Hi. Welcome to the biostatistics Q&A session. I'm Leslie Durden from the University of Washington and Seattle Children's Hospital. So, biostatistics questions are a favorite on the boards because they are questions that are relatively easy to write with a definitely correct answer, and they're really not that hard to answer. So, we will cover lots of examples here. Question 1. In a landmark randomized controlled trial, shock be gone, cut mortality in half among children with multiple organ failure. What is the number needed to treat with shock be gone to prevent one death? So, the correct answer is E, not enough information to determine. The reason for this is that number to need to treat is equal to 1 divided by the absolute risk reduction. Absolute risk reduction is the difference in absolute risk from group A minus the risk in group B. To answer this, we would need the absolute risk in the children in each group, not just the relative risk, which is 0.5. Which brings us to question 2. In the landmark randomized controlled trial, shock be gone, decreased 28-day mortality from 60% to 30%. What is the relative risk of mortality with shock be gone? If you answered C, 0.5, you would be correct. The relative risk is the risk in one group divided by the risk in the other group. Usually, the treatment is the group A in the numerator and the control is the group in the denominator. In this case, the relative risk is 0.3 divided by 0.6, which is 0.5. You have to be careful of which group is the reference. And one way to check yourself is that anything that decreases mortality should have a relative risk of less than 1. Question 3. In a landmark randomized controlled trial, shock be gone, decreased 28-day mortality from 60% to 30%. What is the number needed to treat? So the correct answer is 3.3. To review, number needed to treat, again, is 1 divided by the absolute risk reduction. Absolute risk reduction is the absolute risk in one group minus the absolute risk in the control group. In this case, this is 1 divided by 0.6 minus 0.3, which is 3.33. In another randomized controlled trial in a population with a baseline or control mortality of 10 percent, the relative risk of mortality with shock beyond was again 0.5. In this study, what would be the number needed to treat? If you answered C, you would be correct. The reason for this is that the absolute risk reduction is much smaller than it was in the previous question. The absolute risk reduction here is 10 percent minus 5 percent. We obtain the 5 percent because the relative risk of 0.5 would be multiplied by the control mortality to give you the 5 percent risk of mortality in the treatment group. The difference of these is 5 percent, and 1 divided by 5 percent is 20. Question 5. The chance of being crushed and killed by a properly installed vending machine is 1 in 100 million. However, the risk increases by 100 times if the machine is installed without leveling the legs. How many vending machines would have to be installed without leveling the legs to kill one additional customer on the New Jersey Turnpike? If you answered D, you would be correct. In this example, the risk in the improper installation is 1 in 1 million minus the risk with proper installation, which is 1 in 100 million. If you convert those to decimals, they are very, very tiny numbers, and you can subtract them. 1 divided by a very, very tiny number is a very big number. For question 6, we are going to use a 2-by-2 table based on this information for the next five questions. Meningocorp has developed a test for meningococcal infection. In a study, half of 200 medical student volunteers are infected with Neisseria meningitidis. Of the infected, 90 have a positive test, and of the uninfected, 60 have a positive test. What is the sensitivity of the test? If you answered E, you would be correct. First, we will use the given information to construct our 2-by-2 table. Out of 200 people in the study, half of them were infected, so that gives us the 100 columns of information that we need to construct our 2-by-2 table. Out of 200 people in the study, half of them were infected, so that gives us the 100 column total under disease present and the 100 column total under disease absent. Of the infected, 90 have a positive test, so that goes in that section, and 60 have a positive test out of the uninfected, so that goes in that area, then lets you complete the rest of the table. Using that information, we can calculate the sensitivity of the test. As a reminder, sensitivity is the percentage of people with the disease who have a positive test, or the probability of a positive test given presence of the disease. This is 90 percent, 90 out of 100. Next question, what is the specificity of this test? If you answered 40 percent, you would be correct. Looking at our 2-by-2 table, specificity is the percentage of subjects without disease who have a negative test, or the probability of a negative test given absence of disease. In this case, that's 40 out of 100, or 40 percent. Question 8, same question STEM, what is the positive Question 8, same question STEM, what is the positive predictive value? If you answered 60 percent, so the positive predictive value is if the test is positive, the chance of having a disease, or the probability of disease positive given that the test is positive. In this case, that's 90 out of 150, which is 60 percent. Question 9, what is the negative predictive value? This is 80 percent. Looking at our 2-by-2 table, the negative predictive value is the chance of being disease-free if your test is negative, or the probability that the disease is negative given that the test is negative. In this case, that is 40 out of 50, which is 80 percent. Question 10, what is the positive likelihood ratio of the test? We didn't discuss this in the other slides, so I will go ahead and tell you that the positive likelihood ratio is the sensitivity divided by 1 minus the specificity, or the probability of a person with disease testing positive divided by the probability of a person without disease positive. Using my other notation, the probability that the test is positive given that the disease is positive divided by the probability that the test is positive given that the disease is negative. Based on that, the answer is 1.5. Question 11, the positive predictive value of a test depends on which of the following? The positive predictive value of a test depends on which of the following? Frequency of disease in the population studied, the specificity of the test, the cost of the test, the number of times the test is run, or none of the above. Positive predictive value depends on the frequency of disease. If I have a disease that is 50 percent prevalent, using the example from the prior questions, with a 90 percent sensitivity and a 40 percent specificity, I get a positive predictive value of 60 percent. If the prevalence drops to 9 percent with the same sensitivity and specificity, I now have a positive predictive value of only 13 percent. Question 12, an investigator follows two groups of children in the PICU, one with femoral central lines and one with IJ central lines. She performs weekly ultrasounds to determine whether rates of DVT differ between the two groups. This type of study is best called A. If you chose cohort study, you are correct. The study design here is that children were chosen and then compared based on their exposure status to femoral versus IJ. Exposure status to femoral versus IJ central lines with DVT being the outcome of interest. You would be correct. Question 13, you want to determine if length of stay in children with TBI is associated with admission cortisol level when controlling for presence of an EVD, admission GCS, and age. Which of the following analyses would make the most sense? If you chose linear regression, you would be correct. In linear regression, the dependent variable or the outcome, in this case length of stay, is a continuous variable. That's why this is chosen over logistic regression. In our proposed study, we would want to look at cortisol level, which could be considered the independent variable, and we have multiple potential confounders that we want to adjust for, including prism, EVD, age. The linear regression is going to be the best match for those questions. Question 14, you want to determine if risk of stroke is affected by mode of ECMO, VA versus VB, when controlling for duration of ECMO, heparin dose, PFTT, and ECMO flow rate. For this question, which type of analysis would be most appropriate? In this case, logistic regression is going to be the best approach. You're looking at a binary outcome with multiple potential confounders that you want to adjust for. In this case, the dependent variable or outcome is stroke, which is dichotomous. It either happens or it doesn't, and the independent variables are all the other things that you think might be important that you want to adjust for in your analysis. Question 15, in your study of factors associated with the risk of stroke and ECMO, the logistic regression model output is below. Variables include VV versus VA ECMO, PTT on the day of stroke, and gender, with the beta coefficients and p-values listed. What is the odds ratio for stroke in girls versus boys when controlling for ECMO mode and PTT? In logistic regression, it's important to know that the odds ratio associated with each factor is the exponentiated beta coefficient. In that case, that makes B the correct answer because e to the zero power equals one. Question 16, an investigator wishes to conduct a 14-center RCT of VV ECMO for severe ARDS in children. Which of the following statements is true? The trial cannot be conducted because blinding is not feasible. Every other patient will be assigned to ECMO. Patients who are randomized to ECMO but can't be cannulated because of abnormal neck vessels should be analyzed with the non-ECMO patients because they can't get ECMO. D, if ECMO is found to be superior than all centers in the U.S. who care for children with severe ARDS should use it, or E, none of the above? The correct answer is E, none of the above. A is not the correct answer because some studies can't be blinded. You really wouldn't be able to do a sham ECMO surgery with any kind of IRB approving it. Just because you can't blind doesn't mean that you can't conduct a effective trial. A randomization scheme in which every other patient is assigned to ECMO would actually allow providers to guess whether or not the next patient would be assigned to ECMO. That might influence their behavior as to whether or not to refer them for enrollment. So you wouldn't want to do that. All randomized trials should use an intention to treat analysis where we analyze patients with the group that they are randomized to regardless of whether or not they receive the therapy. If you don't analyze patients that way, it breaks the randomization scheme and can introduce bias into your study. This is a question about external validity or generalizability. Non-ECMO centers may not be able to achieve the same outcomes as ECMO centers, so you certainly wouldn't want to say that based on the results of this study that all centers should suddenly become ECMO centers. Their results would likely be quite different from this RCT. Question 17. On average, the serum cortisol level obtained from a clinical laboratory is very close to the true physiological level in the blood. That would mean that this test has a high degree of what? This means that the test is very accurate. Accuracy reflects the closeness of the value you obtained to the true value, whereas precision is related to how close sequential measures are to each other. Question 18. You want to determine if length of stay in children with TBI is associated with an elevated admission cortisol level. You categorize cortisol level as normal or high. The mean length of stay is seven days and the median is two days. What statistical test would be most appropriate? You're looking here at a skewed data where your length of stay has a mean that's very different from the median, so you know you're going to need a non-parametric test. You are looking at unpaired data because you're comparing children who have high versus low admission cortisol level, not serial cortisol levels in the same set of children, so what you need is an unpaired non-parametric test. The Wilcoxon signed-rank test would be paired non-parametric. The two-way ANOVA with stratification would allow you to use multiple groups and more variables, which you don't need for this situation. The t-tests are both parametric, and linear regression and linear regression is not going to be appropriate because your skewed data is going to fail to meet many of the assumptions needed for linear regression. Question 19. You want to determine if the average blood pressure in children with TBI is associated with an elevated admission cortisol level. You categorize cortisol level as normal or high. The mean blood pressure is 91 and the median is 90. What statistical test would be most appropriate? In this case, a t-test would be the most appropriate. You are looking for a parametric test because the mean versus the median are pretty close together, suggesting that even though we're not looking at the entire distribution of the data, it is likely to be close to a normal distribution, and we're comparing means, so a t-test would be appropriate. Last question. You want to determine if change in blood pressure before and after 3% saline in children with severe TBI is associated with elevated admission cortisol. You categorize the cortisol level as normal or high. Before the 3% saline, the mean systolic blood pressure is 101 and the median is 100. After the 3% saline, the mean systolic blood pressure is 95 and the median is 96. Which statistical test would be appropriate? For this, a paired t-test would be appropriate. You're looking at parametric data, so the means and medians for both the before and after measurement are pretty similar, suggesting that this is likely to be normally distributed data. This is a paired set of data because we are looking at blood pressure in the same set of patients before and after the intervention. Conditional logistic regression would also be an appropriate method for this data. I hope you found these review questions helpful. Good luck, and thank you.

biostatistics

number needed to treat

relative risk

sensitivity

specificity