A theoretical exploration of therapeutic monomania as a physician‐based instrumental variable

Abstract

Purpose

To explore the utility of physician prescribing preference as an instrumental variable.

Methods

Expert (non‐systematic) review of relevant literature on the appropriate selection of instrumental variables and theoretical exploration of individual physician and physician group prescriber preference.

Results

An instrumental variable must satisfy three criteria: (1) It must predict the treatment received (strength of the instrument); (2) it cannot influence the outcome other that through the treatment received (exclusion restriction); and (3) it cannot be influenced by any factor that also influences the outcome (independence assumption). Arguments in favor of prescriber preference as an instrumental variable and suggestions for how to approach specific scenarios that may be encountered are offered.

Conclusions

Prescriber preference, be it of individual physicians or groups of physicians, may, under the right conditions, be powerful instrumental variables. Empiric experimental data are required to determine the appropriateness of combining propensity matching and instrumental variable analysis.

Key Points.

• Instrumental variable analysis, when possible, is a powerful tool for directly addressing unmeasured or unknown confounders.

• An instrumental variable must satisfy three criteria: (1) the strength of the instrument; (2) the exclusion restriction; and (3) the independence assumption.

• The prescribing preferences of individual physicians or of groups of physicians may satisfy these criteria under appropriate circumstances.

• We propose approaches for specific scenarios that might be encountered in research based on our theoretical exploration.

• Empiric validation of the usefulness and appropriateness of a two‐stage adjustment consisting of propensity matching followed by instrumental variable analysis requires empiric validation.

“But, in his narrow‐flowing monomania, …. That was living agent, now became the living instrument.” ‐ Herman Melville, Moby‐Dick

1. INTRODUCTION

In attributing causality, randomized controlled trials (RCTs) are considered the gold standard for their ability to address both known and unknown confounders. When RCTs are not possible, however—be it for ethical, practical, or financial reasons—researchers must attempt to account for confounding by other means. In the design phase, it has been suggested that case‐only study designs may better address unknown confounders.1 Yet, in the analysis phase, nearly all methods, including propensity score (PS) matching and marginal structural models, address only known confounders directly. While it has been suggested that high‐dimension PS models can indirectly account for at least some unmeasured confounding2 and the “missing cause” approach and sensitivity analysis show promise,3, 4 the only well‐validated method that directly addresses unknown confounders outside of RCTs is instrumental variable (IV) analysis.

Herein, we set forth arguments in support of IV analysis and discuss considerations specific to appropriately using the prescription preferences of individual and groups of providers as an IV, including potential pitfalls. We also describe the theoretical benefit of using PS matching prior to IV analysis to ensure that the equipoise population is clearly ascertained and within which the IV assumptions discussed below are most likely to hold. It is important to note that supporting evidence for these hypotheses is currently lacking, but is the subject of ongoing research. The recommendations made therefore represent the considered opinion of the authors regarding what might constitute best practices.

2. THE CASE FOR IV ANALYSIS

The specter of a dreaded unknown confounder is the Achilles heel of observational research. This is particularly true in pharmacoepidemiology where there is the potential for unmeasured confounding by indication. Confounding by indication (and its corollary, confounding by contraindication) refers to any way in which patients receiving a therapy or intervention might be inherently different from those not receiving the intervention by virtue of having been prescribed a treatment for a reason.5 As opposed to RCTs, in observational data, physicians have selected treatments based on everything they know about a given patient. As everything known to treating physicians about their patients is unlikely to be captured in a database, particularly administrative ones, the potential for unknowable (and therefore potentially unaccounted for) confounding persists.

