Flashcards
Lecture 3
Cards in this deck (88)
- 1OUTLIERSOUTLIERS
- 2What are outliers?The individuals scores that can occur in a data file that are quite different from everybody else.
- 3What is the problem with outliers?They can have a disproportionate impact on statistical results
- 4Summarise John's approach to outliersBegins with examining outliers in detail. Asks questions before removing including:
1. What does the outlier represent (true score or process of error)?
2. If it is a true score, is it more appropriate to use a statistical technique that allow us to incorporate that outlier as it is a legitimate score. - 5What does a researcher have to be sure to do regardless of their approach to outliers?Employ transparency regarding their approach in their write up
- 6Recap: Summarise John's approach to outliers1. Delete if it is error
2. Use a statistical technique if it is a true score - 7ASSUMPTIONSASSUMPTIONS
- 8What are parametric procedures?Procedures (t-tests, ANOVA's) that make assumptions that our data have certain characteristics
- 9What happens if your data does not meet these assumptions?Running tests is problematic.
1. Significance levels ( p less than .05) will be distorted e.g., too small or too big)
2. Distortions in significance levels will lead to errors in your decisions. - 10Explain the relationship between assumption testing and screening and cleaningAssumption testing is part of screening and cleaning in preparation for analysing data.
- 11How many assumptions are there?There are many types of assumptions and some are specific to certain types of tests.
- 12Name the three most common assumptions1. Normality
2. Homogeneity of variance
3. Independence of error
Def: every score is separate and unaffected by another score
(doesn't often cause a problem in psychological research) - 13Provide an example of a case scenario whereby independence of error can occur in psychological testingOne group encounters deception (Galvanic skin response scenario)
First group tells second group which affects their score. - 14In what type of design can independence of error become problematic?Repeated measures design because you're measuring a person on multiple occasions. A person's score early on in the study is going to be related to their score later in the study because you're testing the same person.
- 15Define the assumption of normalityAssumes that your data comes from a population that is normally distributed.
The assumption is worded in the context of the population, however, the way we test the assumption is by looking at the sample. - 16Why is normal distribution such an important concept?1. So many variables are normally distributed in nature
2. The normal distribution is central to so much statistical testing - 17What does homogeneity of variance, also called equality of variance assume?Assumes that, if your data is divided into groups, the level of variability in the groups as represented by measures of variability, will be approximately equal.
- 18Provide and example of homogeneity of varianceA study that that includes a clinical group and control group should have approximately the same level of variability.
- 19What are some of the measures of variability?The range, the interquartile range (IQR), variance, and standard deviation.
- 20Name the ways in which normality can be tested1. Visual inspection of histograms and stem and leaf plots
2. Visual inspection of normality and detrended normality plots
3. Normality tests
4. Skewness divided by standard error skewness - 21What is the issue with having multiple ways of testing for normality?Results can be contradictory
- 22What does the visual inspection method rely on?Subjective perception of the images/graphs. Is not objective
- 23What is the purpose of normality and detrended normality plots? Are they objective or subjective?Graphs specifically designed for testing normality
Subjective/ Provide a 'feel' of the data. - 24What are the two objective ways of testing for normality?1. Normality tests
2. Skewness divided by SE skewness - 25Recap: what is the significance level normally set at?Less than 0.05
- 26What are the parameters for skew?Skew < -2 = negative skew
Skew > +2 = positive skew
-2 and +2 are considered within normal range - 27What are the subject ways of testing for normality?1. Visual inspection of histograms and stem and leaf plots
2. Visual inspection of normality and detrended normality plots - 28Are visual inspection of histograms and stem and leaf plots & Visual inspection of normality and detrended normality plots influenced by sample size?No, the graphs will look the same regardless of sample size
- 29Are the two objective ways of testing for normality affected by sample size? 1.Normality tests 2. Skewness divided by SE skewness Why?If you have a really larger sample, even trivial deviations from normally indicate a violation of the assumption. This is essentially a Type I error, nothing is going on but your test tells you something is going on.
- 30Describe the SPSS steps required to obtain information about normality1. Analyse
2. Descriptive stats
3. Explore
4. Add variables to dependent list
5. Include histogram in plots & normality plots with tests - 31Describe the steps to determine normality based onSkewness divided by SE skewness (objective measure)1. Create output
2. Under variable
3. Skewness divided by SD error of skewness - 32What are we looking for with this test?1. If it falls between -2 and + 2, data is within a normal range.
