Data, Analytics and Decision Support

Introduction and context

Infrastructure assets generate large volumes of data: condition inspections, work orders, sensor readings, traffic counts, energy consumption, customer complaints, financial transactions, geospatial coordinates and climate observations. Data alone, however, does not improve infrastructure. Value arises when data is trustworthy, interpreted within context, converted into evidence, and connected to a decision with an accountable owner.

Infrastructure management therefore requires an integrated decision-support capability. This capability links data governance, analytical methods, visualization, forecasting, optimization, geospatial analysis and uncertainty management. It supports decisions such as whether to renew a bridge or maintain it, where to deploy emergency crews, how to prioritize a capital programme, and how to balance cost, resilience, service quality and equity.

A central principle is that analytical sophistication cannot compensate for poor problem definition or unreliable data. A highly accurate model trained on biased, incomplete or incorrectly defined data can produce systematically harmful decisions. Conversely, a simple transparent model may be more useful when data are limited and decisions must be explained publicly.

Learning objectives

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

1. Explain the role of data governance in infrastructure management.
2. Design useful performance indicators and dashboards.
3. Apply basic forecasting and interpret forecast uncertainty.
4. Formulate infrastructure prioritization as an optimization problem.
5. Explain how geospatial analysis supports investment and operations.
6. Distinguish risk, variability and model uncertainty.
7. Evaluate decision-support models critically, including ethical and institutional implications.

1. From operational data to decisions

A decision-support system can be understood as a chain:

Data → information → analysis → insight → decision → action → feedback.

Data are recorded observations, such as a pump vibration measurement of 7.2 mm/s. Information adds meaning, for example identifying that the value exceeds the asset’s alert threshold. Analysis investigates patterns, causes or alternatives. Insight is a decision-relevant interpretation, such as evidence that a bearing is deteriorating. A decision selects an action, such as planned replacement. Feedback records the outcome and improves future decisions.

Infrastructure data usually have six dimensions of quality:

Data quality is decision-specific. A road inventory may be sufficiently accurate for network-level budgeting but inadequate for designing a drainage structure. Managers should define the required level of accuracy, update frequency and confidence before collecting data.

Data governance

Data governance is the allocation of authority, responsibility and controls for data. It includes ownership, stewardship, standards, access rights, security, retention, quality management and auditability. A practical governance arrangement normally identifies:

• Data owner: accountable for the purpose, quality expectations and lawful use of a dataset.
• Data steward: manages definitions, validation and day-to-day quality.
• System custodian: maintains the platform, backups and technical controls.
• Data user: uses data within authorized purposes and reports defects.

A data catalogue should document dataset names, definitions, units, spatial and temporal coverage, update frequency, source systems, known limitations and responsible persons. Master data management is especially important for asset identifiers. If a pump is called P-104 in a maintenance system, 104-P in a drawing and “North station pump” in a spreadsheet, integration becomes unreliable.

Governance must also address access and protection. Critical infrastructure data may reveal vulnerabilities; personal data may appear in travel, complaint or workforce records. Access should follow least-privilege principles, while anonymization, encryption, retention limits and incident response should be proportionate to the risk. Public transparency should be maximized where possible without exposing security-sensitive or personally identifiable information.

2. Performance measurement and dashboards

A performance indicator is a defined measure used to assess progress toward an objective. A good indicator has a clear numerator and denominator, a specified population and period, a data source, an owner and a decision rule.

For example, the proportion of water samples meeting a standard is:

The result is dimensionless and should be reported with the period, location and applicable standard. Invalid or missing samples must not silently be treated as compliant.

Indicators should be organized in a hierarchy:

1. Strategic outcomes: reliability, safety, affordability, resilience and environmental performance.
2. Service-level indicators: interruptions per customer, travel-time reliability or flood protection level.
3. Operational indicators: response time, preventive-maintenance completion and backlog.
4. Diagnostic measures: sensor faults, inspection coverage or data latency.

A dashboard is not merely a collection of charts. It is an interface between evidence and management action. Effective dashboards show actual performance against a target, trend, status, confidence or data-quality warning, and the action required when performance deviates.

Rates, averages and denominators

Indicators can mislead when denominators are ignored. Ten failures in a network of 100 assets is materially different from ten failures in a network of 10,000. A failure rate may be expressed as:

Exposure might be asset-years, vehicle-kilometres, operating hours or customer-years. The chosen denominator must match the mechanism being measured. Comparing raw incident counts across districts with different populations or road lengths can produce inequitable priorities.

