The Use of Matrix Training to Teach Multi-Component Tacting Through Telehealth

What Matrix Training Is and Why It Buys You Free Learning

Children on the autism spectrum often have gaps in how language develops. Your time with each child is short, and staff hours are tight. Approved service hours run out. So the real job is pulling more skill out of every hour you get. Matrix training (a way to arrange teaching targets) was built for that job. It aims at generative responding (new answers you never taught), so some learning comes for free.

The core idea is simple. You take two parts of a skill and put one part on each side of a grid. Every box in the grid is a mix of those two parts. You teach only a few boxes, and most people teach the diagonal. Then you test all the boxes you never taught.

Say you work on noun and verb combos off the VB-MAPP (a common language skills test). Your nouns are cat, dog, and pig. Your verbs are sleeping, running, and jumping. You teach cat sleeping, dog running, and pig jumping until each one is mastered. Then you test the other six pairs with no teaching at all. When the child mixes the parts alone, that is recombinative generalization (mixing taught parts into new pairs). Nothing new got taught; the parts just moved around.

It involves systematically arranging and selecting multi-component instructional targets.

From the talk — Dr. Jessica Osos

Where Matrix Training Shows Up in the Research Base

Matrix training has a real base with autistic children, and most of that work sits in language. It covers listener skills (doing what someone asks you) and tacting (naming what you see). Color and shape pairs are one common set, and noun and verb pairs are another. Any skill built from two or more parts can fit the grid.

It has also been used for play and social skills. That list includes make believe play, pretend play, social play, and peer led work. School skills show up too, such as spelling, telling time, and building sentences.

One study sits under this whole talk. Frampton et al., from 2019, taught color and shape tacts with three by three grids. Colors ran down one side and shapes ran across the other. Those children already knew every color and every shape on its own, but not as a pair. The team also built generalization matrices (grids with no directly taught pair) to push past simple swapping. Teaching one diagonal led to high correct responding on untrained targets for three of the six children. The soft spot was those extra grids.

matrix training is an evidence-based approach that promotes generative responding.

From the talk — Dr. Jessica Osos

Bringing the Procedure to Telehealth: The Study Setup

COVID pushed telehealth into ABA fast, and much of the field has since moved back to in person work. Telehealth still earns its place. Some families simply want it. Some live far from a clinic. Others need the flexibility, or have health reasons to keep visitors out of the home. Before this study, no one had tested matrix training over telehealth.

So the plan was to copy the color and shape procedures and run them online. A second question sat underneath. Did the children learn the real skill of color shape tacting? Or did they just memorize the pairs they were taught?

Three young children on the autism spectrum took part, and each had to meet a few entry skills. They had to attend to materials on a computer or tablet. They had to echo and say at least two word phrases. They had to tact at least nine colors and nine shapes on their own. Sessions ran on Zoom from each child's home, ideally with few distractions. Every target was a plain colored shape on a blank white Google slide.

A caregiver sat in on every session. They ran the preference check, handed over the treat, and redirected the child when the session drifted. Each child worked with three by three grids, colors on one side and shapes on the other. Targets were assigned at random, so no two children had the same set. The design was a nonconcurrent multiple baseline across learners (a staggered single case design). The main measure was the percent of trials correct in each session.

So we're trying to see, can we take it a step further past that recombinant generalization?

From the talk — Dr. Jessica Osos

Inside the Phases: Baseline, Diagonal Training, Post-Test, Maintenance

Every session opened the same way. The caregiver ran a brief MSWO preference check (a quick sort of favorite items). The top three items were noted, and the top one was used through the session. If the child shifted, they moved to the second or third pick. One or two mastered tacts were placed at the start and sprinkled between teaching trials.

Baseline was bare on purpose. A slide came up, the question was what is it, and the child had five seconds. Right or wrong, nothing followed, and the next trial simply came up. Those sessions ran under a minute each.

Diagonal training used a test then teach format. The test part matched baseline but pulled only from the training grid. The teach part ran nine trials, with each of the three diagonal targets presented three times. A correct answer earned praise, and the caregiver delivered the treat. A miss or no answer triggered an error correction (a set way to fix mistakes). Mastery was two sessions in a row at 89 percent or better.

Post-test went back to baseline rules with no feedback at all. To pass, a child needed two sessions at 89 percent or better on each set of targets. That meant the training grid and both generalization grids. Children who fell short got booster sessions, which were just more diagonal training, and then ran post-test again. A maintenance check followed four weeks after the last post-test, in the same quiet format.

