Causal, Experimental and Prescriptive People Analytics

Introduction

People analytics increasingly influences decisions about recruitment, onboarding, compensation, learning, scheduling, retention and employee wellbeing. Descriptive analytics reports what has happened; predictive analytics estimates what is likely to happen; causal analytics asks what would happen if an organization changed a policy or intervention; prescriptive analytics recommends what decision should be taken, subject to objectives and constraints.

The distinction is fundamental. A model may predict that employees who attend leadership training are more likely to be promoted, yet this does not prove that training causes promotion. High-potential employees may be more likely both to receive training and to be promoted. Acting on a predictive association as though it were causal can waste resources, reinforce inequality or produce harmful decisions.

This chapter develops a practical framework for moving from prediction to causal explanation and then to intervention selection. It covers randomized experiments, quasi-experimental designs, causal diagrams, treatment-effect estimation, optimization, simulation and decision criteria.

Learning objectives

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

1. Distinguish prediction, association and causation.
2. Explain potential outcomes, treatment effects and causal estimands.
3. Design and interpret randomized experiments in workplace settings.
4. Select appropriate quasi-experimental designs when randomization is unavailable.
5. Use causal diagrams to identify confounding, mediation and collider bias.
6. Interpret average and heterogeneous treatment effects.
7. Apply optimization, simulation, scenario analysis and decision criteria to people decisions.
8. Evaluate ethical, legal and practical risks in causal and prescriptive analytics.

1. Prediction is not causation

A predictive question is: “Who is likely to leave within six months?” A causal question is: “What would be the effect of increasing flexible-work access on six-month turnover?” The first concerns forecasting; the second concerns an intervention.

Prediction emphasizes out-of-sample accuracy. Causal analysis emphasizes identification: whether the observed data and research design support a defensible claim about a counterfactual outcome. A correlation may arise because:

• a common cause affects both variables;
• the apparent cause is actually a consequence;
• the relationship is produced by selection into treatment;
• measurement or missing-data processes distort the relationship;
• the association is specific to a particular context.

For example, employees who use an employee-assistance programme may have higher absence. It would be incorrect to conclude that the programme increases absence without considering that employees experiencing serious difficulties are more likely to use it.

A useful test is the intervention question: if the organization could change X while holding all else appropriately comparable, would Y change? Causal analysis attempts to construct or approximate that comparison.

2. Potential outcomes and treatment effects

Let Y_i(1) denote the outcome for employee i if exposed to a treatment, and Y_i(0) the outcome for the same employee without treatment. Only one of these outcomes is observed for each employee; the other is the counterfactual.

The individual treatment effect is:

The average treatment effect (ATE) is:

Here, E denotes the population mean. Outcomes must be measured on a defined scale, such as percentage points of retention, days absent, sales units or a wellbeing score. The treatment must also be precisely defined: “flexible work” is not sufficiently specific unless eligibility, duration and permitted arrangements are stated.

For participants who actually receive treatment, the average treatment effect on the treated (ATT) is:

where T = 1 indicates treatment. ATE and ATT may differ when treatment effects vary across employees.

A causal comparison of observed means is valid only under strong conditions. In a randomized trial, treatment assignment is independent of potential outcomes, so the difference in group means estimates the ATE in expectation:

where Ȳ₁ and Ȳ₀ are the observed mean outcomes in treatment and control groups. Randomization, adequate implementation, low differential attrition and a clearly defined outcome are essential assumptions.

3. Causal diagrams

Directed acyclic graphs (DAGs) represent assumptions about causal relationships. Consider training (T), prior performance (P) and promotion (Y):

P ─────→ T ─────→ Y
└────────────────→ Y

Prior performance influences both training assignment and promotion. It is a confounder. Comparing trained and untrained employees without addressing performance may attribute the effect of P to T.

A causal diagram helps distinguish three roles:

• Confounder: a common cause of treatment and outcome. It should generally be controlled, measured or addressed through design.
• Mediator: a variable on the causal pathway, such as training → skill acquisition → promotion. Controlling for it may remove part of the total treatment effect.
• Collider: a variable caused by two variables. Conditioning on it can create a spurious association. For example, conditioning on being selected for a prestigious programme may link motivation and manager support even if they were otherwise unrelated.

