Secondary Logo

Journal Logo

Pharmacoepidemiology

COX-2 Selective Nonsteroidal Anti-inflammatory Drugs and Risk of Gastrointestinal Tract Complications and Myocardial Infarction

An Instrumental Variable Analysis

Davies, Neil M.a; Smith, George Daveya; Windmeijer, Frankb; Martin, Richard M.a

Author Information
doi: 10.1097/EDE.0b013e318289e024

Abstract

Instrumental variable analyses can estimate causal effects in the presence of unmeasured confounding.1–8 Instrumental variable analyses require a variable (instrument) that is associated with treatment, that is unrelated to observed and unobserved confounders of the treatment-outcome relationship, and that affects the outcome only through its effect on treatment.9 Brookhart et al10 proposed physicians’ prescribing preferences as instruments to investigate upper gastrointestinal and cardiovascular11 adverse effects of nonsteroidal anti-inflammatory drugs (NSAIDs).12–14 Physicians’ prescribing preferences as an instrument for NSAID prescriptions have been evaluated among Medicare beneficiaries in Pennsylvania, by a single research team,11,12,15–18 and findings may not replicate across populations.19–21

We conducted an independent application of physicians’ prescribing preferences as an instrument10 and investigated alternative definitions of surrogate instruments, estimation methods, specification tests, and identifying assumptions. As in Brookhart et al10 and Schneeweiss et al,11 we estimated risk differences of two outcomes—upper gastrointestinal complications and myocardial infarction (MI)—comparing selective cyclooxygenase-2 inhibitors (COX-2s) versus nonselective NSAIDs10–12,16,22 and compared our findings with results from randomized controlled trials.23–25 We extended the methods to use multiple previous prescriptions and to specify point-identifying assumptions.26

METHODS

Study Design and Population

Our study was a population-based cohort study nested within the United Kingdom’s Clinical Practice Research Datalink (www.cprd.com). This is a prospectively gathered, anonymized database that holds administrative, clinical, and prescribing records of patients from general practices across the United Kingdom.27 Data are validated, audited, and quality-checked.28,29

We sampled new users of oral NSAIDs (COX-2s or nonselective NSAIDs) older than 60 years registered with 298 “up-to-research-standard” practices between 1 January 1999 (when COX-2s became available) and 31 December 2004 (before regulations restricting the use of COX-2s).27,30,31 Follow-up began with the first NSAID prescription and ended at death, transfer from the practice, outcome occurrence (incident upper gastrointestinal complication or MI), or the study period end, whichever came first. Our license restricted the number of patients we could sample. Therefore, we sampled all patients first prescribed COX-2s and a random sample of 31% of those prescribed nonselective NSAIDs. Patients were eligible if they had a first-time prescription of any oral nonselective NSAID (ibuprofen, diclofenac, naproxen, meloxicam, or other) or COX-2 (celecoxib, rofecoxib, valdecoxib, or etoricoxib) in the study period and were registered for at least 6 months before the first prescription. We restricted our analyses to patients’ first NSAID prescriptions, defined as no NSAID prescriptions in the previous 6 months. We excluded patients whose physicians prescribed NSAIDs to fewer than 10 patients, to increase precision.

Exposures

We defined patients’ exposure as their first COX-2 or nonselective NSAID prescription, identified by all multilex codes in British National Formulary section 10.1.1. We defined patients’ index date using their first NSAID prescription record.

Outcomes

We defined outcomes based on records with codes for upper gastrointestinal complications or MI in the clinical, test, and referral tables, occurring 60, 120, and 180 days, respectively, after the first prescription (eTables 1 and 2; http://links.lww.com/EDE/A665). Our upper gastrointestinal complications’ code lists are based on a modified version of the code lists proposed in Margulis et al.32 Each patient’s date of diagnosis was defined using the event date from the earliest record indicating the outcome. Patients with records indicating upper gastrointestinal (n = 3,556) or MI (n = 993) events before their first prescription were excluded from the analyses for upper gastrointestinal or MI, respectively, but were included in Table 1 to show their distribution by actual prescription and previous prescription. Consistent with previous studies, our primary outcomes were defined to be events within 180 days.10,11 Randomized controlled trials found differences in upper gastrointestinal event rates between COX-2s and nonselective NSAIDs within weeks of starting, persisting for at least 9 months.23,25

TABLE 1
TABLE 1:
Distributiona and Risk or Mean Differences of Observed Confounders by Actual Treatment and the Previous Prescription (n = 56,908)

Potential Confounding Factors

We defined covariables (Table 1) using the most recent value before the first prescription. We replaced missing body mass index (n = 11,418) and blood pressure (n = 10,791) values with the means and included dummy variables for missing values in adjusted models.

