## Lecture: Stratified Covariate Balancing
## Assigned Reading- Stratified Covariate Balancing (use instructor's last name for password) Read► Slides►
- Code for covariate balancing:
## Assignments
- Visually show that Dr. Smith see a different set of patients than his peer group. Show a tree where the nodes are the diagnoses, the consequences are length of stay within the tree branch, and each branch is drawn proportional to the expected length of stay.
- Balance the data through stratified covariate balancing. Graphically show that the weighting procedure of stratified covariate balancing results in same number of different types of patients treated by Dr. Smith or his peer. Switch the tree structure of peer group (but not the length of stay) with Dr. Smith's tree.
- Report the unconfounded impact of Dr. Smith on length of stay
using the common odds ratio of having above average length of stay.
SQL Common Odds►
For calculating the common odds ratio, calculate for each strata the following contingency table:
Outcomes in i ^{th}stratum, i = 1,2, ..., kAbove Average LOS Below Average LOS Dr. Smith (Cases) a _{i}b _{i}Peer Group (Controls) c _{i}d _{i} Then, calculate the common odds ratio using the following equation:
- Reported the impact of Dr. Smith on length of stay using the weighted length of stay. SQL Weighted LOS►
- Using SQL, group the diagnoses into commonly occurring strata.
- Within each strata, calculate the odds of mortality from cancer.
- Calculate the common odds ratio across strata. Common odds
ratio is calculated using the following formula:
Where a, b, c and d are defined as the following:
Outcomes in i ^{th}stratum, i = 1, 2, ...,kDead in 6 months Not Dead in 6 Months Cases with Cancer a _{i}b _{i}Controls without Cancer c _{i}d _{i}
- Conduct sensitivity analysis for the calculated common odds
ratio. Sensitivity analysis is the process of changing one
variable and re-examining the conclusions. Drop one of the 35
comorbidities from the analysis and repeat the entire analysis and
check that 65% of cases are matched to controls. The percent of
cases that are matched is called overlap. It is defined as:
In most problems, one wants to maximize the overlap to be around at least 80%, so that findings can be generalized to the original cases.
Report how the unconfounded and confounded odds of mortality from stomach cancer are different from each other? Data►
The data report the experience of approximately 4,000 patients with various antidepressants: citalopram, bupropion, mirzapine, buspirone, lithium, nortriptyline, sertraline, thyroid, tranylclypromine, and venlafaxine. The objective of this analysis is to find which antidepressant is ideal for a male patient who has PTSD, neurological disorders, and endocrine disease but no other comorbidity. The patient does not have the following covariates: risk of suicide, heart, vascular, haematopoietic, eyes ears nose throat larynx, gastrointestinal, renal, genitourinary, musculoskeletal Integument, psychiatric illness, respiratory, liver, alcohol, amphetamine, cannibis use, opioid use, panic, specific phobia, social phobia, OCD, anxiety, borderline personality, dependent personality, antisocial personality, paranoid personality, personality disorder, anorexia, bulimia, and cocaine use. The following table shows the distribution of the covariates in the data.
The data are reported for a total of 22,254 visits. Visits may be 2 week or more apart. Not every patient shows for every scheduled visit. Organize the data so there is one row for each patient and each antidepressant trial (known in the data as Concat). Note that this field considers combination of antidepressants as a new antidepressant. Ignore the dose of the medication. Patients received multiple antidepressants during these trials until something worked for them. Include each time a new antidepressant was tried as a separate trial. If the patient has taken the antidepressant at any time during the trial, then mark it as 1, otherwise 0. Notice that some patients have taken the medication and others have not. Patients who have not taken a particular medication have taken other medications, so at any time we are comparing one medication to alternative treatments. The medication is considered to have caused the remission if the patient is referred to follow up portion of the study, at any point while taking the medication; i.e. the variable "Treatment_plan_equal_3" is set to 1 while taking the medication . - Clean the data SQL►
- For 3 antidepressants, balance the data using SQL and stratified covariate balancing.
- Describe which of the 3 medications should a patient who has PTSD and neurological disorders take. SQL►
- Check that all variables are positively and monotonely related to prevalence of diabetes in the county.
- Split variables that are not monotonely related to prevalence of diabetes into 2 or more variables that are positively and monotonely related to diabetes
- Calculate the impact of obesity on diabetes while controlling for the other variables
- Calculate the impact of access to food sources on diabetes while controlling for other variables
## MoreFor additional information (not part of the required reading), please see the following links: - Alemi and Amr's original paper on covariate balancing Pubmed►
- Predictors of response to citalopram Read►
- Does citalopram help anxious depressions Read►
- Collapsing strata Read►
This page is part of the course on Comparative Effectiveness by Farrokh Alemi, Ph.D. Home► Email► |
||||||||||||||||||||||||