothing.
6. A Basic Varying-Intercept Model Gives Every Group Its Own Baseline
For student i in school j:
yᵢⱼ = αⱼ + βxᵢⱼ + εᵢⱼ
Each school receives its own intercept αⱼ.
Those school intercepts are then modelled as draws from a population distribution with mean μₐ and between-school standard deviation τₐ.
7. The Population Distribution Creates Shrinkage
A school’s observed mean is noisy.
The population model says school effects are plausibly distributed around a wider system mean.
The fitted school effect becomes a compromise between its own evidence and the population distribution.
This compromise is shrinkage.
8. Small Groups Shrink More
A school with five observed students has a noisy local estimate.
The hierarchical model gives the population distribution more influence.
Its fitted effect moves substantially toward the system mean.
A school with 5,000 students carries much stronger local evidence and shrinks less.
9. Shrinkage Is Adaptive Regularisation
Ordinary ridge regression applies a global penalty structure.
Hierarchical partial pooling regularises group effects according to the amount of information each group contains and the estimated between-group variance.
Weak groups are regularised more strongly.
Strong groups earn more independence.
10. Large Between-Group Variance Produces Less Pooling
If schools truly differ greatly, the estimated τ is large.
The population distribution becomes broad.
Group estimates can remain far from the overall mean without being treated as implausible.
Stan’s partial-pooling examples make this relationship explicit.
11. Small Between-Group Variance Produces More Pooling
If schools look almost identical after accounting for sampling noise, τ is estimated near zero.
The hierarchical model approaches complete pooling.
The data are saying that many apparent school differences were mostly noise.
12. Partial Pooling Corrects the Winner’s Curse of Small Groups
Rank one hundred small schools by raw mean score.
The highest observed mean is likely to be partly lucky.
The lowest is likely to be partly unlucky.
Shrinkage pulls extreme noisy estimates toward the population centre, reducing overreaction to chance extremes.
13. Partial Pooling Does Not Mean “Make Every Group Average”
If one school has abundant evidence of genuinely exceptional performance, it can remain exceptional.
Partial pooling says extraordinary estimates need extraordinary group-level evidence when the group is noisy.
It does not ban heterogeneity.
14. Varying Slopes Let Relationships Differ Across Groups
Maybe study time predicts score more strongly in some schools than others.
A varying-slope model gives each school its own βⱼ and models those slopes as a population.
Hierarchical modelling can therefore represent heterogeneity in relationships, not only baseline levels.
15. Varying Intercepts and Slopes Can Be Correlated
Schools with higher baseline achievement may also show weaker marginal effects of study time.
A multivariate group-level distribution can estimate correlation between αⱼ and βⱼ.
This turns group heterogeneity into a structured covariance problem.
16. Random Intercept and Random Slope Are Traditional Names
Frequentist mixed-model software often calls varying group effects “random effects”.
Bayesian texts often prefer “varying effects” because the parameters themselves are estimated and their distribution is modelled.
The underlying hierarchical idea is the same.
17. “Fixed Versus Random Effects” Has Several Meanings
In mixed modelling, fixed effects often mean coefficients not modelled as exchangeable group draws, while random effects are group-specific deviations from a population distribution.
In panel econometrics, “fixed-effects model” can mean estimating a separate intercept for every unit and controlling all time-invariant unit characteristics.
Terminology must be interpreted in context.
18. Intraclass Correlation Measures How Much Variance Lives Between Groups
In a simple random-intercept Gaussian model:
ICC = τ² / (τ² + σ²)
τ² is between-group variance.
σ² is within-group residual variance.
The ICC measures how similar two randomly selected observations from the same group are because they share the group effect.
19. ICC Connects Hierarchical Models to Design Effects
If clusters contain m observations with intraclass correlation ρ, a rough equal-cluster design effect is:
1 + (m−1)ρ
Even modest correlation can greatly reduce effective independent information when clusters are large.
20. Hierarchical Models Can Explain Why Cluster-Robust SEs Are Larger
Cluster-robust standard errors treat within-cluster dependence as an uncertainty problem.
