7. Multilevel, Panel, and Longitudinal Models

Introduction and Context

Business data are rarely independent observations. Employees belong to teams, teams belong to business units, customers are observed repeatedly, firms are followed across years, and transactions occur within markets, regions, or organizational systems. These arrangements create dependence: observations associated with the same group or person tend to resemble one another more than observations selected from different groups. Conventional regression models that assume independent errors may therefore produce misleading standard errors, inefficient estimates, and inappropriate conclusions.

Multilevel, panel, and longitudinal models provide a principled response to this structure. They allow researchers to distinguish variation occurring within units from variation occurring between units, model dependence explicitly, and investigate how relationships change across contexts and over time. They are especially important in doctoral business research because many organizational and strategic questions are inherently hierarchical or dynamic. Examples include whether leadership practices affect employee performance over time, whether firm-level governance moderates innovation, and how long customers remain active after acquisition.

The central analytical issue is not simply technical. It concerns the level at which a phenomenon exists and the level at which evidence is collected. A firm-level relationship cannot automatically be inferred from employee-level data, and an association observed between firms may not describe changes within a firm. The models in this chapter help preserve that distinction.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Explain dependence, clustering, nesting, repeated measurement, and unobserved heterogeneity.

2. Distinguish multilevel, panel, longitudinal, fixed-effects, random-effects, and mixed models.

3. Specify and interpret random intercepts, random slopes, cross-level interactions, and growth models.

4. Compare pooled regression, fixed-effects, random-effects, and hybrid approaches.

5. Select suitable models for continuous, binary, count, and event-time outcomes.

6. Diagnose common estimation, interpretation, and design problems.

7. Apply these models to managerial and organizational research questions.

8. Evaluate ethical, governance, and practical implications of longitudinal analytics.

Why Ordinary Regression Is Often Inadequate

Suppose employee performance is measured for 500 employees working in 40 departments. If employees in the same department share a manager, incentive system, and workload, their outcomes are not independent. The intraclass correlation coefficient (ICC) expresses the proportion of total variance attributable to differences between clusters. In a simple random-intercept model,

Yij = beta0 + beta1Xij + uj + eij,

where Yij is the outcome for person i in department j, uj is the department-specific deviation from the overall intercept, and eij is the individual-level residual. If the variance of uj is tau-squared and the variance of eij is sigma-squared, then:

ICC = tau-squared / (tau-squared + sigma-squared).

A high ICC indicates substantial within-department similarity. Even a modest ICC can matter when clusters are large, because the effective sample size is smaller than the nominal number of observations. Ignoring clustering commonly causes standard errors to be underestimated and significance tests to become too liberal.

Aggregation also creates a conceptual danger. The average performance of departments may be related to average training expenditure, while the relationship between an individual employee's training and performance may be weak or even negative. This is a form of ecological inference problem. Multilevel analysis retains information at each level and permits separate within- and between-group effects.

The Architecture of Multilevel Data

A multilevel design contains units located within higher-order units. Level 1 might consist of customer-month observations, level 2 of customers, and level 3 of geographic markets. A two-level model may be written as:

Yij = beta0j + beta1jXij + eij,

with

beta0j = gamma00 + gamma01Z j + u0j,

beta1j = gamma10 + gamma11Z j + u1j.

Here, Xij is a level-1 predictor, Zj is a level-2 predictor, beta0j is a group-specific intercept, and beta1j is a group-specific slope. The terms u0j and u1j represent random deviations. The parameter gamma11 is a cross-level interaction: it indicates whether a group characteristic changes the relationship between an individual-level predictor and the outcome.

A random-intercept model allows groups to differ in baseline outcomes but assumes that the slope of X is common. A random-slope model permits the effect of X to vary across groups. Random slopes are substantively valuable when theory suggests that a policy, practice, or intervention works differently under different organizational conditions. They also alter the implied covariance structure and should not be introduced merely as a mechanical refinement.

Centering deserves careful attention. Grand-mean centering subtracts the overall sample mean from a predictor and supports interpretation relative to the average case. Group-mean centering subtracts each group's mean and isolates within-group variation. For example, an employee's training hours can be decomposed into the employee's deviation from the department average and the department's average training level. This decomposition prevents a within-department effect from being confused with a between-department effect.

