Thomas Stephan Juzek

LIN 3042 Applied Linguistics · Fall 2026

Controlling for a confound

Why vitamin takers can look sicker, and what a study means when it says it “controlled for” something. With invented data to play with, then the real data of Romeo et al. (2018).

AI disclaimer. Content guided by Tommie Juzek; built upon his prompting by an AI assistant (Claude Opus 5.5, Anthropic); reviewed and approved by Tommie Juzek.

1 The puzzle

Imagine a study of 200 people aged 30 to 80. Each person either takes a daily multivitamin or does not, and each has a heart risk score: higher is worse. (All the vitamin numbers on this page are invented. What real studies found is in section 8.)

Takes vitamins No vitamins

Figure 1. Every dot is one (invented) person. The black bars are the averages.

The vitamin takers score 11.5 points worse on average. Do vitamins harm the heart? Look at who takes them: the takers are on average 64 years old, the non-takers 46.

Figure 2. Age pushes on both. The question mark is what the study wants to know.

Age is a confounder: it affects who takes vitamins (older people take them more often), and it affects the heart (older hearts are at more risk). So the vitamin group is also the older group, and part of the gap, possibly all of it, is the age gap.

A confounder needs both arrows. If older people were no likelier to take vitamins, or if age did nothing to the heart, age could not open a gap between the two groups.

Controlling for age means asking: among people of the same age, do vitamin takers do better or worse? There are three ways to see how that works. They are the same idea, and they give about the same answer.

2 Way one: compare like with like

The most direct way: cut the sample into age bands, and compare takers with non-takers only within each band. A 65-year-old who takes vitamins is now compared with a 65-year-old who does not.

Takes vitamins (average) No vitamins (average)

Figure 3. Left: everyone together, with the big gap. Right: the same people in five age bands. Within each band, the gap is small and goes either way.

Age bandPeopleTaking vitaminsRisk, takersRisk, non-takersGap

The big overall gap came from who was in each group: the vitamin group is full of older people. Within the bands, that difference is (mostly) gone, and so is the gap. This is what “holding age constant” means, done by hand.

Two problems with doing it this way:

Regression does the same job without cutting the data into bands.

3 Way two: one line per group (regression)

Put age along the bottom and heart risk up the side. Then draw a straight line through each group, with one rule: the two lines must be parallel.

Takes vitamins No vitamins

The slope of the lines says how much heart risk goes up with each year of age. The distance between the lines says how far apart takers and non-takers are at the same age, and because the lines are parallel, that distance is the same at every age. It is the “effect of vitamins, controlling for age”: here 0.0 points, against 11.5 when age is ignored.

Where did the 11.5 come from? The takers are 19 years older on average, and each year adds 0.6 points of risk: 19 × 0.6 = 11.4. Age explains almost the whole gap.

How does the computer find the lines? Of all possible pairs of parallel lines, it keeps the pair that the dots are, all in all, closest to: the smallest sum of squared up-and-down distances (“least squares”). A formula gets there in one step.

The equation, for the curious

risk = b0 + b1 × vitamins + b2 × age

  • vitamins is 1 for a taker and 0 for a non-taker. So for takers the line is lifted by b1; for non-takers it is not.
  • b1 (here 0.0) is the distance between the two lines: the difference between takers and non-takers of the same age.
  • b2 (here 0.6) is the slope of both lines: how much risk rises with each year of age, among people with the same vitamin habit.
  • b0 is where the non-takers’ line would meet age zero. Nobody here is zero, so it means nothing on its own.

A common misreading: “b2 is the effect of age, ignoring vitamins.” No: each number is worked out with the other one held fixed. Ignoring vitamins would be a different regression, with age alone.

Why parallel? Because the question is one number, the gap at a given age. If you think the gap itself changes with age (vitamins help the old but not the young, say), you can let the slopes differ. That is called an interaction, and it answers a different question.

4 Way three: what “partialling out” means

Papers say an effect held “after partialling out” something, or “independent of” it, or “over and above” it. All of these name one recipe in three steps. It gives exactly the same number as the parallel lines.

Takes vitamins No vitamins

After steps A and B, age has had its say twice: nothing in the leftovers can be predicted from age any more. So whatever link is left between the two leftovers cannot come from age. Here the cloud in step C is flat: once age is out, vitamins and heart risk have nothing to do with each other.

The slope in step C is 0.0; the distance between the parallel lines in way two was 0.0. Same number, and not by luck: it is always so (a result known as the Frisch–Waugh–Lovell theorem). With several confounders, the recipe is the same: in steps A and B, let all of them make their prediction together.

Keep these steps in mind: they are exactly what Romeo et al. did, in section 6.

5 Try it

Change the invented world, draw a new sample, and compare the two answers with the truth, which here you set yourself.

