Epidemiology Lecture 7-8 – Confounding and Control of Confounding – Flashcards

Addresses the interpretation of observational data - The presence of confounding weakens the case for a CAUSAL relationship between exposure and outcome - May be multiple confounding factors

Laparoscopic splenectomy patients actually had lower post-op mortality compared to open splenectomy patients However, laparoscopic splenectomy patients also more likely to undergo elective surgery, which has lower mortality

12/874 patients in laparoscopic group died within 30 days of surgery *Incidence 30 day post-op mortality* = 12/874 x 100% = *1.4%* 42 / 907 patients in open group died within 30 days of surgery *Incidence 30 day post-op mortality* = 42/907 x 100% = *4.6%* *Relative risk *= 1.4% / 4.6% = *0.3* comparing laparoscopic to open surgery *Laparoscopic splenectomy associated with 70% lower post-operative mortality compared to open splenectomy* Questions: 1.) Is laparoscopic splenectomy ASSOCIATED w/ lower postoperative mortality? YES: YES, lap splenectomy associated with 70% lower risk of post-op mortality 2.) *Does laparoscopic splenectomy REDUCE post-operative mortality?* Possible. Also possible that other characteristics linked with laparoscopic splenectomy responsible for observed lower post-operative mortality

Observational vs Randomized studies: Observational Study: Laparoscopic surgery patients a little younger, less likely to have emergency surgery, less likely to have malignancy, have better nutrition *Randomization eliminates other characteristics by balancing them*

*Tabulate Potential Confounders as a Function of Exposure Status* - Pts who had Emergency surgery were far more likely to undergo open surgery than lap

*Tabulate Potential Confounders as a Function of Outcome* - Pts who had emergency surgery had a higher rate of death (regardless of exposure) *Confounding factor must be associated with exposure AND outcome*

*Irrelevant to whether or not a factor is confounding* - P-values make an inference to a larger population of similar pts

*Confounding*: factor leads to more exposure and more outcome *Non-Confounding*: exposure leads to factor, factor leads to outcome *Non-Confounding*: Factor not related to both exposure and outcome

1.) Tabulate Potential Confounders as a Function of Exposure Status 2.) Tabulate Potential Confounders as a Function of Outcome 3.) Confounding factor must be associated with exposure AND outcome (Relative Risk) 4.)

Confounding occurs when another factor is responsible for at least part of an association of exposure with outcome - A confounding factor is associated with both exposure and outcome and does not reside on the causal pathway of association

1.) *Use the study data* to decide whether a potential confounding factor is associated with both the exposure and the outcome - Data for association of potential confounder with the study outcome may not be available in a paper - Sometimes have to guess --> 2.) *Use biologic plausibility* to decide whether a potential confounding factor resides on the causal pathway of association

1.) *Is current erythromycin use ASSOCIATED with cardiac death?* Yes, relative risk = 2.01 compared to non-users 2.) *Is current erythromycin use a CAUSE of cardiac death?* - If yes, may considering new warning label for erythromycin Current erythromycin users compared to non users: - A little bit older - More likely to take other medications - More likely to have previous encounters for CV disease

These are all indications for the use of Erythromycin

A special case of confounding in which indications for medication or procedure confound the association - Confounding-by-indication occurs when the indication for a medication or procedure specifically confounds the association - The presence of confounding weakens the case for causality *Solution* = Compare with another drug or procedure w/ similar indication

Incidence rate erythromycin users = 10/5,305 = 1.89 per 1,000 person-yrs Incidence in amoxi users = 8/6,846 = 1.17 per 1,000 person-yrs Relative Tisk = 1.89/1.17 = 1.61 comparing current erythromycin use to current of amoxicillin use ==> Current erythromycin users associated w/ 61% greater risk of death compared to amoxicillin use *Strengthens case for causal association of erythromycin*

Classical evaluation of a confounder: *Asses using Study Data if*: - Associated with the exposure - Associated with the outcome *Asses using own knowledge/biological plausibility if*: - Not on the causal pathway

- Randomization - Restriction - Stratification - Matching - Regression

1.) *Randomization* 2.) *Restriction* prevents confounding by procedure status; However: - Lose 261 study participants - reduces study power - Study results now apply to only elective splenectomies - May still be confounding by other characteristics 3.) *Stratification* 4.) *Matching* 5.) *Regression* - Modern approach for handling multiple confounders; Topic of future lecture

Prevents confounding by procedure status (prevents procedure status from varying in each group)