Averages also conceal distribution. An average emergency response time of 20 minutes could consist of uniformly acceptable responses or a mixture of five-minute responses and severe two-hour delays. Dashboards should therefore use percentiles, ranges or control limits where distribution and reliability matter.

Leading and lagging indicators

Lagging indicators measure outcomes after they occur, such as failures, injuries or service interruptions. Leading indicators measure conditions associated with future outcomes, such as inspection completion, corrosion readings or unresolved defects. Leading indicators support prevention but are not automatically causal. Completing inspections may improve knowledge without improving physical condition unless findings lead to effective work.

3. Forecasting infrastructure demand and condition

Forecasting estimates future values using historical observations, explanatory variables or expert judgment. It supports capacity planning, maintenance scheduling, staffing, procurement and financial planning.

A simple time-series model decomposes an observation into level, trend, seasonality and irregular variation:

Here, y_t is the observed value at time t, ℓ_t is the underlying level, b_t is the trend, s_t is the seasonal component and ε_t is random error. This additive form is suitable when seasonal fluctuations are approximately constant in absolute size. A multiplicative form is more appropriate when seasonal variation grows with the level, but it requires positive observations.

For a moving-average forecast over k periods:

The method is transparent and useful for stable series, but it reacts slowly to genuine changes and does not explain causal drivers. Weighted moving averages or exponential smoothing give greater weight to recent observations.

For a simple linear explanatory model:

y_t is the outcome, x_t is an explanatory variable, β₀ is the intercept, β₁ is the estimated change in the outcome per unit change in x_t, and ε_t is unexplained error. The model should not be interpreted causally merely because two variables are correlated. For example, electricity demand may correlate with temperature, but policy, building stock and tariffs may also matter.

Forecast evaluation

Forecasts must be evaluated using data not used to fit the model. Common measures include mean absolute error:

and mean squared error:

where y_i is the observed value, ŷ_i is the forecast and n is the number of test observations. MAE has the same units as the outcome and is easy to interpret. MSE gives greater emphasis to large errors and has squared units. Percentage measures can be problematic when actual values are zero or close to zero.

Point forecasts should be accompanied by prediction intervals. A 95% prediction interval is not a guarantee that the future value lies inside it; under the model assumptions, approximately 95% of similarly constructed intervals would contain their corresponding future observations. Intervals widen with horizon and should reflect parameter uncertainty, random variation and possible structural change.

Condition forecasting often uses survival concepts. If T is time to failure, the survival function is:

The hazard function describes the instantaneous failure rate conditional on survival to time t. These models require careful treatment of assets that have not yet failed, known as censored observations. Replacing censored observations with arbitrary failure dates biases results.

4. Optimization and prioritization

Optimization selects the best feasible alternative according to an objective and constraints. Infrastructure decisions are rarely about maximizing one number. They involve trade-offs among cost, risk, service, carbon, resilience and distributional fairness.

A linear programme may be written as:

Here, x_i indicates whether project i is selected, b_i is its quantified benefit, c_i is its cost, B is the budget, and Z is total benefit. This formulation assumes additive benefits, known costs, a single budget and binary project choices. Real programmes may require integer or nonlinear models, precedence constraints, workforce limits and interactions between projects.

A risk-based prioritization score can be useful for screening, but it is not automatically an optimization. A common conceptual risk expression is:

The product is meaningful only when probability and consequence are defined consistently and the consequence measure is appropriate. It can undervalue low-probability catastrophic events, non-monetary impacts and unequal burdens. Multi-criteria decision analysis can make such dimensions explicit by scoring alternatives against criteria and applying stated weights. The weights should be tested through sensitivity analysis rather than presented as objective facts.

Multi-year investment planning

A robust investment process typically:

1. defines service and resilience outcomes;
2. identifies candidate interventions;
3. estimates life-cycle costs and benefits;
4. identifies dependencies and constraints;
5. evaluates risk and uncertainty;
6. tests alternative budget scenarios;
7. records the rationale and accountable decision-maker.

Life-cycle cost analysis should include construction, operation, maintenance, renewal, disposal and relevant user or environmental costs. If future monetary flows are discounted at rate r, present value is:

where C_t is the cost or benefit in period t. The discount rate, price basis, analysis period and treatment of residual value must be stated. Financial discounting should not obscure ethical questions about long-term environmental harm or the interests of future generations.

5. Geospatial analysis

Infrastructure is spatial. Location determines service access, hazard exposure, network connectivity, maintenance travel and environmental impact. A geographic information system (GIS) links geometry with attributes and permits analysis across layers.

