Abandonment Rate Modeling and Patience Distributions

Abandonment rate modeling is the discipline of forecasting the proportion of contacts that will terminate before receiving service, and characterizing the distribution of caller patience — the maximum wait time a customer is willing to tolerate before abandoning. Abandonment is not a fixed property of a contact center or a customer population; it is a function of the relationship between wait time and the patience distribution of the arriving population. Understanding this relationship is essential for accurate Service Level forecasting, staffing optimization, and the interpretation of operational performance data. Abandonment modeling is analytically distinct from volume forecasting: while volume forecasting addresses how many contacts will arrive, patience distribution modeling addresses how many of those contacts will survive to receive service under a given wait time regime. The two problems are connected through queueing dynamics and must be addressed jointly in any complete staffing model.

The Patience Distribution Concept

A patience distribution describes the probability that a randomly selected arriving customer will abandon before time $t$ if required to wait. Formally, let $T$ denote the patience of an arriving customer — the maximum time they will wait before abandoning. The patience distribution function is:

$F(t)=P(T\le t)$

representing the fraction of customers who will abandon if made to wait exactly $t$ time units. The complement, $1-F(t)$, is the survival function — the probability that a customer remains in queue after $t$ time units.

Garnett, Mandelbaum, and Reiman (2002) established the foundational theoretical framework for queues with impatient customers, demonstrating that the steady-state behavior of an M/M/N+G queue (Erlang-A) depends critically on the shape of the patience distribution.

Empirical Estimation of Patience Distributions

Direct estimation of patience distributions from operational data faces a fundamental censoring problem: only customers who actually abandon reveal their patience; customers who are served before abandoning have patience greater than their actual wait time, but the true patience value is unobserved. This is a right-censoring problem identical in structure to survival analysis in clinical research.

Brown et al. (2005) provide a landmark statistical analysis of call center data from a U.S. bank, applying survival analysis methods to estimate patience distributions from ACD records. Key findings include:

• The patience distribution is well-approximated by a lognormal distribution in many real-world call centers, though this is not universal.
• Patience varies systematically by time of day, day of week, and customer segment.
• The hazard rate of abandonment is not constant — customers are more likely to abandon in the first 60–90 seconds of wait than in subsequent periods, consistent with psychological models of wait time perception.

Standard estimation approaches include:

• Kaplan-Meier estimator — non-parametric survival function estimation from censored wait-time data.
• Parametric fitting — fitting lognormal, exponential, or Weibull distributions to the observed abandonment times, using maximum likelihood estimation with censoring adjustment.
• Cox proportional hazards model — estimating patience as a function of covariates (time of day, queue type, customer segment) while controlling for censoring.

Abandonment as a Function of Wait Time

The abandonment rate observed in operational reporting is not an intrinsic property of the customer population; it is an endogenous outcome of the staffing and service level regime. This has critical implications for forecasting and interpretation:

• When service levels are high (short waits), abandonment rates are low — but only because customers with low patience never waited long enough to abandon.
• When service levels deteriorate (long waits), abandonment rates rise — because the queue exposes more customers to wait times that exceed their patience.
• Comparing abandonment rates across periods with different average wait times conflates changes in patience with changes in service delivery.

This endogeneity means that forecasting abandonment rate directly — treating it as a stable time series — is methodologically flawed. Accurate abandonment modeling requires: (1) estimating the patience distribution from historical data using survival analysis, and (2) deriving expected abandonment as a function of the projected wait time distribution, which itself depends on staffing.

Queueing Models Incorporating Impatience

The classical Interval Level Staffing Requirements calculation uses the Erlang-C formula, which assumes infinite customer patience — no abandonment. The Erlang-A model (M/M/N+M, with exponential patience) and its generalization (M/M/N+G, with general patience distribution) account for abandonment.

In the Erlang-A model, the steady-state abandonment rate is determined by the intersection of the offered workload, the number of servers, and the patience rate parameter (typically denoted $\theta$). When $\theta$ is large (low patience), abandonment rises steeply with queue length; when $\theta$ is small (high patience), the system behaves closer to Erlang-C.

The practical consequence for staffing models:

• Erlang-C overstates the required number of agents to achieve a target service level, because it assumes all contacts waiting will eventually be served; Erlang-A correctly accounts for the fact that some contacts self-resolve through abandonment.
• However, abandonment is not a desirable substitute for staffing — each abandoned contact represents a failed service event, potentially associated with re-contact, customer churn, and reputational cost.

Implications for Forecasting Practice

Volume Forecast Adjustment for Abandonment

The volume forecast that feeds into staffing calculations should represent offered volume (all contacts that arrive and attempt to reach an agent), not handled volume (contacts that successfully connect). If historical data is sourced from ACD records that exclude abandoned contacts, the forecaster must reconstruct offered volume by adding estimated abandons back to handled contacts.

This requires an estimate of the abandonment rate itself — creating circularity if abandonment is also derived from wait time, which depends on staffing, which depends on the volume forecast. In practice, this circularity is resolved through iterative calculation or simulation.

Re-Contact from Abandonment

Customers who abandon may re-enter the queue, inflating offered volume beyond original arrival volume. The re-contact rate from abandonment varies by customer segment, channel, and the urgency of the underlying need. High-urgency contacts (e.g., fraud disputes, service outages) have near-100% re-contact rates; low-urgency contacts may not re-contact at all.

Patience Variation by Segment and Channel

Patience distributions are not uniform across customer populations. Premium tier customers, customers calling about high-stakes issues, and first-time callers may have systematically different patience profiles than the overall average. For organizations with Multi-Channel and Blended Operations, patience also varies by channel: chat users may have lower patience for queue delays than voice callers, while email senders have higher effective patience given the asynchronous SLA expectation.

Maturity Model Considerations

At Level 2–3 (Foundational/Integrated), abandonment is typically tracked as a reported metric but not explicitly modeled. Staffing calculations use Erlang-C or vendor-default parameters. Abandonment rate is treated as a historical average applied uniformly to future periods.

At Level 3 (Integrated), organizations begin segmenting abandonment by time of day and queue type, and recognize the wait-time dependency. Basic Erlang-A calculations may be introduced.

At Level 4 (Optimized), patience distributions are formally estimated using survival analysis. Staffing models incorporate Erlang-A or simulation-based approaches. Abandonment forecasting accounts for endogeneity and re-contact effects. Segment-specific patience parameters are maintained.

At Level 5 (Adaptive), patience modeling extends to AI and automated contact interactions, where "patience" is replaced by session timeout parameters and retry logic. Human-agent and AI-agent blend models incorporate differential abandonment characteristics.

Related Concepts

• Service Level
• Abandonment
• Interval Level Staffing Requirements
• Demand calculation
• Probabilistic Forecasting
• Forecasting Methods
• Multi-Channel and Blended Operations
• Contact Deflection and Channel Shift Modeling
• WFM Labs Maturity Model