*Advantages*: - Simple and easy to understand - Can control for one or possibly a few major confounders *Limitations*: - Lose study subjects and study power - Reduced applicability of study findings - Difficult to deal with multiple confounders

1.) Separate into 2 groups (emergency vs elective) = *Stratum* - w/in each stratum, procedure status is not confounding 2.) Compare Relative Risk of each group - example: implies that there was some confounding since both RR's are closer to 1 3.) Combine the results to calculate and overall RR - weight the RR by % of the study population in each group - e.g. (RR1*weight1) +(RR2*weight2)

Combined relative risk = RR stratum 1 * weight + RR stratum 2 * weight

Prevents confounding by procedure status 1.) Separate into Stratum based on factor 2.) Compare Relative Risk of each group 3.) Combine the results to calculate and overall RR Combined relative risk = RR stratum 1 * weight + RR stratum 2 * weight

Relative risk Unadjusted = 0.3 Relative risk Adjusted for procedure status = 0.61 (stratification) "After adjustment for procedure status laparoscopic splenectomy associated with 39% lower risk of death compared to open splenectomy" *Could this association still be confounded by other factors?* YES

Divide into more stratum, works best with a really large study group then weight and combine RR across all strata

*Advantages*: - Uses all of the study data - Works well for naturally binary confounding factors (procedure status, gender, past history of diabetes) *Limitations*: - Difficult to deal with multiple confounders at the same time - "Residual confounding" for continuous variables (age, serum albumin level)

For each study patient undergoing laparoscopic splenectomy, find one or more similar pts undergoing open splenectomy with: - Same procedure status (elective vs emergency surgery) - Similar age (for example, within ±3 years

Now analyze the matched data Find matched relative risk = 0.72 "After matching by procedure status and age, laparoscopic surgery associated with 28% lower risk of death compared to open surgery"

Breaks association between confounder and exposure - Varies by Study Design 1.) For each patient, match with another pt according to confounders 2.) Find Matched Relative Risk

*Matching on exposure* - For each exposed person find one or more unexposed people who are the same with respect to the confounding factor(s)

*Matching on outcome* (case and control) - For each case person find one or more control people who are the same with respect to the confounding factor(s)

*Advantages*: - Easy to understand - Possible to deal with multiple confounders *Limitations*: - Must be performed in the Design - before the analysis - May be difficult to control for multiple confounders - Alters generalizibility to matched population

Change to adjusted RR implies that the factor that was adjusted for was a confounder *Alternative definition of a confounding factor = "Meaningful" change in association after adjustment for that factor*

"Meaningful" change in association after adjustment for that factor - No consensus on "meaningful" - 5%, 10%, 20% change proposed - No statistical test for "meaningful" either *Can use 10% for this class*

Laparoscopic surgery associated with 71% lower unadjusted odds of post-operative mortality compared with open splenectomy BUT Post-operative mortality after splenectomy relatively uncommon (Rare) sooo.... - Laparoscopic surgery associated with ≈71% lower unadjusted risk of post-operative mortality compared with open splenectomy

Laparoscopic surgery associated with 61% lower risk of post-operative mortality compared with open splenectomy after adjustment for age, sex, race, indication, ASA class, functional status

After further adjustment for all the rest of that stuff: laparoscopic surgery associated with 61% lower risk of post-operative mortality compared with open splenectomy

These findings imply Some degree of confounding by age, sex, race, indication, ASA class, and/or functional status - No further confounding by a large number of other factors

Model includes adjustment for heart failure: Back to the methods section to check how heart failure was measured - Heart failure assessed using diagnosis codes - Diagnosis codes have poor accuracy for capturing heart failure *What happens when you adjust for a poorly measured confounding factor?* NOTHING = "*Imprecise ascertainment of comorbidity*" - An important consideration when evaluating confounding

Just bc you adjusted for something doesn't mean you actually adjusted for "it" you adjusted for something that represents "it" - and it may respresent it poorly - if you are worried results are still confounded after adjustments, look at how the factor was measured

Part of the lower post-operative mortality observed for laparoscopic splenectomy is likely mediated via less intra-operative bleeding *Over-adjustment* - An important consideration when evaluating confounding

Studies often over adjust by adjusting for things that are part of the causal pathway (mechanisms) - relative risks become falsely LOW - reporting an association that is falsely low, "shoot yourself in the foot"

- Procedures to control for confounding include randomization, restriction, stratification, matching, and regression - Break bonds between confounder and exposure or outcome - Alternative definition of confounding factor is a "meaningful" change in the association after adjustment for that factor - Two important considerations when evaluating confounding include imprecise ascertainment of co-morbidity and over adjustment