Epid 600 Class 6 Case Control Studies

50
EPID 600; Class 6 Case control studies University of Michigan School of Public Health Drug Abuse: A workshop on behavioral and economic research October 18-20, 2004 1

Transcript of Epid 600 Class 6 Case Control Studies

Page 1: Epid 600 Class 6 Case Control Studies

EPID 600; Class 6 Case control studies

University of Michigan School of Public Health

Drug Abuse: A workshop on behavioral and economic research

October 18-20, 2004

1

Page 2: Epid 600 Class 6 Case Control Studies

Three key dimensions to epidemiologic studies

Measures of association Relative measures (relative risks, rates, and odds) Absolute measures (risk and rate differences) Study design Observational Cohort Case-control Cross-sectional Experimental Randomized trial Field trials Group randomized trials Units of analysis Individual Group

2

Page 3: Epid 600 Class 6 Case Control Studies

Three key dimensions to epidemiologic studies

Measures of association Relative measures (relative risks, rates, and odds) Absolute measures (risk and rate differences) Study design Observational Cohort Case-control Cross-sectional Experimental Randomized trial Field trials Group randomized trials Units of analysis Individual Group

3

Page 4: Epid 600 Class 6 Case Control Studies

The world

persons “exposed” persons “unexposed”

4

Page 5: Epid 600 Class 6 Case Control Studies

The case control study

persons with disease persons without disease

5

Page 6: Epid 600 Class 6 Case Control Studies

The case control study

persons “exposed” with disease persons “unexposed” with disease

6

Page 7: Epid 600 Class 6 Case Control Studies

Disease

No disease

E+

E-

E+

E-

Case control studies

7

Page 8: Epid 600 Class 6 Case Control Studies

Disease

No disease Disease No

disease

Exposed a b

Not Exposed c d

E+

E-

E+

E-

Case control studies

8

Page 9: Epid 600 Class 6 Case Control Studies

Disease

No disease Disease No

disease

Exposed a b

Not Exposed c d

E+

E-

E+

E-

Case control studies

9

Page 10: Epid 600 Class 6 Case Control Studies

Disease

No disease Disease No

disease

Exposed a b

Not Exposed c d

E+

E-

E+

E-

Case control studies

10

Page 11: Epid 600 Class 6 Case Control Studies

Disease

No disease Disease No

disease

Exposed a b Not Exposed c d

E+

E-

E+

E-

Case control studies

11

Page 12: Epid 600 Class 6 Case Control Studies

Disease

No disease

Disease No disease

Exposed a b

Not Exposed c d

E+

E-

E+

E-

Case control studies

12

Page 13: Epid 600 Class 6 Case Control Studies

Disease

No disease Disease No

disease

Exposed a b

Not Exposed c d

E+

E-

E+

E-

Case control studies

13

Page 14: Epid 600 Class 6 Case Control Studies

Principles of case control studies

A case control study is conceptually the same as a cohort study but is more efficient The overriding principle is that we select controls that are representative of the population at risk that gave rise to the study cases Cases are more exposed if exposure increases the risk of disease; less exposed if exposure is protective

14

Page 15: Epid 600 Class 6 Case Control Studies

Case control study design

The sample size, number of cases and number of controls (i.e., persons who are not diseased) is determined by the study design However, the exposure has to be assessed retrospectively and the proportions of cases and controls who are exposed are unknown at the beginning of the study

15

Page 16: Epid 600 Class 6 Case Control Studies

Issues in selecting cases

Cases should preferably be new, i.e., incident cases, not existing, i.e., prevalent, ones Remember...Prevalence=Incidence*duration Therefore, factors that influence prevalence influence both whether disease occurs (i.e., incidence) and how long it lasts (i.e., duration) So, if we select prevalent cases, we will not be able to distinguish relative between contribution of factors (exposures) to relative occurrence of new disease and its duration

16

Page 17: Epid 600 Class 6 Case Control Studies

Aside...about prevalent cases

Prevalent cases may be ok if exposure causes rapidly lethal form of disease Of course, in this case there will be very few prevalent cases to choose and this is quite inefficient

17

Page 18: Epid 600 Class 6 Case Control Studies

Where do cases come from?