A multilevel model goes further and estimates how that dependence arises through group-level variation.
Both recognise that row count is not independent-information count.
21. Group-Level Predictors Explain Between-Group Variation
School size.
Neighbourhood deprivation.
Teacher experience.
Hospital teaching status.
These variables live at the group level and can explain part of the distribution of group intercepts or slopes.
22. Cross-Level Interactions Ask Whether Context Changes Individual Relationships
Does individual study time have a stronger association with scores in schools that provide quieter study environments?
An interaction between an individual-level predictor and a school-level predictor tests this cross-level moderation.
Hierarchical models make such context-dependent relationships natural to represent.
23. Centre Predictors According to the Question
Raw student socioeconomic status mixes within-school and between-school variation.
Group-mean centring can separate “this student is higher than classmates” from “this school serves higher-SES students”.
Grand-mean centring answers another question.
Centering is interpretive architecture, not cosmetic preprocessing.
24. Contextual Effects Compare Within-Group and Between-Group Relationships
Within a school, a student’s family income may have one association with score.
Across schools, average family income may have a stronger association because it also proxies neighbourhood and school resources.
Multilevel decomposition prevents those levels from being accidentally conflated.
25. Hierarchical Models Help Avoid the Ecological Fallacy
A relationship among school averages does not automatically describe individual students.
By modelling individual and group levels simultaneously, multilevel models can keep these relationships distinct.
They do not eliminate ecological bias automatically; the correct variables and causal assumptions still matter.
26. Repeated Measures Are Hierarchical Data
Daily observations are nested within students.
A varying-intercept model captures stable person-to-person differences.
Varying slopes can represent different growth rates or treatment responses across people.
27. Growth-Curve Models Turn Time Into a Multilevel Predictor
Measurement occasions are level 1.
People are level 2.
Each person can have an individual starting point and trajectory.
The population distribution describes how those trajectories vary across people.
28. Nonlinear Growth Can Also Be Hierarchical
Learning curves may rise quickly and plateau.
Biological growth may follow logistic or Gompertz shapes.
Parameters controlling asymptote, rate and midpoint can vary by person and be partially pooled.
Hierarchical modelling is not limited to straight lines.
29. Crossed Effects Are Not Purely Nested
Students answer multiple test items.
Items are answered by multiple students.
Neither students nor items are nested inside the other.
A crossed-effects model can include varying effects for both students and items.
30. Language Experiments Often Need Crossed Participant and Item Effects
A reading-time experiment presents many sentences to many participants.
Some people read faster.
Some sentences are intrinsically harder.
Ignoring either source can create pseudo-replication or overconfident inference.
31. Multiple Membership Models Handle Units Belonging to Several Groups
A student may be taught by several teachers.
A patient may receive care from several clinics.
A publication may belong to several research teams.
Multiple-membership models distribute group influence across more than one higher-level unit.
32. Spatial and Temporal Hierarchies Add Structured Group Dependence
Neighbouring regions may resemble one another more than distant regions.
Adjacent time periods may be autocorrelated.
Hierarchical models can combine group effects with spatial conditional-autoregressive priors, Gaussian processes or time-series state structures.
33. Hierarchical Generalised Linear Models Handle Binary and Count Outcomes
Student passes or fails.
Hospital admissions are counted.
Defects occur as rates.
Logistic, Poisson and negative-binomial models can include hierarchical intercepts and slopes just as Gaussian regression can.
34. Shrinkage Happens on the Model’s Parameter Scale
In logistic regression, group effects may be normally distributed on the log-odds scale.
Shrinkage in log-odds does not translate linearly to probability.
Interpret group effects on the scientifically meaningful outcome scale when communicating results.
35. Frequentist Mixed Models Estimate Variance Components
Maximum likelihood or restricted maximum likelihood can estimate fixed coefficients and variance components such as between-school variance.
Empirical best linear unbiased predictors produce shrunk estimates of group effects.
Partial pooling is not uniquely Bayesian.