Defining the Instrument

Our proposed instrumental variable is each patient’s physician’s preference for nonselective NSAIDs or COX-2s.10,33 For each new NSAID prescription, we used the prescription issued to the physician’s previous first-time recipient of an NSAID as a surrogate instrument for each physician’s prescribing preference (based on the entire database, not just our random sample).10,33 This time-varying surrogate instrument reflects changes in preferences over time.10

We improved the precision of the instrumental variable estimates by deriving sets of surrogate instruments (for physicians’ preferences). We constructed seven indicator variables for the number of COX-2 prescriptions written by each physician in their previous 20 NSAID prescriptions.34 We used six indicator variables for one to six previous COX-2 prescriptions and an indicator for more than six previous COX-2 prescriptions. For physicians who issued fewer than 20 previous NSAID prescriptions, we used a count of their observed prescriptions for COX-2s. We also present results using the proportion of the previous 20 NSAID prescriptions that were COX-2s. In a sensitivity analysis, we allowed for within-drug-class heterogeneity in the cardiovascular effects of individual COX-2s and nonselective NSAIDs by estimating the effects of being prescribed rofecoxib, celecoxib, diclofenac, and naproxen independently compared with ibuprofen (the most frequent prescription).23,35 We used seven indicator variables, for the number of previous prescriptions of each NSAID (six indicators for one to six previous prescriptions and one indicator for more than six prescriptions) as a surrogate for physicians’ preferences for individual NSAIDs. For the analysis of individual NSAIDs, we excluded patients whose physicians had fewer than 20 previous NSAID prescriptions, as we could not assume that unobserved prescriptions were the base category, ibuprofen. Finally, we investigated whether the MI results changed after we excluded patients prescribed the nonselective NSAID, diclofenac.

We estimated crude and multivariable-adjusted risk differences of outcomes using conventional regression, comparing patients prescribed COX-2s to those prescribed nonselective NSAIDs. We used conventional regression because we wanted to compare the conventional results with the instrumental variable results under the additive structural mean model. Conventional regression estimates risk differences, and although results from logistic regression models are traditionally reported on the odds ratio scale, inferences for risk differences under logistic and linear regression were similar. We calculated robust standard errors accounting for clustering of patients by physician and the binary (heteroskedastic) outcome36 and used weights to correct for incomplete sampling of patients prescribed nonselective NSAIDs.

We estimated instrumental variable structural mean models, using generalized method of moments estimators,37 reporting risk difference in outcomes, both unadjusted and adjusted for baseline covariables, with standard errors allowing for clustering by physician. With one instrument, generalized method of moments estimators is equivalent to the two-stage least squares estimator.17,38 We used the two-step generalized method of moments estimator for models with multiple instruments,39–41 which was more precise than two-stage least squares regression. The estimating equations, assumptions underlying the instrumental variable analysis, and parameters estimated are described below.

Let Y, X, U**, and Z denote random variables indicating a binary outcome (upper gastrointestinal complications or MI), prescription (COX-2s or nonselective NSAIDs), the unobserved continuous instrument, physicians’ prescribing preferences, and the binary observed surrogate instrument (physicians’ previous prescriptions). The relationships between these variables are denoted in Figure 1. Each patient has two potential outcomes, denoted Y(x), x ∈{0, 1}, where the exposure, x, is an indicator for prescription (COX-2 or nonselective NSAID), and the actual exposures are denoted X ∈ {0, 1}. Four instrumental variable, assumptions must hold to estimate the bounds of average causal effects of prescriptions on outcomes.2 These are as follows: (1) the stable unit treatment value assumption—patients’ potential outcomes are unrelated to other patients’ prescriptions; (2) the exclusion restriction—the surrogate instrument has no direct effect on the outcomes Y(x, z) = Y(x); (3) the independence assumption—the surrogate instrument is independent of the potential outcomes Y(x); and (4) the relevance assumption—a nonzero association of Z and X; E[XZ = 1] − E[XZ = 0] ≠ 0.33

FIGURE 1
FIGURE 1:
Directed acyclic graph of outcomes Y, prescriptions X, instrumental variable the unmeasured physician prescribing preference U* and surrogate instrument physicians’ previous prescriptions Z, and unmeasured confounders U (reproduced from Hernán and Robins33 Figure 2).

When these four assumptions hold, the maximum and minimum values of the average causal effect of prescriptions across the entire population can be calculated (Balke-Pearl bounds), which are consistent with the observed Y, X, and Z42,43:

These bounds are not 95% confidence intervals (CIs), but point estimates of the lower and upper bounds of the causal effect. Point estimates of the causal effect can be identified only under stronger assumptions. We used structural mean models to obtain point identification.33 The additive structural mean model is

where ψ0 is the effect of prescription for patients who are prescribed COX-2s and for whom Z = 0, and ψ1 is the additional effect of prescription for when Z = 1. From the four core assumptions, it follows that the potential outcomes and the surrogate instruments are independent, E[Y(0) Z = 1] = E[Y(0) Z = 0], resulting in the estimating equation E[Y − (ψ0 + ψ1) Z = 1] = E[Y − ψ0Z = 0]. We can identify ψ0 by assuming no effect modification by Z among the treated, ψ1 = 0. The resulting estimand under this assumption is identical to that of the linear structural equation model used by previous authors10:

Under the no-effect-modification assumption, ψ0 is equal to the average effect of prescription on those prescribed, E[Y(1) − Y(0) X = 1]. If no-effect-modification assumption also holds for the untreated, then the average causal effect is identified. However, as the prescriptions are binary, it is impossible for no-effect-modification assumption to hold simultaneously for those prescribed and not prescribed.33 Assuming no effect modification, we further specified multiple previous prescriptions as instruments and used the generalized method of moments estimator to estimate weighted averages of the effect of prescription on those prescribed.38,44