Population-based cases Complete sample of all cases arising in a well-defined population (time and place)

Hospital-based cases Patients admitted to one of several hospitals within a given population or area

Other sources Patients of a medical group, persons enrolled through a screening program, etc.

18

Page 19: Epid 600 Class 6 Case Control Studies

Issues in selecting controls

Controls must belong to the population of origin of the cases The “world” is called the source population However, we are interested in the population at risk Controls must represent the population at risk from which the cases came; this is called the base population If controls are selected correctly, a similar proportion of controls would have developed the disease if they had been exposed to the same exposure as were the cases

19

Page 20: Epid 600 Class 6 Case Control Studies

Where do controls come from?

Population-based controls A random sample of all the disease-free population from where cases came If cases are within subgroups, then that subgroup is the population from where controls must come

Neighborhood controls Similar to cases on some, perhaps not other, factors

Dead people Problematic if exposure in any way leads to death (i.e., exposure associated with control selection)

20

Page 21: Epid 600 Class 6 Case Control Studies

Special case: hospitalized controls

Patients hospitalized for disease unrelated to exposure of interest Only valid if cases/controls come from same population (demographically and geographically) e.g., if controls are patients with myocardial infarction, do not select controls from pathologies (e.g., bronchitis) that may also be associated with smoking Problem is that sometimes we do not know that a particular disease is associated with outcome; hence frequent use of orthopedic cases Advantages of hospital-based controls: convenient, may be representative of population from which cases are selected, may be assessed in much the same way as cases

21

Page 22: Epid 600 Class 6 Case Control Studies

Advantages of case control studies

Efficient for rare diseases Relatively efficient in terms of time and money Can study diseases with long latency period Allow for the evaluation of multiple exposures that may increase risk for a specific disease

22

Page 23: Epid 600 Class 6 Case Control Studies

Disadvantages of case control studies

Cannot directly compute incidence of disease in exposed and non-exposed persons Temporal relationship between exposure and disease may be difficult to establish with certainty Are more prone to errors in selection of cases/controls an in errors pertaining to the collection of information (bias—will be discussed later) Not optimal for rare exposures

23

Page 24: Epid 600 Class 6 Case Control Studies

An example

Is smoking associated with brain cancer?

Cases: All incident cases of brain cancer in Ann Arbor Controls: A random sample of residents of Ann Arbor Exposure assessment: questionnaire about ever smoked Note: exposure assessed independently of case/control status

24

Page 25: Epid 600 Class 6 Case Control Studies

Disease

No disease

E+

E-

E+

E-

Study findings

180

604

140

370

40

234

25

Page 26: Epid 600 Class 6 Case Control Studies

2x2 table

Brain cancer No cancer Total

Smoking 140 370 510

No smoking 40 234 274

Total 180 604 784

26

Page 27: Epid 600 Class 6 Case Control Studies

What measure of associations can we calculate?

We cannot calculate either risks or rates since we do not have a complete population to be denominator for risk nor a complete person time population for rate calculation But we can calculate odds; calculation of odds depends only on numbers of cases and numbers of controls

27

Page 28: Epid 600 Class 6 Case Control Studies

And...

40234

140*234 2.2370*40

, 2.2

140Odds of brain cancer among smokers = 370

Odds of brain cancer among non smokers

Odds ratio of brain cancer

Therefore the odds of brain cancer is times higher among smoke

− =

= =

rs vs. non - smokers

28

Page 29: Epid 600 Class 6 Case Control Studies

Cell phones and cancer: hype or hazard?

There has been persistent concern about the potential carcinogenic effects of electromagnetic radiofrequency fields emitted by cellular phones The vast majority of studies do not show an association between cell phone use and development of tumors Most studies neglect to look at long term users A research group in Israel published results of a population-based case control study to describe the association between cell phone use and parotid gland tumors

Sadetzki et al. Cellular phone use and risk of benign and malignant parotid gland tumors-a nationwide case-control study. Am J Epidemiol. 2007; 167:457-467 29

Page 30: Epid 600 Class 6 Case Control Studies

Cell phone study: set up

1) 2001-2003: Detected new PTG cases of (malignant or benign) in Israel through review of records at all relevant medical institutions in Israel

Cell phone use patterns?