Panel and Longitudinal Models: Similar Data, Different Questions

Panel data observe multiple units across multiple periods. A panel may follow firms annually, countries quarterly, or customers monthly. Longitudinal analysis is a broader term covering repeated observations over time, often with explicit interest in trajectories, developmental processes, or time-varying mechanisms. The distinction is not absolute: a firm-year panel is also longitudinal data, but panel econometrics often emphasizes unit-specific effects and causal identification, whereas longitudinal modeling often emphasizes change and growth.

A basic pooled model is:

Yit = beta0 + beta1Xit + alpha i + lambda t + eit,

where i indexes units and t indexes time. The unit effect alpha i captures stable characteristics such as managerial culture, location, or founder orientation. Time effects lambda t capture shocks common to all units, such as regulation or macroeconomic conditions.

Fixed Effects

The fixed-effects estimator removes time-invariant unit characteristics by transforming the data around each unit's mean:

Yit - Yi-bar = beta1(Xit - Xi-bar) + (eit - ei-bar).

The coefficient is therefore identified by within-unit change. Fixed effects are attractive when unobserved stable characteristics are likely to correlate with predictors. They do not estimate the effects of variables that do not change within a unit, such as industry classification when industry membership is constant. They also require sufficient within-unit variation and are vulnerable to bias from time-varying omitted variables, reverse causality, measurement error, and anticipation of treatment.

Random Effects

Random-effects models treat unit-specific effects as random and assume they are uncorrelated with all predictors in every period. Under that assumption, both within- and between-unit information can be used, and time-invariant predictors can be estimated. The assumption is strong and should be justified substantively rather than selected solely because a statistical test favors it. A Hausman-type comparison can inform the choice, but it is not a substitute for research design and theoretical reasoning.

Correlated Random Effects and Hybrid Models

A useful compromise is to include the unit mean of each time-varying predictor alongside its deviation from that mean. In a hybrid specification,

Yit = betaW(Xit - Xi-bar) + betaB Xi-bar + alpha i + lambda t + eit.

The coefficient betaW represents the within-unit association, while betaB represents the between-unit association. This approach makes the distinction explicit and allows researchers to relax the conventional random-effects independence assumption. It is particularly useful when a substantive argument concerns both changes within firms and persistent differences between firms.

Growth-Curve and Dynamic Models

Longitudinal research frequently asks whether units grow, decline, or follow different trajectories. A linear growth model can be written as:

Yit = eta0i + eta1i time t + eit,

where eta0i is the individual-specific starting point and eta1i is the individual-specific rate of change. At the next level,

eta0i = gamma00 + gamma01Zi + u0i,

eta1i = gamma10 + gamma11Zi + u1i.

The parameter gamma11 indicates whether Z predicts a different growth rate. Time may be coded from baseline, centered at a meaningful event, or represented with nonlinear terms. Quadratic time, splines, piecewise slopes, and latent trajectory approaches are useful when change is not linear.

Longitudinal models must distinguish time-varying covariates from time-invariant covariates. A time-varying predictor may itself be affected by previous outcomes, creating dynamic feedback. Lagged predictors, cross-lagged models, dynamic panel estimators, and marginal structural approaches may be considered, but none automatically establishes causality. A lag does not eliminate confounding simply because it places the predictor earlier in time.

Generalized Multilevel Models and Event History

Not all outcomes are continuous. Generalized linear mixed models extend multilevel logic to binary, ordinal, nominal, and count outcomes. For a binary outcome,

logit[P(Yij = 1)] = beta0 + beta1Xij + uj.

The coefficient beta1 is on the log-odds scale; exponentiating it gives an odds ratio conditional on the random effect. Interpretation should not confuse conditional odds ratios with population-averaged effects, which may differ in nonlinear models.

Count outcomes may require Poisson or negative-binomial models, with an exposure offset when observation periods differ. Overdispersion, zero inflation, and excess structural zeros should be investigated rather than handled automatically with increasingly complex models.

Event-history or survival models analyze time until an event such as employee departure, customer churn, product failure, or acquisition. The hazard function describes the instantaneous event rate among units that have not yet experienced the event. A proportional-hazards model is:

h i(t) = h0(t) exp(beta'X i(t) + ui).

