Can we use population-based longitudinal data to personalize depression treatment? Gregory Simon MD MPH Group Health Center for Health Studies.

Can we use population-based longitudinal data to personalize depression treatment?

Gregory Simon MD MPH

Group Health Center for Health Studies

Outline

• Background on predicting response to depression treatment

• Limitations of randomized trials to identify predictors of response

• Alternative – large observational studies using longitudinal data

• Methodologic and statistical issues in large observational studies

Success of antidepressant treatment

• 35-40% remission with 1st treatment

• 25-30% with 2nd treatment

• 15-20% after 3rd treatment

• Cumulative remission rate: 60-65%

Predicting treatment success

• Moderate ability to predict overall outcome– severity, chronicity, comorbidity, poor

response to previous treatments, etc.

• Poor (actually zero) ability to predict specific or differential response based on:– symptom patterns– biomarkers

• Some support for genetic predictors of adverse effects – less clear for benefits

Core assumption

There are stable characteristics of individuals that predict greater likelihood of good (or bad outcome) with exposure to:

•Active treatment compared to no treatment or inactive treatment

•One active treatment compared to another

Traditional method: search for effect modification in randomized clinical trial

•Random assignment to treatments (comparing active treatments or active treatment to placebo)

•Test for interaction between proposed predictor and treatment assigment

Potential effect modifier:Prevalence = 50%Accounts for 80% of benefitNo effect on untreated prognosis

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

Overall With Without

Placebo

Active Treatment

Observed outcomes

No Remission Remission Total

With Predictor Active Treatment 103 97 200

Placebo 148 52 200

Without Predictor Active Treatment 137 63 200

Placebo 148 52 200

Total 536 264 800

Odds Ratios:With Predictor = 2.68 (1.76-4.08)Without Predictor = 1.31 (0.85-2.02)

Test for interaction: p=.02

Note: This study would cost $5-6 million

Components of placebo response

• Natural history– Stable characteristics– Episode-specific characteristics

• Non-specific benefits of treatment

• Measurement error

We know absolutely nothing about:

• Consistency of placebo response across episodes

• Consistency of response to same or different treatment across episodes

Consistency of response across episodes

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

With Without

Placebo Active Treatment

Alternative: large observational studies using longitudinal data• Use longitudinal data including multiple

treatment episodes per person

• Treatment decision are uncontrolled

• Alternatives for assessing outcome:– Medical records data (proxy measures)– Recall across multiple treatment episodes– Prospective assessment

Episode A1 Episode A2 Episode A3 Episode B1 Episode C1 Episode C2 Episode D1 Episode D2 Episode D3

Patient A Patient DPatient B Patient C

Treatment X Treatment Y

Complex clustered data structure

Pros and cons of large observational studies• Advantages:

– Sample sizes practically unlimited– Recruitment is much more efficient– Multiple episodes per person (can separate

episode-level and person-level variation)

• Disadvantages:– Greater measurement error within episodes– Treatments are not randomly assigned

Distinguishing stable (person-level) and unstable (episode-level) predictors

Patient-Level Predictors Episode-Level Predictors Treatment Characteristics Demographics (sex, race/ethnicity) Pre-treatment sy mptom severity Medication vs. Psychotherapy Genetic variation Pre-treatment episode duration Specific medication or drug class Childhood trauma or abuse Recent stressful events Treatment intensity Family history Current substance use Treatment duration Personality or attachment style Therapeutic alliance

Sources of variance in clinical trials and observational studies

True Response

Nonspecific Tx Effect

Measurement Error

Person-Level Factors

Episode-Level Factors

Biased Tx Assignment

Methodologic questions:

• Use of claims data for proxy outcomes (or at least to identify enriched samples)

• Accuracy of recall for past episodes

• Biased assignment of treatments in later episodes

• Consistency of response across episodes

Feasibility of recruitment

• What proportion of those approached agree to participate in assessments

• What proportion agree to provide genetic material

• How do participants and non-participants differ in:– Demographics– Treatment history– Current mood

Utilization as a proxy for outcome

• Proportion of early discontinuers who reach remission (akin to placebo responders)

• Continued use of original drug as proxy for good response

• Early discontinuation as proxy for adverse effects

• Medication switch or specialty referral as proxy for poor response

Accuracy of recall

• Interested in accuracy of recall for both benefits and adverse effects

• Likely that accuracy of recall decreases with time

• Recall may be influenced by current mood

Biased assignment of treatments

• Likely that good response to a treatment increases likelihood of re-exposure

• May inflate estimates of consistency of good response across episodes

General modeling approach

• Random coefficient regression models to account for clustering of episodes within individuals

• Can consider both general tendency to respond to treatment and tendency to respond to specific treatments

• Consider treatment response as a function of:– Stable person-level characteristics (measured and

unmeasured)– Treatment exposure

Two approaches to biased selection of treatments

• Decompose variation into between-person (i.e. general tendency to respond favorably or unfavorably) and within-person (i.e. tendency to respond specifically to a given treatment)

• Explicitly model selection process

Managing effects of recall error

• Random error – ? overcome with brute force

• Decay in recall over time – may need to censor remote observations

• Effect of current mood state – may need to account for explicitly in models