36. Bayesian Hierarchical Models Put Priors on Population Parameters
Group effects receive a population distribution.
Population mean and variance receive hyperpriors.
The posterior jointly propagates uncertainty about group effects and the population distribution.
Stan’s user guide highlights this hierarchical-prior architecture directly.
37. Hyperpriors Matter Most When There Are Few Groups
With 200 schools, between-school variance is well informed.
With three schools, it is difficult to learn the population variance from data alone.
Stan explicitly warns that hierarchical models with few groups can require stronger hyperprior information.
Nominal row count does not solve a small-group-count problem.
38. Variance Components Near Zero Create Boundary Problems
Between-group variance cannot be negative.
If the estimate lies near zero, ordinary normal approximations can fail.
Likelihood-ratio tests for zero variance components can have non-standard null distributions.
Bayesian priors can stabilise estimation while making their influence explicit.
39. Centered Parameterisation Is Intuitive but Can Create Funnel Geometry
Write αⱼ ~ Normal(μ, τ).
When τ is near zero and group information is weak, posterior geometry can form a narrow funnel.
Hamiltonian Monte Carlo can struggle, producing divergences or slow exploration.
40. Non-Centered Parameterisation Can Improve Sampling Geometry
Write:
zⱼ ~ Normal(0,1)
αⱼ = μ + τzⱼ
The model is mathematically equivalent and computationally much easier in many weak-data hierarchical settings.
Parameterisation becomes computational design.
41. Centered Models Can Be Better When Groups Are Highly Informative
Non-centering is not automatically superior.
When each group has abundant data, centred parameterisation can mix more efficiently.
Modern Stan workflows often choose parameterisation according to information strength.
42. Divergences Are Model Diagnostics, Not Cosmetic Warnings
A Bayesian hierarchical model that produces divergent HMC transitions may not have explored the posterior reliably.
Increasing sampler settings blindly can hide rather than solve bad geometry.
Reparameterisation, prior revision or model simplification may be required.
43. Varying-Slope Correlation Matrices Need Regularising Priors
With several varying slopes, the group-level covariance matrix becomes high-dimensional.
Weak data can produce unstable estimated correlations near ±1.
LKJ-type priors or other structured covariance priors regularise the correlation matrix in Bayesian models.
44. “Maximal Random Effects” Can Be Too Complex for the Data
A model including every possible varying slope and covariance can become singular or weakly identified when group counts are small.
Simplification should follow the scientific design and information available rather than a ritual rule that more random effects are always safer.
45. Singular Fits Tell Us the Data Do Not Support All Variance Dimensions
A frequentist mixed model may estimate one variance component at zero or correlations near ±1.
This can indicate an over-complex random-effects structure relative to available groups.
It is an information warning, not merely a software inconvenience.
46. Prediction for Existing Groups and New Groups Are Different
Predict another student in a school already observed.
The school’s estimated effect is available.
Predict a student in a completely new school.
The new school effect must be drawn from the population distribution.
Prediction uncertainty is therefore larger for unseen groups.
47. Cross-Validation Must Hold Out the Right Level
Randomly hold out students while keeping their schools in training if deployment concerns new students in known schools.
Hold out entire schools if deployment concerns new schools.
The validation split should match the level at which future uncertainty occurs.
See How Cross-Validation Works.
48. Hierarchical Prediction Can Be Calibrated at Several Levels
Are individual outcomes calibrated?
Are school-average predictions calibrated?
Are tail outcomes within schools calibrated?
A model can perform well at one level and poorly at another.
49. Hierarchical Models Naturally Handle Measurement Reliability Differences
Some schools have 20 students.
Some have 2,000.
Raw school means have very different sampling precision.
Partial pooling automatically gives more weight to reliable group estimates and more shrinkage to noisy ones.
50. Meta-Analysis Is a Hierarchical Model of Studies
Each study estimates its own underlying effect with sampling error.
A random-effects meta-analysis models study effects as draws from a distribution with mean μ and heterogeneity τ.