And parents weren't kind of working on color-shaped acting outside of research or during that four-week period.

From the talk — Dr. Jessica Osos

What the Data Showed Across the Three Learners

Baseline told a clear story. Two of the children named the shape and left the color out every time. The third had a few correct blips but stayed low overall. Hold onto that blip, because it matters later.

Diagonal training worked, and it worked fast. One child hit mastery in two sessions, and the other two needed three. All three picked up the taught pairs over a video call, with every bit of home noise in the mix.

Post-test split them apart. One child picked up all the untrained targets with no boosters at all. That is the clean result you hope for. The second child was variable. The third stayed low, even on the untaught boxes inside his own training grid, until a late jump. Both of those learners went back for boosters. The flat pattern fit an extinction effect, since post-test gave nothing after a correct answer. That learner even asked before sessions whether he had to say the color and the shape. The research rules meant the question could not be answered directly.

All three did reach mastery at post-test in the end. At the four week check, two children held their gains across every grid. The third held the training grid but lost both generalization grids. So maintenance held for two of the three learners.

all participants demonstrated some emergence of untrained targets

From the talk — Dr. Jessica Osos

Telehealth Details That Quietly Shape Your Data

Technology shows up in the data whether you plan for it or not. Slides skipped, and sometimes the child saw something other than what the clinician saw. The clearest example came from the variable learner. She kept calling yellow orange and pink purple. Her mother confirmed the colors did look that way on her end. The family moved her to a computer used for graphic design, and correct responding climbed back up. That was a screen problem wearing the costume of a skill deficit.

Home is a busy place. One learner usually had a sibling in the room. That pulled the caregiver's attention away, and sometimes pulled the child's attention too. That meant steady caregiver coaching on when to hand over the treat and how to hold the moment together.

You also give up physical control. You cannot prompt with your hands, and you cannot step in to block an error. The session runs on the relationship you built before you logged on.

The upside is real. Telehealth reaches families who might never get to a behavior analyst, and digital materials cut the prep work down hard. Physical cards for this study would have meant about 36 targets to build, print, cut, laminate, and cut again. That saving holds even if you never run a single session online.

So really making sure that you have a good rapport going into telehealth sessions will set you up for success.

From the talk — Dr. Jessica Osos

A Decision Model for Reading Post-Test Patterns

The Frampton et al. team built a simple decision model, and it is the most portable tool in this talk. You look at the shape of post-test responding, guess what it means, and pick your next move. Three patterns cover most cases.

First pattern: low responding across every grid, including boxes you taught. The likely cause is an extinction effect, since post-test pays nothing for correct answers. The move is more diagonal training, so those responses get harder to wipe out. That is exactly what the booster sessions did in this study.

Second pattern: high responding only on the diagonal targets you directly taught. Mixing has not happened yet, so the learner is holding three memorized pairs. The move is overlap training (teaching extra mixed pairs too), which goes past the diagonal.

Third pattern: high responding across the whole training grid but nothing in the generalization grids. Mixing did happen, but the responding is not fine grained enough to carry to new material. The move is more examples across grids. Teach a diagonal inside a second matrix, then test the last one and see if it finally comes free.

overlap training might be worth exploring for autistic individuals

From the talk — Dr. Jessica Osos

How to Run Matrix Training on Your Own Caseload

Start with components the learner already has alone. In this study every child could tact all the colors and all the shapes on their own. They just had never done it as a pair. Overlap training is the better bet when a grid holds parts the learner does not know yet. Never build a grid from two unknown components. One unknown component can work if you know your learner well.

That opens a nice clinical use. A learner working on pronouns who already tacts actions can get pronouns crossed with actions in a grid. You skip the odd rote drills and keep sentence length matched to real skill. If it is not working, change what you are doing.

You can go past two components. Picture a cube instead of a grid. Published work has crossed object, preposition, and location for phrases like an item on a rod. A play study crossed location, character, and object. It then used video modeling (learning by watching a video) to teach an action plus a phrase in each setting. Two component grids are much better studied, so treat three as the stretch option.

Size is a choice too. This study used three separate three by three grids, which follows the original procedures. One nine by nine grid would create far more untaught combinations to test. That gives a sharper read on whether a real skill actually formed. Also be honest about the limits: no one knows which learner profiles predict success here. The boundaries are unstudied, so try it, watch the data, and use the decision model when the pattern turns odd.

diagonal training is going to be most productive or fruitful for targets with known components

From the talk — Dr. Jessica Osos