Epi Exam 3 Flashcards
null hypothesis (Ho)
- research perspective which states there will be no (true) difference between the groups being compared
- most conservative and commonly utilized
- various statistical-perspectives can be taken: superiority; non-inferiority; equivalency
3 key attributes of data measurments
- order/magnitude
- consistency of scale/equal distances
- rational absolute zero
nominal (dichotomous/binary; non-ranked; Named categories)
- NO order or magnitude
- NO consistency of scale or equal distances
are simply labeled-variables without quantitative characteristics
anything that has 2 categories
ordinal (ordered; rank-able categoreis: non-equal-distance)
- YES order or magnitude
- NO consistency of scale or equal distances
interval/ratio (order and magnitude and equal distances)
Interval: arbitrary zero value (zero doesn’t mean absence)
Ratio: absolute (rational) zero value (zero means absence of measurement value; e.g. physiological parameters)
- YES order or magnitude
- YES consistency of scale or equal distances
e.g. age in years, body weight, height, temperature
usually see units associated with this type of data
discrete data
categorized data
continuous data
continuous evenness of spacing
changing levels of measurement of data
can go down in specificity/detail of data measurement (levels) but never up
descriptive statistics
non-comparative, simple description of various elements of the study’s data
measures of central tendency (dispersion or spread)
- mean/median/mode
- minimum/maximum/range
- interquartile range (IQR)
variance
average of the squared-differences in each individual measurement value (x) and the groups’ mean (xbar)
standard deviation
square root of variance value (restores units of mean)
normally distributed graphical representation of data
- symmetrical
- mean and median are near equal
normal curve: area under the curve
- 1 SD
- 2 SD
- 3 SD
- 1 SD: 68%
- 2 SD: 95%
- 3 SD: 99.7%
positive skewed
- distribution is skewed anytime the mean differs from the median
- when mean is higher than median
- tail pointing to the right
negative skewed
- when mean is lower than median
- tail pointing to the left
kurtosis
- a measure of the extent to which observation cluster around the mean
- for a normal distribution, the value of the kurtosis statistic is 0
positive kurtosis
more clustered
negative kurtosis
less clustered
do not calculate ____ for discrete data
mean
4 questions to select the correct statistical test
- what data level is being recorded: magnitude and consistency of scale
- what type of comparison/assessment is desired: correlation; regression; survival comparison (time); group comparison
- how many groups are being compared: 2 or 3+ study groups
- is the data independent or related/paired [data from the same (paired) or different (independent) study groups]
correlation (r)
- buzz words: correlation; association/relationship
- provides a quantitative measure of the strength and direction of a relationship between variables (-1 –> +1)
partial correlation
(only for interval data/pearson correlation; with assumptions)
-a correlation that controls for confounding variables
correlation test for:
- nominal
- ordinal
- interval
- nominal: contingency coefficient
- ordinal: spearman correlation
- interval: pearson correlation (p>0.05 just means there is no linear correlation, may still be a non-linear correlation)
kappa statistic
correlation test showing relationship of agreement between/consistency of ‘decisions’ ‘determinations’
+1: the observers perfectly ‘classify’ everyone exactly the same way
0: no relationship at all between the observers ‘classifications’ above the agreement that would be expected by chance
- 1: observers ‘classify’ everyone exactly the opposite of each other
regression
- outcome prediction/association
- provide a measure of the relationship between variables by allowing the prediction about the dependent, or outcome, variable (DV) knowing the value/category of independent variables (IVs)
- also able to calculate odds ratio (OR) for a measure of association and control for confounders
regression test for:
- nominal
- ordinal
- interval
- nominal: logistic regression
- ordinal: multinomial logistic regression
- interval: linear regression
survival tests
- compares the proportion of events over time; event-occurrence; or time-to events, between groups (ongoing progression)
- commonly represented by a Kaplan-Meier curve
- e.g. number of days until event occurs
survival test for:
- nominal: Log-rank test
- ordinal: cox-proportional hazards test
- interval: kaplan-meier test
independent nominal data:
- 2 groups of independent data
- 3 or more groups of independent data
- 2 or more groups with expected cell count less than 5
- 2 groups of independent data: (Pearson’s) Chi-square test
- 3 or more groups of independent data: Chi-square test of independence
- 2 or more groups with expected cell count less than 5: Fisher’s Exact test
post-hoc testing for independent nominal data with 3 or more groups of independent data
- bonferroni test of inequality (bonferroni correction)
- bonferroni adjusts the p value for number of comparisons being made
paired/related nominal data:
- 2 groups of paired/related data
- 3 groups of paired/related data
- 2 groups of paired/related data: McNemar test
- 3 groups of paired/related data: Cochran followed by bonferroni to determine where significant difference was found
independent ordinal data:
- 2 groups of independent data
- 3 or more groups of independent data
- 2 groups of independent data: Mann-Whitney test
- 3 or more groups of independent data: Kruskal-Wallis test
- both compares the median values between groups
- if 3+ group comparison significant –> post-hoc test
paired/related ordinal data:
- 2 groups of paired/related data
- 3 or more groups of paired/related data
- 2 groups of paired/related data: wilcoxon signed rank test
- 3 or more groups of paired/related data: friedman test
- both tests compares the median values between groups
- if 3+ group comparison significant –> post-hoc test
ordinal data:
post-hoc tests for 3 or more group comparisons
- student-newman-keul test: all groups must be equal in size
- dunnett test: every group compared to one single control (center of wheel example)
- dunn test: useful when all groups are not the same size
independent interval data:
- 2 groups of independent data
- 3 or more groups of independent data
- 2 groups of independent data: student t-test
- 3 or more groups of independent data: analysis of variance AN(C)OVA
- both tests compares the means of all groups (ANOVA against a single DV)
- if 3+ group comparison significant, must perform a post-hoc test
paired/related interval data:
- 2 groups of paired/related data
- 3 or more groups of paired/related data
- 2 groups of paired/related data: paired t-test (compares the mean values between groups that are related)
- 3 or more groups of paired/related data: repeated measures AN(C)OVA (compares the means of all groups (along with intra- and inter-group variations) of related data against a single DV)
-if 3+ group comparison significant, must perform a post-hoc test to determine where differences is (are)
interval data:
post-hoc tests for 3 or more group comparisons
- student-newman-keul test: all groups must be equal in size
- dunnett test: every group compared to one single control (center of wheel example)
- dunn test: useful when all groups are not the same size
- tukey (slightly more conservative than the Stu.N.K.) or scheffe (less affected by violations in normality and homogeneity of variances-most conservative) tests: compares all pairwise comparison possible; all groups must be equal in size
- bonferroni correction: adjust the p value for number of comparisons being made