University of Missouri · Indiana University Indianapolis

Developmental & CSCW research · EEG, longitudinal infant studies, qualitative fieldwork

Measuring an expectation

Before I asked drivers what they expected from a car, I spent four years asking the same question of people who could not answer it in words.

How a twelve-month-old answers a question

Adult ignores the crying baby 46.33s

Adult goes to the crying baby 42.34s

+3.99s Babies look longer at what they did not expect. Four extra seconds of attention, held across fifteen infants, is the whole of the evidence — and it only appears when both of that infant’s parents were engaged.

The question underneath the field

Attachment research had settled that infants build ~~internal working models~~ — carried-forward expectations about how people will treat them. Almost nothing said how you would recognise one if you were looking straight at it.

119mothers

Ask the brain

EEG. If a mother holds an expectation about responsive care, an insensitive word should register as a violation within half a second.
39families

Follow them forward

Code both parents at six months, then test the infant at fifteen. If early care builds the expectation, the coding should predict the looking.
7of a planned 32

Ask the baby

A remote violation-of-expectation task, rebuilt for Zoom mid-pandemic. Under-powered, and reported as under-powered.
4publications

Peer-reviewed output

Conference posters, a talk, and a CSCW paper. See the list ↓

Study one · asking the brain

Mothers’ neural responding to (in)sensitive caregiving cues

Self-report is a poor instrument for asking someone what they expect of themselves as a parent, because everybody knows the right answer. So we did not ask. Mothers listened to a prime — “If my baby cries, I will be…” — then decided whether a target string was a word. The measure was the N400: a negative deflection about four hundred milliseconds after a word that does not fit.

The control was the part that mattered. Half the primes were attachment-irrelevant — “If my baby crawls, I will be…” — so any effect had somewhere to disappear. A finding that cannot vanish under a control condition is not yet a finding.

32-channel EEG
Attachment Script Assessment
Lexical decision task
Correlational analysis

Schematic, not data

What an N400 is

An idealised parietal waveform, drawn so the measurement in the next figure is legible without a background in ERP research.

Negative is plotted upward, as ERP figures conventionally do. The deeper the deflection at roughly 400 ms, the more the word violated what the listener was primed to expect. The reading is the difference between the two waveforms, not either one alone.

The effect appears only where it should

Correlations between four coded dimensions of mothers’ attachment narratives and their N400 difference, split by whether the audio prime was about the infant’s distress or about something ordinary.

Mothers scoring higher on attachment deactivation — narratives that minimise attachment themes — showed a smaller difference between sensitive and insensitive words, r = −.21, p = .02. Under the non-attachment prime the same correlation collapsed to −.08, n.s. The effect is specific to the parent–child context, which is the claim the control was built to test.

A mother who plays attachment down in her own narrative has a dampened expectation that a crying infant will be answered. Not a self-report, not an observer rating. A measure taken 400 milliseconds after a word, in a task whose stated purpose was to decide whether a string of letters was English.

Study two · the parent nobody coded

Fathers matter too: responsive caregiving and infant expectations

The literature linking early care to infant expectations was almost entirely a literature about mothers. We coded both parents, separately, and then waited nine months to see which coding predicted anything.

Each parent was recorded alone with their six-month-old for five minutes of free play with no toys — a deliberately bare situation, because a toy gives a disengaged parent somewhere to hide. Two coders rated detachment independently. At fifteen months the infant came back for the looking-time task.

Behavioural coding
Longitudinal, 6 → 15 months
Violation-of-expectation
Paired t-tests

Both parents engaged, or the effect does not appear

Mean looking time to each event at fifteen months, split by how the infant’s two parents had been coded at six. Whiskers are one standard deviation.

Where both parents were rated low on detachment, infants looked reliably longer at the adult who walked away, t(14) = 2.77, p = .0149 — the signature of a violated expectation. Where either parent was disengaged, the two events were watched the same amount, t(13) = 0.41, p > .250.

