Two Hundred Excluded Days, and the Year Still Meets the Completeness Requirement

One Day, Two Official Numbers

A fine particle monitor produces one number for a calendar day. Appendix N to 40 CFR part 50 defines it: daily values are the 24-hour average concentrations of PM2.5 mass "from midnight to midnight (local standard time) from suitable monitors." One instrument, one day, one figure.

That figure is then read twice, by two parts of the Code of Federal Regulations, on two very different clocks. The first reading happens within the hour. Appendix G to 40 CFR part 58 converts pollutant concentrations "into a uniform scale from 0-500," and its breakpoint table assigns a 24-hour PM2.5 average of 125.5 to 225.4 µg/m³ the index range 201 to 300, "Very Unhealthy," and anything from 225.5 up the range 301 and above, "Hazardous." The index "must be reported daily for a metropolitan statistical area (MSA) with a population over 350,000," seven days per week.

The second reading takes three years. Under 40 CFR 50.20, the primary annual PM2.5 standard "is met when the annual arithmetic mean concentration, as determined in accordance with appendix N to this part, is less than or equal to 9.0 µg/m³," and the 24-hour standard is met when "the 98th percentile 24-hour concentration" is at or below 35 µg/m³. Those statistics are design values.

The interesting part is not that a smoke day can be taken out of the second reading. It is what "taken out" turns out to mean. The day is not deleted, the year does not become incomplete, and the count that decides which value is the 98th percentile does not shrink. The day acquires a code.

Four Codes, One Monitor Year

EPA's Air Quality System stores the answer to "what was the 98th percentile at this site in this year" more than once. The EPA Air Quality System (AQS) Data Dictionary defines a field called Exceptional Data Type Identifier as "a designation indicating how a summary value is affected by exceptional events," with four allowed values: 0, No Events; 1, All Events Excluded; 2, All Events Included; 5, Concurred Events Excluded.

The companion field, Event Type, says what separates the last two: "Events Excluded means that events occurred but data form them is excluded from the summary. Concurred Events Excluded means that events occurred but only EPA concurred exclusions are removed from the summary."

So for a monitor year touched by a wildfire, three of those codes can carry three different numbers. Code 2 is the year as measured. Code 1 is the year with every flagged day pulled out, whether or not EPA ever agreed. Code 5 is the year with only the agreed exclusions gone. The regulatory determination runs on code 5; a state's own flagging produces code 1, which cannot carry any of the determinations in section 50.14's scope paragraph, because under paragraph (c)(2)(ii) the data stay in until the Administrator concurs.

EPA's public tools expose the same fork. The AirData documentation tells users that a report column "indicates whether there were any flagged data ('NONE'), whether flagged data were included ('INCLUDED') or excluded ('EXCLUDED')," and that "you can select whether you want to include or exclude exceptional events data." The choice is a dropdown, and either number is official.

The Chain

The distance between code 2 and code 5 is a procedure. 40 CFR 50.14, Treatment of air quality monitoring data influenced by exceptional events sets it out.

Scope first. The section applies to "data showing exceedances or violations of any national ambient air quality standard" for a closed list of determinations: designation and redesignation, a nonattainment classification, attainment by a deadline, an attainment date extension, a finding that a state implementation plan is inadequate, and "other actions on a case-by-case basis as determined by the Administrator."

Then the flag. A state gives notice of intent to request exclusion "by creating an initial event description and flagging the associated data that have been submitted to the AQS database," after which the state and the EPA regional office "shall engage in regular communications to identify those data that have been potentially influenced by an exceptional event."

Then the hard stop. Paragraph (c)(2)(ii) keeps a flag from being an exclusion: "The data shall not be excluded from determinations with respect to exceedances or violations of the national ambient air quality standards unless and until, following the State's submittal of its demonstration pursuant to paragraph (c)(3) of this section and the Administrator's review, the Administrator notifies the State of its concurrence by placing a concurrence flag in the appropriate field for the data record in the AQS database." The exclusion is a field in a database, written by the Administrator.

Then the demonstration, filed "after notice and opportunity for public comment," which must establish "a clear causal relationship between the specific event and the monitored exceedance or violation," that "the event was both not reasonably controllable and not reasonably preventable," and that it "was a human activity that is unlikely to recur at a particular location or was a natural event."

