What comparison will you use when evaluating continuous data with a normal distribution?
The means. Use the unpaired or paired t-test.
What comparison will you use when evaluating continuous data without a normal distribution?
The medians. Mann-Whitney WIlcoxon test (unpaired) or signed rank test (paired)
What comparison will you use when evaluating categorical data with ordinal data?
Medians or proportions. Use medians for larger sample size.
What comparison will you use when evaluating categorical data that is not ordinal?
Proportions. Z test/Chi square.
What is a classic case of paired data?
Measuring the same person twice (pre and post measurements)
Cross-over trial. THIS WILL BE ON TEST.
Patients get placebo treatment and drug treatment and reactions are compared. Order is randomized.
Matched case-control study
Measuring two people from similar demographics
Measuring reactions between identical twins
Mann-Whitney Wilcoxon Test
Used to compare medians for data without a standard distribution.
Tests where you are not worried about the distribution. Good with ordinal data. You don't have to worry about outliers. Less powerful than a t-test in normal distributed data.
Wilcoxon signed rank test
Used with paired continuous data that is not normally distributed
difference in proportions/pooled standard error
Chi square test
((observed frequencies-expected frequencies)^2)/ expected
How can you get the Z statistic from the Chi Square statistic?
Take the square root of your Chi Square value
Why do you use a Chi Square test?
Comparing unpaired proportions. To prove that two proportions are the same. Use in RR, OR, AR and prevalence ratio.
Fisher's exact test
Comparing unpaired proportions. Use for 2x2 tables dealing with small samples. Asks how many 2x2 table values are likely to be more extreme than the values you got.
Paired and unpaired tests
McNemar's Chi Square
Determines the ratio of people switching from no symptoms to symptoms. Allows for prevalence comparison.
Comparing unpaired data in 3+ groups
Use analysis of variance for means, Kruskal-Wallis for medians and Chi square for proportions
Multiple comparisons problem
The more t-tests you do with an alpha of .05 accumulates the probability of making a type I error.
Says you are okay with 5% probability of error across the entire test. So in a study of 12 different tests your criteria becomes more strict (5%/12). Increases probability of type II error.
Says you are unlikely to get results due to chance. You could still get your result due to bias. Based on clinical significance, not clinical importance.