The concept of IVs was originally described in a letter by Philip G. Wright to his son in 1926 and first published in a 1928 economic treatise, but it wasn't until 1945 that Olav Reiersøl gave the method its current name,6 and not for another 50 years before it became more widely used in health research, owing in large part to the difficulty in identifying suitable IVs.7

The key step in IV analysis is the identification of an optimal instrument. In the context of pharmacoepidemiology, an instrument is an observable variable that, by virtue of being correlated with the prescribed drug but not with the outcome, except perhaps through its effect on treatment selection, allows one to use it as a proxy for randomization.7, 8, 9 More specifically, the instrument does not itself belong in a regression model as an explanatory variable, but, as it is correlated with the exposure and the endogenous error (but not the outcome)—due to omitted information or measurement error—its inclusion can reduce the error of the regression estimate. An instrument is therefore not a confounder, as a confounder is a variable that directly affects both the intervention and outcome. To be a good instrument, a variable must satisfy three criteria: (1) It must predict the actual treatment received (Strength of the Instrument); (2) it cannot influence the outcome other that through the treatment received (Exclusion Restriction); and (3) it cannot be influenced by any factor that also influences the outcome (Independence Assumption).7, 8, 9 These concepts are described schematically in Figure 1. Ideally, then, an IV would also be unrelated to both known and unknown patient characteristics that might influence treatment.10

Figure 1.

Causal diagram (directed acyclic graph) illustrating the difference between an optimal instrumental variable (IV) and a confounder. The IV is related only to the treatment received and only in that way to the outcome. It does not influence the outcome directly or indirectly by any other means. C₀ is a classic confounder that influences both the treatment received and the outcome. Variable “L” is a poor choice for an instrumental variable owing to an indirect relationship with the outcome via confounder C₁. [Colour figure can be viewed at http://wileyonlinelibrary.com]

3. PRESCRIBING PREFERENCE ESTIMATES AS AN IV

As early as 2003, authors have posited that provider prescribing preferences are, at least in some cases, a stronger predictor of the prescribed drug than the patient's characteristics. {Brookhart,7 #10; Brookhart,11 #9; Schneeweiss,12 #30; Solomon,13 #29} They therefore proposed past prescription behavior as a method of quantifying each care provider's prescription preference (Prescribing Preference Estimate, PPE). Since then, several studies have evaluated the performance of specific methods prescription preference estimation. However, results have varied. Henessey et al14 reported that the PPE based on the previous prescription is a stronger IV than the one based on alternative sets of prior prescriptions, whereas Ionescu‐Ittu et al15 found the opposite, reporting that the proportion of patients that are prescribed a given drug by a specific physician produces a stronger IV with less variance than relying solely on the last prescription. Reasons for these discrepancies include that physician preference may change over time,16 that instruments may have variable performance depending on the number of patients seen by a provider,8 and that the ideal PPE may depend on the treatment studied,17 or the timeframe of the effect of interest.18 The clinical setting may also have an impact on PPE performance as an IV.

Notwithstanding, Davies et al18 found that past prescription behavior was less correlated with patient characteristics than current prescriptions and thus should also be less related to unknown confounders, suggesting that PPE may well be, in many instances, an ideal IV. It would therefore be of interest to establish the ideal clinical research conditions that would support using PPE as an IV.

Only comparisons between two therapies or two classes of therapies should be undertaken. Ideally, the market share of each therapy (ie, rate of prescription) should be reasonably similar between the two therapies, and there should not be differential reimbursement restrictions on the therapies studied. In the case of pharmaceuticals, the patent status (generic or brand name) of the two therapies should also be the same and stable over the period of study. (Favoring comparisons of generic medications has the advantage of minimizing the influence of industry marketing initiatives on physician practice patterns. However, those same industry influences may give rise to monomania for branded therapies.) Like other statistical instruments, PPE will have more statistical power for a given instrument strength when the disease to be treated and outcome are common, when patients are seen by physicians who see a large number of patients, and when the outcome of interest can be assessed over a relatively short period. In addition, in line with general pharmacoepidemiological principles, patients would ideally be “treatment naïve” for the same or a similar condition. Practically, this implies a reasonable time exclusion based on the known duration of treatment effects and the natural history of the illness. Finally, appropriate restrictions of the population under study based on clinical knowledge are essential to ensuring that patients unlikely to be treated with one or either therapy are excluded.

