Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Statistical Analysis of Groundwater Monitoring Data: An Overview of API Publication 4635-1996
Guidance for Environmental Professionals at RCRA Facilities
Scope of API Publication 4635-1996
API Publication 4635-1996, titled Statistical Analysis of Groundwater Monitoring Data at RCRA Facilities, provides a structured framework for evaluating groundwater quality data generated during Resource Conservation and Recovery Act (RCRA) monitoring programs. The document was developed by the American Petroleum Institute to assist environmental professionals in applying appropriate statistical techniques to determine whether contamination has occurred or if existing contamination is migrating.
This publication addresses the entire workflow – from data collection and verification to the selection and application of statistical tests. It is intended for use at hazardous waste treatment, storage, and disposal facilities (TSDFs) and other RCRA-regulated units where groundwater monitoring is mandated. While the guidance is tailored to petroleum industry facilities, many of the statistical principles are equally applicable to broader environmental monitoring programs.
API Publ 4635-1996 has been widely cited in regulatory guidance and remains a reference for statistical groundwater evaluation even as newer resources have been published.
Technical Requirements
Key Statistical Methods
The publication categorizes statistical tests based on data characteristics and monitoring objectives. It emphasizes the importance of verifying underlying assumptions before selecting a test. The major tests covered include:
| Test Category | Method | Data Type / Assumption | Recommended Use |
|---|---|---|---|
| Parametric – Normality assumed | Student’s t-test (one‑sided, two‑sample) | Normally distributed, independent samples, equal variances | Comparing compliance well to background: most powerful when assumptions hold |
| Parametric – Variability | Cochran’s approximation t-test | Normally distributed, unequal variances | When background and compliance wells show different variance |
| Non‑parametric | Wilcoxon rank-sum test (Mann–Whitney) | No normality assumption, ordinal data | Robust alternative when normality fails or sample sizes are small |
| Prediction limits | Normal / lognormal prediction limit | Normally distributed data, future single observation | Identifying exceedances quickly; common in RCRA compliance monitoring |
| Control charts | Shewhart (individuals/moving range) | Approximately normal, process in control | Trend analysis and early warning for shifts in groundwater quality |
Data Quality and Assumption Verification
Before any statistical test is applied, API Publ 4635 requires that the user assess data for outliers, serial correlation, and distributional shape. It recommends the Shapiro-Wilk test for normality and suggests methods for handling outliers (e.g., Grubbs’ test) or transforming data (log transformation). For censored data (non‑detects), the publication introduces substitution methods (e.g., half the detection limit) and more robust techniques such as maximum likelihood estimation.
The use of simple substitution for non‑detects can introduce bias. API Publ 4635-1996 warns practitioners to carefully document the handling of censored data and to consider more sophisticated methods when the proportion of non‑detects is high (>50%).
Implementation Highlights
Effective implementation of the guidance requires a thorough understanding of the site hydrogeology and well construction. The publication stresses that statistical analysis cannot substitute for good conceptual site models – a statistically significant change in concentration does not automatically indicate contamination if the change has a plausible natural explanation.
Practical Workflow
- Data verification – Validate laboratory results, sample chain of custody, and field measurements.
- Exploratory analysis – Generate summary statistics, box plots, and time series plots.
- Assumption testing – Check for normality, variance homogeneity, and independence.
- Test selection – Choose a method from the publication’s decision tree based on data characteristics.
- Application & interpretation – Apply the test at the appropriate significance level (usually α = 0.01 for one‑tailed comparisons to background).
- Documentation – Record all steps, assumptions, and software used to ensure transparency for regulators.
When selecting the significance level, API Publ 4635 recommends balancing Type I and Type II errors. Many RCRA programs use α = 0.01 to reduce false positives in compliance monitoring, but this may lower power. Consider site‑specific risk in consultation with regulators.
