Types of epidemiological studies

 

Written by:

Alain Moren

Jean Claude Desenclos

(March  2006)

 

Reviewed by: Marta Valenciano and Arnold Bosman (oct 2006)

 

 

 

 

Many epidemiologists consider that the studies they are conducting are measurement exercises. Simple studies include measurements of disease frequency which may be expressed as risks, rates, prevalence or odds. More advanced studies will aim at identifying the causes of diseases and the effect of specific exposures on disease occurrence. This achieved by comparing disease frequency between sub groups of a population. This comparison can be expressed as a difference or a ratio. These studies are sometimes called analytical studies and the comparison "effect measure".

 

 

 

 

Measuring effect in various study designs

 

 

As stated above a core function of epidemiologists is to measure the causal effect of an exposure on the occurrence of a disease. To measure a causal effect we would have ideally to compare occurrence of disease in exposed persons to what would have happened in the same persons, at the same time, in the absence of exposure. This is however theoretical since such two measurements, in the same group of persons under study, are not feasible during the same time period.

 

In order to approach this theoretical situation as closely as possible, we will use as unexposed group a population similar to the exposed group but for the exposure. In these two populations (or in 2 subsets of the same population, exposed and unexposed), we will then measure and compare disease occurrence.

 

To compare disease occurrence between exposed and unexposed populations epidemiologists will have either to assign exposure or to observe populations naturally exposed. Assignment of exposure is only ethically feasible when exposure is potentially protective (treatment, vaccine, preventive measures). Observation of accidental or naturally assigned exposures will allow us to study the effect of potentially harmful exposures.

 

To measure the effect of exposure several types of epidemiological studies are available.

 

 

 

Cohort studies

 

In Romans' time, a cohort of legionnaires consisted of a group of soldiers sharing the same military events for a certain period of time. In epidemiology we consider that a cohort consists of people belonging to the same population and sharing similar experience for a defined period of time.

 

Cohort studies involve the comparison of disease incidence over time (risk or rate) between two subsets of a population (two cohorts). One of the 2 cohorts is exposed to a certain characteristic (exposure). The other is not. All other things being equal between the two cohorts but for their exposure. In both cohorts we measure occurrence of disease over the specific study period. However whenever the condition of "all other things being equal" is not met, the comparison might be wrong.

 

The following graph adapted from Rodrigues illustrates occurrence of cases over time in the two cohorts. Initially Ne persons are exposed and Nu persons are unexposed. The number of persons who are disease free decreases over time (shaded area). The number of cases (non shaded area) increases over time but more in the exposed cohort. At the end of the study, respectively Ce and Cu cases have occurred in the exposed and unexposed cohorts. The shaded area represents the cumulative time during which persons were at risk of developing disease in each of the cohorts during the entire follow up period.

Adapted from Rodrigues L et al. Int J Epidemiol. 1990;19:205-13.

 

 

Cohort studies measuring risk (incidence proportion)

 

Cohort studies that measure risk compare occurrence of disease between exposed and unexposed cohorts. The risk (incidence proportion) of disease in those exposed (IPe) and unexposed (IPu) can be computed as follows:

 

IPe = number of cases among exposed / exposed population at risk at beginning of follow-up

In the above example IPe = Ce/Ne

 

IPu = number of cases among unexposed / unexposed population at risk at beginning of follow-up

In the above example IPu = Cu/Nu

 

The absolute effect of exposure on disease occurrence is the risk difference (RD) between the exposed and unexposed cohorts.

 

Absolute effect (RD) = IPe - Ipu  (Also called « absolute risk reduction »)

 

The relative effect of the exposure on disease occurrence can be expressed as the risk difference between exposed an unexposed, divided by (relative to) the risk in unexposed.