4. INDIVIDUAL PROVIDERS: THE PREFERENCE‐MONOMANIA SPECTRUM

Physicians, like anyone, are creatures of habit. These habits might change over time16 depending on the evolution of the scientific literature, clinical practice guidelines, or pharmaceutical industry marketing practices. However, particularly for specific patient populations and over shorter time periods, individual prescribing preferences may be highly stable. We have termed such high levels of prescription stability therapeutic monomania. In a recent informal survey of a convenience sample of Québec general practitioners (JLL), 82% reported always using the same drug when they saw new patients with the following uncomplicated conditions: lower urinary tract infections in women, upper and lower respiratory tract infections, first therapy in essential hypertension, first therapy in light to moderate asthma, and first therapy in type II diabetes.

In contrast to a physician who would always prescribe the same first line therapy for all patients with diabetes, for example, we would describe a physician who prescribes the same first line therapy 82% of the time as having a strong preference, or imperfect monomania. It also stands to reason that, when available, strict monomania is likely to be a stronger IV than preference. However, the stronger the preference (ie, the closer to monomania), the better PPE is likely to perform as an IV.

To this point, PPEs of general practitioners might be better IVs than those of specialists who may be more inclined to individualize therapy. In contrast to general practitioners, who might prefer the efficiency of “routine” first‐line therapies for certain diseases, specialists are more likely to see patients who have either failed previous therapy, are difficult to control, or are otherwise more complex or at higher risk of events.

Moreover, it is possible that not all general practitioner practices are suitable for use as IVs. General practitioners, too, sometimes orient their practices to certain diseases. Clustering of high‐risk cases would violate the third assumption of IV methods.8 Furthermore, as patients' lack of knowledge of a doctor's prescribing preference is a prerequisite for using PPEs as an IV, the possibility of “doctor shopping” would also violate the third assumption.8 It may therefore be necessary to restrict analyses to relatively homogenous groups of physicians.19 We therefore suggest that busy walk‐in clinics or emergency rooms are likely the ideal scenario for using individual physician PPEs as IVs.

As the goal of IV analysis is to identify a variable that will most closely approximate the coin toss of randomization, consider a patient with symptoms of sinusitis who presents to an emergency outpatient clinic for evaluation and treatment. By the nature of the clinic, the patient has no assigned physician. He takes a number and waits. Physicians on duty may well have pre‐established preferences for first‐line treatment of sinusitis, but the patient is unaware and unable to choose by whom he is seen. His number is eventually called, and he is assigned an examination room with a physician. It is the luck of the draw. This “natural” situation is perhaps as close to randomization as we will ever get in an observational setting. In this scenario, PPEs are likely to be strong and valid IVs.

It is important to bear in mind that the optimal method for establishing prescriber preference when using PPEs as an IV is a matter of debate and may change from one study to another for the reasons outlined herein. We suggest that using a relatively rigorous definition of monomania, such as uniformity of the two previous prescriptions with the index prescription, is a reasonable starting point for defining the PPE. It is, however, critical to assess different PPEs in a sensitivity analysis of the IV definition in all studies, such as varying the number of identical prescriptions necessary for assigning monomania (ie, four vs three vs two prior prescriptions).

5. GROUPS OF PROVIDERS: STRENGTH IN NUMBERS

Analogous to the situation with monomaniac prescribers, groups of providers may have similar prescribing behavior within their practice group, but that may differ from the equally homogenous practices of other groups. Indeed, formal practice standardization can be an efficient way of ensuring consistent quality care. However, informal practice standardization may also occur within practice groups, leading to divergent practice “cultures” between groups or institutions, particularly in areas of clinical equipoise. Financial considerations and contractual obligations may also lead to homogenous treatment for certain conditions at different institutions. (This applies particularly to medical devices and other expensive therapies where it might not be economically efficient to offer two seemingly equivalent competitive therapies.) Indeed, Ionescu‐Ittu et al15 found that hospital‐based PPE IVs performed better than individual‐physician–based IVs. However, similar concerns to those with individual providers with regards to case‐mix, specialization, and patients actively seeking out treatment at specific centers for perceived or real differences in care deserve careful consideration when selecting physician groups for study.

