To assess case-only designs for surveillance with administrative databases.
To assess case-only designs for surveillance with administrative databases.
We reviewed literature on two designs that are observational analogs to crossover experiments: the self-controlled case series (SCCS) and the case-crossover (CCO) design.
SCCS views the ‘experiment’ prospectively, comparing outcome risks in windows with different exposures. CCO retrospectively compares exposure frequencies in case and control windows. The main strength of case-only designs is they entail self-controlled analyses that eliminate confounding and selection bias by time-invariant characteristics not recorded in healthcare databases. They also protect privacy and are computationally efficient, as they require fewer subjects and variables. They are better than cohort designs for investigating transient effects of accurately recorded preventive agents, for example, vaccines. They are problematic if timing of self-administration is sporadic and dissociated from dispensing times, for example, analgesics. They tend to have less exposure misclassification bias and time-varying confounding if exposures are brief. Standard SCCS designs are bidirectional (using time both before and after the first exposure event), so they are more susceptible than CCOs to reverse-causality bias, including immortal-time bias. This is true also for sequence symmetry analysis, a simplified SCCS. Unidirectional CCOs use only time before the outcome, so they are less affected by reverse causality but susceptible to exposure-trend bias. Modifications of SCCS and CCO partially deal with these biases. The head-to-head comparison of multiple products helps to control residual biases.
The case-only analyses of intermittent users complement the cohort analyses of prolonged users because their different biases compensate for one another. Copyright © 2012 John Wiley & Sons, Ltd.
Pharmacoepidemiologists who monitor the safety of medical products using healthcare administrative databases are increasingly interested to know when case-only designs can or cannot be used. To address this question, we (i) defined case-only designs in relation to each other; (ii) examined their main strength: self-controlled comparisons; (iii) discussed the major difference among the designs: directionality; (iv) described the range of medical products assessed with these designs in relation to their susceptibility to exposure misclassification; and (v) made recommendations to safety surveillance programs.
The defining feature of case-only designs is that the analysis is restricted to cases, that is, people who experience at least one outcome event. Also in these designs, each individual serves as his or her own self-matched control, hence the term self-controlled.[1-3] The term crossover arises when the analysis is restricted to people who supply both exposed and unexposed person-time and thus ‘crossover’ between two or more exposure levels.[4, 5]
Figure 1 shows the relationship between the self-controlled case series (SCCS) and the case-crossover (CCO) designs, using data from Table 2 in ‘Tutorial in Biostatistics: The Self-Controlled Case Series Method’ by Whitaker et al. The figure shows the times when 10 children (i) received measles–mumps–rubella vaccinations, (ii) passed through a hypothesized 14-day induction period after the vaccination, (iii) then through a hypothesized 21-day effect period (called by Whitaker et al. the prerisk period and the risk period, respectively, terms that are potentially ambiguous when there are nonzero risks in the ‘prerisk’ and ‘postrisk’ periods), and (iv) the day they were diagnosed with meningitis. In the top half, the times are expressed in terms of the child's age in days, and the data are viewed, as in a SCCS, like a cohort of 10 children followed through their second year after birth. In the bottom half, the times are relative to the day of meningitis diagnosis, and the data are viewed retrospectively from the standpoint of the diagnosticians, as in a CCO analysis.
Figure 2 shows the same data realigned, so vaccination date is time zero. A small epidemic of meningitis is visible in the ‘after period’, which is very unlike the incidence in the ‘before period’. This asymmetry of outcomes before and after exposure onset is the focus of sequence symmetry analysis (SSA), an elegantly simple technique for hypothesis screening with large databases.[6, 7] Under the null hypothesis of no direct or indirect causal relation between exposure and outcome, the incidence of outcomes is expected to be symmetric around time zero. Either the exposure or the outcome variable can be set as time zero. If we did an SSA with the outcome date set as time zero, it would look like Figure 1b. Note that the pattern of vaccinations in Figure 1b is an exact mirror image of the pattern of diagnoses in Figure 2.
