This assignment counts for 40% of the total grade for the subject. There are 100 marks for the assignment. Including figures and tables, it should be no more than 10 pages.
Your assignment should be in Times New Roman 12 point font. Please put your student number in the header of each page of every document you submit. Do not put your name on any of the documents you submit. It is to be submitted through the LMS.
You are permitted to discuss this assignment with other students, but you must do the analyses by yourself and write your report without consulting with others.
The data for the assignment are from a randomised controlled trial of patients with tuberculous meningitis. The data were used for one of the interim analyses of the trial after approximately nine months of follow-up since randomisation. Assume that everybody randomised is included in the dataset. The data are different from those used for the publication of the trial results.
The primary aim of the trial was to determine whether patients with tuberculous meningitis had better survival if they were given dexamethasone (a corticosteroid) in conjunction with standard antituberculosis treatment for their disease.
The secondary aim was to determine whether any beneficial effect varied according to the severity of the patient's disease at randomisation as measured by the Glasgow Coma Scale.
Patients were randomised to receive dexamethasone or placebo within strata defined by the hospital in which they were treated and the severity of their disease (Glasgow Coma Scale). The Glasgow Coma Scale has a range of scores from 3 (most severe) to 15 (least severe). Patients were categorised into three groups:
|
Severity
Grade
|
Description
|
|
I
|
Glasgow Coma scale of 15 with no focal neurologic signs (least severe)
|
|
II
|
Glasgow Coma scale of 11 to 14, or of 15 with focal neurologic signs
|
|
III
|
Glasgow Coma scale of 10 or less (most severe)
|
Apart from the hospital and the severity of the disease, the only other prognostic variable included in the data file is each patient's HIV status at entry to the study.
Section A
You have been employed as the trial statistician and are to analyse the data to address the primary and secondary aims of the trial. You are also required to present your results ready for publication in the New England Journal of Medicine.
Please use these headings for your report:
The discussion is limited to the subheadings above. Don't discuss other studies on the topic.
Feel free to consult the CONSORT guidelines and/or publications from similar randomised controlled trials for how to prepare your report. Don't copy the publication from the trial, because that would be plagiarism. (The data are also not the same as in the publication.)
Dallas's word count was about 850 words, and he included three tables and one figure in his report. (You should aim to include no more than 1000 words in your report.)
At the end of this section, please include the Stata commands you used. Do not include a log file. Please edit the commands so that only those essential to reproduce your results are included.
Section B
B1 If you did not have the original data, but you were given a Kaplan-Meier curve, how could you determine whether there was only a little censoring or a lot of censoring?
B2 There is one variable (called "X" below) that does not satisfy the assumptions of Cox regression. For this variable, show an appropriate plot that indicates it violates the assumption and present results of Cox regression analyses exploring how its hazard ratio varies with time since randomisation. Paste the commands you used into your answer (no log file please). Briefly summarise your findings of how the association between X and death varies by time since randomisation.
B3 Split the data by time since diagnosis into five intervals with approximately equal numbers of deaths in each interval. Include the commands you used in your answer (not the log file please). Plot the crude mortality rates according to this new variable. How does the mortality rate vary with time since randomisation?
B4 Use Poisson regression to analyse the split data, fitting an interaction between the new timeband variable (fitted as a categorical variable) and X. Include your commands in your answer (again, no log file). What are the hazard ratios for X within each of the timebands? How do these compare with the hazard ratios from the Cox modelling you did in B2?