The baseline hazard h0(t) is unspecified in the Cox model. Time-varying covariates are permitted, but they must be measured in a way that respects the temporal ordering of the event. Censoring is informative when the probability of being censored is related to the event process; in that case, standard methods can be biased. Competing-risks models are relevant when different event types prevent one another, such as voluntary resignation versus retirement.

A Research Workflow for Dependent Data

Model selection should begin with the data-generating process rather than with a favorite estimator. First identify the observational units, levels of nesting, timing of measurement, entry and exit rules, and possible sources of dependence. Second formulate the estimand: is the question about a within-firm change, a between-firm contrast, an average treatment effect, a trajectory, or a conditional hazard?

Third inspect the structure visually. Plot individual trajectories, cluster means, missingness patterns, and event timing. Estimate a null multilevel model to assess clustering, but do not treat the ICC as the sole criterion for modeling dependence. Fourth specify a theoretically defensible model, including time effects, relevant covariates, and plausible interactions. Fifth assess estimation assumptions and sensitivity to alternative specifications.

Results should report the number of units and observations, cluster sizes, missing-data treatment, fixed and random components, variance estimates, time controls, and the level at which coefficients are interpreted. Confidence intervals are generally more informative than binary significance labels. A technically sophisticated model with vague reporting is not reproducible research.

Illustrative Case: A Multinational Customer-Retention Study

Background

Consider a fictional organization, Meridian Cloud Services, which operates subscription software businesses in 12 countries. The research team wants to know whether proactive customer-success contact reduces cancellation. Customers are observed monthly for up to 24 months. Customers are nested within sales regions, and regions differ in service staffing and market conditions.

Situation and Analytical Problem

A simple comparison shows lower cancellation among contacted customers. However, high-value customers may be more likely to receive contact, and customers who are already disengaging may receive urgent outreach. The data also contain repeated observations and right-censoring because many customers remain active at the end of observation.

Analysis and Approach

The team first constructs a person-month file and defines cancellation as the first subscription termination. A discrete-time hazard model with customer-level random intercepts and region-year effects is estimated. Contact in month t predicts cancellation in month t+1, while prior usage, contract type, tenure, customer size, and calendar month are controlled. Customer-level baseline characteristics are separated from time-varying usage. A complementary fixed-effects linear-probability model is used for monthly cancellation status among customers with adequate within-customer variation.

Because contact is targeted, the analysts do not describe the coefficient as a causal effect without qualification. They conduct sensitivity analyses using lagged contact, alternative definitions of engagement, region-specific trends, inverse-probability weighting, and negative-control outcomes. Missing usage data are examined for informative dropout, and standard errors are clustered at the customer level where appropriate.

Outcome

The principal result is reported as an adjusted association: recent contact is associated with a lower subsequent cancellation hazard, but the magnitude is smaller than the raw comparison suggests. The analysis also indicates substantial regional heterogeneity. In regions with overloaded service teams, contact has little apparent association with retention, whereas in adequately staffed regions the association is stronger.

Learning Points

The case illustrates why timing, censoring, repeated observations, and selection into treatment must be addressed together. It also demonstrates the value of combining a survival framework with multilevel structure and sensitivity analysis. Most importantly, a model does not transform observational association into causal evidence by itself.

Challenges, Limitations, and Common Errors

A frequent error is treating every repeated observation as independent. Cluster-robust standard errors may protect inference in some settings, but they do not solve omitted-variable bias or provide a model for individual trajectories. Another error is using fixed effects when the predictor has little within-unit variation, producing unstable estimates and discarding substantively relevant between-unit information.

Researchers also underestimate the consequences of unbalanced panels. Unbalancedness is not automatically problematic, but dropout may depend on prior outcomes or unobserved characteristics. Multiple imputation, inverse-probability weighting, joint models, or pattern-mixture sensitivity analyses may be appropriate depending on the missingness mechanism and assumptions.

Random-effects models can be misinterpreted as proving that unobserved effects are random. They are modeling assumptions, not facts about organizations. Random-effects distributions may also be misspecified, and random-slope models can become unstable with few clusters. In survival analysis, proportional hazards should be checked; a single hazard ratio may be misleading when effects vary over time.

Ecological and atomistic fallacies remain serious risks. A firm-level coefficient does not describe individual behavior, and an individual-level effect does not automatically characterize firm performance. Researchers should state clearly whether an estimate is within, between, conditional, marginal, or population averaged.