The arrow in Figure 1b, labeled ‘days before diagnosis’, indicates the meaning of the term unidirectional. The standard CCO is a unidirectional design because it looks at exposure frequency only retrospectively, before the time of the outcome. A bidirectional CCO design also includes control time after the outcome. The standard SCCS is bidirectional because it includes unexposed (control) time from both before and after vaccination (two-headed arrow at top of Figure 1a.) The SSA is best viewed as a simplified SCCS in which the unexposed (control) time is only from the period before first exposure; it is therefore unidirectional (arrow at top of Figure 2).
The fundamental commonality of the three designs is that analyses are conditioned on the individual (i.e. one person per stratum) yielding only within-person comparisons. This is a major strength for assessing medical products because we are concerned about potential selection bias or confounding by factors not recorded in healthcare databases. By structuring the comparisons so that each person serves as their own control, we eliminate confounding and selection bias by constant (time-invariant) characteristics, such as chronic regular use of nonprescription drugs, average physical activity, long-term diet, alcohol drinking pattern, habitual health behaviors, tendency to seek professional care (‘medicalization’), long-past health events such as illnesses, vaccinations and injuries, occupation, social support, ethnicity, smoking history, and body mass history.
Self-matching results in strata of three types: (i) individuals always exposed, (ii) individuals never exposed, and (iii) individuals sometimes exposed and sometimes not, that is, who ‘crossover’. Types 1 and 2 (called concordant individuals in CCO literature) automatically drop out of the calculation of the relative risk. For example, self-concordant individuals contribute only zeros to both the numerator and the denominator of the Mantel–Haenszel odds ratio (OR). Thus, merely by deciding to do a highly stratified analysis, we are left with type 3, a subgroup analysis comprising only individuals who cross between exposed and unexposed time, as in a crossover experiment. This is a second commonality among the designs: the study base is restricted to people who cross between levels of exposure.
In analyses involving additional explanatory variables, exposure-concordant patients are retained if they cross between levels of other factors. For example, referring to Figure 1a, Whitaker et al. included child 10 in their two-variable analysis because child 10 did cross between levels of their binary variable for age (less than versus greater than 547 days.)
Of course, a third commonality is the population of analysis is restricted to cases. The restriction can be regarded either as the primary characteristic that causes us to use a self-controlled analysis, or as a secondary consequence of having only one person per stratum. Reasons for preferring to analyze only cases include the following: (i) with fewer patients and fewer data on time-invariant variables, case-only designs help protect data privacy and are computationally efficient, and (ii) sometimes signal refinement requires additional data captured from charts to rule out potential biases, in which circumstances case-only designs would require less data to be captured than case-control studies. Self-controlled analyses would follow of necessity because of the restriction to cases.
In large safety monitoring initiatives that use nationwide or distributed databases, data privacy and efficiency are already established, and the main benefit of case-only designs is a self-controlled analysis to control for unmeasured factors. Accordingly, some investigators regard a case-only design as just a highly stratified analysis of cohort data. This view is illustrated by Fosbol et al. in reporting their study of nonsteroidal anti-inflammatory drugs (NSAIDs) in relation to myocardial infarctions (MIs) and deaths in a Danish cohort of one million people followed 9 years. In their Table 4, the authors present results from CCO analyses (using conditional logistic regression) that have exactly the same population totals as their Table 3 of results from cohort analyses (using Cox proportional hazards models). A benefit of this approach is to reduce the reader's potential confusion by holding the overall context of the analysis constant. However, the reassurance is potentially misleading because there was a shift of the study base when the computer automatically restricted the CCO analysis to the subset of the cohort who crossed between levels of exposure. The meaning of the exposure variable and the operational hypothesis can shift without the investigator realizing it. For illustration, if an analysis of the effect of sex involved stratification by individual, all people whose sex was constant would drop out and the study base would automatically be restricted to people who changed their sex. At the same time, the meaning of the variable ‘male’ would automatically shift to a combination of ‘becoming male’ and ‘the negative of becoming female’. Therefore, we believe case-only designs should not be regarded as just highly stratified analyses of cohort data.
