Screening for autism: agreement with diagnosis.

Doctors gave the toddlers two quick parent forms: the M-CHAT and the SCQ. All kids already had an autism diagnosis from a full team. The goal was to see if the short forms agreed with the full work-up.

Parents filled the forms while waiting at a hospital clinic. Staff then compared each form's score to the child's official diagnosis.

Both forms caught about two-thirds of the kids who truly had autism. They missed one-third, mostly higher-functioning children who spoke in sentences.

Strangely, parents who spoke English as a second language had slightly better accuracy. The authors guess these parents may have under-reported mild quirks, leading to clearer red flags when they did report them.

Jones et al. (2007) ran almost the same check and got much higher hit rates—up to a large share sensitivity. The difference: they only tested kids already sent to a developmental clinic, so the sample was richer in obvious cases. Same tool, different pool, different result.

Marsack-Topolewski et al. (2025) leapfrogs both papers. Their 2025 machine-learning ensemble reaches 97-a large share accuracy across toddlers to adults, blowing past the 63-a large share ceiling of M-CHAT/SCQ. The new tool still needs a laptop, but it shows the old forms are now outdated.

Bong et al. (2021) landed in the middle: adding a 15-minute play observation to a short interview pushed sensitivity to 85-a large share. Their message agrees with C et al.—paper alone under-detects—but shows you can fix the gap by watching the child.

If you screen with only the M-CHAT or SCQ, plan on missing up to one-third of higher-functioning kids. Use the forms to trigger a full assessment, never to rule autism out. For tighter accuracy, add a brief observation module or move to newer digital tools.