Why this is the finding I still cite The interesting result is not that engaged parenting predicts the expectation. It is that the mother’s rating alone did not carry it. Splitting the sample the obvious way — by the parent the field had always measured — would have produced a null and a shrug. The variable that mattered was the one nobody was collecting.

Study three · the one I did not claim

Can babies remember poor caregiving?

My own study, and the one that taught me the most. Twelve-month-olds watched two adults respond to a crying infant — one goes, one keeps tidying — and then saw a still photograph of each face. If the first phase builds an expectation, the second should reveal a preference.

Then the labs closed. I rebuilt the whole protocol for Zoom: consent and the maternal responsiveness questionnaire through Qualtrics, stimuli screen-shared, two observers coding looking time live from the caregiver’s webcam. It worked. It was also slow, and by the presentation date I had seven infants of a planned thirty-two.

Preferential looking
Maternal Responsiveness Questionnaire
Remote testing protocol
Two-observer live coding

Seven infants, and a number I refused to test

Mean looking time at each phase. The plan of analysis — repeated measures ANOVA, paired t-tests, Spearman correlations against the questionnaire — was published alongside the descriptives and left unexecuted.

The gap at familiarisation points the same way as every other study on this page. At test, where a preference between the two faces would have shown up, there was nothing — 0.21 seconds across seven babies is noise, and it was labelled as noise.

The poster showed the descriptives and the analysis plan, and ran neither test. A 5.23-second gap in the predicted direction was sitting right there, and n = 7 with the pandemic as an excuse is exactly the situation in which people report it anyway. Publishing the plan without the p-value is the only part of my academic training I have never had to adapt for industry.

Publications & presentations

2024

Patel, S., Nikkhah, S., Nadkarni, S., Suryawanshi, P., Sureddi, R., Bhoyar, M., et al. Designing for resilient caregiving coordination journeys: prioritizing features across families and phases. Companion Publication of the 2024 Conference on Computer-Supported Cooperative Work and Social Computing.

Peer-reviewed Miller Lab, Indiana University Indianapolis Cited 4×
2021

Bhoyar, M., Luo, Y., Xu, N., vanMarle, K., & Groh, A. Fathers matter too: the significance of responsive caregiving in infant expectations. Poster presented at the virtual mini-conference on father–child attachment.

First author University of Missouri Cited 2×
2021

Bhoyar, M., & vanMarle, K. The development of internal working models in the context of infant attachment. Midwestern Psychological Association annual meeting.

First author University of Missouri
2021

Bhoyar, M., & vanMarle, K. Can babies remember poor caregiving? Undergraduate Research and Creative Achievements Forum, University of Missouri, 21 April 2021.

First author University of Missouri
2021
–22

Developmental Ladder — founder and host. An interview podcast putting parents and caretakers in front of the researchers whose work is about them. Six episodes, twenty to thirty minutes each.

Science communication Watch on YouTube

Google Scholar indexed record

What carried into product research

None of this is transferable as subject matter. What transferred is a set of habits about what has to be true before a difference is allowed to be called a finding — and those are the same four habits I run studies on at Honda.

Study one

The correlation held under the attachment prime and vanished under the control prime.

Became

A result with no condition where it disappears is a description, not a finding.

Decision tests that can fail

Study two

Two coders, blind to each other, agreed at ICC .90 and .96 before a single number was interpreted.

Became

If a second person cannot reproduce the coding, the theme is in my head, not in the data.

Human review on AI-assisted synthesis

Study two

The variable that predicted the effect was the one the field had not been collecting.

Became

The brief names the variable everyone already agrees about. The answer is usually next to it.

I define the problem before I solve it

Study three

n = 7 of a planned 32, presented as descriptive statistics with the analysis plan attached and no test run.

Became

Every recommendation ships with the support behind it and the counterexample that survived it.

A finding is not yet a recommendation

Each of these is checked against a study on this site rather than asserted here. The three commitments · the AI study · e‑Turn.

What this demonstrates

Statistical training that is actually load-bearing, an instinct for the control condition, and four years of practice reporting exactly as much as the data supports — including the times that was nothing at all.

Résumé Google Scholar