A fourth commonality is that both SCCS and CCO analyses are influenced by the same assumptions about the lengths of the induction period and effect period (the difference between the maximum and minimum induction times in the population.) This can be seen from child 5 and child 7. Child 5's meningitis was diagnosed 5 days after the effect period ended. If the authors had chosen a 4-week rather than a 3-week effect period, child 5's diagnosis would have fallen within that period in both SCCS and CCO analyses. Child 7's meningitis was diagnosed on the second day of the effect period. If the authors had chosen a 3-week rather than a 2-week induction period, child 7's diagnosis would have fallen outside the effect period in both the SCCS and CCO analyses.
As a result of these four commonalities among case-only designs, their applicability to safety monitoring of medical products is very similar. They are both suitable for measuring transient effects of accurately recorded exposures on the immediate risk of illnesses with abrupt onset rather than for assessing cumulative effects of long exposures or illness with gradual onset. They both estimate only the within-person transient effect of an exposure event, controlling for any cumulative effect of previous chronic exposure to the same agent (where cumulative effect is defined as any effect of past exposure on the background level of risk in the unexposed times in the window of observation in the case-only analysis.) For safety monitoring programs, the distinction between an adverse event caused by a transient effect of exposure to a medical product and the same type of adverse event caused by a cumulative effect of chronic exposure to the same medical product can be of great importance, especially in the rare situation that the transient and cumulative effects are in opposite directions. For example, a case-only design could yield a relative risk greater than 1, whereas a cohort design with a time-to-event analysis yields a relative risk less than 1 if the medical product has a cumulative protective effect that is greater than its transient harmful effect (e.g. exertion and possibly alcohol consumption can trigger an MI, yet both taken chronically are believed to reduce the risk of MI in the long term.) Both relative risk estimates could be correct because they might measure different biological effects or because they address different operational hypotheses  related to a common biological effect.
Standard CCO designs are unidirectional, right-censored at outcome to avoid, or reduce, reverse causality,  called event-dependent exposure in SCCS literature. A mild form of reverse causality is quite common in drug safety investigations: it is an indirect causal connection between the outcome and the subsequent exposure because of the tendency for outcome-related care (especially hospitalization) to involve review of all the patient's drugs, resulting in some being stopped, possibly just to ‘make room’ for new drugs added. A more serious form is when the outcome directly causes stopping because it is a contraindication. For example, a CCO study of cholinesterase inhibitors (CIs) and risk of hospitalization for bradycardia (a known effect of these drugs) observed 43% of users discontinued CIs after discharge from hospital. By excluding person-time after the outcome, CCOs eliminate a major opportunity for reverse causality. However, they do not eliminate reverse-causality biases that occur before the measured outcome event (e.g. hospitalization), such as within-person protopathic bias and confounding by indication (or contraindication) wherein prodromal signs of the outcome cause (or prevent) initial use of a medical product.
A standard SCCS analysis of the same database on CIs and bradycardia hospitalizations would have included posthospitalization time in assessing each patient's total time exposed and unexposed, which would have biased the relative risk estimate upwards. Sometimes, reverse causality has only a transient effect (e.g. contraindications of vaccination), in which case a bidirectional SCCS design can be used that removes a window of time before exposure from the person-time used for outcome estimation (which is equivalent to excluding that window after an outcome in a bidirectional CCO). However, reverse causality between outcomes and drugs can be prolonged: after hospitalization for bradycardia, prescribing of CIs to some patients would cease permanently.