3) 2001-2003: Go back in time and determine of cell phone use patterns for PTG cases and controls

2) 2001-2003: Controls randomly selected from the National Population Registry; matched to individual cases

Sadetzki et al. Cellular phone use and risk of benign and malignant parotid gland tumors-a nationwide case-control study. Am J Epidemiol. 2007; 167:457-467

30

Page 31: Epid 600 Class 6 Case Control Studies

Cell phone study: cases and controls

Cases: People who were confirmed to have PTG through medical records and verification by a single physician

Controls: People who do not have PTG and are listed in the National Population Registry in Israel

Sadetzki et al. Cellular phone use and risk of benign and malignant parotid gland tumors-a nationwide case-control study. Am J Epidemiol. 2007; 167:457-467 31

Page 32: Epid 600 Class 6 Case Control Studies

Cell phone study: measuring exposure

FOR CASES AND CONTROLS:

Has the participant used a cell phone more than 1x/week for at least 6 months (ever)?

Regular User (exposed)

Not a regular user

(unexposed)

• Number of calls? • Duration of calls? • Use of headsets? • Which side of the head was phone held on? • Urban or rural location?

Sadetzki et al. Cellular phone use and risk of benign and malignant parotid gland tumors-a nationwide case-control study. Am J Epidemiol. 2007; 167:457-467 32

Page 33: Epid 600 Class 6 Case Control Studies

Cell phone study: findings

For the entire group, no increased chance of PGT was observed for ever having been a regular cell phone user (OR=0.87) However, analysis of regular users showed consistently elevated probability of PGT

Cumulative call time (hours) with no hands-free device

Adjusted Odds Ratio

Non Users 1.0 (reference) <=266.3 0.72 266.4-1034.9 1.57 >=1035 1.96

Sadetzki et al. Cellular phone use and risk of benign and malignant parotid gland tumors-a nationwide case-control study. Am J Epidemiol. 2007; 167:457-467 33

Page 34: Epid 600 Class 6 Case Control Studies

It is helpful to think of all case control studies as nested within a population cohort

All case control studies are sampling from a base population, which is the persons at risk in the source population Different forms of control sampling then have implications for how the case control study mimics the underlying cohort “Incidence density sampling” selects from the risk set during the same follow-up period in which cases are identified; that is, the probability of selection is proportional to the time at risk

34

Page 35: Epid 600 Class 6 Case Control Studies

Cell phone study: set up

1) 2001-2003: Detected new PTG cases of (malignant or benign) in Israel through review of records at all relevant medical institutions in Israel

Cell phone use patterns?

3) 2001-2003: Go back in time and determine of cell phone use patterns for PTG cases and controls

2) 2001-2003: Controls randomly selected from the National Population Registry; matched to individual cases

Sadetzki et al. Cellular phone use and risk of benign and malignant parotid gland tumors-a nationwide case-control study. Am J Epidemiol. 2007; 167:457-467

35

Page 36: Epid 600 Class 6 Case Control Studies

The “underlying cohort”

36

Page 37: Epid 600 Class 6 Case Control Studies

Disease cases

37

Page 38: Epid 600 Class 6 Case Control Studies

Incidence density sampling

For every case we select a control from the population risk set during the same follow-up period in which the cases are identified

38

Page 39: Epid 600 Class 6 Case Control Studies

Therefore, in incidence density sampling

Disease Time

Exposed a T1

Not Exposed c To

Total a+c T1+To

IRexp =a

PT1

and IRun exp =c

PT0

aim is to select controls so thatif b is exposed and d is unexposedbd=

PT1

PT0

or bPT1

=d

PT0

or db=

PT0

PT1

therefore

IRR =

aPT1

cPT0

=ac

*PT0

PT1

=ac

*db= OR

so, OR is an unbiased estimate of the IRR in

incidence density sampling

39

Page 40: Epid 600 Class 6 Case Control Studies

Case-cohort sampling

For every case we select a control who is a member of the population at risk; all cases contribute to both numerator and denominator. Note that it is possible that after a person is selected as a control, that person may later become a case 40