This is a two-level hierarchical model.
51. Hierarchical Meta-Analysis Can Add More Levels
Multiple effects can come from one paper.
Papers can come from research labs.
Effects can be nested in outcomes and time points.
Multilevel meta-analysis separates sampling error, within-study dependence and between-study heterogeneity.
52. Hierarchical Models Offer One Answer to Multiple Comparisons
Estimate fifty school effects independently and the most extreme estimates are noisy.
Partial pooling shrinks extremes toward the learned population distribution.
This reduces false extreme discoveries through modelling rather than classical family-wise error control.
It does not provide the same guarantee as Bonferroni or FDR procedures.
53. Partial Pooling Changes the Question From “Which Groups Are Significant?”
Instead of fifty binary significance tests, the model estimates a population distribution and posterior or empirical-Bayes uncertainty for every group.
The question becomes how groups vary and which differences are supported after sharing information.
54. Multilevel Regression and Poststratification Extends Hierarchical Models to Population Estimation
Fit a hierarchical model for an outcome using demographic and geographic predictors.
Predict outcomes for every cell in the target population.
Weight those predictions by known population cell counts.
The Stan User’s Guide describes this multilevel regression and poststratification architecture explicitly.
55. MRP Can Estimate Small Geographic Cells by Borrowing Strength
A national survey may contain only a handful of respondents in one demographic-state combination.
Direct estimation is unstable.
Hierarchical regression partially pools related cells, then poststratification maps predictions back onto population composition.
56. MRP Cannot Repair an Outcome or Population Frame It Does Not Observe
If relevant population groups are absent and no predictors bridge that absence, hierarchical shrinkage cannot manufacture their behaviour reliably.
If survey response depends on unmeasured variables tied strongly to the outcome, poststratification on observed cells may leave residual bias.
Partial pooling reduces variance; it does not abolish selection bias.
57. Hierarchical Models Can Represent Latent Measurement Structure
Test items are nested within domains.
Students differ in latent ability.
Items differ in difficulty and discrimination.
Item-response models can be expressed hierarchically so item and person parameters borrow information across populations.
58. Hierarchical Priors Are Useful for Many Related Regression Coefficients
Suppose an intervention has different effects across age groups.
Instead of estimating each subgroup coefficient independently, model subgroup effects as draws from a common distribution.
This turns noisy subgroup analysis into structured heterogeneity estimation.
59. But Hierarchical Shrinkage Can Hide Real Outlier Groups
One hospital genuinely follows a different clinical protocol.
A normal random-effects distribution treats it as an extreme draw and pulls it toward the centre.
If the hospital belongs to another regime, a mixture model or explicit group predictor may be more appropriate.
Pooling assumes exchangeability after conditioning on modelled information.
60. Exchangeability Is the Hidden Contract of Partial Pooling
Group effects are modelled as draws from a common population because, after conditioning on predictors, their labels are assumed not to carry additional systematic structure.
If urban hospitals and rural clinics come from fundamentally different regimes, pooling them under one exchangeable distribution can be misleading.
61. Robust Hierarchical Models Use Heavy-Tailed Group Distributions
A normal distribution of group effects can overshrink legitimate extremes.
Student-t group distributions allow more heavy-tailed heterogeneity.
Mixture models can represent distinct latent group regimes.
Hierarchical structure itself can be made robust.
62. Missing Data Can Be Modelled Hierarchically
Missing covariates can be imputed using group-aware distributions.
A missing student baseline may be predicted partly from class and school information.
Hierarchical imputation preserves group structure better than a pooled single-level model.
Missingness assumptions remain necessary.
63. Causal Inference With Hierarchical Data Needs Cluster-Level Causal Thinking
Treatment may be assigned to schools rather than students.
Interference can occur within classrooms.
Confounders can exist at student and school levels.
A hierarchical outcome model can improve estimation but does not define the causal estimand or identify the treatment effect by itself.
64. Cluster-Randomised Trials Naturally Produce Hierarchical Outcomes
Schools are randomised.
Students are measured.
