Choosing an analysis method Flashcards
Why is it important to know how to choose an analysis method?
When looking at research papers you should consider if the right method was used or not
Which (2) factors should be considered when choosing an analysis method?
- Purpose of analysis (e.g. comparing groups, assessing agreement between variables OR assessing association between variables)
- Types of variables (continuous, categorical, discrete)
If there are only two groups of data what can they be:
Paired
Unpaired
What is an unpaired comaprison?
Looks at group as whole
e.g. compare salivary buffering capacity between males and females
What is a paired comparison?
Looks at changes in the same person -> can only be done if have 2 sets of measurements linked in some way
e.g. compare salivary buffering capacity before and after recieving a dietry advice leaflet
Is measurement of salivary buffering capacity in twins paired or unpaired?
Paired = 2 measurements from same person of drom different people who are related in some way (e.g. twins or siblings)
If there are more than two groups what can the data/comparison be?
Matched
Unmatched
What is an unmatched comparison?
Comparison of 2+ groups
Each person in sample gets 1 of 3 treatments e.g. salivary fluoride levels after application of fluoride varnish, fuoride rinse or placebo (no treatment)
What is a matched comparison?
Comparison of 2 + groups
3 measurements give to same person at different times
e.g. repeated oral quality of life assessment at different ages (childhood, adolescence and adulthood)
What do we use to chose the appropriate analysis method?
Table
What do parametric methods assume?
Underlying distribution
What does ANOVA for testing multiple groups avoid?
Problems of multiple testing (i.e. t test for each group) = more likely to find an association by chance because smaller groups
When Anova is used for only two groups does it produce the same result as a t-test?
Yes
What is the difference between a paired and unpaired t test?
Unpaired -> based on the differences in the mean
Paired -> mean values of the diferences between groups
What is the advanatge of unpaired t test?
Both groups are not required to be equal (can have either the same no of M & F or can have different numbers of M & F)
What do both t-tests an ANOVA’s assume?
Equal varience assumption across the groups (must be checked to ensure the results are valid)
n.b. if the no in each group is similar t tests and ANOVAs still robust but if very different the test needs to be modified
Which tests are carried out on continuous normal outcomes/ parametric data?
Unpaired & paired t tests
one way & repeated measures ANOVA
What does ANOVA analyse?
Varience
Which tests can be carried out on continuous non normal (cannot be transformed to approximate normality)/ non parametric outcomes?
Mann-Whitney U / Wilcoxon two sample test
Sign test / Wilcoxon signed rank test
Kruskal-Wallis test
Friedman test
Which tests are preferrable parametric or non parametric for continuous data?
Parametric = more powerful
How do non parametric tests work?
Based on ranking observations in order of magnitude
Test rankings
Which types of tests can be used for non parametric categorical outcomes?
Chi squared
Fishers exact test
McNemars test
Ordinal Chi squared test
Cochrane Q test
When should a fishers exact test be used?
If cells in 2x2 table only have small numbers in them (i.e. observed or expected is 5 or less)
When should a chi squared test be used?
Nominal or unordered variable
When should an ordinal chi squared test be used?
if outcome is an ordinal categorical variable
To test the null hypothesis that there is no difference in salivatry fluoride levels between maes and females, which of the following is the most appropriate test and why?
Paired t test
Unpaired t test
ANOVA
Chi squared test
Unpaired t test -> tests 2 independent mean values (assuming the salivayr fluoride levels are or can be transformed to aproximate normality)
NOT:
paired t test-> testing 2 non-independent means
ANOVA-> for testing more than 2 independent means
Chi-squared test-> for testing 2 independent proportions
How do we assess agreement between variables?
Compare measurements mde by 2 or more examiners (i.e. dentist)
Aim is to obtain the same value from the 2 sets of measurememts
n.b. it is possible for 2 sets of measurements to be stronly associated but have low agreement (i.e. 1 examiner may consistently score higher than the other)
= can be extended to compare each of any number of examiners with gold standard
Which tests do we use to assess measures of agreement?
Normal continuous data:
Limits of agreement
Categorical data:
Kappa (unordered data)
weighted kappa (ordered data)
Sensitivity/Specifity
What is a gold standard?
Where method is widely accepted as being best availiable
What is a bland altman plot?
Continuous data
Calculate the difference between two sets of measurements for each subjects and plot against the mean of two measurements
What are limits of agreement?
Limits of agreement =
Mean difference +/- 2 SD of the difference
this can be added to the bland altman plot
n.b. if differences and means are related (e.g. proportional) transform the data before analysis
What are limits of agreement used to judge?
Whether differneces are clinically important
(can obtain confidence intervals)
What is the analysis of agreement in categorical data based on?
Comparing observed proportion of agreement with proportion of agreement that would be expected by chance
Value of Kappa decreases as no. of categories increases (more opportunities for misclassification)
What does the weighted kappa allow for?
Includes partial misclassification into adjacent categories = less of a concern than missclassification into non-adjacent categories!
What does a Kappa score of 1 mean?
Perfect agreement
What does a Kappa score of 0 mean?
No agreement
What are the criteria for assessing magnitude of Kappa?
>0.75 excellent agreement
0.4-0.75 fair to good agreement
<0.4 poor agreement
Why are p values not very useful for expressing Kappa?
What is used instead?
A null hypothesis between agreement and not is unreasonable
Standard errors & confidence intervals
What is sensitivity?
proportion with outcome who are correctly classified by test as a proportion of those who realy do have outcome as determined by the gold standard= indication of accuracy of test in detecting those with outcomes
What is specifity?
Proportion without outcome who are correctly classified by test (indication of accuracy of test in detecting those without outcomes)
When is sensitivity and specifity use to measure categorical data?
In the special cases where there s both a gold standard and diagnostic test which determines if an individual has or has not got the outcome
What does sensitivity and specificity determine?
Assess the accuracy of a dignostic test
How do we measure association betweem variables?
Correlation
OR
Regression
How do we measure correlation?
Pearsons (parametric)
Spearman (non-parametric)
When is pearsons suitable for use?
when you have 2 continuous variables with a normal distribution and of equal importance
When do we measure regression?
When variables are not of equal importance
Parametric & non-parametric (rarely used)
The choice of statistical analysis method depends on which of the following?
- The purpose of the analysis
- The types of variables
- Both of these
Both
Which of the following is the odd one out?
- Discrete data
- Ordinal data
- Binary data
- Nominal data
- discreet data -> it is the only one that is not categorical
Height measurements on males and females are which of the following?
- Paired data
- Unpaired data
Unpaired
A paired t test is based on which of the following?
- Mean difference
- Difference in means
Mean difference
Non-parametric tests should always be used in reference to parametric tests… True or False?
False
Parametric tests are based on actual values = more statistical power
The Chi-squared test is based on which of the following?
- Observed values
- Observed and expected values
- Expected values
- Observed and expected values
It is used to see if there is an association between 2 factors e.g. gender and group
Is correlation an appropriate method for assessing agreement?
No
Need bland-altman plot to determine agreement
Can be correlated but not agreed
Limits of agreement can be used to assess whether the differences between two sets of continuous measurements are:
- Clinically acceptable
- Statistically significant
- Clinically acceptable (do they disagree to the extent it is a problem?
What is the maximum value possible for a kappa statistic?
1
Regression is an appropriate analysis method in which of the following situations…?
- Two variables of equal importance
- One variable can be identified as an exposure and one as an outcome (of differing importance)
2.
Which is the first data check that should be carried out before deciding whether a parametric or non-parametric method should be used?
Normal distribution
What do the standard errors of Kappa statistics indicate?
Precision of the estimates