Some event types get a presumption on one element. For wildfires, "provided the Administrator determines that there is no compelling evidence to the contrary in the record, the Administrator will determine every wildfire occurring predominantly on wildland to have met the requirements identified in paragraph (c)(3)(iv)(D)." For high wind dust, EPA "will accept a high wind threshold of a sustained wind of 25 mph" in thirteen named states, again "provided this value is not contradicted by evidence in the record at the time the State submits a demonstration." Both provisos do real work: the presumption is rebuttable, and covers one element out of five.

The clocks sit outside the criteria. Under 40 CFR 58.15, Annual air monitoring data certification, the head official of the monitoring agency certifies that "the previous year of ambient concentration and quality assurance data are completely submitted to AQS," in a letter "due by May 1 of each year." Table 2 to section 50.14 then ties the event clock to the designation calendar: where state recommendations for a new or revised standard are due August through January, "the Initial Notification deadline will be the July 1 prior to the recommendation deadline."

The public number, meanwhile, was final on the day. Appendix G lets an agency discharge the whole daily duty by pushing raw data out: "submitting hourly data in real-time from appropriate monitors ... to the EPA's AirNow (or future analogous) system satisfies the daily AQI reporting requirement." The AirNow documentation says of that stream: "These data are not fully verified or validated and should be considered preliminary and subject to change," and "official regulatory air quality data must be obtained from EPA's Air Quality System (AQS or AirData)." We have followed this shape before, in the gap between a pressure event and the bacteria result that confirms it.

The Rank Selector Does Not Move

Now follow one number down. For the 24-hour standard the statistic is the annual 98th percentile, and appendix N to 40 CFR part 50, Interpretation of the NAAQS for PM2.5 does not interpolate it. It picks a rank out of a sorted list. Table 1 to the appendix maps the annual number of creditable samples to the rank that counts: 1 to 50 samples, the 1st maximum; 51 to 100, the 2nd; 101 to 150, the 3rd; on up to 351 to 366 samples, the 8th maximum.

That puts the weight on the definition of creditable samples, which does two jobs in one definition. Creditable samples "are daily values in the combined site record that are given credit for data completeness. The number of creditable samples (cn) for a given year also governs which value in the sorted series of daily values represents the 98th percentile for that year." One count decides whether the year is usable and which rank is the answer.

Take a continuous monitor reporting a value every day of a 365-day year. The count sits in the 351 to 366 band, so the 98th percentile is the 8th maximum. Now suppose EPA concurs on exclusion for the three highest days. The series used for the regulatory comparison holds 362 values. Does the count that chose the rank drop too?

Here it would be easy to declare the regulation silent. It is not silent. The appendix ties the capture rate to "valid reported sample values," and EPA guidance then says the rest outright, in a table cell. In Additional Methods, Determinations, and Analyses to Modify Air Quality Data Beyond Exceptional Events, the April 2019 memorandum from the director of EPA's Air Quality Assessment Division, Table 2 states: "Valid monitoring data are to be counted towards data completeness even if the data have been affected by an event and even if the data have been excluded from a design value or other calculation."

Counted toward completeness means counted as creditable. The count stays at 365, the rank stays at the 8th maximum, and the pool it is applied to has shrunk. The 8th maximum of the 362 retained values is the 11th-highest value the monitor actually measured. Excluding three days moves the percentile three places down the sorted list, not one, because the selector was set by a count the exclusion never touched.

Two Hundred Days at One Station

The scale this reaches is not hypothetical. In a July 2014 technical support document, EPA Region 9 reviewed Hawaii's demonstration for volcanic emissions at the Kona monitoring station, AQS site 15-001-1012, covering "2 days in 2011, 200 days in 2012, and 66 days in 2013" against the 2012 annual PM2.5 standard. EPA's finding: "based on the weight of evidence, EPA concurs that the exceedances were caused by volcanic exceptional events."

Two hundred days in one calendar year at one station. Set that against the completeness rule in appendix N: "A year meets data completeness requirements when quarterly data capture rates for all four quarters are at least 75 percent," where the capture rate is "the number of creditable samples for the quarter divided by the number of scheduled samples for the quarter."