If the no-effect-modification assumption is violated, the instrumental variable estimand equals:

Therefore, if ψ1 ≠ 0, the estimand is a biased estimator of effect of prescriptions on those prescribed. The size of bias is a decreasing function of the strength of association of Z and X, and the proportion of the patients with Z = 0 who were not prescribed COX-2s. For example, if E[XZ = 0] = 0, we would identify ψ0 + ψ1 = E[Y(1) – Y(0) X = 1, Z = 1]. Furthermore, the estimand is equal to 0 under the null of no effect of prescription, ψ0 = ψ1 = 0.

Alternatively, Hernán and Robins33 showed that the instrumental variable estimand identifies a weighted average of causal effects under an assumption of a monotonic relationship between physicians’ preferences U* and prescriptions. Monotonicity requires that if a patient was prescribed a COX-2 by a physician with a preference U**= u**, they would also have been prescribed a COX-2 had they attended a physician with a similar or higher preference for COX-2s, U* ≥ U*, and vice versa. The instrumental variable estimand identifies a weighted average of the average treatment effect MTP(U*), defined as:

This is the local average causal effect for patients who would be prescribed COX-2s by physicians with prescribing preference U*, but not by physicians with a lower preference v. The specific group of patients for whom this parameter applies is unknown.45

To investigate the assumptions defining our surrogate instruments, we estimated associations of previous prescriptions and observed baseline covariables using linear regression. We estimated the association of previous prescriptions with actual prescriptions using linear regressions of exposures on previous prescriptions and report the partial r2 and partial F tests. Larger partial F statistics imply stronger associations of previous prescriptions and exposure.46 The tests were robust to heteroskedasticity and clustering of patients by physician and use a small sample adjustment. We tested whether results were robust to adjusting for physician fixed effects, using one indicator variable per physician. Conditioning on physician fixed effects means that even if there were observed or unobserved time-invariant differences in physicians’ preferences for other gastroprotective drugs (such as proton-pump inhibitors) or differences among physicians’ patients, the fixed effects results will be asymptotically unbiased.

We used the methodology described by Brookhart and Schneeweiss26 to estimate the prevalence difference ratio:

where C indicates a potential risk factor. If the prevalence difference ratio is smaller than E[XZ = 1] − E[XZ = 0], then the asymptotic bias because of C will be smaller using instruments than the actual prescription.26 We also applied a nonparametric bootstrap and report the proportion of samples in which the inequality holds.47 We estimated the strength of the surrogate instrument in subsamples by covariables.26 If prescriptions have different effects in these subgroups, and the previous prescriptions are more strongly associated within subgroups, then the instrumental variable estimand will reflect a weighted average of the effects of treatment in patient subgroups affected by their physicians’ preferences. Patients who are rarely treated, such as those with contraindications, will contribute little to the results.

The analysis was performed in STATA,48 using the command BPBOUNDS43 for calculating the Balke-Pearl bounds, IVREG249 and option gmm2 for instrumental variable analyses, and XTIVREG2 and option gmm2 for instrumental variable analyses, adjusted for physician fixed effects. For the examples using multiple previous prescriptions, we estimated test statistics for weak instruments, the null hypothesis of which is that the instruments are jointly weakly associated with the exposure.46 Both the partial F test and the test of weak instruments used a small sample adjustment. We did this because if the instruments are only weakly associated with exposure, then instrumental variable inferences can be misleading.

RESULTS

We sampled 62,933 patients prescribed oral NSAIDs within the period from 1 January 1999 to 31 December 2004 inclusive. We omitted 391 patients missing physician identifiers, 1596 patients not issued prescriptions by physicians, and 4038 patients whose physician treated fewer than 10 patients. This left 56,908 patients, of whom 15,396 were prescribed COX-2s and 41,512 were prescribed nonselective NSAIDs. Of these, 3556 patients had records indicating a diagnosis of upper gastrointestinal complications and 993 had records indicating a diagnosis of MI before their first prescription. These patients were excluded from the upper gastrointestinal complication and MI results, respectively. We sampled 1710 physicians, who each prescribed NSAIDs to an average of 33.3 patients in our study.

The most commonly prescribed COX-2 was rofecoxib (52% of COX-2s), followed by celecoxib (41%); the remaining COX-2s were etoricoxib (7%) and valdecoxib (1%). The most common nonselective NSAID was ibuprofen (47%), followed by diclofenac (32%), naproxen (7%), meloxicam (4%), and others (10%). In the 6 months following the first NSAID prescription, patients prescribed nonselective NSAIDs had a median of 28 days of medications and those prescribed COX-2s a median of 30 days, consistent with a previous study.50

There were 143 incident upper gastrointestinal and 158 incident MI events within 180 days of first NSAID prescription. The upper gastrointestinal and MI incident rates were 5.4 and 5.7 per 1000 patient-years of follow-up, respectively, similar to reported rates of 4.03 and 4.57 per 1000 patient-years for upper gastrointestinal complications and MI, respectively, in patients older than 65 years in British general practices51,52 and rates of 4.8 and 5.0 per 1000 person-years for upper gastrointestinal bleeding and MI, respectively, in patients older than 55 years in the Netherlands.53,54