2. If it is greater than +2, it is positively skewed.
3. If it is less than - 2, it is negatively skewed.
A figure greater than -2 and +2 is POTENTIALLY problematic - 33Why is the standard error of skewness the same across our variables?Because it is based on sample size
- 34Recap: what is wrong with the metric?Affected by sample size. If the sample size is large, even trivial deviations from normality will come up as a problem.
- 35What are the two tests of normality produced by SPSS (objective)1. Kolmogorov-Smirnov
2. Shapiro-Wilk - 36How do you read: 1. Kolmogorov-Smirnov 2. Shapiro-WilkUsing sig. Level.
If sig. level is less than 0.05, that indicates that the assumption of normally has been violated. - 37What is the issue with the below tests of normality? 1. Kolmogorov-Smirnov 2. Shapiro-Wilk Why?Affected by sample size. If the sample size is large, even trivial deviations from normality will come up as a problem.
All significance tests are sensitive to sample. If you have a larger sample, you are more likely to get a significant result. - 38What are normality plots? (subjective)A graphical technique for assessing whether or not a data set is approximately normally distributed.
- 39What are the two normality plots called?1. Normality plot
2. Detrended normality plot - 40What are we looking for in 1. Normality plot 2. Detrended normality plot ?1. Dots to sit as closely possible to the straight line (the closer the fit, the more normal the data is).
2. Looking for randomness and spread all over the place not a pattern - 41Recap: four methods to ascertain normality. Provide a summaryGraphs are subjective and are not influenced by sample size
Tests are objective, but are influenced by sample size. With a large sample, even trivial deviations from normality will indicate a violation of the assumption - 42What is the paradox of normality testing?Normality is the least important assumption underlying parametric tests.
Statistical tests are very robust (resistant) to even gross deviations from normality, particular when you have a large sample. - 43Why is this the case? Why should we be cautious when using the tests?Normality becomes less of a problem as sample size increases. The tests that are testing for normality don't work for the situations where normally isn't a problem.
If we have a decent sample size (e.g., 50+), the reality is normality is less likely to be a problem. However, tests are useful for sample sizes less than 50. - 44Recap: what is the central limit theorem?(1) Statistical tests are based on sampling distributions
(distributions of sampling means).
(2) Sampling distributions will always be normal, regardless of the shape of the underlying population distribution.
(3) They will get more normal as the sample size increases. - 45When is normality a problem?Normality becomes more problematic when:
1. The sample size is small
2. When the deviation from normality is strong - 46What is considered to be a small sample size?There's no hard and fast rule. We have to factor in others considerations such as deviations from the norm in conjunction with the sample size.
- 47What's John's approach for testing for normality?Focuses on graphics unless the sample size is less than 50, in which case the tests are used.
- 48What needs to be considered before 'dealing with' violations of normality?1. The sample size
2. The deviation from normality
3. The nature of the skew
4. The four sources of information
(information needs to be consolidated to make a decision) - 49How can you 'deal with' violations of normality?Use non-parametric tests. However, you are limited with what you can do with them and they're not particularly statistically powerful.
A better option is to...
Carry out data transformation. - 50What is data transformation?Applying a simple mathematical operation to the data to deal with violations of assumptions
(massaging it into normality) - 51Name some common data transformationsChanging currencies
kilo to pound
Percentile to raw score - 52Explain how data transformations preserve aspects of the dataData transformations do not change the ordinal properties of raw scores. For example, if you score high on the original measure, you'll also score high on the transformed measure.
- 53Is there a set data transformation to be used in situations?No, it's a process of trialling what works.
- 54What are some common data transformations?1. Log 10
2. Square root
3. Reciprocal - 55If you get an exam questions that indicates a sample of 100+, asking if you would complete a data transformation, what would be the answer?No, as the sample size is large enough that this is unlikely to be required (e.g., output for GHQ) unless the skew is extremely dramatic (e.g., output for smoking).
- 56What are the steps to complete a log 10 transformation (of the original variable)? Computation of base 10 log1. Transform
2. Compute variable
3. Give name
4. Highlight variable
5. Formula under numeric expression lg10(variable you want transformed abbreviated)
6. OK - 57What does the compute variable function allow?Allows us to create new variables from old variables
- 58What are the steps to complete square root transformation (of the original variable)? Computation of square root1. Transform
2. Compute variable
3. Give name
4. Highlight variable
5. Formula under numeric expression sqrt(variable you want transformed abbreviated)
6. OK - 59What are the steps to complete reciprocal transformation (of the original variable)?1. Transform
2. Compute variable
3. Give name
4. Highlight variable
5. Formula under numeric expression 1/(variable you want transformed abbreviated)
6. OK - 60What is a reciprocal transformation?When you divide the scores into 1/
1 divided by score. - 61Homogeneity of varianceHomogeneity of variance
- 62Recap: Homogeneity of varianceRefers to the assumption that when you have groups, the level of variability in the groups is not profoundly different.
