PBS05 SAQ stuff

Observation

Structured play → predefined tasks with the same opportunities e.g. MSEL which assesses early cognitive and motor abilities
Strengths - standardised and easy to code, good for testing capacity
Limitations - low ecological validity, relies on subjective interpretation, adult behaviours must be controlled for

Naturalistic observation → observes everyday behaviour with minimal interference
Strengths - high ecological validity, good for observing general behaviour
Limitations - child is influenced by caregiver’s behaviour, child may behave differently in front of the experimenter, may be difficult to determine the direction of influence

Semi-structured → planned tasks but with more experimental freedom e.g. NBAS measures early attention and responsiveness
Strengths - standardised but flexible, more developmentally appropriate (infants cannot follow strict instructions)
Limitations - still partly subjective

Systematic (predefined rules of what to look for → objective and comparable, but may miss important things if they are not predefined) vs. unsystematic (experimenters decide what is important when it happens → high ecological validity, but more subjective)

Coding

Global ratings → overall judgements using a scale

State coding → moment-by-moment coding (looks at the state at specific moments rather than counting behaviours)
Strengths - can look at dyadic interactions and interaction dynamics
Limitations - requires strict definitions

Frequency coding → counts behaviours within a timeframe
Strengths - good for behaviour with a clear onset/offset
Limitations - bad for continuous behaviours

Micro-level coding → concrete rules for looking at discrete behaviour

Interval sampling → divides observation into intervals and marks whether a behaviour is present/absent
Strengths - good for asynchronous interactions
Limitations - oversimplifies behaviour

Automated measures e.g. LENA
Strengths - removes observer bias

Reliability and biases

Inter-rater reliability → the agreement between independent observers and ensures findings reflect real behavioural variation (poor IRR can be due to ambiguous categories, fatigue, etc.)

Cultural context e.g. in some cultures, parents wait for infants to signal needs

Observer effects → parents behave how they think they should behave, coders become tired and less consistent

Why might we need alternative methods to observation?

Infants cannot verbally report, many abilities are latent, and brain developmental can precede behaviour

What are the challenges of studying infant brains?

Infants do not stay still, they have a short attention span, they show rapid anatomical change, and an absence of behaviour does not necessarily translate into an absence of capacity

What is fNIRS?

Shining near-infrared light into the scalp and measuring changes in light absorption to infer changes in oxygenated/deoxygenated Hb in the cortex

Strengths - silent, tolerant to movement, can be used when infants are awake

Limitations - correlational and indirect, poor spatial and temporal resolution

Why do we need global developmental science?

WEIRD populations are not representative, and tools may not be culturall valid e.g. MSEL (this is because behaviour depends on language, cultural norms, and experience) e.g. the BRIGHT project looked at whether developmental tools in the UK would work in the Gambia and they found that they are not culturally neutral

What is mediation and moderation?

Mediation → how X affects Y through a third variable Z
Direct effect - association is not explained by Z, indirect effect - the effect of X un Y through Z

Moderation → when or for whom X affects Y (this requires large samples and the moderator must precede the predictor)

What are longitudinal and cross-sectional studies?

Longitudinal → tracks changes over time (which is useful because developmental unfolds over time and different individuals have different developmental trajectories)

Cross-sectional → looks at different people across time

What is correlation and regression?

Correlation → association does not equal casuation (for causal interpretation, we need temporal order, no confounding variables, the correct functional form, and X measured without systematic bias)

Regression → predicts outcomes from variables (multiple regression separates unique variance and shared variance)

What is ANOVA?

Looks at the differences in mean values across groups

Between-subjects → different people in each group
Within-subjects → the same people across conditions
One-way vs. two-way

ANOVA assumes independence, normality, and equal variances

What are multilevel models (MLM) (a type of regression)?

They are useful for longitudinal and nested data

Their data points are not independent

Structure → level-1 (within-person, time), level-2 (between-person), and level-3 (context that the child belongs to, allows us to test contextual moderation, higher-level nesting)

Measurement and bias

Latent variables → constructs that cannot be directly observed, but are inferred from observable measures (observed variables are noisy indicators)

Measurement error → the observed score is the true score plus the measurement error, measurement error weakenes correlation and can make effects look smaller than they truly are

Stability bias → later mediator often correlates with earlier versions of itself (we give too much influence to the later mediator and must control for earlier versions of the mediator)

NHST

Asks how probable the observed data (or more extreme data) would be if the null hypothesis were true

If p < 0.05, we reject H0, but if p > 0.05, we fail to reject H0

Limitations - it gives a binary decision based on an arbitrary threshold, it does not compare H0 directly against H1, it cannot provide evidence for the null, it requires sampling and analysis to be planned in advance

Frequentist vs. Bayesian interpretation of probability

Frequentist → probability refers to the proportion of times something would happen in the long run

Bayesian → probability means degree of belief (you start with a prior belief, then update it using the data, giving a posterior belief)

What is Bayes theorem?

Updates beliefs by combining prio beliefs with new evidence

Prior - what was believed before seeing the data (this can be informed or uninformed)
Likelihood - how probable the observed data are under a given hypothesis or parameter value
Posterior - updated belief after combining the prior with the likelihood

What is the likelihood principle?

All the evidence the sample provides about a parameter is contained in the likelihood function (once you have observed the data, the evidence is entirely in the data you actually got, not in the data you could have got but did not)

Bayesian hypothesis tests are based on this (Bayesian methods focus on the observed data)

Frequentists approaches also consider the data that could have occurred but did not

Bayes factor

BF compares how well the data supports one hypothesis over another
BF10 = evidence for H1 relative to H0
BF01 = evidence for H0 relative to H1

BF10 > 1 favours H1
BF10 < 1 favours H0
BF between 1/3 and 3 are inconclusive

BF is not the same as posterior odds unless H1 and H0 started with equal prior probability
BF = what the data says, posterior odds = what you believe after combining the data with your starting beliefs

BF strengths compared to p-values

They can provide direct evidence for H0, they compare H0 and H1 directly, they give graded rather than purely binary evidence, they allow for incorporation of previous knowledge through priors

Bayesian t-test

Frequentists t-test asks if there were really no difference, how surprising would this data be

Bayesian t-test asks does the data support no difference or a real difference more (differs from NHST because it explicitly evaluates both H0 and H1 and places a prior distribution over effect size

Optional stopping

This means checking the data as they come in and deciding whether to continue collecting more data or not

Under NHST - this is a problem because repeated cheking of data inflates false positive rates

Under Bayesian methods - allows for optional stopping because inference depends on the evidence in the observed data