The observed baseline covariables were strongly associated with prescriptions (Table 1). Patients prescribed COX-2s visited their physician more frequently and were prescribed more drugs (including gastroprotective drugs, warfarin, or glucocorticoids). They were also more likely to be women and older; to have a history of hospitalization, arrhythmia, renal complications, ischemic heart disease, upper gastrointestinal complications, gastritis, and endoscopy; to participate in less moderate or heavy exercise; and to have higher blood pressure and body mass index. In contrast, the surrogate instrument (the physicians’ previous prescriptions) was associated with age, hospitalization in the previous year, prevalent upper gastrointestinal complications, and gastroprotective drug use. When adjusted for physician fixed effects, only two covariables (age and gastroprotective drug use) remained associated with the previous prescription. The surrogate instruments defined using 20 previous NSAID prescriptions were jointly associated with six covariates and adjusting for physician fixed effects reduced the number of associations to two (eTable 3; http://links.lww.com/EDE/A665). All combinations of surrogate instruments, prescriptions, and outcomes had one or more events (Table 2), and each previous prescription was strongly associated with actual prescription (Table 3), particularly when using seven surrogate instruments, rather than just one.

TABLE 2
TABLE 2:
Treatment and Outcomes of Patients Prescribed COX-2s or Nonselective NSAIDs Registered to a General Practice Research Database Eligible Practice, by Previous Class of Drug Prescribed, 1999–2004
TABLE 3
TABLE 3:
First-stage Association of Previous Prescriptions and Exposure to COX-2s (n = 53,352)

The cumulative incidence of upper gastrointestinal complications by actual prescription issued was similar among patients prescribed COX-2s and nonselective NSAIDs (Figure2). However, patients whose physician previously prescribed a COX-2 had a lower incidence of upper gastrointestinal complications (Figure 3).

FIGURE 2
FIGURE 2:
Cumulative incidence of upper gastrointestinal complications by actual prescription.
FIGURE 3
FIGURE 3:
Cumulative incidence of upper gastrointestinal complications by physicians’ previous prescriptions unadjusted for covariates or prescription year.

eTable 4 (http://links.lww.com/EDE/A665) shows the maximum and minimum values (Balke-Pearl bounds) of the average causal effects, under a minimum set of assumptions.42,43 The bounds for the average causal effect of COX-2 prescriptions on both upper gastrointestinal complications (−7.65 to 64.48 per 100) and MI events (−8.09 and 63.43 per 100) within 180 days were wide and consistent with large protective or harmful effects.

Tables 4 and 5 show point estimates and 95% CIs for the risk differences of incident upper gastrointestinal complications and MI, by COX-2 versus nonselective NSAID prescriptions, from the ordinary least squares analysis and the instrumental variable analysis under a stronger set of assumptions (either no effect modification or monotonicity). The unadjusted and adjusted ordinary least squares estimates provide little evidence of an association of COX-2s with upper gastrointestinal complications (Table 4). In contrast, the instrumental variable estimates using one previous prescription imply 0.46 (95% CI = −0.15 to 1.07) fewer upper gastrointestinal complications within 180 days per 100 patients prescribed COX-2s. When we used the instruments based on 20 previous prescriptions, the estimates were marginally attenuated and more precise.

TABLE 4
TABLE 4:
Instrumental Variable and Conventional Multivariable Regression Estimates of Risk Differences of Incident Upper Gastrointestinal Complications per 100 Patients Prescribed COX-2s Compared with Nonselective NSAIDs (n = 53,352)

Conventional regression provided little evidence of associations of COX-2s with incident MI within 180 days (Table 5). The instrumental variable estimates using one previous prescription were imprecisely estimated, and neither a clinically meaningful increased nor a decreased risk of MI could be ruled out. The results based on 20 previous prescriptions were more precise, implying 0.74 (95% CI = 0.28 to 1.19) fewer MI events per 100 patients prescribed COX-2s. The estimates were less precise after adjusting for physician fixed effects or averaging the 20 previous prescriptions (eTable 5; http://links.lww.com/EDE/A665). Naproxen was associated with 0.62 (95% CI = 0.30 to 0.93) fewer MI events within 180 days, relative to ibuprofen (using 20 previous prescriptions to define the instruments) (eTable 6 http://links.lww.com/EDE/A665). The instrumental variable analysis (using 20 previous prescriptions) also suggested protective effects of rofecoxib and celecoxib compared with ibuprofen prescriptions for MI events within 180 days. When we excluded patients prescribed the nonselective NSAID diclofenac, we found evidence, using 20 previous prescriptions, that patients prescribed COX-2s had higher rates of MI in the 60 days following first prescription, compared with those prescribed other nonselective NSAIDs (eTable 7; http://links.lww.com/EDE/A665).

