by Emily Oster
Nothing in the data will answer these questions. But we can make more progress if we diversify our approach.
In the field of statistics, there are at least two broad approaches. The first is “frequentist statistics,” which approaches learning about relationships in data using only the data we have. The second is “Bayesian statistics,” which tries to learn about relationships by starting with a prior belief about the truth, and using data to update it.
To give an example in this context, let’s say we have a well-run study that showed that kids who watch SpongeBob SquarePants are much more likely to be able to read at the age of two, and that this is the only study on this topic. In the world of frequentist statistics, you’d be forced to conclude that SpongeBob is a great learning tool.
For a Bayesian, this conclusion is less clear. Before seeing the data, we are very unlikely to think that SpongeBob can teach two-year-olds to read. Observing the data should make us more likely to think this relationship is real, but if we start out very skeptical, we should remain quite skeptical even after seeing the data.
A Bayesian approach is to think about how to incorporate other things you know—or think you know—about the world into your conclusion along with the data.
Why is this relevant here? I think we have some prior beliefs on this topic. There are only thirteen or so waking hours in the day for kids. If they spend eight of those hours watching TV, there is not enough time to do pretty much anything else. It seems very unlikely that this won’t have some negative impacts.
On the other hand, it is hard to imagine that watching an hour a week of Sesame Street or Dora the Explorer will lower your child’s IQ, or have much of any effect on them in the long run.
You can subject the iPad to similar logic. A two-year-old who is on an iPad all day: likely bad. A half hour of math games twice a week: probably not bad.
When we start from this point, the data—while sparse—looks a lot more helpful, since it actually provides a lot of information about precisely the things we have less intuition about (what’s known in the Bayesian approach as “having a weaker prior”).
For example, I don’t have much intuition about whether young children can learn from videos. The data—which indicates they cannot—is therefore very informative and useful. Similarly, although I have a good sense that watching eight hours of TV a day is bad and an hour a week is fine, I have less intuition about “normal” watching—in the realm of, say, two hours per day. For this question, Jesse’s work is quite informative, since it looks at precisely this magnitude of exposure and shows there is no impact.
If I want to map out the whole relationship between test scores and any amount of TV, I am still not done, but I can start to use the combination of my priors—my beliefs before I saw the data—and what we do see from the data to fill in where I was most uncertain.
This starts to give us a sense, as well, of where more studies might be most useful. Many kids use apps on iPads or tablets for some time every day. We basically have no research on this, and it’s not something about which one is likely to have very good intuition. I could believe that this is good—there are many very neat apps for math and reading. I could also believe it is bad—you’re not really learning, you’re just tapping around.
Finally, our intuitions should be informed by the economic idea of “opportunity cost of time.” If a child is watching TV, they are not doing something else. Depending on what that “something else” is, TV watching may be better or worse. Many studies of this emphasize that (for example) your kid can learn letters or vocabulary from Sesame Street, but they are better at learning those things from you. That’s almost certainly true, but it is less obvious to me that this is the alternative. Many parents use TV to take a break, get their breath, make a meal, do some laundry. If the alternative to an hour of TV is a frantic and unhappy parent yelling at their kid for an hour, there is good reason to think the TV might actually be better.
The Bottom Line
Your zero- to two-year-old cannot learn from TV.
A three- to five-year-old can learn from TV.
It is worth paying attention to what they are watching.
The evidence is sparse overall. When in doubt, use your “Bayesian priors” to complement the data.
15
Slow Talking, Fast Talking: Language Development
When I was twenty-two months old, my parents (both also economists—I know, I know) were at a cocktail party, and my mother struck up a conversation with a visiting professor, Katherine Nelson. Her field was child language development, and my mother mentioned that she had a daughter (me) who talked a lot, especially alone in her crib prior to falling asleep. Professor Nelson got very excited and asked if my mother would be willing to consider recording my crib speech, for research. Indeed, she would.
For the next eighteen months or so, my parents recorded me most nights and provided the tapes to Professor Nelson and her team of researchers. Early on, my mother transcribed many of the tapes to try to make sense of my poor diction. This large corpus of tapes and text—some of it of me talking alone, some talking with my parents—provided a trove of data for researchers studying how children acquire language. They were interested in questions like, does the concept of the future develop for kids before the concept of the past? There were papers, academic conferences, and eventually a book of collected research papers on the tapes.
(The irony of both being the subject of a book like this and also writing one is not lost on me.)
This book—Narratives from the Crib1—came out when I was about nine. I have a vivid memory of coming home from school and finding an advance copy on the table in the porch. I opened it, eager for some insights into my younger self, but sadly found it somewhat lacking in that department. This was a dry academic book—a set of papers written by linguists analyzing verb form and sentence structure. I remember reading some of the funnier quotes from me and putting it aside.