Ethics, Governance, and Responsible Use

Longitudinal business data can reveal sensitive patterns about employees, customers, and suppliers. Governance should specify legitimate purposes, access controls, retention periods, data minimization, and procedures for correcting inaccurate records. Linking data across periods can create detailed behavioral profiles even when each individual measurement appears harmless.

Automated retention or performance systems may disadvantage groups if historical decisions reflect discrimination or unequal access to resources. Fairness assessment should therefore examine error rates and consequences across relevant groups, while respecting privacy and legal requirements. Analysts should document variable construction, exclusions, model changes, and uncertainty. Managers should not treat a predicted hazard or random effect as a moral judgment about an individual.

Digital Tools and Emerging Developments

Modern statistical software supports mixed-effects, generalized mixed, fixed-effects, Bayesian hierarchical, survival, and joint models. Reproducible scripts, version-controlled data pipelines, and model registries improve auditability. Bayesian hierarchical models are increasingly useful when there are many groups with sparse observations because partial pooling can stabilize group estimates. Machine-learning methods can improve prediction, but prediction accuracy does not replace identification of within-unit effects or explanation of dependence.

Emerging work combines causal inference with panel data, synthetic controls, staggered interventions, dynamic treatment regimes, and heterogeneous treatment-effect estimation. These methods require careful attention to treatment timing, anticipation, interference, and positivity. In large-scale operational data, real-time event streams create opportunities for dynamic risk prediction, but they also intensify concerns about surveillance, drift, and automated decision-making.

Best Practices for Doctoral Research

A defensible study should:

• define the unit of analysis and the estimand before choosing the estimator;
• represent nesting, repeated measurement, and time explicitly;
• separate within-unit and between-unit variation where they have different meanings;
• justify fixed or random effects using theory, design, and sensitivity analysis;
• inspect missingness, censoring, cluster size, and temporal ordering;
• report variance components, uncertainty, and substantive effect sizes;
• test plausible nonlinearities and time interactions without indiscriminate specification searching;
• compare results across defensible alternative models;
• distinguish prediction from causal explanation; and
• preserve privacy, reproducibility, and transparent documentation.

Chapter Summary

Multilevel, panel, and longitudinal models are designed for data in which observations are clustered, repeated, or ordered through time. Their central contribution is to model dependence and distinguish within-unit change from between-unit difference. Random-intercept and random-slope models represent hierarchical heterogeneity; fixed-effects models remove time-invariant unit characteristics; random-effects and hybrid models use both within- and between-unit information under explicit assumptions. Growth models describe trajectories, generalized mixed models handle noncontinuous outcomes, and survival models analyze time to events while addressing censoring and changing risk.

Sound analysis begins with the research question and data-generating process. Researchers must define the estimand, inspect dependence and timing, assess missingness and heterogeneity, and interpret coefficients at the correct level. No estimator alone resolves confounding or establishes causality. Ethical governance is essential when longitudinal analytics affect employees, customers, or other identifiable groups.

Key Terms

Intraclass correlation coefficient (ICC): The proportion of outcome variance attributable to differences between clusters.

Multilevel model: A model that represents observations nested within higher-level units.

Random intercept: A group-specific deviation from the overall intercept.

Random slope: A group-specific deviation in the effect of a predictor.

Cross-level interaction: An interaction in which a variable at one level modifies an effect at another level.

Panel data: Repeated observations on multiple units across time.

Longitudinal data: Observations collected over time to study change or temporal processes.

Fixed effects: Unit-specific effects removed or estimated while allowing correlation with predictors.

Random effects: Unit-specific effects modeled as random and conventionally assumed independent of predictors.

Within effect: The association arising from changes within the same unit.

Between effect: The association arising from differences among units.

Hybrid model: A model that includes within-unit deviations and unit means to estimate distinct effects.

Growth-curve model: A longitudinal model describing initial status and rate of change.

Generalized linear mixed model: A mixed model for binary, count, ordinal, or other noncontinuous outcomes.

Hazard: The instantaneous event rate among units still at risk.

Censoring: Incomplete observation of event time because the event is not observed during the study.

Competing risks: Mutually exclusive event types that prevent one another from occurring first.

Unobserved heterogeneity: Relevant differences among units that are not directly measured.

Partial pooling: Estimation that combines group-specific information with the overall population distribution.