 

                                Risk difference                      IPe-IPu            IPe         IPu

Relative effect =     ------------------------       =        ------------ =     --------  -  ---------- =  RR - 1

                                Risk in unexposed                                 IPu                  IPu         IPu

 

Where RR is the risk ratio defined as:

 

                                IPe

Risk Ratio =            -------

                                IPu

 

Example 1

 

Cases of gastroenteritis according to consumption of food X, nursing home A

 

Consumption of food X    Population at risk  Cases         IP                 Risk Ratio      Relative effect

Yes                                      150                          60                0.4               4                      3

No                                       100                          10                0.1

 

One can express the result by saying that the relative effect of consuming food X is 3 which would suggest a 300% increased risk of gastroenteritis among exposed. One can also express the results by saying that the risk of disease is 4 times higher in the exposed cohort than in the unexposed cohort.

 

Thus the relative effect is the risk ratio minus 1. Since the relative effect is RR - 1, epidemiologists frequently refer to RR as a measure of relative effect without subtracting 1. The term "relative risk" is very popular among epidemiologists even if, as mentioned above, it is not a measure of relative effect but rather a risk ratio.  When using the relative risk that way we have to remember that a value of 1 corresponds to an absence of effect.

 

 

 

Cohort studies measuring incidence rates

 

The computation of effects with incidence rates is similar to calculation of effects from incidence proportions (risk).  The incidence rate of disease in exposed (IRe) and unexposed (IRu) can be computed as follows:

 

IRe = number of cases exposed / sum of person-time at risk among exposed population

IRu = number of cases unexposed / sum of person-time at risk among unexposed population

 

A rate difference can be computed:   Rate difference  =  IRe - IRu

 

The relative effect of the exposure on disease occurrence can be measured by computing the rate ratio minus 1.

 

                                 IRe

Relative effect  =     -------  - 1

                                 IRu

 

The rate ratio is:

 

                                 IRe

Rate ratio   =           ------- 

                                 IRu

Example 2

 

*** cancer cases and person-years of observation for women with tuberculosis repeatedly exposed to multiple x-ray fluoroscopies and unexposed women with tuberculosis

 

Radiation        Person-years     ***     Rate            Rate ratio        Rate            Relative effect

exposure                                    cancer    /10000 p-y                          difference

_______________________________________________________________________________ 

Yes                  28 010                 41            14.6             1.86                 6.7               0.86    

No                   19 017                 15              7.9                                                                        

_______________________________________________________________________________

Total                47 027                 56            11;9                                                

 

Source: Boice & Monson; Rothman, Epidemiology, an introduction, p 50

 

One can express the result by saying that the relative effect is 0.86 which would suggest an 86 % increased rate of *** cancer among exposed. One can also express the results by saying that the rate of *** cancer is 1.86 times higher in the exposed cohort than in the unexposed cohort.

 

 

Risk-odds ???

Case control studies

 

In cohort studies the denominator represents the exposure experience of the source population. If it is the exposure status as observed at the beginning of the cohort we will compute a risk. If we allow for exposure to vary overtime we will compute a rate which takes into account the time spent by each individual in the exposed and unexposed cohorts over time.

 

The main constraint in cohort studies is the necessity to collect information on exposure from large populations (to have denominators for the exposed and unexposed cohorts). We will see below that instead of collecting exposure information from the entire study population we can use a sample of it to calculate or estimate the risk ratio or the rate ratio. In other words, instead of using the entire cohort denominator we will use a sample of it. This sample is also frequently called a control group and it is used to represent the exposure experience of the source population.

 

The rationale behind using a sample of the denominator comes from the following formula for risk and rate ratio which can alternatively be expressed as follows:

 

For risk

                                Ce/Ne                                          Ce/Cu

RR  =                     ----------                   =                 ----------

                                Cu/Nu                                          Ne/Nu

 

 

For rates

 

                                Ce/PTe                                        Ce/Cu

RR =                 -------------           =             -----------------

                                Cu/PTu                                        PTe/PTu

 

 

From the above formula we already see that if we take an unbiased random sample of Ne and Nu the ratio of exposed to unexposed (Ne/Nu)  will not be modified and therefore the risk ratio will remain the same (Ne/Nu or a sample of it gives the same risk ratio if sampling is unbiased). The same applies if we use person years at risk (PT). The concept will be further explained below.