TABLE 5
TABLE 5:
Instrumental Variable and Conventional Multivariable Regression Estimates of Risk Differences of Myocardial Infarction per 100 Patients Prescribed COX-2s Compared with Nonselective NSAIDs (n = 55,915)

Physicians’ previous prescriptions were more strongly associated with patients’ prescriptions among those with risk factors for upper gastrointestinal complications, such as previous diagnoses of gastritis, gastroprotective drug prescription, or endoscopy (eTable 8; http://links.lww.com/EDE/A665). We found less evidence of differences in associations for MI-specific risk factors, such as previous diagnosis of arrhythmia or ischemic heart disease. The prevalence difference ratio indicated that the instrumental variable estimators are likely to have lower asymptotic bias than conventional analysis for confounding because of upper gastrointestinal complication risk factors: in over half of the bootstrap samples, the prevalence difference inequality held. In contrast, the prevalence difference inequality did not hold for most of the MI risk factors. We found no evidence that the results differed when we used an identical code list to that validated by Margulis et al32 (eTable 9; http://links.lww.com/EDE/A665), when we excluded patients who were missing body mass index or blood pressure measures (eTables 10 and 11; http://links.lww.com/EDE/A665), or when we excluded patients with more than 1 year’s but less than 2 years’ registration before the index prescription (eTables 12 and 13; http://links.lww.com/EDE/A665).

DISCUSSION

We found little evidence from conventional regression that COX-2s were associated with incident upper gastrointestinal complications or MI. The instrumental variable results suggested that patients prescribed COX-2s had fewer incident upper gastrointestinal complications after first prescription, consistent in direction with the VIGOR25 and CLASS24 randomized controlled trials (Table 6),10 although our instrumental variable estimates were less precise than our conventional analysis. Our data suggest that previous conventional observational studies may have underestimated the gastroprotective effects of COX-2s.50,51,55

TABLE 6
TABLE 6:
Comparison of Instrumental Variable Estimates from Our General Practice Research Database Analysis and from the Study of Brookhart et al with Estimates from Randomized Controlled Trials (Risk Difference Comparing COX-2s to Nonselective NSAIDs for Upper Gastrointestinal Complications and Myocardial Infarction per 100 Patients)

Our conventional and instrumental variable analyses using a single previous prescription found no evidence of differences in rates of MI, but our results using multiple previous prescriptions suggested that patients prescribed COX-2s had a reduced risk of MI at 120 and 180 days after first prescription. In contrast, meta-analyses of trials found an increased risk of cardiovascular outcomes in participants allocated to rofecoxib compared with controls (naproxen, placebo, or other nonselective NSAIDs).23,35 However, the difference in risk of MI between COX-2s versus nonselective NSAIDs, as drug classes, is unknown.35 A previous observational study, which used prescribing preferences as instruments, could not exclude null or beneficial COX-2 class effects compared with nonselective NSAIDs.11 Most patients prescribed nonselective NSAIDs in our sample were prescribed ibuprofen or diclofenac, and only 7% were prescribed naproxen. A network meta-analysis of randomized controlled trials found patients allocated to diclofenac had a similar risk of MI as those allocated to celecoxib and higher rates of cardiovascular death than those allocated to rofecoxib or celecoxib.35 When we excluded diclofenac from the nonselective NSAID comparator group, we found weak evidence that patients prescribed COX-2s had more MI events shortly after their first prescription than those prescribed other nonselective NSAIDs (eTable 7; http://links.lww.com/EDE/A665).

Our data have three advantages relative to the results from Brookhart et al.10 First, our database records prescriptions written by general practitioners and most patients were eligible for free prescriptions, so over-the-counter NSAIDs may be better measured than in the United States (although not completely, as discussed below).10 Second, our database includes all computerized prescriptions written by physicians. This may increase the instrument strength and reduce prescription measurement error.

Point identification of causal effects depends on strong assumptions that affect the interpretation of the results. Brookhart et al10 identified average effects of prescriptions using linear structural equation models. They assumed a constant effect of prescription, which is logically impossible with binary outcomes.33 We calculated Balke-Pearl bounds using a minimal set of assumptions, but these were noninformative. The no-effect-modification assumption cannot be tested and constrains the functional form of the data-generating process, and it may or may not hold for binary data.33,38 If no-effect-modification assumption holds, then our results reflect the average causal effect of COX-2 prescriptions on those prescribed COX-2s. If no-effect-modification assumption is violated, then our results would be a biased estimate of ψ0 + ψ1 = E[Y(1) − Y(0) X = 1, Z = 1]. Given the strength of the instruments and the proportion exposed, the size of this bias is of the order of 0.25 ψ1. If no-effect-modification assumption is violated but a monotonicity condition for the continuous physicians’ prescribing preferences holds, then the instrumental variable estimand identifies a weighted average of average treatment effects MTP(U*).33