Compliance Notes
Although API Publ 4635-1996 is not a regulatory requirement, it has been formally adopted or referenced by several state environmental agencies as an acceptable methodology for demonstrating compliance under RCRA. The U.S. EPA’s Statistical Analysis of Groundwater Monitoring Data at RCRA Facilities (EPA 530-R-93-003, 1992) – on which the API publication builds – remains a central guidance document. When following API Publ 4635, the following compliance points are critical:
- Comparison to background: The publication uses a fixed baseline approach. Updating background data after a release has been detected is generally not allowed because the background data set would then be contaminated.
- Multiple comparisons: When many constituents or wells are tested, the probability of at least one false positive increases. The publication discusses Bonferroni correction and other multiplicity adjustments.
- Annual reporting: Results of statistical tests are typically submitted as part of the facility’s annual groundwater monitoring report. Regulators expect a clear description of the statistical method and all assumption checks.
- Software validation: If commercial or open‑source statistical software is used (e.g., ProUCL, R, Minitab), the facility should ensure that the software implements the methods exactly as described in the publication. Some agencies require proof of validation.
Failure to follow an approved statistical method – or applying a method incorrectly – can lead to regulatory rejection of the monitoring results and may trigger enforcement actions, including increased sampling frequency or mandatory corrective measures.
Since its publication in 1996, the field of environmental statistics has evolved. API Publ 4635-1996 does not cover newer methods such as regression on order statistics for non‑detects or Bayesian approaches. However, for facilities operating under a legacy permit that references this publication, continued use of its methods is often required unless a revision is formally approved by the regulating authority.
Q: Which statistical test does API Publ 4635-1996 recommend for routine compliance monitoring?
A: The publication recommends the Student’s t-test (one-sided, two-sample) for comparing a compliance well to a background well when data are normally distributed and variances are equal. If assumptions are not met, a non‑parametric alternative such as the Wilcoxon rank-sum test is suggested. For prediction limit approaches, a normal prediction limit is often used for a single future observation.
A: The publication recommends the Student’s t-test (one-sided, two-sample) for comparing a compliance well to a background well when data are normally distributed and variances are equal. If assumptions are not met, a non‑parametric alternative such as the Wilcoxon rank-sum test is suggested. For prediction limit approaches, a normal prediction limit is often used for a single future observation.
Q: When should I choose a non‑parametric test over a parametric test?
A: Non‑parametric tests (e.g., Wilcoxon rank-sum) should be chosen when the data violate normality, even after transformation, or when sample sizes are too small to reliably assess the distribution shape (n < 10 per well). They are also preferable when the data contain many non‑detects that are handled by substitution of ranks.
A: Non‑parametric tests (e.g., Wilcoxon rank-sum) should be chosen when the data violate normality, even after transformation, or when sample sizes are too small to reliably assess the distribution shape (n < 10 per well). They are also preferable when the data contain many non‑detects that are handled by substitution of ranks.
Q: What is the minimum sample size required to apply the methods in API Publ 4635-1996?
A: The publication recommends at least four independent sampling events per well (four quarterly samples) before applying parametric tests. For non‑parametric tests, six or more independent samples are recommended to achieve adequate power. Fewer samples may be used provided the user is aware of the increased risk of error.
A: The publication recommends at least four independent sampling events per well (four quarterly samples) before applying parametric tests. For non‑parametric tests, six or more independent samples are recommended to achieve adequate power. Fewer samples may be used provided the user is aware of the increased risk of error.
Q: Can I update my background data set after a release has been confirmed?
A: No. Once a release has been detected at a compliance well, the background data set used for comparisons must remain unaltered to avoid introducing bias. The publication specifically advises against including data collected after contamination is suspected. Alternative procedures are described when background wells themselves become impacted.
A: No. Once a release has been detected at a compliance well, the background data set used for comparisons must remain unaltered to avoid introducing bias. The publication specifically advises against including data collected after contamination is suspected. Alternative procedures are described when background wells themselves become impacted.
Reference: API Publication 4635-1996, Statistical Analysis of Groundwater Monitoring Data at RCRA Facilities. Washington, D.C.: American Petroleum Institute.
Footer: © 2026 All rights reserved.