1

Lecture 4Study design and bias in

screening and diagnostic tests

• Sources of bias :– spectrum effects/subgroup analyses – verification/workup bias– information (review) bias

• Critical assessment of studies:– e.g., STARD criteria

2

Bias

• What is it?– Bias in a measurement vs bias in the result of a

study

• Selection vs information bias• What does it mean in studies of screening

and diagnostic tests?• Difference between bias and effect

modification?

3

Reducing bias

• Studies of diagnostic tests give variable results

• Biassed studies generally overestimate sensitivity/specificity

• STARD criteria proposed to improve quality of these studies

4

Spectrum effect: bias or modification?

• Sensitivity and specificity are not innate characteristics of a test, but vary by study population – e.g., by age, sex, comorbidity– e.g, exercise stress testing: worse performance

in women than men– Study population should be representative of

population in which test will be used

5

Design implications

• Investigate test performance in sub-groups

• Report characteristics of study population

6

Verification/work-up bias

• Results of test affect intensity of subsequent investigation– e.g., risky or expensive follow-up

• Selection or information bias?

• E.g. Exercise stress test and angiography– effects?– solutions?

7

Example of verification/work-up bias

• VQ (ventilation/perfusion scanning to detect pulmonary embolism – Positive scan -> angiography– Studies with selective referral of patients:

sensitivity = 58%– Study (PIOPED) with prospective investigation

of all patients: sensitivity = 41%

8

Information/review bias

• Examples:– Diagnosis is not blind to test result– Diagnosis is made with access to other clinical

information – Knowledge of results of follow-up used in

interpretation of screening test

• Effects?• Solutions?

– (NB: raw test performance vs “real-world” situation)

9

Other sources of bias

• Indeterminate test results:– How do they affect results?– Solutions?

• Context:– Interpretation varies with changes in disease

prevalence

• Criteria for positivity– Technical advances, operator experience

10

Optimal design

• Cohort vs case-control?

• Prospective cohort with blind evaluation

• Case-control:– Sources of bias?

11

Example for discussion

• Seniors in emergency department (ED):– High risk of functional decline, death etc.– Needs usually not recognized at ED visit

• Objective: Development and validation of tool to identify “high-risk” seniors in ED (need more careful assessment and follow-up)

• Methods?

12

RESULTS: ISAR development

Adverse health outcome defined as any of following during 6 months after ED visit

• >10% ADL decline

• Death

• Institutionalization

13

0

10

20

30

40

50

60

ADL decline Death Institution Any

%

DischargedAdmitted

Adverse outcome by disposition

14

Scale development

• Selection of items that predicted all adverse health events

• Multiple logistic regression - “best subsets” analysis

• Review of candidate scales with clinicians to select clinically relevant scale

15

Identification of Seniors At Risk (ISAR)

1. Before the illness or injury that brought you to the Emergency, did you need someone to help you on a regular basis? (yes)

2. Since the illness or injury that brought you to the Emergency, have you needed more help than usual to take care of yourself? (yes)

3. Have you been hospitalized for one or more nights during the past 6 months (excluding a stay in the Emergency Department)? (yes)

4. In general, do you see well? (no)

5. In general, do you have serious problems with your memory? (yes)

6. Do you take more than three different medications every day? (yes)

Scoring: 0 - 6 (positive score shown in parentheses)

16

0

20

40

60

80

0 1 2 3 4 5-6

ISAR SCORE

%

DischargedAdmitted

Any adverse outcome by ISAR score and disposition

17

Predictive validity of ISAR scale

• AUC and 95% CI– Overall (n=1673): 0.71 (0.68 – 0.74)– Admitted to hospital (n=509): 0.66 (0.61 –

0.71) – Discharged (n= 1159): 0.70 (0.66 – 0.74)– Similar results by informant (patient vs proxy)

• Next steps?

18

Second study

• Multi-site randomized controlled trial of a 2-step intervention using ISAR + nurse assessment/referral

• Study 2 population had lower % ISAR +ve than study 1 population – implications for sensitivity, specificity, AUC,

LR, DOR?

19

Area Under the curve (AUC) for concurrent validity criteria

Detection of depression

at baseline

Study 2

OARS: Study 1

Severe functional

impairment

OARS: Study 2

SMAF: Study 1

AUC (95% confidence interval)

0.5 0.6 0.7 0.8 0.9 1.0

20

Area Under the Curve(AUC) for predictive validation criteriaamong patients discharged from ED

Increase in depressivesymptoms

Study 2

10+ community healthcenter visits/5 months

Study 2

11+ hospital days/ 5 months

Study 1

Study 2

2+ ED visits/ 5 months

Study 1

Study 2

Adverse health outcome

Study 1

AUC (95% confidence interval)

0.5 0.6 0.7 0.8 0.9 1.0

21

Other predictive measures in elderly• Pra screening tool (widely used in US HMOs):

– AUC values of 0.61 - 0.71 for prediction of hospital utilization or functional decline (Coleman, 1998)

• Hospital Admission Risk Profile (HARP)– AUC of 0.65 for prediction of nursing home

admission (Sager, 1996)• Comorbidity indices (diagnosis and medication-based

measures from administrative data):– AUC values of 0.58-0.60 for emergency

hospitalization (Schneeweiss, 2001)