The 62% Agreement

No - it is not interpretable, and the number is not the only problem.

• Different measurement systems. IOA requires two observers to use the same response definition and the same measurement system across the same observation period, independently. Partial-interval recording and event recording answer different questions: partial interval credits an interval as 'occurred' if the behavior appears for any part of it, and it systematically overestimates relative to a duration measure and cannot be converted into a count. Dividing one total by the other compares two different quantities.
• Wrong agreement formula. Smaller/larger total (a total-count IOA) is only valid when both observers produced frequency counts.
• Small sample. One session tells you about that session, not about the reliability of the measurement system.

Fix. Retrain and recalibrate: agree on one operational definition with examples and non-examples, have both observers run a practice video to criterion, then collect live IOA using the identical system. If partial interval is retained, compute interval-by-interval agreement (agreements ÷ agreements + disagreements × 100). If frequency is retained, use total or exact-count IOA and keep session length fixed. Schedule IOA on at least 20–33% of sessions, aim for 80% or better, retrain whenever agreement falls below that, and stay alert to observer drift and reactivity to being watched.

Not on this evidence base. Three baseline points with a clean ascending trend (4.2 → 5.8 → 8.1) are unstable. If you start treatment on an ascending baseline, any subsequent change is ambiguous: a drop could be the intervention, a continued rise could be the pre-existing trend, and you will not be able to show experimental control - which is exactly what the funder is asking for.

Before anything else, validate the data. A trend that steep can be real, or it can be observer drift, a definition change, reactivity to being observed, or a recording error. Check who collected each session and how.

Then choose deliberately.

• If the behavior is not dangerous, extend baseline a few sessions to see whether a stable level emerges.
• If treatment cannot wait because of risk or authorization timelines, use a design that tolerates a trend - multiple baseline across settings or behaviors, or a changing-criterion design - and treat the projected trend as the comparison, documenting that rationale.
• Loop in the BCBA before the protocol starts, and get the design decision into the record, because it is a decision the funder may audit.

What I would not do is let 'it's getting worse' alone drive the start date. Urgency is a reason to check the data faster, not a reason to skip the check.

Collapse to one system. Pick the system that fits the question being asked - for stereotypy that occurs in bouts, duration-per-occurrence or partial interval is usually more informative than a raw count - and make everyone use it, including any new staff.

Standardize the mechanics. Fixed session length, fixed timer, fixed data sheet, same observer training video, and a written definition posted at the data station.

Build in routine agreement checks. A standing IOA schedule (e.g., every fourth session, plus one unannounced), computed on the same system, recorded in the same file as the primary data, and graphed.

Graph and review weekly. Plot trend lines, not just dots. Bring the graph to supervision so the conversation is about the pattern rather than the last number.

Document the retraining. When agreement falls below criterion, record what was retrained and when agreement recovered. That record is what makes the data defensible later.