 

 

We have generally speaking three major ways to select a sampled control group which reflect three ways to estimate exposure experience in the source population.

 

1 - controls are randomly selected from the population present (at risk) at the beginning of the study (Ne and Nu in the above graphic). The related study design is called a case cohort study. The exposure measured reflects exposure status at the beginning of the cohort.

 

2 - controls are selected proportionally to the person-time contributed by exposed and unexposed cohorts (PTe and PTu). The related study design is called a density case control study. The exposure measured reflects the varying exposure of people at risk along the cohort.

 

3 - controls are selected  from people who are still free of the disease at the end of the study period (Ne-Ce and Nu-Cu). We will call the related study design a traditional case control since it is the design most frequently used. The exposure measured reflects the exposure experience or status of people still free of disease at the end of the cohort.

 

Modify the order? And start with traditional case-control studies?

Case cohort studies

 

In case cohort studies we aim to achieve the same goal as in cohort studies but more efficiently using a sample of the denominators of the exposed and unexposed cohorts. Properly conducted case cohort studies provide information that should mirror what could have been learned from a cohort study.

 

We will call "source population" the population which gives rise to cases. The source population includes exposed and unexposed cohorts and in that source population we could have conducted a cohort study comparing risk or rates of disease between exposed and unexposed cohorts.

 

If instead we decide to do a case cohort study, we will include the same cases, and classify them as exposed or unexposed. But, instead of getting exposure information from all individuals constituting the denominators of exposed and unexposed cohorts, we will only use a sample of them. The purpose of this sample is to estimate the relative size of exposed and unexposed components of the source population (the proportion of exposed in the source population at the beginning of the cohort).

 

To do so we select a random sample from the entire source population. If that sample is unbiased (sampling done independently from exposure status) we expect (disregarding sampling variation) the distribution of exposed and unexposed persons in the sample to reflect the exposure distribution in the source population at the beginning of the cohort. This is an important aspect of case cohort studies. The sample should be representative of the population giving rise to cases (the source population) regarding exposure.

One way to imagine case cohort studies is therefore to think of them as nested within cohorts of exposed and unexposed people. Any case cohort study could be thought off as nested from the source population. The sample group (control group) is a sample of the denominator present at the beginning of the cohort.

 

From a cohort study measuring risk of disease in exposed and unexposed cohorts we can draw the following results table:

 

Exposure                                    cases                              Population at risk      IP             Risk Ratio             

Yes                      a                      Ne                                   a/Ne                a/Ne /  c/Nu                        

No                       b                      Nu                                   c/Nu

 

If instead of studying the entire denominators of exposed and unexposed we were sampling them (let's say 10%) we would have the following table:

 

Exposure                Cases                     Sample from source population

 

Yes                          a                              Ne/10     

 

No                           b                              Nu/10                     

 

 

Obviously from the above table the risk of disease cannot be computed since denominators sampled from exposed and unexposed cohorts are only a sampling fraction of these two populations. However, if risk can no longer be computed for exposed and unexposed, the risk ratio remains the same. If in the risk ratio calculation we replace the denominators by the 10% samples representing them, we obtain the same value for the risk ratio.

                                                a

                                                ---                            a/Ne       

                                                Ne/10

Risk Ratio =                            ---------       =           ---------

                                                b

                                                ---                            b/Nu

                                                Nu/10

 

When the sample is randomly selected from the source population the risk ratio computed using the sample equals the risk ratio computed within the entire cohorts.

 

Since we are randomly selecting controls from the source population as it was at the beginning of the study (before disease occurrence), it may happen that persons who will later become a case will be selected as controls. Therefore some persons may appear both in the case and control groups. This should not come as a surprise. In a cohort study cases are counted in the numerator and denominators of exposed and unexposed. The same applies to case cohort studies since we use a sample of exposed and unexposed people of the source population.  We are not concerned by the disease status of the control group but by its exposure status. The aim of the control group is to properly reflect the exposure in the source population and this source population originally includes people who will later become cases. Excluding future cases would lead to overestimating the risk ratio, this particularly when disease occurrence is high.

