Table 1.
Sources of Confounding in Observational Studies of Prevention Related to Patient Health Behaviors and Functional Status, and Potential Methods to Adjust for Them
| Type of confounding | Source of confounding | Design approaches to reduce confounding | Methods for adjustment in observational studies |
|---|---|---|---|
| Healthy-user effect | Patients who use preventive tests or therapies are more likely to pursue other health-seeking behaviors | o Identify an active comparator group of subjects who initiated a different preventive therapy rather than non-users to serve as a control group13 | o Adjust for use of preventive services unrelated to the outcome or other healthy behaviors |
| o Use an instrumental variable approach36 | o Adjust with high-dimensional propensity score to capture proxies for health-seeking tendency33 | ||
| Healthy-adherer effect | Patients who adhere to preventive tests or therapies are more likely to pursue health-seeking behaviors | o Use a new user design and analyze results on a intention-to-treat basis38 | o Adjust for adherence to medications unrelated to the outcome34 |
| Confounding by functional status or cognitive impairment | Patients with poor functional or cognitive status are less likely to pursue health-seeking behaviors | o Use an active comparator group | o Adjust for markers of dementia (e.g., dementia medications) or poor functional status (nursing home stays)17 |
| o Restrict the study population to subjects who are similar in having evidence of recent preventive service use17 | o Adjust with high-dimensional propensity score | ||
| o Use an instrumental variable approach | |||
| Confounding by selective prescribing | Physicians are less likely to order preventive treatments for frail or acutely ill patients or those with lower perceived functional or cognitive ability | o Use an active comparator group | o Adjust for patient “frailty”17 |
| o Restrict the study population to subjects who are similar in having evidence of recent preventive service use17 | o Adjust for unmeasured confounding using high dimensional propensity score3 or instrumental variables36 | ||
| o Use an instrumental variable approach |