Inference must respect the school-level assignment while modelling student-level outcomes.
Multilevel models can improve precision and describe heterogeneity, but randomisation remains the causal foundation.
65. Treatment Effects Can Vary Hierarchically
An intervention may work differently across schools.
Model school-specific treatment effects as a population of effects.
Partial pooling reduces noise in school-by-school estimates while quantifying genuine heterogeneity.
Prediction intervals for a new school can then answer whether the treatment is likely to help elsewhere.
66. Education Is Almost Unavoidably Hierarchical
Questions within tests.
Tests within students.
Students within classes.
Classes within schools.
Schools within systems.
A flat average can mix all these levels into one number.
Multilevel modelling allows each layer to retain its own variation and relationships.
67. The Hostile Test: Rank Tiny Schools by Raw Means
One school has six students and mean 95.
Another has 600 students and mean 88.
The tiny school is ranked first without acknowledging enormous sampling noise.
Partial pooling would shrink the six-student estimate much more strongly.
68. The Second Hostile Test: Pool Groups That Are Not Exchangeable
Private specialist hospitals and rural primary clinics are put in one random-intercept distribution without explanatory predictors.
The model shrinks both toward one centre.
The grouping model hides two structurally different systems.
Partial pooling only makes sense after the exchangeability story is defensible.
69. The Third Hostile Test: Few Groups, Weak Variance Estimate
A study has 10,000 observations from four schools.
The analyst treats the between-school variance as precisely estimated because the row count is huge.
The relevant information for the group distribution is still four groups.
Hyperprior or small-sample uncertainty matters enormously.
70. The Fourth Hostile Test: Random Student CV for New-School Deployment
Students from every school appear in both train and validation folds.
The model uses learned school effects to predict held-out students.
Deployment is actually to entirely new schools.
The validation task was easier than the real task.
71. The Fifth Hostile Test: Shrinkage Mistaken for Evidence That Groups Are Similar
Posterior group estimates look close together.
The report concludes the schools are intrinsically similar.
But the closeness partly reflects shrinkage imposed because local estimates were noisy.
Interpret the estimated population variance and raw information alongside pooled estimates.
72. Primary School: Hierarchical Thinking Begins as “Things Belong to Groups”
Students are in classes.
Classes are in schools.
If one class has only three students and scores unusually high, we should be less certain that the class is truly exceptional than if thirty students score unusually high.
Hierarchical thinking says a small group should listen to itself—but also to what we know about the larger family of groups it belongs to.
73. Secondary School: Compare Complete, No and Partial Pooling
Plot raw means for ten classes with different class sizes.
Then compare:
- one common mean for everyone;
- ten unrelated class means;
- ten partially pooled class means.
The movement of small classes toward the overall distribution makes shrinkage visible.
74. JC and University: Hierarchical Models Become Probability Models for Variation Across Levels
At higher levels, learners should reconstruct:
- level-1 units;
- grouping variables;
- complete pooling;
- no pooling;
- partial pooling;
- varying intercepts;
- varying slopes;
- between-group variance;
- ICC;
- group-level predictors;
- cross-level interactions;
- nested and crossed effects;
- frequentist variance components;
- Bayesian hierarchical priors;
- centred and non-centred parameterisations;
- new-group prediction.
75. Where Hierarchical Models Fit in the eduKateSG “How Works” Landscape
- How Regression Works — owns the general conditional-modelling framework.
- How Standard Errors Work — owns dependence-aware uncertainty when clustering is treated as a variance problem.
- How Regularisation Works — owns shrinkage as complexity control; hierarchical models supply adaptive group-level shrinkage.
- How Bayesian Inference Works — owns priors, hyperpriors, posterior computation and uncertainty.
- How Meta-Analysis Works — applies hierarchical modelling to study effects.
- How Causal Inference Works — owns causal identification that multilevel outcome models cannot create.
- How Cross-Validation Works — owns level-appropriate held-out evaluation.
