What happens when you teach children to make inferences while reading?

Once children are fluent decoders, the most frequent problem in reading is poor comprehension due to a failure to make inferences. Even seemingly straightforward anaphoric inferences can elude these students: they might read “Bob gave Tamisa some of his snack because she was hungry” and still be unsure of the referent for “she.” Other inferences require bridging information from long term memory, and these are still more challenging. For example, “Kevin said he was cold. Zeke gave him his coat.” Even if these two sentences are understood, each on its own, deeper comprehension entails making the (probably accurate) inference that Zeke gave Kevin the coat because Kevin said he was cold, which requires knowing that putting on a coat is something one does when cold.

Educators have sought to improve inferencing. In some cases they can teach students reading comprehension strategies: create a summary, for example, or create a graphic organizer. The task provides some structure that will prompt the student to make the necessary inferences. Alternatively, students might be taught more directly to make inferences, usually by instruction to elaborate on what they read in the text,  to use cues in the text that provide clues to inferencing (e.g., words like “because,” or “so,”), and to monitor their comprehension.

​A new meta-analysis of inference instruction (Elleman, 2017) shows that it’s quite effective, but carries some important caveats…ones that I’ve mentioned before.

The overall mean effect size on comprehension of inference instruction was a healthy g = .58. Elleman also evaluated separately comprehension due to inferences, and literal comprehension, and observed a difference moderated by ability. 

As the figure shows, both skilled and less skilled readers improve in inference-making ability, but inference instruction provides a boost to understanding of things stated explicitly in the text only for less skilled readers.
​
Still more interesting to me was the report that the amount of instructional time had no impact on the effectiveness of the intervention, which I’ve shown in the graph below, compiled from a data table in the paper.

Each dot represents a study condition. Two things are notable: first, most studies entail rather little practice, but nevertheless show a large effect. Second, more practice does not lead to a larger effect.

​This finding is important because it provides an important clue to the mechanism by which this instruction helps. Practice usually helps, especially early in training. The curve looks like this

Performance improves with practice, and the curve is negatively accelerating. You get the most bang for the buck from practice early in training. The data from this meta-analysis don’t show either effect. It’s true that the range is relatively small…most studies use very little practice, so it’s harder to observe any effect of practice. That still doesn’t explain why you get such a big effect with very little practice.

​This failure to observe a practice effect is more understandable if the relationship of practice and performance for inference generation look more like this: 

The plot would look like this if what kids learn during inference instruction is easy to learn, and easy to implement, but carries a one-time benefit. For example, this instruction might prompt children to better understand the importance of making the effort to coordinate meaning across sentences. It might teach them to look for cues to inference possibilities like the word “because,” or time cues like “later.”

As Elleman notes, theories of reading have centered on the knowledge from long-term memory needed to make inferences (Kintsch, 1998) and/or the working memory capacity needed to hold information in mind simultaneously so meanings can be compared (Daneman & Carpenter, 1980). Neither of these is going to change over the course of 10 hours of instruction.

And that’s the practical implication of this meta-analysis. There’s a big benefit to inference instruction, but all of the benefit accrues rapidly. There seems to be no point in spending extended classroom time on the practice.

Similar results have been observed in meta-analyses of studies examining the impact of reading comprehension strategy instruction. Gail Lovette and I (2014) reported on nine meta-analyses of typically developing readers and of readers either identified with a reading disability, or at risk. In each case, the pattern was the same: a large effect with few hours of instruction, and no benefit to more instruction. (This piece was published as a commentary in Teachers College Record and seems to have disappeared from the website. Email me if you’d like a copy.)

This interpretation is also consistent with theories of reading comprehension. Inferences are situation specific: you can’t really teach how to make inferences because the inference to be made depends on the content of the text. Rather, you can teach them to seek and use some cues (probably a causal connection here) and you can teach them that it’s important to make inferences. But making them requires broad background knowledge in long term memory.

That is an essential goal to improve reading comprehension, but it is a goal requiring years of planning, not hours.