I didn’t really look at the book again until Penelope was getting to the same age. And this time it was in service of the perennial parental neurosis: comparing your child to others. I scoured the book to try to figure out how Penelope and I compared. The earliest quote in the book is, “When Daddy comes I put that there then eat my breakfast and Daddy make my bed,” from twenty-two months and five days. Was Penelope saying things like that at a similar age? It was hard to tell—I pushed my mother: “Did I really say that, or was that just what you thought I said?” Needless to say, she could not remember. (Or so she claimed.)
Communicating with one another—talking, signing, writing—is among the things that make us most human. The moment your child stops having to cry and point desperately at the refrigerator and can instead say, “Milk, please” (or even just “MILK!!”) is one in which you can start to see glimmers of a person in there. We usually remember our children’s first words (Penelope: “shoes”; Finn: “Penelope [Puh-Puh]”), and early on many of us will probably admit to counting just how many they have.
Talking is also a natural point of comparison—of your children to other children, of your children to each other, and (in my case) of your children to yourself. I was warned before I had Finn that this problem is especially acute if you have a daughter first, followed by a son.
“Boys are slower with language,” warned my more delicate friends. Some less delicate ones said, “You’ll think your son is stupid.” People whose children were born in the opposite gender order told me how brilliant they thought their daughter was.
Figuring out how your child compares with others is not, in fact, straightforward. As with physical milestones, doctors tend to focus on identifying children for early intervention. At the two-year-old doctor visit, it is common to be asked whether the child has at least twenty-five words they say regularly. At fewer than this, it may be appropriate to bring in some ou
tside help to figure out what is wrong. But this is a cutoff to indicate a problem, not a measure of the average or anything about the range. The average child has more than twenty-five words at age two. But how many more?
Most pediatrics books have similar approaches—they warn you when to be concerned, but don’t give a sense of the full distribution.
Even with the full distribution, there are other questions: Does it matter? Is talking early a marker of anything later? Both of these questions have answers—the first a bit more satisfying than the second—we just have to go to the data.
THE DISTRIBUTION OF WORDS
In principle, it seems like it would be straightforward to collect data on how many words children say. Specifically, you could just count them. And it’s true that when a child is very small—when they have five or ten or twenty words—probably parents could remember most of them if asked. But this procedure can break down as children talk more and more. Let’s say your child says four hundred words, some of them used frequently and some infrequently. Will you really remember them all?
A related problem in comparisons is how to count words that are specific to your child. For example: At just over two, Finn became obsessed with a song entitled “Bumblebee Variety Show,” written by the local Music Together instructor, Jen. We played it on repeat every time we were in the car. He liked to sing it loudly—in the car with the music, in his crib, in the bath.
The primary lyrics in this song are “Bumblebee variety show.” Technically, then, he could say this, although he pronounced it as one word: bumblebeevarietyshow. So: When counting words, should I think of him as knowing the word variety? He certainly would not use it in a sentence, nor did he think of it as a separate word. So, probably not. But then should I count bumblebeevarietyshow as a single word? This seems more plausible. But still, it’s not even clear he thought of this as a word as opposed to just a noise. Also, it is actually not a word.
Researchers get around both of these problems—recall and the comparison set—by using a standardized measure of vocabulary size from a consistently used survey. The commonly used one is the MacArthur-Bates Communicative Development Inventory (MB-CDI).
The MB-CDI is administered to parents (Want to do it yourself? See the endnotes).2 The vocabulary portion lists 680 words in various categories—animal sounds, action words (bite, cry), body parts, etc. Parents check off all the words they have heard their child say, giving them a count of vocabulary size on these words.
For kids above sixteen months, the survey uses words and sentences; for those younger than that, there is a separate form for words and gestures.
This approach to vocabulary size works well for two reasons. First, by listing the words and asking about them rather than asking parents to remember, parents are less likely to forget words. I may not be able to recall without prompting that my son knew the word shovel, but once it is mentioned, I may remember an incident in which he asked for one. Second, by looking at the same words for every kid, it is much easier to compare across children.
An obvious downside to this approach is that it will understate speaking ability for children who know a lot of unusual words but miss some common ones. For example, one of the words on the list is Coke; if your children do not drink soda, they may not know this word. Similarly, children in Hawaii may be less familiar with the word sled.
This problem is most acute as you get to ages where children know most of the words. It may not really be feasible to distinguish between a child who says 675 of the words and one who says 680. For children who know fewer words, these small differences balance out—one child knows sled, another knows beach.