Common spatial operations include:

• Buffering: identifying assets within a specified distance of a road, river or hazard.
• Overlay: combining layers to find areas satisfying multiple conditions.
• Network analysis: calculating routes, travel times, connectivity or catchments.
• Spatial interpolation: estimating values between observations, subject to assumptions about spatial continuity.
• Hotspot analysis: identifying unusually concentrated incidents relative to a defined baseline.

Coordinate reference systems matter. Distances and areas should generally be calculated in an appropriate projected coordinate system rather than directly in latitude and longitude, whose angular units do not represent a constant ground distance. Boundary changes, geocoding errors and positional accuracy should be documented.

A simple accessibility measure can be defined as the minimum network travel time from population location j to a service facility:

where t_ji is travel time from location j to facility i. The interpretation depends on the travel mode, time of day, network condition and whether capacity or service quality is considered. A map showing a facility within a five-kilometre radius may exaggerate accessibility if a river, restricted crossing or steep terrain separates users from it.

Spatial analysis also raises equity concerns. Aggregated areas can conceal vulnerable groups, a problem known as ecological fallacy. Results may change when boundaries change, known as the modifiable areal unit problem. Analysts should test multiple spatial scales and report uncertainty rather than imply that a map is neutral.

6. Uncertainty, risk and robust decisions

Uncertainty is unavoidable because infrastructure systems, data and future conditions are imperfectly known. Important categories include:

• Aleatory variability: inherent randomness, such as fluctuating demand.
• Epistemic uncertainty: incomplete knowledge, such as uncertain deterioration rates.
• Measurement uncertainty: error in sensors, surveys or inspections.
• Model uncertainty: uncertainty about whether the chosen model represents reality.
• Scenario uncertainty: uncertainty about future policy, climate, technology or behaviour.

Risk analysis should distinguish the probability distribution of outcomes from the decision-maker’s tolerance for those outcomes. Expected value is useful for comparing repeated or diversified decisions, but not sufficient when losses are catastrophic, irreversible or unequally distributed.

Monte Carlo analysis propagates uncertainty by repeatedly sampling uncertain inputs and calculating outputs. If project cost is C and service benefit is B, the simulation produces a distribution of outcomes rather than a single estimate. Inputs must have defensible distributions and dependencies; assuming independence when costs and delays are correlated can materially understate risk.

Sensitivity analysis changes one or more assumptions to identify influential drivers. Scenario analysis examines coherent alternative futures, such as high growth, constrained funding or severe climate exposure. Robust decision-making seeks options that perform acceptably across many plausible futures, rather than optimizing for one forecast that may be wrong.

A decision tree can represent sequential choices. At a chance node, branches carry probabilities that sum to one; at a decision node, the manager selects an action. Probabilities should be updated when credible new evidence arrives. The tree should not be treated as a substitute for judgment: its structure reflects assumptions about which events matter and which choices are available.

7. Artificial intelligence and advanced analytics

Machine learning can classify defects in images, detect anomalies in sensor streams, forecast demand and prioritize inspections. Its use is appropriate when a sufficiently large, representative and governed dataset exists and when the predicted output is linked to a real decision.

Key controls include separation of training and test data, protection against leakage, evaluation across asset types and demographic or geographic groups, monitoring for drift, and human review of high-consequence decisions. A model that performs well overall may fail on rare asset classes or under changed operating conditions. Explainability is especially important where residents, workers or contractors may be affected.

AI should augment, not obscure, professional accountability. Managers remain responsible for defining objectives, approving interventions, validating outputs and communicating limitations. Automated recommendations should include provenance, confidence and an escalation path.

8. Illustrative case: renewal planning for a municipal water network

Background

This illustrative case concerns a city with 1,200 km of water mains. Breaks have increased, but the annual renewal budget is fixed. The utility holds work-order records, pipe age and material data, pressure readings, customer complaints, road locations and census-based vulnerability indicators. Records use inconsistent asset identifiers.

Problem

The utility’s previous method ranked pipes by age alone. This produced visible renewal activity but did not consistently reduce service disruption. Older pipes in low-consequence locations sometimes displaced younger pipes serving hospitals or areas with limited alternative supply.

Analysis

The utility first established a common asset identifier and audited missing coordinates, duplicate records and inconsistent failure codes. It defined the decision as selecting a five-year renewal programme subject to annual budget, crew capacity and road-coordination constraints.

An interpretable deterioration model estimated the probability of failure from age, material, diameter, soil conditions and prior breaks. The team tested the model on later records and reported calibration, not merely classification accuracy. Consequences included repair cost, expected interruption duration, traffic disruption, water loss and service criticality. Vulnerability was included as a separate criterion, with a documented policy rationale rather than hidden inside a technical score.