For argument's sake, let us consider two hospitals (Figure 2). Hospital A consistently prescribes treatment X at discharge to all patients presenting with a certain disease. Hospital B, on the other hand, consistently prescribes treatment Y for the same disease. Each group of prescribers could be considered therapeutic monomaniacs. On the assumption that the patients treated at both hospitals are part of the same population, the situation could be considered analogous to cluster randomization based on the hospital where a given patient seeks treatment. In this scenario, one can appreciate how IV methodology could also be employed to account for unmeasured confounders of the relationship between treatment X versus Y and outcomes.

Figure 2.

Representation of two theoretical hospitals exhibiting “therapeutic monomania” for different treatments [Colour figure can be viewed at http://wileyonlinelibrary.com]

However, while therapeutic monomania in individual prescribers might be difficult to establish, perfectly monomaniac treatment cultures in groups of providers may be even rarer depending on the treatment or pathology studied. Consider Figure 3, in which hospital B is replaced by hospital C where most (90%), but not all, patients with the same disease are managed with treatment Y (Figure 3).

Figure 3.

Representation of two theoretical hospitals; one exhibiting “therapeutic monomania” for treatment X and the other near‐monomania for treatment Y [Colour figure can be viewed at http://wileyonlinelibrary.com]

The reasons for receiving treatment X at hospital C may be varied, including a minority of “rogue” practitioners who prefer treatment X despite the prevailing local culture, a minority of patients deemed unsuitable for treatment Y by all practitioners, or the spurious influence of pharmaceutical industry marketing at various points in time. It may not be discernible within a dataset which influence lead to this minority treatment population at hospital C. As such, it may be difficult to determine how to best statistically address this nearly, but imperfectly, monomaniac hospital.

With individual prescribers, one may simply exclude practitioners that do not fit whatever definition of “monomania” one sets. However, the number of practice groups from which one may choose is necessarily a smaller pool than that of individual prescribers, and it might not be possible to analyze only perfectly monomaniac hospitals. Moreover, it is possible that perfectly monomaniac hospitals also differ from imperfectly monomaniac centers in unmeasured ways.

One solution to Figure 3 is to simply ignore the imperfection and apply an IV methodology as for Figure 2. Retaining all patients at hospital C would be analogous to an intention‐to‐treat (ITT) analysis of an RCT with imperfect compliance. This solution might lead to a dilution of the treatment effect, or, conversely, if the 10% of patients at hospital C receiving treatment X are substantially different, lead to a spurious exaggeration of a treatment effect. The risk of substantial bias appears small to us when there is not more than 10% practice divergence.

Another simple solution is to “transform” hospital C into a center with a perfectly monomaniac practice, as in hospital B, by excluding the minority treatment population from the analysis and applying IV methods. This is analogous to a per‐protocol (PP) analysis of a prospective trial and may yield a better estimate of the real‐world treatment effect. In a PP analysis, only patients receiving the randomly assigned treatment are analyzed, and those deviating from the assigned treatment are excluded from the analysis. Exclusion of these patients in an IV setting may well be appropriate if the minority population was substantially different from the majority population. However, there is nonetheless a risk of introducing bias if there is an equivalent unidentified subset of patients in hospital A that we cannot so easily exclude. Again, we believe that this risk is low if the divergent population is relatively small (≤10%).

However, both solutions (either ignoring or excluding the minority population at hospital C) might introduce significant bias if the size of the divergent population is moderate or large (Figures 4 and 5). Either excluding or ignoring the minority population is unlikely to be an optimal solution in these scenarios. However, foregoing IV methods opens the door to confounding.

Figure 4.

Representation of two theoretical hospitals; one exhibiting “therapeutic monomania” for treatment X and the other exhibiting a strong preference for treatment Y. However, as treatment X is prescribed 20% of the time, Hospital D might not be considered a “monomaniac” hospital [Colour figure can be viewed at http://wileyonlinelibrary.com]

Figure 5.