Limitations

First, our instrumental variable results are less precise than the conventional analysis. This means that for many of our outcomes, we find little evidence of systematic differences between our instrumental variable results and the conventional results. We could improve the precision of these results if we had more data. Second, physicians’ preferences would not be valid instruments if they are related to the patients’ baseline characteristics. However, the observed upper gastrointestinal risk factors were more balanced across the previous prescriptions than actual exposure (Table 1). Third, we cannot rule out residual confounding by unobserved variables. However, because the associations of observed risk factors are smaller for the previous prescription than the actual exposure, the proposed instruments may also be less strongly associated with the potential unobserved confounding factors than the actual prescription. Fourth, if previous prescriptions were weakly associated with exposure, then bias because of association of surrogate instruments and confounders can be magnified.33 The F-statistic indicated that the instrumental variable results did not suffer from weak instruments.46 Fifth, a lower proportion of patients were prescribed COX-2s in the United Kingdom than the United States. This suggests that British physicians had higher thresholds for prescribing COX-2s. Therefore, relative to Brookhart et al,10 it is likely we identified our results from patients with greater upper gastrointestinal risk factors. These marginal effects of prescription may not reflect the marginal effects of prescriptions in the population not prescribed COX-2s.

Finally, we do not know if patients took their prescriptions, as our data recorded only the issuing of prescriptions; we took no account of duration or treatment intensity, and we had no information on over-the-counter aspirin or ibuprofen use. Hence, we identified effects of prescription (intention-to-treat), rather than effects of actually complying with treatment. Finally, although we could not directly validate our outcomes, incidence rates were similar to previous studies and our code list for upper gastrointestinal complications was similar to a validation study that had found 70% positive predictive value for peptic ulcer codes.51–54,56

In conclusion, we found some evidence, using instrumental variable analysis, that patients prescribed COX-2s had fewer upper gastrointestinal complications shortly after first prescription—which is consistent in direction with randomized controlled trials.51 Our instrumental variable analysis using multiple previous prescriptions (but not when using a single previous prescription) found that patients prescribed COX-2s had a lower risk of MI compared with those prescribed nonselective NSAIDs (largely ibuprofen and diclofenac). It has been hypothesized that confounding by indication may explain why some observational studies have been contradicted by subsequent randomized controlled trials.57–60 These results suggest that causal modeling using instrumental variables may address unmeasured and residual confounding in administrative data.57,61,62 Although the use of instrumental variables is growing in the medical literature, it is not widespread outside Mendelian randomization applications.6,63 Limitations of instrumental variable analysis are its imprecision and the large sample sizes required, as well as the unverifiable assumptions required for point identification. However, the increasing quantity and quality of administrative data may offer a promising source of information to precisely estimate these models.64,65

ACKNOWLEDGMENTS

We are grateful to participants at the Society for Epidemiological Research Congress in Seattle 2010 and Society for Social Medicine in Belfast 2010 for their comments and to the patients and general practitioners participating in the Clinical Practice Research Datalink and to their staff for help and advice, particularly Tarita Murray-Thomas for preparing the data extract. Thanks to Miguel Hernán, Alan Brookhart, and an anonymous referee for comments on an earlier draft.