 

Case cohort studies are a reasonable way to conduct case control studies when disease if frequent, when people in the source population have been followed for the same length of time or for short periods of time, and when exposure does not change over time.

 

 

 

Density case control studies

 

In cohort studies, incidence rates are sometimes called incidence density rates. By similarity we will call "density based sampling" a sampling method in which the sample used as controls will represent the person time experience of exposed (Pte) and unexposed (PTu) cohorts in the source population. Thus the probability of any person from the source population to be selected in the sample is proportional to his/her person-time contribution to the denominators of the incidence rates in exposed and unexposed cohorts.

 

Incidence (density) rates in exposed and unexposed cohorts of the source population can be expressed as follows:

 

                a

IRe =        ----

                PTe

 

                b

IRu =        ----

                PTu

 

Where "a" is the number of cases exposed, "b" the number of cases unexposed, "PTe" the total person-time accumulated by exposed persons and "PTu" the total person-time accumulated by the unexposed group.

 

If instead of studying the entire denominators of person time being exposed and unexposed we were sampling them (let's say 10%) we would have the following table:

 

Exposure                Cases                     Sample from source population

 

 

Yes                          a                              PTe/10   

 

No                           b                              PTu/10                   

 

 

 

Obviously from the above table the incidence rate of disease cannot be computed since person time denominators sampled from exposed and unexposed are only a sampling fraction of these two populations. However, if incidence rate can no longer be computed for exposed and unexposed, the rate ratio remains the same. If in the rate ratio calculation we replace the person time denominators by the 10% samples representing them, we obtain the same value for the rate ratio.

 

 

 

 

                                                a

                                                ---                            a/PTe     

                                                PTe/10

Rate Ratio =                            ---------       =           ---------

                                                b

                                                ---                            b/PTu

                                                PTu/10

 

In this study design the sample (control group) is randomly selected from the person time experience of the source population. As a consequence the rate ratio computed using this sample is equal to the rate ratio computed within the cohort study done with the entire person time denominators of the source population.

 

The next issue is obviously about how to select a sample and make sure it represents the person-time experience of the exposed and unexposed cohorts in the source population. It is in fact quite simple. Each time a case occurs, an individual (or several) is randomly selected from the source population which is still free of the disease at the time of the case onset. This is sometimes called prospective case control study. A mathematical explanation of this rational can be found in Laura Rodrigues (reference).

 

For each person contributing time in the source population experience, the time that this person is eligible to be selected in the sample is the same time during which she is also eligible to become a case if the disease should occur. Selecting an individual at the time of disease onset in a case leads us to select a sample among people still at risk and therefore proportionally to their time participation so far in the study. People who have left, are dead or who are already cases cannot be selected from that time on. This is also meaning that a selected individual who is still at risk of disease can later become a case in the study.

 

Let us suppose that Boise and Monson had decided to do a density case control study instead of a person-time cohort study. They would have identified the 56 cases that occurred in the two cohorts and selected a sample series of 470 women. The sample series group should be sampled so that the person time distribution of the sample mirrors the person time distribution of the source population. If randomly selected and unbiased, this would give us 280 exposed and 190 unexposed in the sample (59.6 % of the sample is exposed which is equal to the proportion of exposed person time in the source population, 28010 / 47027).

 

 

Example 3

 

Cases and sample selected from *** cancer cases and person-years of observation for women with tuberculosis repeatedly exposed to multiple x-ray fluoroscopies and unexposed women with tuberculosis

 

Radiation                                                Person-years ***             Rate                Rate ratio        Sample   Rate ratio

exposure                                                cancer            /10000 p-y      source                                    sample

                                                                                                                population

______________________________________________________________________________________   

Yes                              28 010                 41                    14.6                 1.86                 280  1.86        

No                               19 017                 15                    7.9                                           190

______________________________________________________________________________________

Total                            47 027                 56                    11;9                            

 

 

Source: Boice & Monson; Rothman, Epidemiology, an introduction, p 50

 

 

The controls represent person years at risk experience among exposed and unexposed. Controls are selected concurrently from those still at risk when a case occur. A person selected as a control can later become a case, the opposite is not possible since a case is no longer at risk of developing disease (with a non recurrent disease). A control which later becomes a case is kept in both groups.