- 63What do we call the violation of the assumption of homogeneity of variance?Heterogeneity of variance
- 64Is homogeneity of variance serious?Yes.
Violations are definitely more problematic - 65What impacts on how problematic it is & what does this refer to?Becomes more a problematic if the sample is 'unbalanced'. The more unbalanced, the more problematic. If you have a violation of the assumption plus an unbalanced design, there can impact statistical results.
An unbalanced sample refers to the situation where you have unequal group sizes. - 66What is the term used to describe equal and unequal numbers in groups?A balanced & unbalanced design
- 67What test do we use to test homogeneity of variance assumption & is it impacted by sample size?Levene’s test & is quite reliable even with larger sample sizes.
- 68What are the steps to carry out Levene’s test?1. Analyse
2. Descriptive stats
3. Explore
4. Enter the dependent variable into the dependent list
5. Enter the grouping variable into the factor list
6. Power estimation under plots - 69When testing for normality, do you break down the test into groups or conduct it overall?Conduct overall
- 70When testing for homogeneity, do you break down the test into groups or conduct it overall?Break into groups
- 71How do you read Levene’s test?1. Look at first row
2. If the sig. level is less than 0.05, thats a problem as it indicates that the assumption has been violated. If above, the assumption has been met. - 72What does a violation of the assumption tell us & where can we find more information about that?That the level of variability between the two groups is significantly different.
Comparing groups using the main statistics for variability
1. Standard deviation
2. Range - 73If the standard deviation and range is greater in one group vs another, what does that tell us?Greater variability in the group
- 74When does the issue of variability often come up in psychological research.When comparing a clinical group (e.g., group with schizophrenia) vs a control group. The clinical group is likely to have a lot more variability.
- 75Where can we see the difference between the two group graphically?Box and whisker plots
(e.g., whisker length, outliers, IQR) - 76How do you fix homogeneity of variance?1. Strategies used for normality may work for homogeneity
2. Using a power transformation - 77What is a power transformation & where can we find the information to complete this?Raising to the power of a particular number
Using the fine print on the spread vs level plot [power transformation no.] - 78What are the steps to complete a power transformation1. Transform
2. Compute variable
3. Name
4. ** (means raise to the power) followed by number obtained from the spread vs level plot
5. OK - 79What are some caveats re transformations?(1) They are no “magic bullet”.
(2) They can’t cope with zeroes or negative numbers (can't take square/log10 etc). Need to add a constant.
(3) They are unpredictable (e.g., a transformation designed to deal with normality can fix up homogeneity of variance and vice versa)
(4) Some data is “untransformable” - 80What do you do if transformation don't work?1. Work with your original data using parametric procedures that don't rely on the assumptions
2. Use procedures anyway but draw to the attention of the reader that you're aware it is problematic & to interpret with caution - 81How do you add a constant in the case of O's1. Transform
2. Compute variable
3. Add target variable
4. Add variable plus a constant (e.g., beck + 10) into numeric expression
Results: Produces a new variable with 10 added to every variable. - 82How do you use the transformed data?Only use it for the null hypothesis significance testing because that's why we've done the transformation. We did the transformation to deal with an assumption that underlies statistical tests.
Original variable is used for descriptive stats and graphs - 83Why do you use the original variable for descriptive stats and graphs?1. Makes sense to the reader (conveys info better)
2. Assumptions have nothing to do with the descriptive stats - 84What do you do with a scale or continuous variable with extreme skew that you need to transform but cannot? (e.g., cigarette smoke)Convert it into something more ordinal or categorical
Examples:
1. Smokers or non-smokers
2. Three groups (non smokers, light smokers and heavy smokers depending on the numbers). - 85What command allows you carry out data manipulation such as computing new variables from old variables, adding things, and completing transformations?Compute command
- 86What is data recoding?A process by which you take the original variable and you change or recode the original value.
- 87Why do we recode data?1. T1 EN scores were coded into two groups of high and low Neuroticism
2. T2 when a likert scale scale was reversed to mirror wording
3. Reducing numbers of groups (e.g., recoding education due to limited numbers in groups) - 88During what part of the process is recoding conducted?During screening and cleaning as part of EDA.
Information
- Description
- Flashcards on Lecture 3, created by Samar Castelli on 15/06/2018.
- Created by
- Samar Castelli
- Views
- 0
- Rating