If concurred exclusions cost a site its creditable samples, an exclusion that size would be self-defeating. Quarters would fall under 75 percent, years would go invalid, and the design value would collapse for want of data rather than come down. The 2019 table is what prevents that: the days leave the calculation and stay in the count. The regulatory statistic then rests on a year that is complete by rule and incomplete in fact, and the difference between those two descriptions is precisely the set of days someone demonstrated and EPA agreed to.

Where the Substitution Test Answers the Follow-Up

One more place the question could hide. Appendix N does not treat an incomplete year as fatal. Section 4.1(b) ends with a proviso: "Furthermore, where the explicit 75 percent data capture and/or 11 sample minimum requirements are not met, the 3-year annual PM2.5 NAAQS DV shall still be considered valid if it passes at least one of the two data substitution tests stipulated below." A design value at or below the standard can be validated by the maximum quarterly value test, which fills the holes with high values from the site's own record and checks whether the result still meets the standard. That route has a floor of its own. The substitution is "permitted only if there is at least 50 percent data capture in each quarter that is deficient of 75 percent data capture in each of the three years under consideration," and the appendix adds: "If any quarter has less than 50 percent data capture then this substitution test cannot be used."

The obvious worry is that an excluded smoke day could return as one of those substitutes, undoing the exclusion through the back door. The appendix answers this in two places. The maximum quarterly value data substitution test "substitutes actual 'high' reported daily PM2.5 values from the same site (specifically, the highest reported non-excluded quarterly value(s) (year non-specific) contained in the combined site record for the evaluated 3-year period) for missing daily values." Non-excluded. That is one lock, in a parenthesis in a definition list. The other sits in the operative text, where the procedure tells the agency to identify, for each deficient quarter, "the highest reported daily value for that quarter, excluding state-flagged data affected by exceptional events which have been approved for exclusion by the Administrator." The back door is shut in both places.

That is the reading habit this area demands. The rule text says exclusion happens; the definitions and the guidance tables say what exclusion does downstream. The same lesson turned up when we traced where an ambulance response interval starts, and again in the exclusion rules that let a utility's reliability average drop its worst days out of the annual figure.

Where This Doesn't Apply

None of this reaches the number the public saw. Appendix G to part 58 contains no exceptional events provision. Its only exceptions to the reporting duty are that a pollutant staying below index 50 for a season or year "may be excluded from the calculation of the AQI," that an area whose index values all stay below 50 for a year may report at its discretion, and that a real-time feed to AirNow satisfies the daily requirement. Naming the cause is optional: "causes for unusually high AQI values" sits among the items a report "may also contain" but is "not required" to.

The machinery also does not reach every value that drives an index. For particulate matter, appendix G allows that "PM measurements may be used from monitors that are not reference or equivalent methods," while appendix N builds design values only from suitable monitors under part 58. Some concentrations behind a hazardous-category day were never eligible for a design value, so there is nothing to exclude.

Section 50.14 reaches only data showing exceedances or violations, and only the determinations in its scope paragraph. Which standard is at issue decides what a day has to look like: paragraph (b)(7)(ii) directs the Administrator to accept a state's comparison of a 24-hour concentration to the level of a standard for the same pollutant "with a longer averaging period," and to do so "without the State having to demonstrate that the event caused the annual average concentration of the pollutant to exceed the level of the NAAQS with the longer averaging period." A day that is no 24-hour exceedance can still come out of an annual determination, which is the posture the Kona demonstration was in. Trends work, health research and emergency response use whichever code the analyst selects.

Finally, the arithmetic above is the 24-hour path. The Kona demonstration addressed the annual standard, where the statistic is an arithmetic mean and the rank table plays no part. Each pollutant has its own appendix and its own selection rules.

What the Codes Are For

The design value and the index are not competing estimates of one quantity. The index answers what the air was like where people were breathing it. The design value answers whether a control program is working, which is why the rule lets a state argue that a volcano says nothing about its controls. Both readings are defensible. Quoting one and describing it as the other is not.

So the check on any PM2.5 figure has three parts. Which code produced it, 0, 1, 2 or 5. Whether a flag in the record was ever concurred, because a flag alone changes nothing under paragraph (c)(2)(ii). And how many days went into the count as against how many went into the value, because appendix N uses the first number to choose a rank and the second to supply it. A year can be complete by rule with two hundred of its days pulled out of the answer, and the record says so plainly, as far down as the definitions.