 

Using density sampling allows us to compute a rate ratio which is equal to the rate ratio we would have obtained if a cohort study had been conducted to compare rates between exposed and unexposed cohorts in the source population. Density case control studies require an analysis matched on time of disease onset and control selection.

 

 

 

Traditional case control studies

 

The two above study designs used respectively as denominators a sample of the exposed and unexposed source population or a sample of their person time experience.

 

Let suppose that we now are at the end of the follow up period and have respectively Ce and Cu cases, and Ne-Ce and Nu-Cu persons still free of disease (non cases), in the two cohorts. If the disease is rare it is clear from the graph that persons free of disease at the end of the study period reflect the exposure experience of the source population. If the disease is frequent, exposure among persons free of disease at the end of the study may be lower than in the source population (since exposure increases the risk of disease).

 

If the disease is rare, similarly to what we did in the case cohort study, we can use a sample of non cases at the end of the study period to estimate the risk ratio. Using non cases to estimate the source population exposure experience is the principle of traditional case control studies.

 

Let's call "c" and "d" respectively the number of exposed and unexposed in the sample. If sampling is done independently from the exposure status we would expect:

 

c              Ne-Ce                     Ne

---   =       ---------          =        ----           if the disease is rare

d              Nu-Cu                     Nu

 

 

or equivalently

 

c                              d

---------    =              ---------

Ne-Ce                      Nu-Cu

 

If the above is true the risk ratio estimated from a traditional case control study can be represented as:

 

IPe                           a/Ne-Ce                  a              Nu-Cu                     a              d

----           =              ------        =              --     x      ----           =              --     x      -- 

IPu                           b/Nu-Cu                  b              Ne-Ce                     b              c

 

The quantity ad/bc is the odds ratio. It represents the ratio of the odds of disease among exposed divided by the odds of disease among unexposed.

 

However if the disease is not rare a large part of Ne/Nu is represented by future cases who are more likely to be exposed than non cases. Consequently, the odds ratio may dramatically overestimate the risk ratio.

 

Ne-Ce                                                      Ne

---------     may not be equal to              ------

Nu-Cu                                                      Nu

 

To illustrate this point let's now move to the example of a food borne outbreak in a nursing home with 200 residents and 74 cases of gastroenteritis. The epidemic curve is consistent with a point common source of infection and example 4 shows the results of a retrospective cohort study. It suggests that the risk of gastroenteritis is 3.4 times higher among residents who consumed a specific food item compared to those who did not (Ref.).

 

Example 4: Occurrence of gastroenteritis among residents of nursing home A according to consumption of a specific food item.

 

Specific food item              Total           Cases            Risk               Risk ratio

____________________________________________________________________

Yes                                      60             44                  73.3%            3.4

No                                      140             30                  21.4%            Reference

____________________________________________________________________

Total                                   200             74                  37.5%           

 

 

Let's suppose investigators would have preferred to conduct a traditional case control study (case - non cases study) rather than a retrospective cohort.

 

Example 5: Consumption of a specific food item among cases and various samples of residents of a nursing home

_________________________________________________________________________________

Consumption       Cases               50 %sample               OR             50% sample         RR

                                                   of non cases                                of source

                                                                                                      population

_________________________________________________________________________________

Yes                     44                         8                           10.1              30                      3.4

No                       30                       55                                                70                     

_________________________________________________________________________________

 

Using as controls a 50% sample of the non cases the odds ratio equals 10.1, overestimating the risk ratio by a factor of three. This should not come as a surprise, though. When selecting controls from non-cases, and since the disease is frequent (the overall risk of gastroenteritis is 37.5%), the control group is no longer representing the distribution of exposure in the source population. The frequency of exposure in the control group selected from non cases is 7.3% and was 30% in the source population.