The backdoor principle states, informally, that a set of pre-treatment variables can be adjusted for when it blocks all non-causal paths from treatment to outcome and does not include descendants of treatment. Diagram-based reasoning must be supported by institutional knowledge; software cannot discover all causal assumptions from data alone.

4. Randomized experiments

In a randomized controlled trial (RCT), eligible employees, teams or locations are assigned by chance to treatment or control. Randomization balances observed and unobserved characteristics in expectation and therefore removes systematic selection bias.

Design choices

A people-analytics experiment should specify:

1. Unit of assignment: employee, team, department, site or manager.
2. Treatment: exact intervention, dosage and implementation period.
3. Control condition: no intervention, usual practice or an alternative intervention.
4. Outcome: operationally meaningful primary measure and measurement window.
5. Assignment probability: often one-half, but unequal allocation may be justified.
6. Power and minimum detectable effect: the smallest effect worth detecting with acceptable error probabilities.
7. Analysis population: intention-to-treat (assigned condition) or treatment-on-the-treated (actual receipt).

Cluster randomization is appropriate when employees influence one another or when treatment is delivered by manager or site. It reduces contamination but requires more clusters because observations within a cluster are correlated.

Worked example: randomized scheduling intervention

An organization randomly assigns 200 comparable employees to a scheduling-support programme and 200 to usual practice. After three months, average absence is 4.8 days in treatment and 5.6 days in control. The estimated effect is:

Thus, assignment to the programme is associated with a reduction of 0.8 absence days over three months. A confidence interval and a pre-specified statistical test are needed before judging precision. The practical question is also important: if the programme costs £120 per employee, the organization should compare its cost with the value of 0.8 fewer absence days, while considering implementation and unintended effects.

Randomization does not automatically solve every problem. Non-compliance, attrition, spillovers, poor treatment fidelity and outcome measurement changes can weaken inference. Intention-to-treat estimates the effect of offering or assigning the programme; it is often the most policy-relevant estimate. Per-protocol analyses may be more vulnerable to selection bias because actual participation is not randomized.

5. Quasi-experimental designs

When randomization is infeasible or unethical, quasi-experimental methods seek a credible comparison using institutional rules, timing or naturally occurring variation.

Difference-in-differences

Suppose one region adopts a retention policy while another does not. Let Y_T,post be the treated region’s post-policy outcome, with analogous notation for treated pre-policy, control post-policy and control pre-policy outcomes:

The key assumption is parallel trends: absent the policy, treated and control regions would have followed comparable average trends. This cannot be proved definitively, but pre-treatment trends and placebo tests can increase credibility. Staggered adoption requires careful methods because conventional two-way fixed-effects models can behave poorly when treatment effects vary over time or across units.

Regression discontinuity

A regression-discontinuity design uses a threshold rule. Employees scoring at least 70 on a development assessment receive coaching; those below 70 do not. If employees cannot precisely manipulate scores and those just above and below the threshold are comparable, the discontinuity in outcomes estimates a local treatment effect.

The estimate applies near the cutoff, not necessarily to all employees. Analysts should inspect the running variable, test for manipulation, justify bandwidth choices and report sensitivity to model specification.

Instrumental variables

An instrumental variable Z affects treatment T but affects outcome Y only through T, and is not related to unmeasured causes of Y. For example, random offer rotation might influence programme participation. The exclusion restriction is demanding and usually cannot be verified from data alone. Instrumental-variable estimates commonly identify a local average treatment effect for individuals whose treatment status responds to the instrument.

Matching and weighting

Propensity-score matching or weighting can balance measured pre-treatment covariates. A propensity score is:

where X is a vector of observed covariates. These methods do not remove bias from unmeasured confounders and should not be treated as substitutes for randomization. Balance diagnostics, overlap and sensitivity analysis are essential.

6. Estimating and interpreting treatment effects

Analysts should distinguish statistical significance, practical significance and distributional impact. An effect may be precisely estimated but too small to justify implementation; a valuable effect may be uncertain because the sample is small.

Heterogeneous treatment effects ask “for whom and under what conditions does the intervention work?” Effects may differ by tenure, job family, baseline risk or manager support. Subgroup analysis should be theory-driven or pre-specified where possible. Searching many subgroups increases false discoveries.