GIS analysis identified schools, hospitals, emergency routes and areas with limited network redundancy. A multi-year optimization model considered project cost, expected risk reduction, dependencies and coordination with planned road works. Scenarios were run for baseline funding, a 20% reduction and a climate-related increase in demand.

Approach and outcome

The final programme combined high-risk replacements, targeted pressure management and additional inspection where model confidence was low. It reserved a small contingency for emerging failures and required post-project monitoring. The utility did not claim that the model predicted every break. Instead, it used the model to compare options transparently and retained engineering review for unusual conditions.

After two years, reported interruptions declined in critical service zones, and coordination reduced repeated road excavation. The principal improvement was not a single algorithm; it was the integration of governance, spatial context, uncertainty analysis and accountable decision rules.

Learning points

1. Asset identity is a foundation for analytics.
2. Age is a predictor, not a complete decision rule.
3. Consequence and equity must accompany probability.
4. Validation must use future or held-out data.
5. A model should reveal uncertainty and support, not replace, professional judgment.

9. Implementation, governance and professional practice

An analytics programme should begin with decisions, not software procurement. A useful implementation sequence is:

1. Define the decision, stakeholders, time horizon and authority.
2. Specify indicators, data definitions and minimum quality requirements.
3. Establish ownership, access controls, lineage and retention.
4. Build a reproducible data pipeline with validation checks.
5. Develop a transparent baseline model before adding complexity.
6. Validate against out-of-sample data and operational experience.
7. Conduct sensitivity, equity, privacy and security assessments.
8. Pilot with users, document decisions and monitor outcomes.
9. Review models after changes in assets, policy, climate or behaviour.

Documentation should state the purpose, population, exclusions, assumptions, transformations, model version, performance, limitations and approval date. Reproducibility requires preserved source data or snapshots, code or configuration, and a record of human interventions.

Ethically responsible infrastructure analytics must address fairness, privacy, accessibility and accountability. A technically efficient investment can still be unacceptable if it systematically shifts risk to a disadvantaged community. Consultation should be meaningful: affected groups should understand the decision, have an opportunity to challenge assumptions and receive an explanation of how evidence was used.

Challenges and limitations

First, data are often incomplete precisely where risks are greatest: remote assets, informal settlements or undocumented historical infrastructure. Second, models may encode past institutional bias. Third, dashboards can encourage target chasing, especially when staff are evaluated on indicators that are easy to improve but weakly connected to outcomes. Fourth, interoperability failures create duplicated and contradictory records. Fifth, forecasts can fail under structural breaks such as pandemics, extreme weather or regulatory change. Finally, optimization may produce mathematically efficient but politically or ethically unacceptable allocations.

These limitations do not justify abandoning analytics. They require proportionality, transparency, triangulation with professional knowledge and explicit recognition of what is not known.

Chapter Summary

Infrastructure decision support converts governed data into choices about investment, maintenance and operations. Data quality, asset identity, lineage and access controls are prerequisites for credible analysis. Performance dashboards should connect indicators to targets, trends, distributions, data quality and action, while distinguishing leading from lagging measures.

Forecasting methods range from transparent moving averages to explanatory and survival models. Forecasts must be evaluated on held-out data and accompanied by uncertainty intervals. Optimization makes objectives and constraints explicit, but risk, equity, resilience and non-monetary consequences require careful formulation and sensitivity testing. GIS adds spatial context through buffering, overlays, networks and accessibility analysis, while analysts must address coordinate systems, aggregation effects and positional uncertainty.

Risk-informed practice distinguishes variability, incomplete knowledge, measurement error, model uncertainty and future scenarios. Simulation, sensitivity analysis and robust decision-making help prevent false precision. AI can extend analytical capability, but validation, fairness, explainability, monitoring and human accountability remain essential. The strongest infrastructure organizations integrate technical models with governance, institutional judgment and transparent public value decisions.

Key Terms

Data governance: Authority, rules and controls for managing data assets.

Data lineage: A traceable record of data origin, transformations and uses.

Data steward: Person responsible for practical data quality and definitions.

Performance indicator: Defined measure used to assess progress toward an objective.

Dashboard: Decision-oriented visual interface presenting performance evidence.

Leading indicator: Measure associated with future performance or risk.

Lagging indicator: Measure of an outcome after it has occurred.

Forecast: Estimate of a future value or event.

Prediction interval: Range intended to contain a future observation at a stated level under model assumptions.

Survival function: Probability that an event time exceeds a specified time.

Optimization: Selection of the best feasible alternative under stated objectives and constraints.

Constraint: Requirement limiting the alternatives available to a model.

