1/36
chapter #6
raw score distribution
shape individual observations in your sample
sampling distribution
distribution of a statisti
across many hypothetical repeated samples
homoscedasticity/homogeneity of variance
samples should come from populations with the same variance
can affect:
parameters
nhst
assessing homogeniety of variance
graphs
levene’s tests
if variances in different groups are the same
solutions
robust tested
welch’s t
welch’s f
adjusted se
how to read levene’s test
evaluates null that group variances are equal
small p → evidence that the variances differ
does not tell you whether the difference is important
large samples = trivial variance differences statistically signifiant
reading the homogeneity figure: left panel
each group has similar vertical spread
group means differ
variability is comparable
reading the homogeneity figure: right panel
vertical spread changes across groups
later groups = more variable than earlier groups
independence
errors in model shouldn’t be related to each other
boxplots and outlier flags
uses iqr range
values more than 1.5 x iqr below q1 or above q3 are often flagged as outliers
spotting normality (4)
test observed data
central limit theorem
graphical displays
value of skews/kurtosis
kolmogrov-smirnov test
central limit theorem
if n> 30, the sampling distribution is normal anyway
graphical displays
p-p plot (or q-q plot)
histogram
values of skew/kurtosis
0 in a normal distribution
convert to z (by dividing value by SE)
q-q plot
asks whether observed quantiles follow the pattern expected under normality
systematic curvature is more informative tiny random deviations
kolmogrov-smirnov test
test if data differs from a normal distribution
don’t use:
in large samples → small differences will be significant
in small samples → large differences won’t be significant
how to read q-q plot scales: approx. normal
points lie close to the reference line without systematic curvature
how to read q-q plot scales: positive skew
points show a characteristic bend bc the upper tail extends farther than a normal distribution would predict
how to read q-q plot scales: negative skew
curvature appears in the opposite direction
how to read q-q plot scales: heavy tails
both ends depart from the reference line
more extreme observations than expected under normality
how to read q-q plot scales: light tails
ends bend in the opposite pattern
fewer extreme observations than expected
residual plots: assumptions met
residuals form an unstructured cloud around zero
roughly constant vertical spread
residual plots: heteroscedasticity
spread increases or decreases with fitted values
fan/funnel pattern
residual plots: non-linearity
avg resiudal follows a curve
straight line model systematically misses part of the relationship
residual plots: both problems
curved mean pattern plus changing dspread
reducing bias
analyse with robust methods
trim the data
windsorizing
transform the data
analyse with robust methods
bootstrapping
methods based on medians and trims
trim the data
delete a certain amount of scores from the extremes
windsorizing
substitute outliers with the highest value that isn’t an outlier
transform the data
applying mathematical function to the scores
log transformation (log (xᵢ)
reduce positive skew
compresses large values much more than small values
square root transformation (√xᵢ)
also reduces positive skew
stabilizing variance
compresses high values but more gently
reciprocal transformation (1/xᵢ)
divi 1 by each sore
reduce impact of large scores
reverse the scores
compresses large values
can reverse ordering depending on the exact formula used
should you transform?
identify the actual problem
whether chosen model is robust to that problem at your sample size
if transforming, choose method that makes sense
check model again after transforming
report what you did
trimming the data
removes a fixed percentage of observations from each tail before calculating a statistic
trimmed mean is less sensitive to extreme obs
sd
spread of individual scores
if n goes se will go down → if ss gets larger, se will decrease
se of the mean
s/ √n