A special form of reverse-causality bias, aptly named immortal-time bias, arises because death eliminates a patient's future opportunity for all exposures. It occurs when a cohort is defined in the middle of the follow-up period, as happens in the standard SCCS and SSA. To enter the cohort, one must have at least one exposure event and therefore must have survived the period before exposure onset. This forbids death in the before period but not in the after period. For example, if the outcome in Figure 1 were death, child 1 would not have survived to be vaccinated. Like child 10, child 1 would contribute nothing to a single-variable (only exposure) analysis of the effect of vaccination; the estimate would be biased upwards by underestimating prior mortality.
Two ways to deal with immortal-time bias, if the period before first exposure is to be retained in the analysis, are (i) to exclude deaths also from the after period (in which case, the standard SCCS and SSA cannot be used to assess risks of fatal outcomes) and (ii) to include deaths in the before period, which could be done only in rare situations when it is known that the deceased would have been exposed at a known point if they had survived (e.g. vaccinations at a certain age when almost every child is vaccinated). The simplest way to eliminate immortal-time bias is to exclude the period before first exposure. Although this unidirectional approach was used in the first SCCS, it is not used in the tutorial on the standard SCCS. Recently, Farrington et al. proposed a generalization of the SCCS method to cope with multiple exposures, ‘the key … principle [being] we start with the last observed pre-event exposure and work back through the exposures’. This statement corresponds to Figure 1b, which suggests that the method could share some of the limitations of the unidirectional CCO, such as exposure-trend bias. Kuhnert et al. presented a unidirectional SCCS in which only the exposure period following the last vaccination is considered, which allows them to consider fatal outcomes.
Unidirectional CCOs are susceptible to exposure-trend bias because the control window always precedes the case window.[15, 16] If exposure to a medical product is growing rapidly in the source population, the case window will be more exposed than the control window, especially if those windows are long or far apart. Initially, when a safety surveillance program is monitoring a new medical product, exposure-trend bias could be a major concern. If the new medical product is used chronically (persistently without interruption) by most patients, then in the initial period of follow-up of an inception cohort, the only self-discordant individuals in a unidirectional CCO would be starters (the discordant-exposed cases). To be a discordant-unexposed case, a person would need to be a previous user who stopped, and initially these would be few, if usage of the new medical product were continuous. This temporary shortage of discordant-unexposed patients would mean the initial discordant pair ratio (an estimate of the relative risk) would be spuriously very high. It would approach the true relative risk as the population approaches a steady state of starting and stopping. A steady state is reached almost immediately if the new medical product is used only briefly, for example, a vaccine.
A bidirectional CCO was first developed to deal with exposure-trend bias in studies of the health effects of air pollution, a setting where there is no possibility of reverse causality: air pollution levels are not affected by rates of hospitalization. A bidirectional CCO includes control windows after the outcome so that if control windows are sampled symmetrically from the left and right of meningitis in Figure 1b, a linear background trend in exposure cancels out. In air pollution studies, effects on fatal outcomes still can be assessed because, although a patient's death eliminates their individual future exposure, it does not affect the population's future exposure to air pollution, which is an ecological or group-wide exposure. It is not clear whether there is an analogous ecological exposure in pharmacoepidemiology, other than the sudden withdrawal of a drug from the market. Bidirectional CCOs of fatal outcomes would work in these situations.
Another way to deal with exposure-trend bias is the case–time–control design. It is a unidirectional CCO plus a unidirectional time-matched noncase group (i.e. a traditional matched control group sampled from the population that produced the cases). Exposure ORs are calculated the same way in the case group and the noncase (control) group, and the latter's OR is considered an estimate of the exposure-trend bias in the former's OR. Dividing the case OR by the control OR gives an adjusted OR that has little, or at least less, exposure-trend bias. In response to the concern that noncases might have different exposure trends than cases, the case–time–control design has been adapted using future cases as present controls, an adaptation called the case–case time–control design.
The standard SCCS is bidirectional and therefore much less susceptible to exposure-trend bias. In both the standard and unidirectional SCCS, the time trend in exposure probability can be included in the model. SSA is very susceptible to exposure-trend bias because of its unidirectional reference period; the unexposed control time always precedes the time of first exposure.