Page 41: Epid 600 Class 6 Case Control Studies

Therefore, in case cohort sampling

Disease No disease Total

Exposed a b a+b

Not exposed c d c+d

Total a+c b+d a+b+c+d

Rexp =a

a + b and Run exp =

cc + d

given that controls are selected from base population, then the proportionof exposure vs. non exposure among controlsrepresents proportion of exposure vs. non - exposurein the population, that is bd=

a + bc + d

or db=

c + da + b

therefore

RR =

aa + b

cc + d

=ac

*c + da + b

=ac

*db= OR

41

Page 42: Epid 600 Class 6 Case Control Studies

Summary of controls in a case control study

In case control study, the only statistic is ac

*db

, or a * dc * b

Therefore, if control selection is such that db=

PYOun exp

PYOe xp

then OR is an unbiased estimate of IRR

and if conrol selection is such that db=

total unexposedtotal exposed

then OR is an unbiased estimate of RRDisease No

disease Time Total

Exposed a b PYOexp a+b

Not exposed c d PYOexp c+d

Total a+c b+d PYOtotal a+b+c+d 42

Page 43: Epid 600 Class 6 Case Control Studies

And if...(old-fashioned)

controls are

selected from those who are no

longer cases at

the end of the study

Here we are overestimating the risk ratio because, at end of study, the proportion of exposure among those who are controls is less than population (assuming a positive exposure-disease association)

When disease is “rare”, this is not much of an issue since there are very few cases so the proportion of exposure among controls approximates population anyway

43

Page 44: Epid 600 Class 6 Case Control Studies

Sampling fractions

Exposed Unexposed

Cases Non-cases Cases Non-cases

Cases Controls

Exposed Unexposed Exposed Unexposed

F1 F2 F3 F4

Sample

Target population

a b c d

44

Page 45: Epid 600 Class 6 Case Control Studies

Case control study as sampling from a cohort study

Disease No disease

Total

Exposed a b a+b

Not exposed

c d c+d

Total a+c b+d a+b+c+d

1 4 1 4

3 2 3 2

1 4

3 2

*

* * * *( * )* * * *( * )

***

a d (exposed cases * unexposed non - cases)Cohort OR c* b (unexposed cases* exposed non - cases)

a F d F ad F FCase control ORc F b F cb F F

F FCase control OR Cohort ORF F

Case control OR C

=

= =

=

= 1 4

3 2

* 1*

F Fohort OR if F F

=

45

Page 46: Epid 600 Class 6 Case Control Studies

The “underlying cohort”, 2x2 table

Brain cancer PYO

Smoking 140 55,360

No smoking 40 35,060

Total 180 90,420

46

Page 47: Epid 600 Class 6 Case Control Studies

And...

4035,060

2.24035,060

, 2.2

140IR of brain cancer among smokers = 55,360

IR of brain cancer among non smokers

14055,360Odds ratio of brain cancer

Therefore the incidence rate ratio of brain cancer is ti

− =

= =

mes higher among smokers vs. non - smokersTherefore, when controls are sampled from person time of observation of population, IRR = OR

47

Page 48: Epid 600 Class 6 Case Control Studies

A note about number of controls

We do not need to sample as many controls as there are cases Selecting 4 controls for every case (i.e., 4:1 control: case selection) improves statistical power, i.e., the ability of a study to detect associations There is little statistical advantage to selecting more than 4 controls for every case

48

Page 49: Epid 600 Class 6 Case Control Studies

Special topic: Exchangeability of odds ratio

Disease No disease Total

Exposed a b a+b

Not Exposed c d c+d

Total a+c b+d a+b+c+d

exp1

exp

, exp exp

aaa bodds of being a case among osed a b

a bc

cc+dodds of being a case among un osed c d1-c+d

aa* dbrelative odds of being a case comparing to un =c b* c

d

odds of being exposed among cas

+= =−

+

= =

=

1

1

exp ,

aaa ces = a c

a cb

bb dodds of being exposed among controls = b db d

aa* dcrelative odds of being osed comparing cases to controls =b b* c

d

+ =−

+

+ =−

+

=49

Page 50: Epid 600 Class 6 Case Control Studies

Comparing cohort and case control studies

Cohort Case control

Complete source population experience tallied

We sample from source population and its experience

Can calculate incidence, risk, and relative incidence

Can calculate OR, which, under proper sampling conditions, mimics either RR or IRR

Convenient for many diseases Convenient for many exposures

Prospective or retrospective Prospective or retrospective

50