Six files, 209,321 programs, three join keys: turning PPD:2026 into answers your provost actually asks

The question comes in a Monday email. "How many of our programs would fail STATS?" One sentence. Forty hours of analyst time, minimum, if it's the first time you've opened PPD:2026.

PPD:2026 is the Program Performance Data release the U.S. Department of Education published in early 2026 to inform the AHEAD negotiated rulemaking. It covers 209,321 unique programs across 5,096 institutions and roughly 20.6 million Title IV students. It is, on paper, an exceptional dataset, the closest thing IR teams have ever had to a complete program-level outcomes file with federal earnings benchmarks baked in.

On disk, it is six separate Excel files. They share three join keys, but those keys have to be exactly right or you silently lose rows. They use two different earnings horizons, two different cohort definitions, two different dollar-year bases for debt versus earnings, and a 4-digit CIP code while the actual STATS test will eventually operate at 6-digit CIP. Every time the provost rephrases the question (different filter, different scenario, different peer comparison), the wrangling starts over.

This piece walks through where the hours actually go. If you have an analyst running PPD:2026 today, this is what their week looks like. If you're looking for the strategic frame instead (which programs to triage, what the 50% institutional trigger is), start with the exposure-map playbook and the 50% institutional trigger explainer.

Every file in PPD:2026 keys on the same three columns: opeid6 (six-digit institutional OPEID), credlev (credential level: certificate, associate, bachelor's, master's, doctoral), and cip4 (the four-digit Classification of Instructional Programs code). All six files join one-to-one on the combination.

The "one-to-one" is the part that fools people. If the join is on opeid6 alone, multiple credential-level rows collapse incorrectly and a single program inherits another program's earnings. If it's on opeid6 + cip4 only, a certificate and an associate's degree in the same field merge into one row. If credlev is coded inconsistently between files, and in user-uploaded crosswalks it often is, entire programs drop silently.

The Earnings Premium test is one formula, graduate earnings minus a benchmark, but the benchmark isn't one number. PPD:2026 carries six different benchmark columns, and the correct one depends on credential level, in-state vs national, field of study, and which version of the test you're running (2023 FVT/GE or OBBBA-aligned). Picking the wrong column quietly inflates or deflates the failure rate. The OBBB-to-Earnings-Premium math walkthrough covers the statutory side of why these benchmark columns exist; this section covers the operational side of picking the right one.

PPD:2026 carries two different earnings variables, and they produce materially different pass/fail results. md_earn_ne_p3_1516 is median earnings of all completers, working and non-working, measured three tax years after exit, pooled across the 2014–15 and 2015–16 cohorts. md_earn_wne_p4 is median earnings of working completers only, measured four tax years after exit, pooled across the 2017–18 and 2018–19 cohorts.

The 2023 FVT/GE test uses the first. STATS uses the second. The two are not interchangeable. Excluding non-working completers raises the median (people earning $0 are dropped). Adding a year of career development raises it again. The combined effect, depending on the field, is a 10 to 25 percent lift in measured graduate earnings. Programs that fail under the 2023 definition routinely pass under the STATS definition, and vice versa.

The chart is illustrative (the exact lift varies by field and state), but the point is structural. If your analyst pulls the wrong earnings variable, the answer to "how many of our programs would fail" can be off by a factor of two. PPD:2026 includes both ge_overall_2023_fail (combined 2023 FVT/GE failure) and ge_overall_algn_fail (OBBBA-aligned combined failure) precisely so the dataset can show the policy shift side by side. Comparing the two for your own programs is one of the most useful analyses the dataset enables. It is also one of the easiest to get wrong if the join keys or benchmark columns are off.

PPD:2026 applies layered privacy protections that disproportionately suppress data for small certificate programs, the same programs most likely to fail the Earnings Premium test. If you don't account for the suppression rules explicitly, your at-risk count is structurally biased downward.

PPD:2026 reports at the 4-digit CIP level. ED has signaled that the actual STATS test, when results are published in February 2027, will operate at the more granular 6-digit CIP level. The difference matters more than it sounds.

Two programs that share a 4-digit CIP can land in different 6-digit codes, for example, two health-administration certificates with different specialisations. At CIP-4 they aggregate into a single PPD:2026 row; at CIP-6 they will be measured separately, with separate cohorts and separate Earnings Premium tests. A program that looks safe in PPD:2026 because it's averaged with a stronger sibling may fail on its own at CIP-6. A program that looks at-risk in PPD:2026 may be dragged down by a sibling and pass on its own.

This is not a flaw in PPD:2026. ED was explicit that the dataset is informational and not the final implementation file. But it means any analysis built strictly on PPD:2026 has a known accuracy ceiling, and the CIP-4-to-CIP-6 unrolling is one of the things institutions will have to do themselves as the official February 2027 publication approaches.

In AIR's 2023 survey on FVT/GE implementation, 72 percent of responding members reported their IR or IE office held institutional responsibility for the compliance reporting. The widely-cited pain points were tight timelines and a lack of clear communication from ED's FVT/GE guidance. Both of those land directly on PPD:2026 today. (For the policy lineage from FVT/GE into STATS, see the STATS vs FVT/GE vs PPD explainer. For the final FVT/GE submission timeline, see the FVT/GE final reporting year guide. For how the sector is asking the Department to soften this in the final rule, see our analysis of all 8,796 NPRM public comments.)

The wrangling pattern this piece walks through (six-file join, benchmark-column selection, definition-swap awareness, suppression handling, CIP-4-to-CIP-6 unrolling) is not a one-time setup. It repeats every time the provost asks a slightly different question. Filter to programs in our college of allied health. Compare our cosmetology mix to community-college peers in our state. Model what happens if median earnings drop 5 percent. Each rephrasing rebuilds the join, re-selects the benchmark, re-applies the suppression rules. Each rephrasing is most of a week.

That recurring weekly cost is what burns out IR analysts and what makes the question feel impossible at the cabinet level. The work is not hard. It is just slow, repetitive, and unforgiving of small errors in the join.

Clema's STATS (FVT/GE) AI Agent reads PPD:2026 directly. The join keys, the benchmark matrix, the earnings-definition pair, the suppression rules, the CIP-4 caveat: all of it is built into how the agent answers questions, not into a query the analyst has to write.

You ask: how many of our programs fail the OBBBA-aligned test. The agent joins the right files on opeid6 + credlev + cip4, picks the composite benchmark column, applies fail_obbb_cip2_wageb, returns the count with the underlying program list. You ask the same question against the 2023 FVT/GE methodology to see the policy-shift delta; the agent swaps to ge_overall_2023_fail and re-runs. You ask for the rollup by college, then by department, then by Title IV concentration. Three questions, three answers, fifteen minutes total. The wrangling pattern is constant and the agent has it memorised.

That is the actual difference. Not faster joining. Not better visualisation. The conversational frame (provost asks, analyst asks the agent, analyst answers) instead of the project frame. Once the data is approachable in conversation, the program-level exposure-map playbook and the 50% institutional trigger become questions you can actually answer this quarter, not next year.

AIR, Program Accountability Reporting resource center. U.S. Department of Education, Federal Student Aid Partners, FVT/GE knowledge center. FVT/GE Frequently Asked Questions.