 

If instead we had done a case cohort study and chosen a 50% random sample of the source population, the sample (if unbiased and ignoring random variation) would be likely to provide the same proportion of exposed (30%) than in the source population. The risk ratio obtained (3.4) would again be similar to the risk ratio observed in the cohort study.

 

 

 

Conclusion

 

Cohort studies allow to directly measuring risk or rate of disease occurrence and their related ratio in subgroups of a population (exposed and unexposed). Case control studies do not allow measurement of risk or rates. They however allow estimation of the risk ratio and the rate ratio. The selection of the control group is a crucial step of the study. The following table summarises the type of measures and controls selection as described in the above chapter.

 

Measuring risk, rate and odds ratios in a case control study, using various sampling methods for controls (Source: Rodrigues L et al. Int J Epidemiol. 1990;19:205-13)

 

Measure                  Definition       Case control                                              Controls

                                                        Formulation               Design                    sampled from

 

1 - Risk ratio          Ce/Ne              Ce/Cu                          Case cohort           Total study population

                                Cu/Nu              Ne/Nu                                                          present at beginning of study

 

2 - Rate ratio          Ce/PTe            Ce/Cu                          Density                   People at risk

(incidence              Cu/PTu            PTe/PTu                      case control           at time of case disease onset

density ratio)

 

 

3 - Odds ratio        Ce / (Ne-Ce)   Ce/Cu _________       Traditional             People disease free throughout

                                Cu / (Nu-Cu)   (Ne-Ce) / (Nu-Cu)      case control           study period

 

 

 

Case cohort

Controls are randomly selected from the population present (at risk)

at the beginning of the study (Ne and Nu)

Case cohort studies are a reasonable way to conduct case control studies when people in the source population have been followed for the same length of time.

 

 

Density case control

Controls are selected proportionally to the person-time contributed by exposed and unexposed cohorts (PTe and PTu).

A matched analysis on time of selection is necessary to give an unbiased estimate of the incidence density ratio.

Example: prolonged outbreak of hepatitis C in a dialysis unit where 3 controls per cases are sampled among those at risk at the same time as the case occur

 

 

Traditional case control

Controls are selected from people who are free of the disease at the end of the study period (Ne-Ce and Nu-Cu).

The OR is a good estimate of the risk ratio if the disease is rare.

 

 

These parallels in thinking between case-control and cohort studies help to clarify the principles of control selection and illustrate the importance of viewing case-control studies simply as cohort studies with sampled denominators.

 

When facing difficulties in recruiting appropriate controls in a case control study one should think at the cohort that could have been done instead. Identifying the source population, the disease frequency and the type of exposures will then guide the selection of a control group representative of the exposure experience of the source population.

 

 

 

What design should be used and when?

 

Traditional case control studies are an easy and very convenient way to conduct epidemiological studies when the disease is rare. Because of its simplicity it is the most popular method. It has been extremely useful to epidemiologists in the past 50 years. Provided that the disease is rare the odds ratio provides a good estimate of the risk ratio. However it should not be used when disease incidence is high. This particularly applies to investigation of food borne outbreaks with very high incidence.

 

Case cohort studies are not very popular. Their concept in not well understood to the point that some journals would reject a case cohort study on the reason that the control group includes cases. Case cohort studies are a very suitable design when disease incidence is high. They provide a direct estimate of the risk ratio. They are not suited when exposure changes over time (if  exposure is measured at the beginning of a follow up period and differs from the overall exposure experience during the entire study period).

 

Density case control studies are suited for estimation of rate ratios (incidence density rate ratios). They are simple to conduct. In fact, to select a control among persons still free of disease, at the time a case occurs, is common practice frequently called prospective case control study. It provides a good estimate of the rate ratio. Density case control studies are suited when unequal length of follow up occurs for study members.

 

Advantages and disadvantages of cohort and case control studies.

 

Many text books have described advantages and disadvantages of cohort and case control studies. The following table summarises usual comments.