Life-cycle cost: Cost incurred across acquisition, operation, maintenance, renewal and disposal.

Geographic information system: System for storing, analyzing and displaying spatial data.

Spatial overlay: Combination of geographic layers to evaluate coincident conditions.

Aleatory variability: Inherent randomness in a process or outcome.

Epistemic uncertainty: Uncertainty caused by incomplete knowledge.

Sensitivity analysis: Examination of how results change when assumptions vary.

Model drift: Decline in model performance as conditions or data change.

Multi-criteria decision analysis: Structured comparison of alternatives across several criteria.

Review Questions

Short Answer Questions

1. Why is data lineage important in infrastructure decision support?
2. Distinguish a leading indicator from a lagging indicator and give one example of each.
3. What is the difference between a forecast point estimate and a prediction interval?
4. Why should distance calculations in GIS use an appropriate coordinate reference system?
5. Distinguish aleatory variability from epistemic uncertainty.

Long Answer Questions

1. Explain how data governance supports reliable infrastructure investment decisions. Include ownership, stewardship, quality, security and lineage.
2. Compare moving-average forecasting, explanatory regression and survival analysis in terms of application, assumptions and limitations.
3. Explain how an infrastructure prioritization model can combine cost, risk, resilience and equity without hiding value judgments.
4. Discuss the role of GIS in network accessibility and hazard-informed asset management, including spatial limitations.
5. Develop a governance and validation framework for using AI to prioritize infrastructure inspections.

Analytical / Application Questions

1. A utility has 12 monthly demand observations and wants to forecast month 13 using a three-month moving average. Explain the calculation, identify the required observations, and discuss when the method could be misleading.
2. A road agency must select projects under a fixed budget, but project benefits are uncertain and some projects serve highly vulnerable communities. Propose an objective, constraints, uncertainty analysis and equity safeguards.
3. A dashboard reports an improving average emergency response time, while the 95th-percentile response time is worsening. Diagnose possible explanations and recommend additional indicators and management actions.

Multiple Choice Questions with Answers

Question 1

Which role is primarily accountable for defining the purpose and quality expectations of a dataset?

A. Data owner

B. System custodian

C. Data user

D. External auditor

Correct Answer: A. Data owner

Question 2

A three-period moving-average forecast for month 4 uses which observations?

A. Months 1, 2 and 3

B. Months 2, 3 and 4

C. Only month 3

D. Months 1 and 4

Correct Answer: A. Months 1, 2 and 3

Question 3

Which measure has the same units as the forecasted variable?

A. Mean squared error

B. Mean absolute error

C. Coefficient of determination

D. Percentage error only

Correct Answer: B. Mean absolute error

Question 4

In a binary project-selection model, what does x_i = 1 usually represent?

A. Project i is selected

B. Project i has zero cost

C. Project i has failed

D. Project i is delayed by one period

Correct Answer: A. Project i is selected

Question 5

Why can a straight-line buffer around a facility misrepresent accessibility?

A. It ignores network barriers and travel routes

B. It always overestimates population size

C. It cannot display asset attributes

D. It converts all data into financial data

Correct Answer: A. It ignores network barriers and travel routes

Question 6

Which is an example of epistemic uncertainty?

A. Random hourly fluctuations in demand

B. Incomplete knowledge of a pipe’s deterioration rate

C. A known deterministic tariff

D. A fixed asset coordinate

Correct Answer: B. Incomplete knowledge of a pipe’s deterioration rate

Question 7

What is the principal purpose of testing a forecast on held-out data?

A. To measure performance on observations not used for fitting

B. To guarantee the forecast is correct

C. To eliminate all random variation

D. To increase the historical sample automatically

Correct Answer: A. To measure performance on observations not used for fitting

Question 8

Which statement about risk is most appropriate?

A. Probability alone fully describes risk

B. Expected loss can be represented as probability multiplied by consequence, subject to appropriate definitions

C. Risk is always equal to asset age

D. A low-probability event can be ignored

Correct Answer: B. Expected loss can be represented as probability multiplied by consequence, subject to appropriate definitions

Question 9

Which practice is most important when deploying an AI model for high-consequence infrastructure decisions?

A. Removing all human review

B. Using only the overall accuracy score

C. Monitoring performance, bias, drift and uncertainty

D. Hiding the model’s limitations from users

Correct Answer: C. Monitoring performance, bias, drift and uncertainty

Question 10

What does a prediction interval communicate?

A. The exact future value

B. A range intended to contain a future observation under stated assumptions

C. The historical average only

D. The project budget constraint

Correct Answer: B. A range intended to contain a future observation under stated assumptions