The first reason that case-only designs complement cohort designs is because intermittent users complement continuous users and continuous nonusers, together comprising the entire population. Continuous users and nonusers are the purest subgroups from the standpoint of an investigator of a cohort study, whereas the intermittent users are problematic, like patients who do not adhere to protocol in a randomized controlled trial. In contrast, an investigator of a case-only self-controlled design is interested in intermittent users more than continuous users and nonusers.
A second reason is that the biases in analyzing intermittent users complement the biases in analyzing continuous users. We have seen that intermittent users enable self-controlled designs that eliminate time-invariant confounding. However, this often comes at a price: greater potential for bias from exposure misclassification when dispensing date in healthcare databases is not a good measure of the timing of self-administration. Another source of greater susceptibility to exposure misclassification occurs when using the discordant-pair ratio to estimate the OR: error in one of the paired observations robs information from the other observation because the pair becomes falsely self-concordant and drops out. Continuous users enable better exposure classification, but the price is greater potential bias because of unmeasured confounders and selection factors. Among intermittent users, the degree of exposure misclassification, of course, depends on the nature of the medical product.
Table 1 lists medical products that have been assessed by case-only designs, ranked approximately by their brevity of use and effect periods and the accuracy of data on exposure timing. Brevity is related to accuracy of exposure timing because unknown stopping times can be better guessed from starting times when an exposure is normally brief, for example, an antibiotic. If stopping times could be determined accurately, long exposures might be preferred because they might yield more exposed cases and greater statistical power. However, longer exposures allow more time for time-dependent covariates to cause bias by influencing both the stopping time and the risk of outcome, as happens when comorbidities affect adherence to prescriptions.
|Medical product||Windows (days)||Adverse events||Design||Reference|
|A. Professionally administered|
|Acellular pertussis vaccine, DTaP||0, 1–3||Seizures||SCCS||Huang et al.1|
|DTP||0–3, 4–7, 8–14||Febrile convulsions||SCCS||Farrington et al.2|
|DTP||0–3, 4–7, 8–14||Convulsions||SCCS||Gold et al.3|
|DTP, diphtheria–tetanus (DT/Td)||0–3, 4–7, 8–14||Convulsions||SCCS||Andrews et al.4|
|DTP/Hemophilus influenzae||0–3, 0–7||Fever and convulsions||SCCS||Ward et al.5|
|Penta-/hexavalent, multidose vaccines||3||Unexplained sudden unexpected death||SCCS||Kuhnert et al.6|
|MenC||0–3, 4–7, 8–14||Convulsions||SCCS||Andrews et al.4|
|MenC||0–3, 0–7||Fever and convulsions||SCCS||Ward et al.5|
|MenC||0–27||Idiopathic thrombocytopenia purpura||SCCS||Andrews et al.4|
|MenC||0–30;0–60;0–180||Relapse in nephrotic syndrome||SCCS||Taylor et al.7|
|MMR||6–11||Convulsions||SCCS||Musonda et al.8|
|MMR||6–11, 15–35||Convulsions||SCCS||Andrews et al.4|
|MMR||6–11,15–35||Convulsions and encephalitis||SCCS||Ward et al.5|