What vitamins really do

Comparison ignoring age

Comparison controlling for age

Takes vitamins No vitamins

Figure 6. Arrow thickness follows your settings. The lines are the parallel lines of way two; the numbers are in heart risk points, plus means worse.

Things to try

  1. Set either age slider to zero. Do the two comparisons still disagree?
  2. Make vitamins really protect (set them to −5) and push both age sliders up. What does the comparison that ignores age say?
  3. Tick health-consciousness. Does controlling for age still find the truth?
  4. Push “older people take vitamins more” all the way up and press New sample a few times. What happens to the controlled comparison?
  5. Set the study to 50 people and press New sample a few times.
What you should see
  1. No: with either arrow gone, age is no longer a confounder, and both comparisons land near the truth. A confounder needs both arrows.
  2. Vitamins look harmful, although they help: the age gap swamps the real effect and flips its sign. Controlling for age gets it back.
  3. No. Controlling for age removes age, and only age. The health-conscious bias stays in, and vitamins look protective although they do nothing. You can only control for what you measured.
  4. It jumps about. Now nearly all old people take vitamins and nearly all young people do not, so there are hardly any people of the same age in both groups to compare. The comparison rests on the few who are left, and on the straight line.
  5. Both comparisons wobble from sample to sample. A small study can miss a real effect, or show one that is not there.

6 The real thing: Romeo et al. (2018)

Romeo and colleagues recorded two days of language at home for 36 children aged 4 to 6, then scanned the children’s brains while they listened to stories. They counted, among other things, conversational turns: back-and-forth exchanges between the child and an adult. Their claim: children with more turns had more activation in Broca’s area, “independently of SES” (socioeconomic status, measured as parents’ education and family income).

SES is to their study what age was to the vitamins: it could push on both.

Figure 7. The same shape as Figure 2. In their data, SES goes with more turns (r = .34 for education, .37 for income) and with higher verbal scores (r = .69 and .48; pp. 704–705).

Their data are public (Harvard Dataverse, released into the public domain under CC0), so we can redo their analysis here. Every number below is calculated in your browser from their file, and the numbers match the paper.

Figure 3 of the paper, decoded

The paper’s Figure 3 (p. 705) has a y-axis called “Composite Verbal Score (Standardized Residual Controlled for SES)”. Residual is the statistician’s word for leftover. So this is step B of the recipe: each child’s verbal score, minus what their parents’ education and income predict.

Some phrases from the paper, translated:

A small difference between their Figure 3 and the After picture

Their Figure 3 takes the leftover of the verbal score only and keeps the raw turns along the bottom. Strictly, the partial r = .43 printed on it belongs to the version where both are leftovers, as in the After picture here. With their own leftover column, the correlation is .40. The two pictures look almost the same.

Figure 5 of the paper, decoded

The paper’s Figure 5 (p. 706) does the same, but for the brain, and it does it everywhere at once. The brain image is cut into tens of thousands of tiny cubes (voxels). At each one, the authors correlated turns with activation and coloured the voxel if the correlation passed a strict threshold. Panel a uses no covariates. Panels b to e repeat the whole map after partialling out (b) education and income, (c) verbal and nonverbal scores, (d) adult words and child utterances, and (e) all six.

Here is a toy version, with a made-up brain of a few hundred points instead of tens of thousands:

Figure 9 (toy). Front of the brain to the left. Coloured: correlation of turns with activation passes p < .01 with 36 children; darker is stronger. In this toy, activation in the front patch tracks turns, and also, over a wider area, SES. Before SES is partialled out, both colour the map; after, only what tracks turns over and above SES.

So the patch shrinks when covariates are partialled out, and what survives is the part that the covariates cannot account for. That is what you see from panel a to panel e in the paper: the patch gets smaller, but a patch remains each time. “Remained significant” means some voxels still passed the threshold.

Your turn: the covariate lab

The paper’s maps are voxel by voxel. Their data file has one number per child for Broca’s area as a whole (activation averaged over the region as anatomy defines it, the region used in the paper’s Figure 6). So here we can partial out the covariate sets of the paper’s Figure 5 for that one number, or for the verbal score. Tick covariates, or use the buttons.

Figure 10. Each dot is one child. With covariates ticked, both axes show leftovers: what the covariates do not predict.

Figure 11. Turns and Broca’s area activation (the whole region), with the covariate sets of the paper’s Figure 5. Dot: the correlation of the leftovers. Line: its 95% interval, the range of values these 36 children are compatible with. Our calculation from their open data; the paper does not report these numbers.

For the region as a whole, the correlation shrinks from .53 with no covariates to .16 with all six partialled out, and the interval then includes zero. Both this and the paper’s panel e can be true at once: a map tests thousands of voxels and shows the ones that pass, while one number for the whole region picks nothing. But it shows what “remained significant” does not say. It does not say the relationship stayed as strong. Here, most of it went away.