Representation of two theoretical hospitals; one exhibiting ”therapeutic monomania” for treatment X and the other not exhibiting a strong preference for either treatment X or Y [Colour figure can be viewed at http://wileyonlinelibrary.com]

>We propose that combining PS and IV methodologies is a promising solution if used correctly. It is critical to view IV and PS methods as separate and complementary solutions in this scenario to be applied sequentially or in parallel, as it has been shown that embedding an IV in the PS model can lead to increased bias and regressor inconsistency.20, 21 While Myers et al22 suggest that the introduced bias is minor in comparison with the correction for confounding afforded by IV adjustment, we propose that a two‐step method of PS matching followed by IV adjustment, if appropriate, might avoid this potential problem altogether.

PSs are generated by creating a regression model for treatment assignment as opposed to outcome. Matching patients based on their likelihood of treatment with X or Y is then possible.23 An important principle for propensity matching, however, is that of “common support.” That is, for matching to be effective (ie, both identifying the population of clinical equipoise while preserving sufficient statistical power), there must be a substantial number of patients who received either treatment X or Y who would have had a statistical likelihood of receiving the other treatment. Conversely, patients with a zero likelihood of receiving treatment X should not be compared with patients who indeed received treatment X, as they are fundamentally a different population, and similarly for treatment Y.

From this point on, we define the population of interest as the population for whom there is evidence of clinical equipoise in the data set, that is, the equipoise population. On the assumption that observed practice (prescription of treatment X or treatment Y) represents medically reasonable practice, the exclusion of subpopulations without common support (ie, the non‐overlapping tails of treatment propensity with a zero likelihood of receiving the other treatment) would leave only the subpopulation for whom clinical equipoise could be established based on the available data. This concept is important in that it suggest that PS matching, in addition to clinically guided exclusions and restrictions, can help one hone in on the population where true clinical equipoise exists. In Figure 5, if the 40% of patients at hospital E would have received treatment X regardless of where they were treated, there is no clinical equipoise in this subpopulation, and their exclusion is appropriate. However, without PS methodology, it is difficult to know who at hospital A should also be excluded for not being part of the population of interest.

While not the typical use of PS methodology, if the two hospital populations are of the same clinical population, common support should exist between them. We could therefore identify the population of patients at hospital A who could have received treatment Y had they instead presented to hospital E and excluding those patients who would have been managed with treatment X at both hospitals. In this way, we would be appropriately “transforming” both hospitals A and E into two monomaniac centers with equipoise populations fit for comparison with IV methodology.

Because residual unmeasured confounding is likely even after propensity matching, we recommend then using IV methodology for analysis of treatment in the transformed (matched) monomaniac center populations, provided that near‐perfect monomania (≤10% treatment divergence) was obtained. However, IV analysis is likely not appropriate if substantial divergence (ie, low levels of prescriber preference <80%) persists.

Consider Figure 6, in which we compare two nonmonomaniac hospitals. It is conceivable that the minority population in hospital F (20% receiving treatment Y) would have also all received treatment Y had they presented to hospital E instead. These patients would therefore not represent the equipoise population and would be excluded by PS matching because of their 100% likelihood of receiving treatment Y at either center. Similarly, if the same were true of the minority population at hospital E, they would also appropriately be excluded by propensity matching, generating two “monomaniac” centers with regards to the equipoise population. PS matching can therefore be an effective tool for both restricting the analysis to only the population of interest and providing a robust IV to address residual unmeasured confounding.

Figure 6.

Representation of how propensity matching might be able to “transform” two apparent non‐monomaniac centers into monomaniac centers for a restricted population. In this scenario, both centers are in fact monomaniac for the subpopulation for which there is clinical equipoise (i.e. common support) [Colour figure can be viewed at http://wileyonlinelibrary.com]

Alternatively, consider now that only a portion of the minority populations at both centers would have received the same treatment regardless of the hospital to which they presented (Figure 7). Those patients would be excluded, as above. Similarly, a portion of the majority populations at either center might also be excluded for the same reason. The remaining majority and minority populations at either center all have a nonzero likelihood of treatment with either therapy and therefore represent the equipoise population of interest.

Figure 7.