|MMR||6–11, 15–35||Febrile convulsions, aseptic meningitis, purpura||SCCS||Farrington et al.2|
|MMR||6–11||Febrile convulsions||SCCS||Gold et al.3|
|MMR||1–30, 31–60||Gait disturbance||SCCS||Miller et al.9|
|MMR||42||Aseptic meningitis||CCO||Ki et al.10|
|MMR||0–42||Idiopathic thrombocytopenia purpura||SCCS||Miller et al.9|
|MMR||15–35||Idiopathic thrombocytopenia purpura||SCCS||Gold et al.3|
|MMR||0–42||Idiopathic thrombocytopenia purpura||SCCS||Andrews et al.4|
|MMR||1–30, 31–60, 61–91||Invasive bacterial infection||SCCS||Miller et al.11|
|Oral rotavirus vaccine||0–2,3–7,8–14,15–21||Intussusception||SCCS||Murphy et al.12|
|OPV||0–13, 14–27, 14–41||Intussusception||SCCS||Andrews et al.13|
|OPV||0–7, 8–14,15–21,22–28,29–35,36–42||Intussusception||SCCS||Galindo Sardiñas et al.14|
|OPV||3–7,8–21,14–27,28–41||Intussusception||SCCS||Cameron et al.15|
|Influenza vaccine||2,14||Asthma||SCCS||Kramarz et al.16|
|TIV||0, 1–3, 4–7, 8–14, 15–28||immunization-related adverse events and multiple other diagnoses||SCCS||Mullooly et al.17|
|TIV||0–2,1–2,0–7,1–21,1–42||Guillian–Barré syndrome and multiple other diagnoses||SCCS||Greene et al.18|
|TIV||0–3, 1–14,15–42||Gastritis/duodenitis and multiple other diagnoses||SCCS||Hambidge et al.19|
|TIV||0–2,1–3,1–14||Impetigo and multiple other diagnoses||SCCS||France et al.20|
|Influenza vaccine||1–14, 15–28, 29–59…||Acute MI||SCCS||Gwini et al.21|
|Influenza vaccination||1–2, 3–14||Asthma, COPD exacerbations||SCCS||Tata et al.46|
|Parenteral inactivated influenza vaccine||1–30, 31–60, 61–91||Bell's palsy||SCCS||Musonda et al.8|
|Influenza vaccine—inactive nasal form||1–30, 31–60, 61–91||Bell's palsy||SCCS||Stowe et al.22|
|Influenza vaccine—inactive nasal form||1–30, 31–60, 61–91||Bell's palsy||SCCS||Mutsch et al.23|
|Influenza; tetanus; pneumococcus||1–3,4–7,8–14,15–28, 29–91||MI and stroke||SCCS||Smeeth et al.24|
|DTP, HBV, HIB, OPV, MMR||1–7,8–14,15–30,31–44||Wheezing||SCCS||Mullooly et al.25|
|DTP, DTaP, hepatitis B, or any vaccine||42||Immune hemolytic anemia||SCCS||Naleway et al.26|
|Hepatitis B, tetanus, Influenza, other vaccines||0–30;0–60;0–90||Multiple sclerosis relapse||CCO||Confavreux et al.27|
|Hepatitis B vaccine||0–60, 61–365||Central nervous system demyelinating events||SCCS||Hocine et al.28|
|MMR||1 yr; 2yr||Autism||SCCS||Taylor et al.9|
|MMR||5 yr||Autism||SCCS||Farrington et al.29|
|Colonoscope||7; 28||Ulcerative colitis exacerbation||CO||Menees et al.30|
|Quinolones, sulfonamides, azoles (with warfarin)||0-5,6-10,11-15,16-20||Gastrointestinal bleed||CCO||Schelleman et al.31|
|Antibiotics||0-15;16-30;31-60||Flare of inflammatory bowel disease||CCO||Aberra et al.32|
|Macrolides and fluoroquinolones||0-28; 0-168||Ventricular arrhythmia and cardiac arrest||CCO/TC||Zambon et al.33|
|NSAIDs||1, 3, 6||Diarrhea||CCO||Etienney et al.34|
|NSAIDs||0-28||Hepatitis||CCO||Lee et al.35|
|NSAIDs||0-30||MI, death||CCO||Fosbøl et al.36|
|NSAIDs||0-30||MI, stroke, cardiovascular death||CCO||Fosbøl et al.36|
|NSAIDs||0-30||Death, MI or second hospitalization for heart failure||CCO||Gislason et al.37|
|NSAIDs||0-30||Death or reinfarction after MI||CCO||Gislason et al.38|
|NSAIDs||0-30||Stroke||CCO||Chang et al.39|
|NSAIDs||0-90||Gastrointestinal bleeds||CO||Biskupiak et al.40|
|Benzodiazepines, other psychotropics||1*; (first 1-14)||Motor vehicle accident||CCO||Barbone et al.41|