A simple regression adjustment may be written as:

where Y is the outcome, T is treatment, X is a vector of pre-treatment covariates, α is an intercept, τ is the adjusted treatment coefficient, β is a coefficient vector and ε is an error term. In a randomized experiment, adjustment can improve precision when specified appropriately; in observational data, causal interpretation additionally requires no unmeasured confounding, correct model structure or a valid alternative identification strategy.

Uncertainty should be reported using confidence intervals, standard errors or credible intervals, rather than a point estimate alone. Analysts should also report missingness, attrition, treatment take-up, balance, exclusions and deviations from the analysis plan.

7. From causal effects to prescriptive analytics

Causal evidence estimates consequences; prescriptive analytics selects an action. A recommendation should connect:

1. decision variables, such as number of training places or schedule changes;
2. objective function, such as expected contribution, retention or wellbeing;
3. constraints, such as budget, capacity, legal requirements and fairness;
4. uncertainty, including effect estimates and demand variation;
5. implementation and monitoring plans.

Optimization

Suppose intervention j has decision variable x_j, expected benefit b_j, cost c_j, and capacity u_j. A simple linear programme is:

Here B is the available budget. This model assumes additive expected benefits, known or adequately estimated coefficients, divisibility or suitable integer decisions, and no omitted constraints. If interventions are indivisible, x_j may be binary. Real workforce decisions may require fairness constraints, minimum service levels and interaction terms.

Causal effects should inform b_j, not merely predictive risk scores. If an intervention reduces turnover probability by 0.03 for a defined group and replacement cost is £10,000, its expected gross benefit per treated employee is approximately £300, before programme cost and uncertainty. This calculation assumes the estimated effect transports to the target group and the cost estimate is appropriate.

Simulation and scenario analysis

Simulation represents uncertainty in arrivals, turnover, treatment response, capacity or costs. Monte Carlo simulation repeatedly samples plausible values from specified distributions and records outcomes. Results should be summarized using expected value, percentiles, probability of exceeding a target and downside risk.

Scenario analysis changes coherent assumptions together: for example, low, central and high adoption combined with different effect sizes. Sensitivity analysis changes one assumption or parameter at a time, while stress testing examines adverse conditions. These approaches do not predict the future with certainty; they clarify consequences of assumptions.

Decision criteria

Under uncertainty, managers may use:

• Expected value: choose the action with the highest probability-weighted outcome.
• Maximin: choose the action with the best worst-case result.
• Minimax regret: minimize the largest loss relative to the action that would have been best after the state is known.
• Risk-adjusted value: penalize volatility or downside exposure.
• Value of information: assess whether additional research is worth its cost.

The appropriate criterion depends on organizational risk tolerance, reversibility, employee welfare and legal obligations. A high expected return is not sufficient when a policy creates severe low-probability harms.

8. Illustrative case: flexible work and retention

Background. A fictional technology firm observes higher resignation among employees with long commutes. Leaders propose a flexible-work policy but worry about productivity and unequal access.

Problem. Commute length predicts resignation, but it is not necessarily a causal lever. The firm needs to estimate whether offering two additional remote days per week changes retention and performance.

Analysis. A DAG suggests that role type, manager support, caregiving responsibilities and prior performance affect both eligibility and outcomes. The firm therefore avoids conditioning on post-policy productivity or self-selected participation. It identifies roles where remote work is operationally feasible and defines outcomes before launch: six-month resignation, validated performance measures, absence and an employee wellbeing scale.

Approach. Within eligible teams, employees are randomly offered the policy for six months, with usual practice as control. Team-level assignment is used to limit spillovers. The primary analysis follows intention-to-treat. A pre-specified subgroup analysis examines whether effects differ by commute category, while monitoring workload and promotion outcomes for unintended inequality.

If randomization is blocked by collective bargaining, the firm could use a phased rollout and difference-in-differences, after checking pre-policy trends and documenting why rollout timing was unrelated to anticipated outcomes. A prescriptive model then compares policy packages under budget and manager-capacity constraints. It includes estimated treatment effects, confidence intervals and scenarios for adoption.

