1/28
Looks like no tags are added yet.
Variance
How spread out observations in a dataset are
average of squared differences from the mean
PPDAC cycle
Problem, plan, data, analysis, conclusion
Discrete Variables
whole numbers
categorical data
dichotomous variables
continuous variables
numerics that can be fractions
ex. age, distance
nominal data
categorical
no order, no quantitative distinction
categories
ordinal data
categorical
order matters
difference between categories is not necessarily consistent
Interval data
Continuous
order matters
values indicate quantity, not categories
equal intervals between values
NO TRUE ZERO - even if 0 exists it does not indicate a lack of the variable
ex. temperature
Ratio Data
continuous
order matters
values indicate quantity, not categories
equal intervals
true zero exists
ex. Weight
Plots and best data for them
bar chart - categorical data
histogram - continuous data, good for seeing distribution
column/bar plot - grouped continuous data
box plot - grouped continuous data
violin plot - grouped continuous data
scatterplot - relationships between continuous variables
line graphs - data that changes over time
Screening data
check accuracy of data
check relevant statistical assumptions
univariate outliers
an observation that is extreme for one variable
detected by
plotting data and finding them visually
computing z-scores (observations with a z-score of more than ±3.29 are too many standard deviations away from the mean)
multivariate outliers
an observation that is extreme in multiple variables
CAN BE MULTIVARIATE WITHOUT BEING UNIVARIATE - these are sneaky and remain in data set even after handling all univariate outliers
only occurs in data with a pattern
these data are too distant from the pattern which makes them outliers
but they’re not too extreme on any one variable to count as outliers
detected by:
calculating Mahalanobis Distance
Mahalanobis Distance
distance of an observation from the centroid point (center) of the data
corrects for variance in the dat so that values that are far from the mean in the expected direction are less likley to be categorized as outliers
dealing with outliers - first steps
set rules for what counts as outlier beforehand
address univariate outliers first
check if outliers is a data entry error
Dealing with outliers - statistical approaches
transform data (ex. log transform) to reduce overall variance
winsorize - change outliers score into next-most extreme score in the data set so it still represents a high score in the same direction without messing with statistical tests
use non parametric tests to analyze (rank based)
dealing with outliers - trimming strategies
omit cases if confident they come from a different population than the desired one
trim certain percentage of cases from high and low tails
just trim outliers (careful with this one)
The Normal Distribution
Many statistical tests assume data is normal shaped
when this is not true, statistical tests lose power
skew = 0
kurtosis = 0
Skew
symmetry/asymmetry of distribution
positive - data shows a longer tail on high end of scale
negative - data shows a longer tail on the low end of scale
Kurtosis
peakedness/flatness of distribution
positive - sharp peak in the middle, indicates concentration of observations around that point
negative - peaks at the extremes of the scale, indicates observations are concentrated at highest and lowest values
Testing for Normality of Distribution
test whether skew and kurtosis differ significantly from zero using z-scores
Use Q-Q plot - compares observed values against expected values of a normal distribution
Q-Q Plot

Correcting Normality
transform variable
winsorize/trim
choosea statistical method that fits the shape of data
use non-parametric tests
Linearity
Assumption that variables have a linear relationship
necessary for many statistical tests
visualize with bivariate scatterplot\
nonlinear data must use alternative analyses
homogeneity of variance
assumption that variance of DV for all levels of a categorical IV is similar
Homoscedasticity
assumption that variance of DV is similar at all points of a continuous IV
multicollinearity
when variables are too closely correlated
indicates conceptual redundancy between variables, one can be deleted
causes problems especially in regression
hard to tell what is actually causing the effect
convergence problems
makes interpreting role of individual predictors difficult
Dealing with Multicollinearity
Remove the less important variable
combine them if they measure the same construct
Types of missing data
Missing completely at random (MCAR)
no pattern to missing data, existing data is representational of desired population
Missing at Random (MAR)
pattern of missingness depends on a variable but not the primary relevant one
Missing not at Random (MNAR)
missing data depends on a variable that is central to the research
can seriously bias results
Dealing with missing data
listwise/pairwise deletion - delete the whole participant if some vital data is missing from their observation or if too much data is missing
estimate/impute missing data - use knowledge of previous research
mean substitution - impute mean value for missing value bc it wont significantly change data distribution
regression - use complete cases to create regression model to predict missing values