1/35
Looks like no tags are added yet.
statistical model
Outcomei = model + errori
i = each person i
model = expected value (prediction for personi)
errori = expected - observed value (personi deviaiton from prediction)
hypothesis
a specific, testable and falsifiable prediction about the world that data could contradict
always come in a duo that must cover all possibilities and must not overlap
null hypothesis
prediction that there is no effect/difference (always tested against since prediction is precise i.e. know what data would look like)
alternative hypothesis
prediction that there is an effect/difference (can be one or two-sided)
two-sided
alternative hypothesis direction unspecified
one-sided
alternative hypothesis direction specified
to assess significance half p-value (given effect is in predicted direction)
only use with prior directional expectation
dependent variable
independent variable
nominal
category with no order (e.g. sex)
ordinal
categories with order (e.g. mild/moderate/severe)
interval
continuous variable with equal steps and no true zero (e.g. °C)
ratio
continuous variable with equal steps and true zero (e.g. reaction time)
mean
the sum of all scores divided by N
in statistical testing, usually the simplest model of the data
variance
average squared deviation of observations from the mean (i.e. how spread out the data is, in squared units)
standard deviation
typical distance of a score from the mean, in the original units of the data (square root of variance)
skew
measures the symmetry of a distribution (0 = symmetric)
positive (right) skew: most scores are low → long tail to the right
negative (left) skew: most scores are high → long tail to the left
kurtosis
measures how heavy the tails of a distribution are (i.e. how many extreme scores there are)
positive: heavy tails and a sharp tall peak (more extreme scores)
negative: light tails and a flat broad peak (fewer extreme scores)
platykurtic
negative kurtosis with light tails a flat peak
leptokurtic
positive kurtosis with heavy talks and sharp tall peak
parameter
true value in the population:
mean (μ)
standard deviation (σ)
variance (σ²)
regression coefficient (β)
estimate
observed values as best guess of the population true value:
mean (x̄)
standard deviation (s)
variance (s²)
regression coefficient (b)
sampling variability
natural differences between estimates from different samples, caused by chance in who ends up in each sample
sampling distribution
histogram of an estimate taken from the same sample an infinite amount of times (typically normal and centred on the true population mean)
standard error
the SD of the sampling distribution of an estimate (i.e. how much the estimate differs between samples)
the bigger the sample the smaller the value
confidence interval
range of plausible values for the true population value, built around the parameter estimate (i.e. where does the true population value probably lie)
estimate ± 1.95 x SE
the bigger the sample the smaller the SE and the narrower the range
repeat a study many time and 95% of intervals will contain the true population value
systematic variance
variation explained by the model or manipulation (signal)
outcome = model + error
unsystematic variance
variation not explained by the model (noise/error)
outcome = model + error
test statistic
standardised ratio of the systematic and unsystematic variance in a sample (summarises how far the data deviates for H0 prediction)
test statistic = signal / noise = estimatediff / SEdiff
follow a known distribution that can be used to estimate whether the effect is significant
p-value
probability of observing a test statistic at least this extreme, assuming that the null hypothesis is true (i.e. assuming no result, how often would chance alone give a result this big or bigger)
small sample = little power (might miss effect)
big sample = oversensitive (might overestimate/flag minor effects)
for one-sided tests divided by two, given that effect is in predicted direction
parametric test
test that assumes a normal distribution and uses its parameters:
independent observations
interval/ratio outcome
normality of sampling distribution
homogeneity of variance
more power when assumptions hold
non-parametric test
test that works with ranks/counts instead of parameters and makes no distributional assumption:
independent observations
underlying continuous distribution
effect size
a standardised measure of how big an effect is, independent of sample size (can compare across studies that used different measures)
p-value tells you whether there is an effect, this value tells you how big the effect is (no overesimtation)
pearson’s r
measure of how strongly two variables are association (range: -1 to 1 with sign = direction, size = strength)
small = .1
medium = .3
large = .5
effect size for correlation, paired t-test, contrast/post-hoc
variance explained
proportion of the total variation in the outcome (DV) that the model or predictor(s) can account for (or proportion of variance two variables share → square of r)
effect size for regression
cohen’s d
differences between two group means in SD units (i.e. how much sample means overlap)
small = .2
medium = .5
lage = .8
effect size for independent t-test
intra-cluster correlation
proportion of the total variance in the outcome that is due to differences between clusters (the clustering in the data)
high (dependency) = observations from the same cluster are very similar (overlap of information → smaller effective sample)
low (independent) = cluster membership tells you little, so the observations carry unique information
effect size for multi-level analysis