Outcome. In this illustrative case, suppose the estimated offer effect is a reduction in resignation but no reliable average change in performance. The firm implements the policy in eligible roles, adds manager training and reviews outcomes quarterly. It does not extend the policy automatically to ineligible roles; instead, it conducts a separate operational assessment.

Learning points. A predictive relationship between commuting and resignation did not establish that remote work would solve the problem. The experiment estimated the effect of an offer, not merely the effect among enthusiastic users. Prescriptive selection required costs, capacity, equity and uncertainty in addition to the causal estimate.

9. Ethics, governance and responsible implementation

People analytics affects livelihoods and dignity. Governance should include:

• informed and proportionate use of employee data;
• purpose limitation, data minimization, access control and retention rules;
• transparency about experimentation where feasible;
• review by ethics, privacy, employee-representative and legal functions;
• monitoring for disparate impact and unequal treatment effects;
• secure handling of sensitive attributes and linkage keys;
• a process for challenge, correction and appeal.

A/B testing may be inappropriate where withholding a proven safety or wellbeing intervention creates material harm. Consent, legitimate organizational purpose and applicable employment, privacy and anti-discrimination law must be considered. Fairness is not guaranteed by predictive accuracy or causal validity: an intervention can have a positive average effect while burdening a protected group.

AI can assist with heterogeneous-effect discovery, optimization and simulation, but automated recommendations require human accountability. Models should be documented with data provenance, estimand, assumptions, version history, validation results and monitoring thresholds. Drift, policy changes and behavioral adaptation can make historical effects unreliable.

10. Best-practice workflow

A defensible causal-prescriptive project follows these steps:

1. Define the decision, treatment, population, outcomes and time horizon.
2. Draw a causal diagram and document assumptions.
3. Prefer randomization when feasible and ethically acceptable.
4. If not, select a quasi-experimental design that matches the institutional setting.
5. Pre-specify the primary estimand, analysis and success criteria.
6. Check implementation, balance, overlap, missingness and attrition.
7. Estimate effects with uncertainty and test plausible heterogeneity.
8. Translate effects into benefits, costs, constraints and risks.
9. Compare interventions through optimization, scenarios or simulation.
10. Pilot, monitor, audit and revise without treating one study as permanent truth.

Chapter Summary

Prediction forecasts outcomes, whereas causal analysis estimates the consequences of interventions. Potential-outcomes notation defines counterfactual outcomes and treatment effects; ATE and ATT answer different questions. Randomized experiments provide strong identification when assignment, implementation and measurement are sound. Quasi-experimental designs—including difference-in-differences, regression discontinuity, instrumental variables and matching—can be valuable but depend on explicit assumptions.

Causal diagrams expose confounding, mediation and collider bias. Treatment effects should be reported with uncertainty and examined for meaningful heterogeneity. Prescriptive analytics uses causal estimates within optimization, simulation and scenario frameworks, while accounting for budgets, capacity, fairness and risk. Ethical governance requires transparency, privacy protection, employee safeguards and human accountability. The central professional discipline is to separate what the data predict from what an intervention is likely to change.

Key Terms

Causal effect: Change in an outcome attributable to an intervention.

Counterfactual: The unobserved outcome that would have occurred under an alternative treatment condition.

Treatment: A defined policy, programme, exposure or action applied to a unit.

Control: The comparison condition against which treatment is evaluated.

ATE: Average treatment effect across the defined target population.

ATT: Average treatment effect among units that received treatment.

Confounder: A common cause of treatment and outcome.

Mediator: A variable through which treatment affects an outcome.

Collider: A variable caused by two variables, whose conditioning can create bias.

DAG: Directed acyclic graph representing causal assumptions.

Randomized controlled trial: Experiment assigning units to conditions by chance.

Intention-to-treat: Effect of assignment or offer, regardless of actual take-up.

Difference-in-differences: Comparison of changes over time between treated and comparison groups.

Regression discontinuity: Design exploiting a treatment threshold.

Instrumental variable: Variable affecting treatment but satisfying assumptions for causal identification.

Propensity score: Probability of treatment conditional on observed covariates.

Optimization: Selection of actions that maximize or minimize an objective under constraints.

Simulation: Repeated analysis of outcomes under sampled uncertain conditions.

