oral exam Flashcards
Explain the difference between a risk ratio and an odds ratio.
Risk ratio is the likelihood of developing a condition (disease) under certain conditions (exposed or control). Calculated by adding rows together. AR = [A/(A+B)] & AR=[B/C+D)]
RRRatio or RR reduction (AR/AR)
Odds ratio is the probability of having the condition or not. calculated by adding columns Odds= [A/A+C)]
OR = AD/BC
Name one type of model you could run if you were comparing two groups and your dependent variable was dichotomous.
A chi square test will compare two dichotomous variables
A statistical model that can compare dichotomous outcome variables is a logistic regression
What is the difference between ratios of relative risk and absolute risk? Why is the latter important to always report when possible?
Absolute risk is the likelihood of developing the condition under a specific set of circumstances
relative risk is risk that the experimental group incurred in comparison to the control group.
Comparing the relative or increaesd risk back to the baseline absolute risk puts the change in risk in proper proportion. If the baseline incidence of a condition is very low, even a high relative risk will only shift the absolute risk a smaller amount. For example AR is 6% with 50% RR then the new risk is 7.5%; whereas if the AR was 20% and a RR of 50% then the new risk is 30%.
RR= ARe/ARc
where ARe= A/(A+B) and ARc = C/(C+D)
In a clinical trial with more than one time-point (e.g. baseline, 6 weeks, 6 months), what does it mean to adjust for multiple comparisons? What is the name of a test you could use to adjust for multiple comparisons?
Each time the analysis is run, in this case for different time points, the likelihood of a type I error or false + finding increases. So a correction is run to accomodate this. A conservative test such as the Bonferroni post hoc test will correct for this.
Conservative tests for confirmation trails; liberal tests for explanatory trials
If two therapists are independently measuring 100 knees to determine if each knee maximum flexion ROM is either above or below 90 degrees, and the agreement between them is equal to a Kappa of 0.92, what does this mean?
Kappa is a measure of agreement between raters. kappa takes into account the agreement beyond chance. In this case because there are two raters the statistic used is Kohen’s Kappa, for 3 or more raters it is a weighted kappa. .92 means almost perfect agreement.
.81-.1.00 is almost perfect agreement .61-.80 substantial .41-..60 moderate .21-.40 fair .01-.20 slight
When you have missing data in a trial, you want to know if the data is missing at random (or if there is a specific pattern. You run a Little’s Missing Completely at Random (MCAR) test which gives you a value of P=0.382. What do you infer from these results?
Missing completely at random means that there is no relationship between missing data and any of the values. In this case we reject the null hypothesis which indicates the data is missing completely at random.
If you are comparing costs of care between two treatment groups in a trial, which is likely the better model to use and why? A T-Test or a generalized linear model?
A generalized linear model is more flexible in that the data does not have to be normally distributed and in this case cost data is unlikely to be so. A gamma distribution (scale data) generalized linear model is more appropriate as cost is
When considering the types of patients you want to include in your study, you have to think about heterogeneity in your design and in your statistical plan. What does this mean?
Heterogeneity refers to differences between the groups. From a study design perspective sound research design
in setting up study conditions and and randomization can minimize this.
Statistically this is measured with Levene’s test for homogeneity of variance. The null hypothesis is that there is no difference between groups, variance is equal. So p>.05 means equal variances assumed
What is a confidence interval?
The range of values in which statistical confidence that the true values lies. For example a 95% CI means the range of values that the test mean will fall into 95 of 100 trials.
CI estimates the precision of the data, smaller CI means more precise estimates.
Also can measure the magnitude of the difference between groups.
CI that cross the summary statistic are not significant
if group CI overlap by < 25% then they. are probably significant
What does it mean when the confidence interval crosses 0 or 1?
For ratio’s such as odds ratio, risk ratio then crossing 1 means no significant difference
for comparison measures when CI crosses 0 then no significance
When you are assessing predictor variables for your prediction model, you have to consider multicollinearity between the variables. What does this mean?
multicollinearity indicates that the independent variables are related to each other; only performed with multiple linear.
r value is the correlation statistic–> .2
no more than 10% increase in st. error as new variable added to the equation
What does it mean when you have wide confidence intervals for the treatment effect?
A wide confidence interval indicates less precision, greater uncertaintity in the results. This may be due to a small sample size.
meta analysis indicates greater heterogeneity in studies.
If CI crosses the mean of the other test statistic then. it is not significant
Why is it better to see the confidence intervals of a point estimate than it is to see the p- value?
p values can only distinguish if there is a statistically significant difference. CI can identify
- significance (cross 0,1)
- direction
- strength of the effect (width).
For randomized clinical trials, why is it controversial to put the p-value for the differences between baseline variables in your Table 1?
Some journals require and others don’t. The argument is that proper randomization will decrease variablity in groups and so if there are differences these are by chance and even if there is a difference, it does not change how the study is conducted.
What is the difference between prediction and causation?
Prediction studies
In a prediction analysis, the goal is to develop a formula for making predictions about the dependent variable, based on values of the independent variables. Regression studies
In causation analysis, the independent variables are regarded as causes of the dependent variable. RCT’s
Causation indicates that one event is the result of the occurrence of the other event; i.e. there is a causal relationship between the two events. Also referred to as cause and effect. In a causal analysis, the independent variables are regarded as causes of the dependent variable. (ADDED 8/11/2020: RCT’s aim to prove causality)
(Causation is not to be confused with correlation! Correlation indicates the amount which two variables move together.)
Prediction is the act of forecasting what will happen in the future. A prediction is specific to the study and experiment that you design to test your hypothesis. It’s the outcome you would observe if your hypothesis were supported.
In a prediction study, the goal is to develop a formula for making predictions about the dependent variable, based on the observed values of the independent variables
Prediction is simply the estimation of an outcome based on the observed association between a set of independent variables and a set of dependent variables. Its main application is forecasting. Causality is the identification of the mechanisms and processes through which a certain outcome is produced.
“
When planning a clinical trial with 2 comparison groups, if you increase the magnitude of the effect size you expect to see between two groups, what does that do to your sample size estimate (how does it affect your power)?
The sample size estimate will decrease in order to meet the expected power level. If the sample size remains the same there will be increased power in the study.
power = 1- beta (beta is “c” in 2 x2 table)
sample size n=30 p=.05
80% power indicates willing to acept 20% type II (false -) and 5% type I false + rate
90% power indicates 10% type II and 5% type I
If you set your expected effect size for treatment between two groups to be large, but the effect size between the groups ends up being only moderate, what will your findings likely be?
The study may be underpowered as the sample size will not have been large enough resulting in a potential type II error–rejecting the null hypothesis when in fact there was a difference.
may not reach statistical significance
A cohen’s d effect size most often refers to ratio-level data (e.g. difference in means). How would you measure the effect size for dichotomous outcomes?
Odds ratio or risk ratios
OR = column totals; odds of developing the condition A/(A+C)
RR= row totals; risk of getting disease based on exposure or not; ; A/(A+B)
partial eta2 is the effect size
What does it mean to have a negative correlation? Provide an example.
A negative correlation means as one variable increases in magnitude the second decreases.
-1 is perfect negative correlation 0 = no correlation; . 5 low and .3 neglible
Example: heating costs in summer in Alaska. As temperature increases, there is less heat used.
increase in exercise time decrease in body fat percentage
Explain why plotting your data before making any decisions about which analysis to use is an important first step.
Plotting the data allows to look for any patterns in missing data points, check for normality and see outliers.