Many people have completed this form. Much of this is in service of research. Some is in service of evaluating children for developmental delays or simply to satisfy curious parents. Regardless of the reason, the developers of this survey have a website where results can be uploaded. And from this, we can get a first answer to the question of the distribution of words. The graph below was created out of their data—the horizontal axis is the age, and the vertical axis is the count of words as scored in the survey.
The lines in the graph show “quantiles”—basically, the distribution of words at each age. Take, for example, age 24 months. This data says that the average child—that’s the 50th percentile line—at 24 months has about 300 words. A child at the 10th percentile—so, near the bottom of the distribution—has only about 75 words. On the other end, a child at the 90th percentile has close to 550 words.
For younger children, these surveys and data focus on both words and gestures (i.e., signs). The graph on this page shows similar data for children aged eight to eighteen months on this metric. One main takeaway from these graphs is the explosion of language after fourteen or sixteen months. Even the most advanced one-year-old has only a few words. At eight months, virtually no children have any words or gestures.
I was interested to note this, given my mother-in-law’s continual insistence that Jesse said the word fishy at six months.
The website for this data is publicly accessible3 and has the capacity to make all sorts of graphs—they can show you the data broken down by parental education or birth order (later children talk more slowly), for example, and they have similar data for other languages and for counts of words children understand in addition to being able to speak. It is worth noting here that kids who are bilingual—that is, their parents or caregivers speak to them in two different languages—tend to be slower to talk, although when they do, they can speak both languages.
Perhaps the most interesting of these splits is by gender, given the general impression that boys develop more slowly. This is, indeed, borne out in the data. The graphs on this page separate out boys and girls, and we can see that boys have fewer words at all points in the distribution. At twenty-four months, for example, the average girl has about fifty more words than the average boy. By thirty months, the most advanced boys and girls are similar, but there are still large differences at other points in the distribution.
This data provides some useful norming, but it is important to be cautious about where it comes from. It is not (for the most part) nationally representative data. There are many more parents with college or graduate degrees in these data points than you would see in the overall population. This means these figures are likely to overstate the average among all children. Having said that, they give you something beyond a general guideline about when to be worried, and also provide reassurance that there is a significant range in this distribution at all young ages.
DOES IT MATTER ANYWAY?
We all enjoy navel-gazing about our own children, so knowing where your child falls in this distribution may be simply a fun fact. But virtually everyone learns to talk. It is natural to wonder, though, whether these early differences do predict any long-term differences. Do children who learn to talk earlier learn to read earlier? Do they do better in school later?
There are certainly counterexamples to this idea—stories of extremely precocious children who didn’t talk until very late but were reading at eighteen months. And there are also supportive stories: early talkers who also turned out to be unusual in other ways. But examples like this, in either direction, do not tell us anything about the relationship on average.
To echo a refrain from throughout this book, this is difficult to learn about given other relationships in the data. Language development is clearly associated with parental education. But parental education is also associated with many other outcomes, including early reading and later test scores. What we’d really like to ask is whether early language development is a marker of later things, conditional on what we know about the parents. But our information about parents in the data is likely to be incomplete. As a result, the studies that I’ll talk about are likely to overstate the relationship between early talking and later outcomes.
There are basically two questions y
ou could ask here: Can you take anything from your child being either a very strong early talker or a very delayed one? And, assuming your child is in the middle of the distribution, does it matter where they are? Are there any later-life differences between a two-year-old who is at the 25th percentile of the distribution versus the 50th, or the 75th?
The largest and most rigorous studies of this focus on whether children who are abnormally late talkers are also delayed in other ways later.
In a series of studies, a researcher named Leslie Rescorla recruited a set of thirty-two delayed talkers from twenty-four to thirty-one months old.4 The children in this delayed cohort—nearly all boys—had an average of twenty-one words at this age. Based on the previous graphs, this is way below average. She recruited a sample of comparison children with similar characteristics but with normally developing language skills.
Notably, this study followed the children—or at least most of them—to much later ages, up to age seventeen. At older ages, researchers looked at verbal abilities, test scores, and similar outcomes.5
The results provide a mixed bag of evidence. On one hand, the group with delays in talking did seem to have slightly worse outcomes on the tests even much later. Their IQ scores at age seventeen were lower than the comparison group. On the other hand, these children were not especially likely to score very poorly—none of them scored in the bottom 10 percent on IQ tests at seventeen, for example, despite having been in the bottom 10 percent of talkers.
This basic result—that there is a correlation, but the predictive power is limited—is consistent across many studies. Some of the studies are much bigger. For example, a paper reporting on six thousand children in the Early Childhood Longitudinal Study found that limited vocabulary at twenty-four months predicts verbal skills through the age of five, but again, most of the children were well in the normal ranges later.6