Module 5 Practice Questions Flashcards
What is data fishiness?
Properties of data or statistical tests that suggest potential problems
What are the three assumptions to evaluate?
- Normality
- Homogeneity of variance
- Independence of observations
Using an NHST approach, what are the statistical tests that can assess normality?
Kolmogorov-Smirnov test
Shapiro-Wilk test
These tests compare our distribution of data to a normal distribution of data. If data does not differ from normal distribution than null is true
Using an NHST approach, what are the descriptive statistics that can assess normality?
Skew
- tells us how asymmetrical our distribution is
Kurtosis
- tells us the prevalence of extreme scores in tails
- a certain number of extreme scores is normal
- too much = positive kurtosis, heavy tails
- too little = negative kurtosis, light tails
What are the limitations of statistical tests of the assumption of normality? (There are three)
Role of sample size
- tolerates violations of normality in small sample sizes (this is because power is low, fail to detect even big violations)
- very sensitive to violations of normality in large sample sizes (detect even the slightest violations)
Logic of test is flawed
- We should not be asking if the deviation from normality is 0 BUT instead is the deviation from normality large
- This is because it is unlikely to be perfectly 0
Does not take into account type of non-normality
- Different types of deviation will be more problematic than others
Which is worse, positive or negative kurtosis?
Positive kurtosis tends to be more problematic than negative kurtosis since it produces larger distortions
Using a graphical approach, what are the two visual displays that can assess normality?
q-q and p-p plots
What is a problem with using the visual display approach?
Element of subjectivity
- Easy to judge good versus bad
- Difficult to judge in ambiguous situations
____________ approach makes more sense than ________ approach
descriptive; NHST
Using an NHST approach, what are the statistical tests that can assess homogeneity of variance
Levene’s test
Hartley’s variance-ratio test
F-max test
Similar to the assumption of normality, what are some drawbacks to these statistical approaches of the assumption of homogeneity of variance?
Role of sample size (same as normality)
Logic is not right - should be asking if there is a big enough difference (same as normality)
Describe the descriptive approach to testing the assumption of homogeneity of variance and provide examples
Take largest variance and smallest variance and compute ratio
3: 1 has been advocated as a threshold
- this means that as long as largest variance is not 3 times bigger than smallest variance, test is OK
What is the difference between a qq plot and a normal qq plot? What are they each used for? What does it mean if the scatterplot line is straight with a slope of 1? For qq plots what should the intercept be equal to? What does it mean when the slope of the line is different from 1?
Normal q-q plot
- Scatterplot where your dataset is on Y axis (DV) and generate a normal distribution of data for x axis (IV)
- If data are normal dots should cluster together in a straight line
- A graphical display used to assess normality
q-q plot
- One condition is on the x axis the other on y axis
- Intercept should be equal to difference between two means
- Slope of 1 indicates variances are equal to one another and data points cluster around straight line
- Slope different from 1 indicates unequal variance and data points represent a cloud all very spaced out
- A graphical display used to assess homogeneity of variance
What are the results of positive and negative correlations among data points?
Positive correlation among data points = inflated alpha rates
Negative correlation among data points = inflated beta rates
How do we evaluate the assumption of independence of observations?
Interclass correlation
- A value of 0 indicates independence exists
- A value that is not 0 indicates a violation
What are the four methods for addressing violations of each assumption?
- Use alternative statistical procedures that don’t require the specific assumption
- common across all 3 assumptions - Transform data to normalize the distribution
- unique to assumption of normality - Identify and remove outliers
- common across assumption of normality and homogeneity of variance - Evaluate level of measurement
- common across assumption of normality and homogeneity of variance
What is an outlier? What are they often responsible for?
Outliers are extreme scores in dataset
They are often responsible for violations of normality and homogeneity of variance
What are some quick ways to weed out outliers?
Generate histograms
Generate normal q-q plots
Compute standardized residuals or studentized deleted residuals
What do tails look like in q-q plots as a result of outliers?
Steep (aka thick) tails indicate more extreme values than what is acceptable
What are the two primary approaches to dealing with/ responding to outliers?
They are “trimmed” or “capped” to most extreme acceptable value
- reduces disproportionate impact of observation
They are treated as missing data
- eliminates impact of observation all together
There are 3 perspectives for dealing with outliers
Minimalist perspective
- Data set should be minimally altered
- Distributions should have some extreme values
- Getting rid of outliers or altering them can create its own distortions
Maximalist perspective
- Routine to alter or delete outliers
- Hard to interpret results with outliers
- Outliers create violations of assumptions
Intermediate perspective
- Justifiable with clear rules and procedure as well as high thresholds for outlier status
What are the 4 levels of measurement?
Nominal, ordinal, interval, ratio
Which levels of measurement are appropriate for t-tests and ANOVAs?
Has been argued that t-tests and ANOVA are only meaningful when DV has at least an interval level of measurement
How do rating scales fit into this? What are the basic guidelines for dealing with quasi-interval data?
Some data can be ambiguous with respect to level of measurement (i.e., are traditional 5-point or 7-point rating scales ordinal or interval)?
Thus, rating scales have been deemed “quasi-interval”
Standard statistical procedures can function reasonably well for quasi-interval data if a sufficient number of response categories are provided and distributional assumptions are reasonably well satisfied (less than 5 points is problematic, 5 points is ambiguous, 7 points is sufficient)
