|Home | About | Journals | Submit | Contact Us | Français|
Increasing interest in end users’ reactions to health information technology (IT) has elevated the importance of theories that predict and explain health IT acceptance and use. This paper reviews the application of one such theory, the Technology Acceptance Model (TAM), to health care. We reviewed 16 data sets analyzed in over 20 studies of clinicians using health IT for patient care. Studies differed greatly in samples and settings, health ITs studied, research models, relationships tested, and construct operationalization. Certain TAM relationships were consistently found to be significant, whereas others were inconsistent. Several key relationships were infrequently assessed. Findings show that TAM predicts a substantial portion of the use or acceptance of health IT, but that the theory may benefit from several additions and modifications. Aside from improved study quality, standardization, and theoretically motivated additions to the model, an important future direction for TAM is to adapt the model specifically to the health care context, using beliefs elicitation methods.
Health information technology (IT) research often focuses on IT design and implementation [1-4], but perhaps not enough on how clinician end users react to already implemented IT. Of course, there is more to health IT “success” than designing or purchasing a reasonably functional technology. Many recent reports of the unintended consequences of health IT [5-13] show that the fit between IT and the clinical work system will lead intended end users to accept or reject the IT, to use it or misuse it, to incorporate it into their routine or work around it [14-19]. In other words, “whether an information system is ‘successful’ or not is decided on the workfloor…” . Numerous cases of underuse, resistance, work-arounds and overrides, sabotage, and even abandonment are evidence for this claim [21-28]. Yet, many more studies of health IT are about its adoption—that is, whether clinics, hospitals, or clinicians have purchased and installed IT, and why—than about end-user reactions—e.g., how and why implemented ITs are used. Simon et al  noted this trend with respect to studies of electronic health records (EHR):
Most studies have focused on EHR adoption. Relatively little attention has been paid to the capabilities of those systems and the degree to which physicians with EHRs actually use them. Electronic health records have great potential to improve quality and safety in health care…but these improvements will occur only if clinicians have access to key functions in EHR systems and use them regularly.
There are, of course, numerous exceptions to implementation-focused research [30, 31]. Perhaps most notable is the application of the Technology Acceptance Model (TAM) to the prediction and explanation of end-user reactions to health IT . That this parsimonious theory of IT acceptance and use has penetrated the health IT literature is not surprising. In industries outside of health care, TAM is somewhat of a gold standard, if not a paradigm of its own . As much as 10% of the space allocated to Information Systems publications is claimed by TAM research . Reviews of the most basic version of the theory routinely find that it accounts for 30 to 40% of IT acceptance, despite its relative simplicity [34-38].
With over twenty studies testing TAM in health care and dozens more empirical and theoretic health IT papers mentioning the theory, TAM is increasingly portrayed as a fitting theory for the health care context. Yet, the TAM is not a model developed specifically in or for the health care context. If used in its generic form, TAM may not capture—or indeed may contradict—some of the unique contextual features of computerized health care delivery. Thus, determining whether TAM is a fitting theory for health care is an empirical question requiring critical examination. Accordingly, the objectives of the present work are (1) to undertake a comprehensive, critical review of studies of TAM with respect to health IT, with particular focus on how, if at all, those studies account for the health care context; (2) to judge the efficacy of TAM and its worth as a theory of health IT acceptance and use; and (3) to propose avenues of change and extension to the theory in an effort to improve its usefulness to researchers, designers, and health care decision-makers.
TAM was developed in the 1980’s, in light of concern that workers were not using ITs made available to them [39, 40]. Its originators reasoned that the key to increasing use was to first increase acceptance of IT, which could be assessed by asking individuals about their future intentions to use the IT. Knowing the factors that shaped one’s intentions would allow organizations to manipulate those factors in order to promote acceptance, and thus increase IT use. Early TAM research discovered that only three factors were needed to explain, predict, and presumably control acceptance.