One covariate set needs a second look: (c), the children’s own verbal scores. If conversations build verbal skill, and verbal skill shows in Broca’s area, then partialling out verbal scores removes part of the very effect the study is after. That makes (c) a strict test rather than an unfair one, but it is a reason not to control for everything you can.

What “controlling for SES” cannot rule out

  1. SES was measured roughly. Parents’ education is a scale from 1 (high school or less) to 5, and income is one number. Partialling out a rough measure removes only part of what SES does; the rest stays in the leftovers, where it can still produce a correlation. In an illustration by Westfall & Yarkoni (2016), ice cream and swimming-pool deaths go together (r = .49) because both rise with the heat. Partial out people’s rough ratings of how hot it felt, and a correlation remains (.33). Partial out the measured temperature, and it vanishes (−.02).
  2. Nobody measured the parents’ own language ability, or the genes parents and children share. Verbally able parents may have more conversations with their children and pass on verbal ability. SES does not capture this, and nothing in the data does. The study cannot tell “conversation builds the brain” from “able parents have able children who also get more conversation”.
  3. Turns take two. A turn needs the child to answer. A child who is good at language, for whatever reason, may draw adults into more back-and-forth: then ability drives turns, not only the other way round. The authors reply (p. 708) that child utterances, a measure of how talkative the children were, related only weakly to scores and brain activation. That rules out mere chattiness. It does not rule out an able child who keeps a conversation going without saying more.
  4. Thirty-six children. The intervals are wide: the partial r of .43 is compatible with anything from about .11 to .67. Each covariate costs a little more precision.
  5. “Remained significant” is not “unchanged”. As Figure 11 shows, a relationship can lose most of its strength and still pass a threshold somewhere.
  6. The wording. The title says turns are “associated with” brain function. The abstract ends: “results suggest that conversational experience impacts neural language processing over and above SES or the sheer quantity of words heard”. “Suggest” is careful; “impacts” is a claim about cause. Controlling narrows down the alternatives; it does not turn a correlation into a cause. The direct test is an experiment: change how much some families, chosen at random, converse with their children, and see whether the children’s brains and scores change.

7 What controlling for something cannot do

The clean way round all of this is random assignment: if a coin decides who takes vitamins, age, health-consciousness and everything else end up balanced on average, measured or not. Our priming demo did this: which words came after a related word was decided at random for each of you.

Where randomising is impossible, studies combine tools. Weisleder & Fernald (2013) used two: restriction (they studied only low-SES families: “By focusing on differences within a homogeneous group of disadvantaged families … we reduced variability in these confounding factors”, p. 2150) and adjustment (maternal education was “controlled in all analyses”, p. 2144).

8 Vitamins in real life

The vitamin numbers on this page are invented to make the logic visible. Real studies, briefly:

So in real data two confounders pull in opposite directions: age makes vitamin takers look worse, health-consciousness makes them look better, and the second is the one researchers worry about most. That is the sandbox with the health-consciousness box ticked: controlling for age alone would make vitamins look protective. The randomised trials settle it, because the coin flip balanced both.

Sources
  • Bailey, R. L., Gahche, J. J., Miller, P. E., Thomas, P. R., & Dwyer, J. T. (2013). Why US adults use dietary supplements. JAMA Internal Medicine, 173(5), 355–361. doi:10.1001/jamainternmed.2013.2299
  • Loftfield, E., et al. (2024). Multivitamin use and mortality risk in 3 prospective US cohorts. JAMA Network Open, 7(6), e2418729. doi:10.1001/jamanetworkopen.2024.18729
  • Mishra, S., Stierman, B., Gahche, J. J., & Potischman, N. (2021). Dietary supplement use among adults: United States, 2017–2018 (NCHS Data Brief No. 399). National Center for Health Statistics. doi:10.15620/cdc:101131
  • Sesso, H. D., et al. (2012). Multivitamins in the prevention of cardiovascular disease in men: The Physicians’ Health Study II randomized controlled trial. JAMA, 308(17), 1751–1760. doi:10.1001/jama.2012.14805
  • Sesso, H. D., et al. (2022). Multivitamins in the prevention of cancer and cardiovascular disease: The COcoa Supplement and Multivitamin Outcomes Study (COSMOS) randomized clinical trial. American Journal of Clinical Nutrition, 115(6), 1501–1510. doi:10.1093/ajcn/nqac056
  • Westfall, J., & Yarkoni, T. (2016). Statistically controlling for confounding constructs is harder than you think. PLoS ONE, 11(3), e0152719. doi:10.1371/journal.pone.0152719
  • The course readings: Romeo et al. (2018), Psychological Science, 29(5), 700–710; Weisleder & Fernald (2013), Psychological Science, 24(11), 2143–2152; Linck, Kroll & Sunderman (2009), Psychological Science, 20(12), 1507–1515.

Romeo et al. (2018): the article · their data (Harvard Dataverse, CC0) · LIN 3042 course page