Value of information: Expected benefit from obtaining additional information before deciding.

Treatment-effect heterogeneity: Variation in causal effects across people or contexts.

Review Questions

Short Answer Questions

1. Why does a predictive association between training and promotion not by itself establish causation?
2. Define the average treatment effect and average treatment effect on the treated.
3. What is the parallel-trends assumption in difference-in-differences?
4. Distinguish a mediator from a confounder.
5. Why can intention-to-treat be more policy-relevant than a per-protocol estimate?

Long Answer Questions

1. Explain how a randomized workplace experiment should be designed, implemented and evaluated.
2. Compare difference-in-differences, regression discontinuity and instrumental variables, including a major assumption for each.
3. Explain how a DAG can guide covariate selection and prevent collider bias.
4. Describe how causal estimates can be incorporated into an optimization model for intervention allocation.
5. Discuss ethical and governance requirements for prescriptive people analytics.

Analytical / Application Questions

1. A manager assigns coaching to employees with the lowest performance scores. Coaching participants later improve less than nonparticipants. Identify possible causal explanations and propose a credible evaluation design.
2. Two regions have retention rates of 80% and 82% before a policy. Afterward, the treated region is at 86% and the comparison region at 84%. Calculate the difference-in-differences estimate and interpret it, subject to its assumptions.
3. An intervention costs £150 per employee and is estimated to reduce turnover probability by 0.02. If replacement cost is £9,000, calculate expected gross benefit per employee and state what additional information is needed before implementation.

Multiple Choice Questions with Answers

Question 1

Which question is causal rather than predictive?

A. Which employees are likely to resign?

B. What is the probability of absence next month?

C. What is the effect of flexible work on resignation?

D. Which teams have the highest forecast turnover?

Correct Answer: C. What is the effect of flexible work on resignation?

Question 2

In the potential-outcomes framework, Y_i(1) represents:

A. The observed outcome only when employee i is untreated

B. Employee i’s outcome if exposed to treatment

C. The average outcome for all employees

D. The treatment assignment probability

Correct Answer: B. Employee i’s outcome if exposed to treatment

Question 3

Which variable is generally a confounder?

A. A post-treatment mediator

B. A common cause of treatment and outcome

C. A random measurement error

D. A variable caused by both treatment and outcome

Correct Answer: B. A common cause of treatment and outcome

Question 4

The key identifying assumption in a standard difference-in-differences design is:

A. Treatment effects are identical for every employee

B. The treated group is larger than the control group

C. Treated and control groups would have followed parallel trends without treatment

D. Every employee complies with treatment

Correct Answer: C. Treated and control groups would have followed parallel trends without treatment

Question 5

What does intention-to-treat estimate?

A. The effect of assignment or offer, regardless of take-up

B. The effect only among perfect compliers

C. The effect after controlling for a post-treatment mediator

D. The effect among employees who refuse treatment

Correct Answer: A. The effect of assignment or offer, regardless of take-up

Question 6

A regression-discontinuity design is based primarily on:

A. A random sample from several countries

B. A treatment rule tied to a cutoff

C. A forecast generated by a neural network

D. Matching treated employees to managers

Correct Answer: B. A treatment rule tied to a cutoff

Question 7

If an intervention reduces turnover probability by 0.02 and replacement cost is £9,000, its expected gross benefit per employee is:

A. £18

B. £180

C. £450

D. £9,020

Correct Answer: B. £180

Question 8

In an optimization model, a budget limit is an example of a:

A. Causal estimand

B. Constraint

C. Mediator

D. Counterfactual

Correct Answer: B. Constraint

Question 9

Which practice can create collider bias?

A. Adjusting for a pre-treatment confounder

B. Randomly assigning treatment

C. Conditioning on a variable caused by treatment and another cause of outcome

D. Reporting a confidence interval

Correct Answer: C. Conditioning on a variable caused by treatment and another cause of outcome

Question 10

Why should prescriptive people analytics include scenario analysis?

A. To eliminate all uncertainty from causal estimates

B. To compare consequences under different coherent assumptions

C. To replace outcome measurement with expert opinion

D. To guarantee that an intervention is fair

Correct Answer: B. To compare consequences under different coherent assumptions