Hierarchical Models own one precise canonical job: represent variation and dependence across related levels by modelling group-specific parameters as draws from shared population distributions, thereby enabling partial pooling, multilevel prediction and explicit estimation of heterogeneity.
76. What This Article Does Not Claim
- Hierarchical models are not only Bayesian; frequentist mixed models also produce partial pooling.
- Partial pooling does not mean forcing all groups to be equal.
- More hierarchy is not automatically better; overly complex varying-effect structures can be weakly identified or singular.
- Large row counts do not substitute for having enough higher-level groups.
- Groups must be plausibly exchangeable after conditioning for pooling to make sense.
- A normal group-effects distribution can overshrink genuine regime outliers.
- Hierarchical modelling does not automatically repair selection bias or measurement bias.
- A multilevel outcome model does not create causal identification.
- Validation must hold out the level that will actually be new at deployment.
- Shrunk group estimates should not be interpreted without acknowledging the population variance and amount of local information.
77. A Compact Hierarchical-Model Audit
- What are the observational units?
- What groups or levels do they belong to?
- Are the groups nested, crossed or multiple-membership?
- What dependence does group membership create?
- Would complete pooling be too rigid?
- Would no pooling be too noisy?
- What parameters should vary by group?
- Should intercepts, slopes or both vary?
- How much between-group variance is estimated?
- What is the ICC?
- How much does each group shrink?
- Does shrinkage appropriately reflect group sample size?
- Are group-level predictors available?
- Are cross-level interactions scientifically meaningful?
- Should predictors be group-mean or grand-mean centred?
- Are group effects plausibly exchangeable?
- Would a heavy-tailed or mixture group distribution fit better?
- Are there enough groups to estimate variance components?
- Do hyperpriors matter because group count is small?
- Is centred or non-centred parameterisation more computationally appropriate?
- Are divergences or singular-fit warnings present?
- Does prediction target existing groups or new groups?
- Does cross-validation hold out the correct level?
- Does the model separate within-group from between-group relationships?
- If causal, does the design identify the effect independently of the hierarchical outcome model?
78. Frequently Asked Questions
What is a hierarchical model?
A hierarchical or multilevel model represents data at multiple related levels and models group-specific parameters as arising from shared population distributions, allowing dependence and between-group variation to be estimated directly.
What is partial pooling?
Partial pooling combines group-specific evidence with information from the wider population of groups. Small noisy groups are shrunk more strongly toward the population distribution, while large informative groups retain more of their local estimate.
What is the difference between complete pooling and no pooling?
Complete pooling assumes all groups share the same parameter. No pooling estimates every group independently. Hierarchical partial pooling estimates group differences while treating the groups as related draws from a population distribution.
Why do hierarchical models shrink estimates?
Shrinkage reflects uncertainty in noisy group estimates. A small group provides limited local information, so the population distribution contributes more strongly. Large groups provide stronger evidence and therefore shrink less.
Are random effects and hierarchical effects the same?
They often refer to the same multilevel mechanism in mixed-model terminology: group-specific effects are modelled as draws from a population distribution. “Varying effects” is often a clearer description because the group parameters themselves are estimated.
79. Authoritative Research Corridor
- Stan User’s Guide 2.38 — Multilevel Regression and Poststratification
- Stan Case Study — Hierarchical Partial Pooling for Repeated Binary Trials
- rstanarm — Hierarchical Partial Pooling
- eduKateSG — How Bayesian Inference Works
- eduKateSG — How Regression Works
- eduKateSG — How Meta-Analysis Works
Final Thought: Hierarchical Models Let Small Groups Speak Without Pretending They Know as Much as Large Ones
There are two easy mistakes.
Erase every group difference.
Or believe every group difference.
Complete pooling commits the first.
No pooling risks the second.
Hierarchical modelling builds a conversation between the local group and the population it belongs to.
The local data say, “this is what happened here.”
The wider population says, “this is how much groups like you usually vary.”
The fitted estimate listens to both.
Partial pooling is statistical humility with memory: trust each group in proportion to the evidence it carries, and let the larger family of groups lend support where local information is thin.