REFERENCES

1. Chen Y, Briesacher BA. Use of instrumental variable in prescription drug research with observational data: a systematic review. J Clin Epidemiol. 2011;64:687–700
2. Angrist JD, Imbens GW, Rubin DB. Identification of causal effects using instrumental variables. J Am Stat Assoc. 1996;91:444–455
3. Angrist JD, Lavy V. Using Maimonides’ rule to estimate the effect of class size on scholastic achievement. Quarterly J Econ. 1999;114:533–575
4. Earle CC, Tsai JS, Gelber RD, Weinstein MC, Neumann PJ, Weeks JC. Effectiveness of chemotherapy for advanced lung cancer in the elderly: instrumental variable and propensity analysis. J Clin Oncol. 2001;19:1064–1070
5. Timpson NJ, Lawlor DA, Harbord RM, et al. C-reactive protein and its role in metabolic syndrome: Mendelian randomisation study. Lancet. 2005;366:1954–1959
6. Davey Smith G, Ebrahim S. What can Mendelian randomisation tell us about modifiable behavioural and environmental exposures? BMJ. 2005;330:1076–1079
7. Klungel OH, Martens EP, Psaty BM, et al. Methods to assess intended effects of drug treatment in observational studies are reviewed. J Clin Epidemiol. 2004;57:1223–1231
8. Martens EP, Pestman WR, de Boer A, Belitser SV, Klungel OH. Instrumental variables: application and limitations. Epidemiology. 2006;17:260–267
9. Angrist JD. Estimation of limited dependent variable models with dummy endogenous regressors. J Bus Econ Stat. 2001;19:2–28
10. Brookhart MA, Wang PS, Solomon DH, Schneeweiss S. Evaluating short-term drug effects using a physician-specific prescribing preference as an instrumental variable. Epidemiology. 2006;17:268–275
11. Schneeweiss S, Solomon DH, Wang PS, Rassen J, Brookhart MA. Simultaneous assessment of short-term gastrointestinal benefits and cardiovascular risks of selective cyclooxygenase 2 inhibitors and nonselective nonsteroidal antiinflammatory drugs: an instrumental variable analysis. Arthritis Rheum. 2006;54:3390–3398
12. Rassen JA, Schneeweiss S, Glynn RJ, Mittleman MA, Brookhart MA. Instrumental variable analysis for estimation of treatment effects with dichotomous outcomes. Am J Epidemiol. 2009;169:273–284
13. Schneeweiss S, Seeger JD, Landon J, Walker AM. Aprotinin during coronary-artery bypass grafting and risk of death. N Engl J Med. 2008;358:771–783
14. Schneeweiss S, Glynn RJ, Avorn J, Solomon DH. A Medicare database review found that physician preferences increasingly outweighed patient characteristics as determinants of first-time prescriptions for COX-2 inhibitors. J Clin Epidemiol. 2005;58:98–102
15. Rassen JA, Mittleman MA, Glynn RJ, Alan Brookhart M, Schneeweiss S. Safety and effectiveness of bivalirudin in routine care of patients undergoing percutaneous coronary intervention. Eur Heart J. 2010;31:561–572
16. Rassen JA, Brookhart MA, Glynn RJ, Mittleman MA, Schneeweiss S. Instrumental variables I: instrumental variables exploit natural variation in nonexperimental data to estimate causal relationships. J Clin Epidemiol. 2009;62:1226–1232
17. Brookhart MA, Rassen JA, Wang PS, Dormuth C, Mogun H, Schneeweiss S. Evaluating the validity of an instrumental variable study of neuroleptics: can between-physician differences in prescribing patterns be used to estimate treatment effects? Med Care. 2007;45(10 suppl 2):S116–S122
18. Brookhart MA, Rassen JA, Schneeweiss S. Instrumental variable methods in comparative safety and effectiveness research. Pharmacoepidemiol Drug Saf. 2010;19:537–554
19. Wang PS, Schneeweiss S, Avorn J, et al. Risk of death in elderly users of conventional vs. atypical antipsychotic medications. N Engl J Med. 2005;353:2335–2341
20. Schneeweiss S, Setoguchi S, Brookhart A, Dormuth C, Wang PS. Risk of death associated with the use of conventional versus atypical antipsychotic drugs among elderly patients. CMAJ. 2007;176:627–632
21. Hernán MA, Wilcox AJ. Epidemiology, data sharing, and the challenge of scientific replication. Epidemiology. 2009;20:167–168
22. Rassen JA, Brookhart MA, Glynn RJ, Mittleman MA, Schneeweiss S. Instrumental variables II: instrumental variable application-in 25 variations, the physician prescribing preference generally was strong and reduced covariate imbalance. J Clin Epidemiol. 2009;62:1233–1241
    23. Jüni P, Nartey L, Reichenbach S, Sterchi R, Dieppe PA, Egger M. Risk of cardiovascular events and rofecoxib: cumulative meta-analysis. Lancet. 2004;364:2021–2029
    24. Silverstein F, Faich G, Goldstein J, et al. Gastrointestinal toxicity with celecoxib vs nonsteroidal anti-inflammatory drugs for osteoarthritis and rheumatoid arthritis: the CLASS study: a randomized controlled trial. JAMA. 2000;284:1247–1255
    25. Bombardier C, Laine L, Reicin A, et al. Comparison of upper gastrointestinal toxicity of rofecoxib and naproxen in patients with rheumatoid arthritis. New Engl J Med. 2000;343:1520–1528
    26. Brookhart MA, Schneeweiss S. Preference-based instrumental variable methods for the estimation of treatment effects: assessing validity and interpreting results. Int J Biostat. 2007;3:Article 14
    27. Anon. . GPRD. The database. 2010 www.gprd.com/products/database.asp Available at: http://www.gprd.com/home/home.asp Accessed 20 October 2010
    28. Herrett E, Thomas SL, Schoonen WM, Smeeth L, Hall AJ. Validation and validity of diagnoses in the General Practice Research Database: a systematic review. Br J Clin Pharmacol. 2010;69:4–14
    29. Khan NF, Harrison SE, Rose PW. Validity of diagnostic coding within the General Practice Research Database: a systematic review. Br J Gen Pract. 2010;60:e128–e136
    30. Wheeler BW, Metcalfe C, Gunnell D, Stephens P, Martin RM. Population impact of regulatory activity restricting prescribing of COX-2 inhibitors: ecological study. Br J Clin Pharmacol. 2009;68:752–764