Representation of a scenario in which propensity matching fails to reveal underlying monomania for the equipoise population. In this scenario, use of instrumental variable methodology would be problematic [Colour figure can be viewed at http://wileyonlinelibrary.com]

However, in this scenario, the hospitals are no longer perfectly monomaniac, and the appropriateness of using IV methods at this stage is debatable. If the minority population remains substantial at either center after PS transformation, standard multivariable regression (MVR) analysis is probably the most appropriate next step. However, IV analysis might still be appropriate if the residual minority population is small in both centers (≤10% to 20%), so long as the limitations of ITT‐ or PP‐type analyses are acknowledged. It is also perhaps indicated to perform sensitivity analyses comparing standard MVR, ITT IV, and PP IV methodologies.

Finally, we strongly recommend using PS methods as a check in all analyses, even when comparing two seemingly monomaniac hospitals, as the appearance of therapeutic monomania may in fact be a by‐product of the unique patient populations treated at different centers. For example, if practitioners at hospital A had instead seen the patient population of hospital B (Figure 2), they might have also have prescribed exclusively treatment Y. PS matching would in such a case be expected to identify a lack of common support between two disparate populations that are therefore inappropriate for comparison and prompting the consideration of other centers for analysis.

6. RECOMMENDATIONS

Our recommendations for addressing both perfectly and imperfectly monomaniac hospitals are outlined in Table 1 according to the proportion of patients not treated with the primary therapy of the institution. PS matching prior to IV methods may ensure bias minimization in all cases. It should be noted in scenarios VI and VIII that physician‐based IV methods might not be appropriate depending on the residual divergent population in hospital B. Again, these recommendations represent the opinion of the authors, and not all recommendations have been empirically validated. Additionally, comparing PS and IV methods in isolation to one another may also serve as a form of sensitivity analysis for the magnitude of effect of unmeasured confounding.

Table 1.

Predicted impact of methodology on the risk of bias due to unmeasured confounding

Scenario	Divergent Treatment Population Hospital A, %	Divergent Treatment Population Hospital B, %	Proposed Method	Bias Risk
I	0%	0%	IV	Minimizeda
II	0%	0%	PS + IV	Minimized
III	0%	10%	ITT or PP IV	Low
IV	0%	20%	PS + ITT or PP IV	Minimized
V	0%	20%	ITT or PP IV	Moderate
VI	10%‐20%	10%‐20%	PS ± ITT or PP IV	Minimized
VII	0%	40%	PS ± ITT or PP IV	Minimizedb
VIII	0%	40%	ITT or PP IV	High

Abbreviations: ITT, intention‐to‐treat; PP, per‐protocol; PS, propensity score.

^aThe risk of bias is minimized only assuming both hospital populations represent the equipoise population of interest. We recommend propensity score matching for this reason in all cases.

^bWe do not recommend using IV methods if the residual divergent treatment population after PS matching is >20% and caution should be used if the population is 11% to 20% divergent with acknowledgement of the shortcomings of either an ITT or PP instrumental variable analysis.

7. CONCLUSIONS

In conclusion, IV analysis can be a powerful tool for addressing unmeasured confounding in pharmacoepidemiogic studies, and physician‐based IVs exploiting strong prescription preferences are particularly promising. Perfect therapeutic monomania (extreme preference, 100%) is perhaps the ultimate IV for both individual and groups of providers. Near‐perfect monomania (≥90%) is likely a reasonably strong IVs, and low PPEs (<80%) are likely poor IVs. When strict and robust monomania definitions are possible for individual prescribers, they should be preferred. We strongly recommend PS matching prior to IV analysis when studying groups of providers to (1) ensure that there is common support and (2) to attempt to “transform” centers with relatively weak treatment preferences into centers with strong preference (≥80% to 90%) for the equipoise population of interest. We do not recommend physician‐based IV analysis for centers that have lower than 80% treatment preference after PS matching.

CONFLICT OF INTEREST

None known.

ETHICS STATEMENT

The authors state that no ethical approval was needed.

Potter BJ, Dormuth C, Le Lorier J. A theoretical exploration of therapeutic monomania as a physician‐based instrumental variable. Pharmacoepidemiol Drug Saf. 2019;29(S1):45–52. 10.1002/pds.4757