To arrive at the model, its originators adapted the Theory of Reasoned Action (TRA), a general social- psychological/behavioral theory that had been proven useful for understanding a variety of behaviors such as voting, exercise, and condom use . As was customary for adapting such theory to new contexts , a preliminary study took place to determine what would be the appropriate variables to include in order to understand IT use behavior . The variables that were selected, and formed the first version of TAM, are defined in Table 1 and are graphically depicted in Figure 1a. The most proximal antecedent to IT use is behavioral intention to use it (BI), and this is now commonly what is meant when one refers to acceptance [40, 44, 45], although another common conceptualization of acceptance is end-user satisfaction [46, 47]. Because BI is thought to reliably predict actual use, and the latter is difficult to measure, BI is sometimes the only measured outcome of interest in a study of TAM [48, 49]. BI is influenced by one’s attitude toward using the IT (ATT). Attitude, in turn, has two determinants: perceived usefulness (PU) and perceived ease of use (PEOU). Additionally, PU is specified to have an independent effect on BI, and PEOU has an effect on PU. The theoretical and practical importance of all of these seemingly esoteric relationships is discussed later in this review. TAM is a theory that has gone through a number of changes. For example, an update called TAM2  (Figure 1b) removed the ATT component from the model, which originally mediated some of the influence of PU and PEOU. TAM2 also added a variable meant to capture the social influence (e.g., from colleagues or bosses) that compels end users to positively evaluate and accept IT, called subjective norm (SN). Even more recently, an impressive effort to unify the IT acceptance literature resulted in the Unified Theory of Acceptance and Use of Technology (UTAUT), a theory with obvious resemblance to TAM . UTAUT (Figure 1c) incorporates PU into a performance expectancy construct, PEOU into effort expectancy, and SN into social influence. New to the UTAUT, but not to IT acceptance research generally , is the modeling of facilitating conditions as one determinant of BI. UTAUT is a new but promising theory; early tests of UTAUT explained an impressive 70% of the variance in BI and about 50% in actual use.
Finally, some studies use early psychological models of behavior on which TAM was originally based [41, 42, 53]. These models are the Theory of Reasoned Action (TRA), or, more often, its successor, the Theory of Planned Behavior (TPB). The latter is depicted in Figure 1d. Variables from TAM and the other mentioned theories are defined in Table 1. As reviewed elsewhere [33, 34, 54, 55], many other revisions have been suggested to TAM and some past revisions have been questioned, but the three TAM descendents described above, and the general TPB, are the most commonly used.
Below we review studies in health care that used TAM and related models described above. In reviewing TAM’s application to health care settings, we cannot overlook the excellent recent review by Yarbrough and Smith  of various studies of physicians’ acceptance and use of health IT, which included 4 datasets that have been statistically analyzed using TAM as a theoretical base.1 That review also contains a discussion of TAM and a proposed research model that builds on TAM, as do recent writings by Holden and Karsh [28, 56], Ammenwerth et al , and others [58-62]. The present review differs from Yarbrough and Smith’s in several ways. First, it includes only those studies that have quantitatively assessed TAM or a small family of related models. Studies testing only Rogers’ Innovation Diffusion Theory , for example, are not included in the review. Second, here we review a larger set of 16 data sets analyzed in 22 studies, including ones that were missed by the earlier review and ones that have been published since. Third, the present review includes studies of non-physicians, including studies of nurses, technicians, pharmacists, physician assistants, and physiotherapists. Fourth, and perhaps most importantly, the conclusions and suggested future directions that we will present below differ substantially from those of Yarbrough and Smith.
Studies included in the review had to be published on or before July 2008. They had to quantitatively test relationships between variables specified by TAM. This criterion excluded studies that measured TAM variables but not the relationships between them [64-71] and qualitative studies that used TAM as a framework [72, 73]. Unlike Yarbrough and Smith , who reviewed qualitative studies of TAM, we were interested specifically in studies of TAM, a model of quantitative relationships between variables. Nevertheless, qualitative TAM studies can be informative. Day et al.’s qualitative interview study of hospice providers’ use of videophones showed that providers perceive videophone technology as useful (improved communication, better access to care) but not easy to use (low technical quality, difficult to use for patients), and that ease of use appeared to be a more dominant contributor to intentions to use . Karsh et al.’s study of primary care physicians’ and nurses’ perceptions of using error reporting systems revealed perceptions of usefulness (improved care, suggestions for improvement) and ease of use (easy and quick to use, minimal extra workload, good instructions), as well as subjective norms (colleagues, supervisors, internal and external organizations) .
We also excluded studies that did not use TAM as a theoretical framework but tested one or more relationships specified by TAM [74-76] because those studies do not represent tests of the model. For example, a study that only examined the relationship between PEOU and PU would not have been included. Studies had to focus on health IT use or acceptance (i.e., behavioral intention to use). This excluded studies of nurses’ acceptance of tasks or processes [77, 78] and studies that used TAM to predict IT adoption by an organization . We defined health IT on the basis of convergence between international definitions of health IT and health/medical informatics. Two illustrative definitions of health IT are:
Based on such definitions, we restricted the review to studies of technologies that digitized information for the purpose of delivering (direct) patient care. This excluded studies of TAM that focused on web-based learning or online training courses [82-85], clinical decision support knowledge authoring tools , and adverse event reporting systems . The end users of the health IT had to be health care professionals providing medical care. This excluded studies of health IT use by patients [88-90] or non-medical health care providers such as social workers . Those exclusions were made so as to keep our reviewed studies in the same general context. That is, in all of the included studies, the health IT would be affecting the actual work of the clinicians by affecting the care they deliver. Had we included studies of health IT not used for patient care or not used by clinicians, it would have increased the heterogeneity and made drawing conclusions more tenuous.