    31. Dieppe PA, Ebrahim S, Martin RM, Jüni P. Lessons from the withdrawal of rofecoxib. BMJ. 2004;329:867–868
    32. Margulis AV, García Rodríguez LA, Hernández-Díaz S. Positive predictive value of computerized medical records for uncomplicated and complicated upper gastrointestinal ulcer. Pharmacoepidemiol Drug Saf. 2009;18:900–909
    33. Hernán MA, Robins JM. Instruments for causal inference: an epidemiologist’s dream? Epidemiology. 2006;17:360–372
    34. Abrahamowicz M, Beauchamp ME, Ionescu-Ittu R, Delaney JA, Pilote L. Reducing the variance of the prescribing preference-based instrumental variable estimates of the treatment effect. Am J Epidemiol. 2011;174:494–502
    35. Trelle S, Reichenbach S, Wandel S, et al. Cardiovascular safety of non-steroidal anti-inflammatory drugs: network meta-analysis. BMJ. 2011;342:c7086
    36. White H. A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity. Econometrica. 1980;48:817–838
    37. Clarke PS, Windmeijer F. Instrumental variable estimators for binary outcomes. The Centre for Market and Public Organisation Working Papers. 2010;239
    38. Clarke PS, Windmeijer F. Identification of causal effects on binary outcomes using structural mean models. Biostatistics. 2010;11:756–770
    39. Wooldridge J Econometric Analysis of Cross Section and Panel Data. 2002 Cambridge, MA MIT Press
    40. Greene W Econometric Analysis. 2002 Upper Saddle River, NJ Prentice Hall
    41. Hansen LP. Large sample properties of generalized method of moments estimators. Econometrica. 1982;50:1029–1054
    42. Balke A, Pearl J. Bounds on treatment effects from studies with imperfect compliance. J Am Stat Assoc. 1997;92:1171–1176
    43. Palmer T, Ramsahai R, Didelez V, Sheehan N. Nonparametric bounds for the causal effect in a binary instrumental variable model. Stata J. 2011;3:345–367
    44. Angrist JD, Imbens GW. Two-stage least squares estimation of average causal effects in models with variable treatment intensity. J Am Stat Assoc. 1995;90:431–442
    45. Dawid AP. Causal inference without counterfactuals. J Am Stat Assoc. 2000;95:407–424
    46. Stock J, Yogo M. Testing for weak instruments in linear IV regression. National Bureau of Economic Research Technical Working Paper Series. 2002;No. 284 Available at: http://www.nber.org/papers/t0284.pdf Accessed 1 April 2010
    47. Efron B. Better bootstrap confidence intervals. J Am Stat Assoc. 1987;82:171–185
    48. StataCorp. Stata Statistical Software: Release 12. College Station, TX: StataCorp LP; 2011
    49. Baum CF, Schaffer ME, Stillman S ivreg2: Stata module for extended instrumental variables/2SLS, GMM and AC/HAC, LIML and k-class; 2010
    50. van Staa TP, Leufkens HG, Zhang B, Smeeth L. A comparison of cost effectiveness using data from randomized trials or actual clinical practice: selective Cox-2 inhibitors as an example. PLoS Med. 2009;6:e1000194
    51. Hippisley-Cox J, Coupland C, Logan R. Risk of adverse gastrointestinal outcomes in patients taking cyclo-oxygenase-2 inhibitors or conventional non-steroidal anti-inflammatory drugs: population based nested case-control analysis. BMJ. 2005;331:1310–1316
    52. Hippisley-Cox J, Coupland C. Risk of myocardial infarction in patients taking cyclo-oxygenase-2 inhibitors or conventional non-steroidal anti-inflammatory drugs: population based nested case-control analysis. BMJ. 2005;330:1366
    53. de Torbal A, Boersma E, Kors JA, et al. Incidence of recognized and unrecognized myocardial infarction in men and women aged 55 and older: the Rotterdam Study. Eur Heart J. 2006;27:729–736
    54. van Soest EM, Valkhoff VE, Mazzaglia G, et al. Suboptimal gastroprotective coverage of NSAID use and the risk of upper gastrointestinal bleeding and ulcers: an observational study using three European databases. Gut. 2011;60:1650–1659
    55. van Staa TP, Smeeth L, Persson I, Parkinson J, Leufkens HG. What is the harm-benefit ratio of Cox-2 inhibitors? Int J Epidemiol. 2008;37:405–413
    56. McMahon AD. Approaches to combat with confounding by indication in observational studies of intended drug effects. Pharmacoepidemiol Drug Saf. 2003;12:551–558
    57. LaLonde R. Evaluating the econometric evaluations of training programs with experimental data. Am Econ Rev. 1986;76:604–620
    58. Lawlor DA, Davey Smith G, Ebrahim S. Commentary: the hormone replacement-coronary heart disease conundrum: is this the death of observational epidemiology? Int J Epidemiol. 2004;33:464–467
    59. Lawlor DA, Davey Smith G, Kundu D, Bruckdorfer KR, Ebrahim S. Those confounded vitamins: what can we learn from the differences between observational versus randomised trial evidence? Lancet. 2004;363:1724–1727
    60. Egger M, Schneider M, Davey Smith G. Spurious precision? Meta-analysis of observational studies. BMJ. 1998;316:140–144
    61. Weiss NS. The new world of data linkages in clinical epidemiology: are we being brave or foolhardy? Epidemiology. 2011;22:292–294
    62. Stürmer T, Jonsson Funk M, Poole C, Brookhart MA. Nonexperimental comparative effectiveness research using linked healthcare databases. Epidemiology. 2011;22:298–301
    63. Davey Smith G, Ebrahim S. “Mendelian randomization”: can genetic epidemiology contribute to understanding environmental determinants of disease? Int J Epidemiol. 2003;32:1–22
    64. Hernán MA. With great data comes great responsibility: publishing comparative effectiveness research in epidemiology. Epidemiology. 2011;22:290–291
    65. Chandra A, Jena AB, Skinner JS. The pragmatist’s guide to comparative effectiveness research. J Econ Perspect. 2011;25:27–46

    Supplemental Digital Content

    © 2013 by Lippincott Williams & Wilkins, Inc