PU699 Unit 5 Data Analysis Section Example

Nineteen Calls Became Thirty-Eight Without More Closures, at Two Beaches Out of Four: A Unit 5 Data Analysis Section Testing a Morning Prediction Model Against Yesterday's Result on Two Seasons of a Composite City's Beach Data

[Student Name]

Purdue University Global

PU699: Master of Public Health Capstone

Unit 5 Data Analysis Section

[Instructor Name]

[Date]

Port Alder's beach records, weather data and model are composite; the methods, the bacterial standard and the research cited are real.

Data and Methods

Records: E. coli results for every weekday sample at the four public beaches, Memorial Day to Labor Day, in 2024 and 2025, 140 beach-days per beach; the advisory log; 24-hour rainfall from the airport station; daily mean discharge of Mill Creek from the stream gauge; turbidity and wave height from the offshore buoy; and wind direction. Each predictor is available before 8 a.m., so the model could be run before the morning posting. A day counts as an exceedance when its result is over the 235 single-sample value used in the current protocol. Similar morning models have supported advisories at several Great Lakes programs (Francy, 2009). Two rules are compared. The current rule, persistence, treats the previous sampling day's result as today's. The model is a logistic regression predicting the chance of exceedance from rainfall, creek discharge, turbidity, wave height, onshore wind and a term for each beach. To avoid testing it on the data that built it, the model was fit to 2024 and used to predict 2025, then fit to 2025 and used to predict 2024. Its cutoff was set so that its false advisories across both seasons would not exceed those of the current rule, which keeps the comparison fair: any gain must come from catching more exceedances, not from posting more advisories. The analysis script is written in R with a fixed random seed of 2026, and the appendix holds the script, a synthetic file with the same structure as the real one and the expected output.

Counts

1. Mill Creek Beach: 31 exceedances in 140 beach-days.

2. North Pier Beach: 12 in 140.

3. Lakeside Beach: 9 in 140.

4. Alder Point Beach: 12 in 140.

5. All beaches: 64 in 560, 11.4 percent.

Twenty-four of the 31 Mill Creek exceedances followed more than 10 millimeters of rain in the previous 24 hours. At Alder Point, only 4 of 12 did.

Comparison of the Two Rules, Validated

1. Exceedance days with an advisory in place: current rule 19 of 64, 29.7 percent; model 38 of 64, 59.4 percent.

2. Advisories on days that met the standard: current rule 42 of 496; model 43 of 496.

3. Both rules correct on 15 exceedance days; current rule only on 4; model only on 23; neither on 22.

4. Difference in exceedance days caught: 29.7 percentage points, with a bootstrap 95 percent interval of 15.6 to 42.2 from 2,000 resamples. An exact McNemar test on the 27 days where the rules disagreed gives p below 0.001.

The model caught twice as many exceedance days as yesterday's result while posting one more false advisory over two summers.

By Beach

1. Mill Creek: current rule 9 of 31; model 23 of 31. False advisories 14 and 9.

2. North Pier: 4 of 12 and 7 of 12. False advisories 10 and 11.

3. Lakeside: 2 of 9 and 5 of 9. False advisories 8 and 9.

4. Alder Point: 4 of 12 and 3 of 12. False advisories 10 and 14.

Sensitivity Checks

Three checks ask whether the result would survive different analytic choices. To begin with, the cutoff. Moving it so the model posts ten fewer false advisories than the current rule still leaves it catching 33 of 64 exceedance days, against 19. Second, the predictors. Dropping turbidity and wave height, which come from a single offshore buoy that was out of service for nine days in 2025, reduces the model's catch to 34, so most of the gain comes from rainfall and creek discharge, the two most reliable inputs. Third, the threshold. Repeating the comparison with a lower single-sample value of 190 colony-forming units changes the counts but not the pattern: the model gains at Mill Creek and loses at Alder Point. None of the three checks reverses the main finding, and the second shows the model does not depend on the least dependable instrument.

Interpretation

The gain is concentrated at Mill Creek Beach, where most exceedances follow rain and the model's weather variables carry real information. North Pier and Lakeside improve modestly, on small numbers that make the beach-level gains uncertain. At Alder Point, the model does worse than the current rule on both counts. That beach's exceedances mostly occur on dry days, consistent with a source in the foreshore sand and the gull colony rather than runoff, a pattern described at other Great Lakes beaches (Whitman & Nevers, 2003). This is the beach-by-beach variation the Chicago comparison found, where models helped some beaches and not others (Nevers & Whitman, 2011).

What the Analysis Cannot Show

It cannot show that swimmers would have been healthier, because the program has no illness data and the analysis measures only how well advisories match the water. It cannot show that the model will perform as well in the next season, since two seasons are a short record and a wet summer could change the balance. It cannot show the effect of weekends, which were never sampled. And it relies on predictors measured by other agencies, any of which could change instruments or stop reporting.

How the Model Will Be Kept Honest

If adopted, the model should be refit each winter with the latest season added and validated on the newest season before use, and its daily calls should be logged beside the next day's results so the department can see whether its performance holds. A predictive model is a claim about the future, and the program should test that claim every summer rather than trust the first fit.

Privacy and Data Custody

No individual data are involved; all records are environmental. The synthetic file in the appendix reproduces the structure of the real data so a reviewer can rerun the analysis without the department releasing its records before the board has seen them.

Summary

At the same false-advisory count, a morning model would have roughly doubled how often an over-standard day had its warning posted across the two seasons, with the gain concentrated at Mill Creek Beach and absent at Alder Point. The protocol should use the model where it helps and keep the current rule where it does not.

References

Francy, D. S. (2009). Use of predictive models and rapid methods to nowcast bacteria levels at coastal beaches. Aquatic Ecosystem Health and Management, 12(2), 177-182. https://doi.org/10.1080/14634980902905767

Nevers, M. B., & Whitman, R. L. (2011). Efficacy of monitoring and empirical predictive modeling at improving public health protection at Chicago beaches. Water Research, 45(4), 1659-1668. https://doi.org/10.1016/j.watres.2010.12.010

Whitman, R. L., & Nevers, M. B. (2003). Foreshore sand as a source of Escherichia coli in nearshore water of a Lake Michigan beach. Applied and Environmental Microbiology, 69(9), 5555-5562. https://doi.org/10.1128/AEM.69.9.5555-5562.2003