|Benzodiazepines, antipsychotics, all||2**||Falls||CCO||Neutel et al.42|
|Benzodiazepines||0-7||Motor vehicle crashes||CCO||Hébert et al.43|
|Benzodiazepines, zoplicone, zolpidem||1*||Motor vehicle accident||CCO||Yang et al.44|
|Psychotropics||0-7||Motor vehicle accident||Both||Gibson et al.45|
|Tricyclic and SSRI antidepressants||0-7,8-14,15-21,22-28||MI||SCCS||Tata et al.46|
|SSRI antidepressants, ASA, NSAIDs||1*||Gastrointestinal bleeds||CCO||Dall et al.47|
|Antipsychotics||0-7||Stroke||SCCS||Pratt et al.48|
|Antipsychotics||0-35,36-70,71-105||Stroke||SCCS||Douglas et al.49|
|Bupropion||0-7||Sudden death||SCCS||Hubbard et al.50|
|CIs||0-90||Bradycardia||CCO/TC||Park-Wyllie et al.51|
|Statin||0-84;0-182||Myopathy, myalgia||CCO||Molokhia et al.52|
|Antihypertensives||33 months||Depression (initiation of antidepressant)||SSA||Hallas et al.53|
|Angiotensin-converting enzyme inhibitors||~6 months||Lupus exacerbation||CCO||Duran-Barragan et al.54|
|Acitretin||0-20||Vulvovaginal candidiasis||CCO||Sturkenboom et al.55|
|Inhaled anticholinergics, corticosteroids, and beta agonists||0-30||Stroke||SCCS||Grosso et al.56|
|Folic acid antagonists||60||Birth defects||CCO/TC||Hernández-Díaz et al.57|
|Chinese herbs (prescribed)||14,21,30,60,90||Hepatitis||CCO||Lee et al.58|
|Ephedrine, caffeine||1*; (also: 0-90)||Cardiovascular events||CCO||Hallas et al.59|
|Many drugs||2-32;2-62;2-92;2-122||Central nervous system dysfunction||CCO||Wang et al.60|
|Many drugs||90||Psoriasis vulgaris hospitalizations||CCO||Cohen et al.61|
|Isotretinoin, minocycline||365||Antidepressant prescription||SSA||Hersom et al.62|
Vaccines and episodic use of medical devices (e.g. colonoscopy) are ranked at the top of Table 1 because the exposure timing is well documented by the health professionals who administer them. Although we found no report of a case-only design for assessing a periodically injected monoclonal antibody, such a report would have been ranked near the top of Table 1. The SCCS method is particularly suited to situations when the investigator can accurately classify each person's time in the effect period (the window when a population is susceptible to adverse effects caused by the use of a medical product at a specific time), which is why the SCCS method is routinely used both for active prospective surveillance and retrospective investigations of vaccines. Exposure misclassification in vaccine assessments arises mainly from overly exclusive or inclusive time windows for the effect period, which results from mistaken assumptions about the minimum and maximum induction times or other aspects of the distribution of induction times in the population.
Ranked second are antibiotics because they are normally used briefly and infrequently, and the timing of use is usually immediately after dispensing. However, many people stop taking antibiotics before the full course of tablets is complete. Therefore, exposure misclassification increases with days since dispensing. Unlike vaccines, but like colonoscopy, antibiotics are usually prescribed in response to medical problems, and occasionally in anticipation of a medical event (e.g. imminent surgery). Therefore, within-person confounding by indication and reverse-causality bias are more likely in assessments of antibiotics than vaccines.