 

 

 

Cohort studies

Case control studies

Suited for rare diseases

No

Yes since starting wit a set of cases

Suited for rare exposures

Yes since starting with exposure status

No

Allows for studying several exposures

Difficult but examples exists

(Framingham study)

Yes

Allows for studying several outcomes

Yes

No

Disease status easy to ascertain

Sometimes difficult

Easier since starting point of the study

Exposure status easier to ascertain

Yes since starting point of the study.

Except for retrospective cohorts

Sometimes difficult.

Information biases.

Allows computation of risk and rates

Yes

No

Allows computation of effect

Computation of risk ratio

and rate ratio

Estimation of risk ratio, rate ratio

from odds ratio

Allows studying natural history of disease

Yes

Easier to show that cause precedes effect.

More difficult

Temporality between cause and effect difficult to establish

Based on existing data sources

Difficult

Yes but access to information sometimes difficult

Easiness to find a reference group

Usually not difficult to identify an unexposed population

No

Major potential biases when selecting a control group

Sample size

Large

Small

Cost

Elevated

except if retrospective cohorts

Smaller

Time required

Long, sometimes very long except if retrospective cohorts

Shorter

Follow up

Difficult, loss to follow up

No follow up

Logistics

Heavy

Many staff, large data sets

Long duration

Easier

Concept

Easy to understand

Difficult to understand particularly if case cohort or density case control study

Ethical issues

Major if studying risk factors.

Interruption of study if exposure shown to be harmful.

Need for intermediate analysis.

None since outcome already happened.

 

 

 

 

Case cross over design

 

 

Among cohort designs, cross over studies are intervention studies in which the same group of people is exposed to two different interventions in two separate periods of time. This requires that the effect of the intervention is short enough not to impact on the effect of the second intervention and that a time gap between the two interventions is respected.

 

Case cross over studies are the case control version of crossover studies. This concept was introduced by Maclure in 1991 (Am J Epidemiol 1991;133:144-53). In a case cross over design all subjects are cases and exposure is measured in two different periods of time. The general principle is to find an answer to the question: "Was the case-patient doing anything peculiar and unusual just before disease onset?" or "Did the patient do anything unusual compared to his routine?". The assumption is that if there are triggering events, these events should occur more frequently immediately prior to disease onset than at any similar period distant from disease onset.

 

In case cross over studies, instead of obtaining information from two groups (cases and controls), the exposure information is obtained from the same case group but during two different periods of time. In the first period exposure is measured immediately before disease onset. In the second period exposure is measured at an earlier time (supposed to represent background exposure in the same person). Exposure among cases just prior disease onset is then compared to exposure among the same cases at an earlier time. Each case and its matched control (himself) are therefore automatically matched on many characteristics (age, sex, socio economic status, etc.).

 

To illustrate that point Maclure used the following example. Let suppose we study the role of heavy physical activity in the occurrence of myocardial infraction (MI). Using a case cross over design we could document exposure to heavy physical activity among cases in the hour immediately preceding MI. We would then document exposure to heavy physical activity among those same cases at another earlier time.

 

The following figure illustrates periods of exposures taken into account in a case cross over study.

Source: Adapted from Jean Claude Desenclos, InVS, France

 

In the above figure the period immediately before onset is called the « current » period and the other period "the reference period". The two periods are separated by a "wash out period" in order to avoid that exposure in the reference period is mixed with exposure in the current period. The reference period of exposure is used to reflect average exposure experience among cases. Case 1 was unexposed in current period (just prior to onset) and exposed in the reference period. Case 2 was exposed just prior onset and unexposed in the reference period. Case 3 was exposed in both periods and case 4 in none.

 

From the above we should consider that the same case and its 2 periods of exposure constitute a matched pair. Cases 1 and 2 are discordant pairs and cases 3 and 4 concordant. This is why with a case cross over design a matched pair analysis is required. Only discordant matched pairs will be used in the analysis (see chapter on matching for rational).

 

In addition some characteristics of exposure and outcome are noteworthy.

 

Exposure should change over time in the same person and over short period of time.

 

Exposure should not be changing in a systematic way over time. In the example of physical activity let' suppose we have documented exposure in the hour immediately before onset and that we have documented reference exposure two days before at the same time. This would not be appropriate if physical activity occurs in a systematic timing (every second day at the same time).