Review Questions

Short Answer Questions

1. Why can ignoring clustering lead to misleading statistical inference?

2. What does the ICC measure in a random-intercept model?

3. How does a fixed-effects estimator identify an effect in panel data?

4. Distinguish a random intercept from a random slope.

5. What is the meaning of censoring in event-history analysis?

Long Answer Questions

1. Compare fixed-effects, random-effects, and hybrid panel models, including their assumptions and interpretive consequences.

2. Explain how centering choices affect the interpretation of multilevel predictors.

3. Describe the specification and interpretation of a longitudinal growth-curve model.

4. Discuss the challenges of modeling missingness and dropout in longitudinal research.

5. Explain why a sophisticated multilevel or survival model does not by itself establish causality.

Analytical / Application Questions

1. A researcher observes employees nested within departments over three years and wants to know whether leadership training changes performance within departments. Design an appropriate model and identify the key assumptions.

2. A customer-churn dataset contains monthly records, right-censoring, region-level staffing differences, and targeted retention calls. Propose an analysis strategy and explain how selection bias would be addressed.

3. A firm-level study finds that organizations with higher average innovation spending have higher productivity, while within-firm increases in spending are unrelated to productivity. Explain how a hybrid model could represent these results and why the distinction matters.

22. Multiple Choice Questions with Answers

Question 1

A researcher observes employees nested within departments. Which model component allows departments to have different baseline performance while retaining a common predictor slope?

A. Random intercept

B. Random slope

C. Fixed time effect only

D. Competing-risk term

Correct Answer: A. Random intercept

Question 2

In a panel fixed-effects model, the coefficient on a time-varying predictor is primarily identified by:

A. Differences between industries

B. Changes within the same unit over time

C. The overall sample mean only

D. Units that never change

Correct Answer: B. Changes within the same unit over time

Question 3

A high intraclass correlation coefficient indicates that:

A. Most variation is explained by measurement error

B. Observations within clusters are relatively similar

C. There is no need to model clustering

D. All predictors are exogenous

Correct Answer: B. Observations within clusters are relatively similar

Question 4

Which specification best separates within-firm and between-firm effects of a time-varying predictor?

A. Including only the current predictor value

B. Including only firm fixed effects

C. Including the deviation from the firm mean and the firm mean itself

D. Removing all time observations with missing values

Correct Answer: C. Including the deviation from the firm mean and the firm mean itself

Question 5

In a growth-curve model, a random slope represents:

A. Variation in the measurement scale

B. Differences among units in their rate of change

C. A time-invariant predictor that cannot be estimated

D. The probability of censoring

Correct Answer: B. Differences among units in their rate of change

Question 6

For a binary repeated-measures outcome, which model is generally appropriate?

A. Generalized linear mixed model with a logit link

B. Ordinary least squares without dependence adjustment

C. A linear growth model with no error term

D. A pure fixed-effects model for every categorical variable

Correct Answer: A. Generalized linear mixed model with a logit link

Question 7

In survival analysis, right-censoring occurs when:

A. The event occurs before observation begins

B. The event is not observed before the observation period ends or the unit leaves observation

C. Two events occur simultaneously for every unit

D. The hazard is constant over time

Correct Answer: B. The event is not observed before the observation period ends or the unit leaves observation

Question 8

Why might a random-effects model be inappropriate even when it produces efficient estimates?

A. It cannot include any time-varying predictors

B. Its conventional independence assumption may be implausible

C. It always removes time-invariant variables

D. It cannot model clustered observations

Correct Answer: B. Its conventional independence assumption may be implausible

Question 9

A customer receives a retention call because the firm anticipates imminent cancellation. Using the call as an ordinary predictor may create:

A. Reverse causality or treatment-selection bias

B. Perfect randomization

C. A guaranteed proportional-hazards assumption

D. An increase in the ICC by definition

Correct Answer: A. Reverse causality or treatment-selection bias

Question 10

Which statement best distinguishes prediction from causal explanation in longitudinal analytics?

A. A predictive model automatically estimates a causal effect

B. A lagged predictor always eliminates confounding

C. High predictive accuracy does not by itself identify what would happen under intervention

D. Causal models do not require temporal ordering

Correct Answer: C. High predictive accuracy does not by itself identify what would happen under intervention