Case-only designs to investigate NSAIDs are more susceptible to exposure time misclassification than antibiotics because NSAIDs are not normally prescribed as a course; many patients take them sporadically. Immediately after dispensing, the probability of NSAID use is high, but choosing a cutoff date when usage has probably stopped is more difficult than with antibiotics. Therefore, overall relative risk estimates from case-only designs are more questionable for NSAIDs than for antibiotics.
The more uncertainty there is about when patients were exposed to a product, the more selective the investigator must be about what people and what times to include. For example, we can exclude sporadic users and do a case-only analysis restricted to people with a series of regularly spaced dates of NSAID dispensing spanning several months, preceded and followed by long periods with no dispensing of NSAIDs.
Some oral medication regimens involve scheduled intermittent use. For example, etidronate for osteoporosis is commonly given orally for 2 weeks followed by a 3-month period of nonuse. Patterns of etidronate dispensing dates would permit reasonable guesses as to when the patients were more likely and less likely to be taking etidronate, so a case-only design could be feasible.
Some psychotropic medications (e.g. drugs for anxiety) are taken sporadically in response to fluctuating symptoms, so the days when the patient is exposed are mostly unknown and case-only designs are challenging and sometimes infeasible. Feasibility is improved by the use of comparator drugs. For example, in an investigation of benzodiazepine use in relation to vehicle collisions, other psychotropic drugs served as comparators; the association was greatest for benzodiazepines. If psychotropics are taken with regularly spaced dispensing dates spanning several months and this period of probably continuous usage is preceded and followed by long periods with no dispensing, then a case-only design would be feasible.
Many cardiovascular medications are prescribed as lifelong therapies and would not be amendable to case-only designs if all patients were adherent. However, stoppers are common enough that case-only designs have proven to be possible, although the reason for stopping might be an unmeasured contraindication.
The more selective we are about subsets of people to include, the more selective we are inclined to be about subsets of times to include. Consequently, as we move down Table 1 and exposure timing becomes more inaccurate, the more attractive are matched CCO designs. If outcome time is known more accurately than exposure times, it makes sense to choose the outcome as time zero and inspect the patterns of exposure data in a case window and a matching control window. Also, inspecting data for potential within-person confounding by factors that coincide with both the outcome and the immediately preceding exposure event is probably easier to do when the outcome event is chosen as time zero. By analogy, within-person confounders would probably be easier to visualize in Figure 1b than that in Figure 1a simply because of the way the data are aligned.
One message from Table 1 is that SCCS and CCO designs are complementary. The SCCS design is preferable at the top of the table and the CCO design becomes more preferable as we move down. There is no obvious cutoff point. In the middle, some investigators (whom we might call ‘lumpers’) would prefer to keep all the observation time in the analysis and deal with threats to validity by including additional terms in statistical models, as in SCCS designs. Other investigators (whom we might call ‘splitters’) would prefer to handle threats to validity by restriction/selection, as in CCO designs. Other factors that played no role in Table 1 rankings, for example, time-varying confounding, would also influence investigators' preferences.
The further down Table 1 we go, the more chronic users there are in the population and the more we regard cohort designs as primary and case-only designs as secondary. Also, the further down we go, the more we rely on head-to-head comparisons, which in case-only designs entails examining multiple medications in relation to one class of outcome. In reviewing these studies, we repeatedly found that comparing relative risks for different drugs, particularly similar active comparators or ‘negative control’ drugs that are expected to have no effect, was helpful for assessing potential biases. Therefore, safety surveillance programs should anticipate investigators will want to assess multiple comparator products as controls in case-only investigations that are primarily intended to assess a single medical product.
The authors declare no conflict of interest.
This work was supported by the Mini-Sentinel which is funded by the Food and Drug Administration through the Department of Health and Human Services contract number HHSF223200910006I. The authors thank Kerry Patriarche for assistance with reviewing papers. This review of published papers was not reviewed by an ethics committee.