 

Exposure should have a short term effect. Duration of exposure effect should be shorter than average time between two routine exposures in the same individual. The effect of a first exposure should have stopped before the next exposure.

 

Induction time between exposure and outcome should be short.

 

Disease must have an abrupt onset. Case cross over are not appropriate if the exact date/time of onset is not available or if abrupt onset does not exist (some chronic disease).

 

Several reference time periods can be used to document average exposure among cases. In that instance, an average of time being exposed is computed and compared to exposure just prior disease onset. The efficiency of the case cross over method increases with the number of reference periods included.

 

As in any case control study the capacity to properly document exposure should be identical in the two periods of time. In case cross over designs information biases are a sensitive issue.

 

Even if confounding is controlled since a case is its own control, within-person confounding can occur. In the example of heavy physical activity and MI, another factor (anger) may be linked both to exposure (heavy physical activity) and outcome (MI).

 

 

Case cross over and food borne outbreaks.

 

Case cross over design was sometime used by epidemiologists to try to identify a food item as the vehicle for a food borne disease outbreak. Several of the above listed points merit to be challenged. A recall (exposure) period of around three days may be too large to use this design. In addition food habits (average exposure) do not happen randomly in an individual. Finally, comparing consumption of a potentially infected food item in the "current" period to average consumption of a similar un-infected food item in the reference period does not relate to the same exposure. Consumption of a food item could be identical in the current and reference time periods and still only the food item in the current period was contaminated.

Case to case study design

 

 

Case to case studies are types of case control studies used when the disease of interest can be sub-classified in two or several groups that have specific risk factors. In a case to case study cases with a particular sub-type of a disease are compared to cases with another subtype. For example during a listeriosis outbreak, cases with the outbreak sub-type would be compared to sporadic cases (the controls).

 

Some assumptions are made. Non outbreak cases (the controls), if infected with the outbreak subtype would have been classified as cases. They come from the same population which gave rise to outbreak cases. They represent exposure (e.g. food consumption) in the source population for outbreak cases. This is probably the major issue. Are sporadic cases of listeriosis representing food consumption in the general population? This may not always be true. Non epidemic cases may be more likely to be exposed than the overall source population. We may therefore underestimate the odds ratio.

 

Some advantages lie with case to case studies. Cases are readily available. Since all subjects in the study are sick there also may be less differential recall between cases and controls.

 

Case to case studies may be a convenient design when information is available for the sub class of cases used as controls. However, as in any case control study, investigators need to be very cautious and verify that exposure in the control group reflects accurately exposure in the source population for cases.

 

 

 

References

 

Cornfield (1961

 

Greenland (1982), Thomas DC. On the need for the rare disease assumption in case-controls  studies. Am. J. Epidemiol. 1982;116:547-553.

 

Hernandez-Diaz S et al. Am J Epidemiol 2003;158:385-391

 

Miettinen OS. Estimability and estimation in case(referent studies. Am. J. Epidemiol. 1976;103:226-235.

 

Mittleman MA, Maclure M, Robins JM. Control sampling strategies for case-crossover studies: an assessment of relative efficiency. American Journal of Epidemiology 1995;142(1):91-8.

 

Maclure M. The case-crossover design: a method for studying transient effects on the risk of acute events. American Journal of Epidemiology 1991;133(2):144-153.

 

Mittleman MA, Maldonado G, Gerberich SG, et al. Alternative approaches to analytical designs in occupational injury epidemiology. American Journal of Industrial medicine 1997;32(2):129-41.

 

Rodrigues L et al. Int J Epidemiol 1990;19:205-13.

 

Rothman KJ; Epidemiology: an introduction. Oxford University Press 2002, 73-93

Smith (1984)

 

Suisa S. The case-time-control design. Epidemioogy. 1995;6:248-253.

 

Greenland S. Confounding and exposure trends in Case-cross-over and case-time-control designs. Epidemiology. 1996; 7231-239.

(Which refutes conclusions of the Suisa's article).