We excluded unpublished dissertations . Studies had to be published in a language spoken by the authors and accessible to the authors through the University of Wisconsin-Madison library and its comprehensive electronic subscription and interlibrary loan systems. Studies that were excluded because they did not meet our criteria may still provide useful insight for understanding health IT use and acceptance, and should be prime candidates for broader reviews in the future.
Following Yarbrough and Smith , we conducted searches of the online database PubMed/MEDLINE with the keywords “Technology Acceptance Model,” “TAM,” “TAM2,” “UTAUT,” and “Universal Theory of Acceptance and Use of Technology” and the ABI/INFORM Global database with the same keywords plus “health*,” “physician*” and “nurs*” in order to narrow the search to health care settings. Obvious non-health care/non-technology results were removed first, and the abstracts of all remaining results were read. Full versions of all articles that were possibly relevant were retrieved and read. For each retrieved article, we conducted a search of references that might meet inclusion criteria, and retrieved those articles.
For this review we purposely decided not to conduct a quantitative meta-analysis. As others who have reviewed health IT research have pointed out, a meta-analysis would have been inappropriate because of the significant heterogeneity among studies in terms of sample characteristics, specific technologies studied, and functions of those technologies [93, 94]. However, we did provide information on the percentage variance explained in the dependent variable (e.g., R2), for readers to evaluate and compare. Those values must be interpreted with caution, because of the substantial variability in studies.
Efforts to apply TAM to health IT date back to the late 1990’s, beginning with studies by Hong Kong researchers testing the TAM and, subsequently, different versions of TAM and TPB, in a sample of 408 surveyed physicians with access to telemedicine IT [32, 48, 49, 95]. Their findings were disappointing, and they asserted that TAM was a poor fit for physician acceptance of health IT, perhaps because of professional differences between physicians and other workers who use IT . In total, 16 datasets have been analyzed, and TAM has fared much better in later tests. Reviewed studies are summarized in Table 2 and are discussed below.
Of the 16 data sets analyzed using TAM and related models, 7 analyzed data from all-physician samples. The physician specialties covered were endoscopy , pediatrics , disability care , and some mix of specialists and general practitioners [32, 99-101]. In a few studies, participants were nurses [102, 103], and physio-  or occupational  therapists. Five studies collected data from a mix of physicians, nurses, pharmacists, and medical technicians [106-110].
Interestingly, studies of TAM in health care have been conducted in a wide variety of countries. Several studies were set in Taiwan [103, 109, 110], Hong Kong [32, 48, 49], the UK [96, 104], Australia [102, 111], and the mainland US [98, 99, 106, 107] and Hawaii . Others were set in Belgium , Finland , and Quebec .
There was variation in the health IT that served as the focus of the study, including telemedicine technology , picture archiving and communications systems [108, 109], and computerized provider order entry (CPOE)  (see Table 2 for full list). In health care and health IT domains, there is currently much interest in clinician use of electronic medical records (EMR) for patient care. Surprisingly, only one set of studies focused on EMR , although it was only a trial system that was temporarily evaluated by study participants. Additionally, a study that is not here considered as a study of TAM focused on the perceptions of 51 physicians and 51 midwives toward a planned EMR system . The study used TAM as a theoretical framework to formulate survey questions and then to interpret results, but did not test the relationships between TAM variables.
The studies mentioned did not test a single, uniformly specified model of relationships between TAM variables. Most studies began with TAM as a conceptual framework and removed [32, 102, 104] or added predictor variables [49, 96, 98, 100, 101, 103, 106, 107, 109]. In some cases, added variables were treated as independent predictors of intention to use or actual use of health IT, and in other cases added variables were used to predict TAM variables such as PEOU and PU. For example, several studies tested the effect of health IT compatibility on behavioral intentions [98, 100], whereas others tested the effect of compatibility on PEOU and PU [48, 49, 106], and yet others tested both kinds of effects [103, 105, 109]. The list of added variables includes various perceived IT system characteristics such as how well the system performs [96, 107, 112] and how relevant it is to one’s job , personal characteristics of users [98, 105, 106], characteristics of the organization such as readiness for IT or technical support [98, 105, 106, 109], and psychological variables such as ownership  and trust . One study tested TAM2 , two tested UTAUT [108, 110], and two tested some hybrid of UTAUT, TAM and TPB [99, 105]. Sometimes, the researchers compared TAM to another model, for example, TPB [48, 102].
Studies varied slightly in how key variables were conceptualized, as depicted in Table 3. Perceived usefulness (PU) was defined as health IT use leading to enhancement or gains in job performance. Notably, no study used a broader definition of usefulness, despite the notion that health IT might be useful for not only enhancing performance, but also making performance easier and more satisfying, increasing efficiency and lowering costs, improving quality and safety of care, and more [113-115]. Whereas definitions based on those of Davis’s original TAM [39, 40] focused on personal performance, those based on UTAUT  did not, but they also did not explicitly refer to performance benefits for other members of the care team, referring specialists, or patients’ families. It was unclear from either definitions whether usefulness referred to enhanced performance process (e.g., fewer steps, more information for decision making) or enhanced performance outcomes (e.g., faster care, more accurate decisions) .
Perceived ease of use (PEOU) was either defined as the lack of (physical or mental) effort or simply as “ease of use.” As with PU, the definitions were somewhat limited (e.g., they do not account for the difficulty of low-effort but highly repetitive tasks) and generic (e.g., not referring to specific components of usability ).
Subjective norm (SN), or social influence, was consistently defined with respect to the opinions of important others about an individual’s use of health IT. Although this definition follows the TPB conceptualization of how social influence is directly exerted, the definition ignores other ways that social factors indirectly influence behavior, for example through a culture of health IT use or when people are influenced by vicariously observing the actions and attitudes of others around them, as opposed to being told directly to use or not use health IT . Another open question is whether the social influence is having its effect through compliance, identification, or internalization processes , a point that was addressed in only one reviewed study .
Perceived behavioral control (PBC), or facilitating conditions, was defined variously as perceptions of existing (a) infrastructure, (b) internal and external resource constraints, or (c) skills, resources, and opportunities necessary to use the system. It is not surprising that definitions differed, given some of the disagreement and variation in how the construct is defined in the psychological literature [120, 121].
To sum up so far, it is a testament to the strength of TAM and related theories that the definitions were similar. At the same time, using TAM and UTAUT definitions resulted in a limited and generic way of thinking about constructs such as usefulness and social influence. Why such definitions are problematic despite being seemingly consistent is most clearly illustrated by looking at how key study constructs were measured in reviewed studies.
How constructs were operationalized differed greatly between studies, as can be seen in Table 4. Each construct was typically assessed with four or five items. PU was most commonly measured by asking about how much the health IT was “useful to the job” or to a specific task, how much it increased productivity, or how much it increased job effectiveness. Those measures followed the conventional definition of PU as enhancing (personal) job performance, although the productivity and effectiveness dimensions dominated, compared to other useful health IT use outcomes that could be measured, such as efficiency, quality, and so on. Indeed, some studies asked questions specifically about the following outcomes of use: quicker task completion; easier work; increased quality of care or quality of work; improved efficiency; more accurate or more objective accomplishment of tasks; support of critical tasks; increased chances to get a raise; greater control over work; better evidence-based decisions; and improved patient care. Some of those dimensions of usefulness referred to the process of performance (e.g., easier work process) and others to the outcome (e.g., better quality of care).
PEOU was variously assessed on these dimensions: easy to use; clear and understandable; easy to become skillful with system; easy to get the system to do what one wanted; easy to learn to operate; flexible; requiring low mental effort; easy to do what one wants when using the system; easy to do tasks when using the system; clear; understandable; not demanding of much care and attention; easy navigation; and tasks were easy to remember. It is notable that some of those dimensions reflect specific aspects of usability such as learnability and memorability . Despite the PEOU definitions given in Section 3.4 above, only two studies had questions mentioning effort specifically [96, 97].
SN measures were similar to one another in that almost all asked about the degree to which some referents thought the clinician should use the system. What varied widely was the referent’s identity, from “People whose opinions I value” to “People who are important to me” to “Pediatricians who influence my behavior” to “Colleagues,” “Supervisors,” and “Subordinates,” as well as other similar permutations. Thus, the direct route of social influence—through others having an influential opinion about another’s health IT use—was almost always the same, but the source of social influence varied in specificity (e.g., “important others” versus “important other pediatricians”) and type (e.g., “colleagues” versus “subordinates”). Notably, one study  used two questions—about the helpfulness and support of the hospital and management—that did not appear to follow from the study’s stated definition of social influence. In general, it was not uncommon that a study’s definition for a construct did not match its operationalization.
PBC measures varied as well. Some measures asked about the availability of knowledge or technical assistance for using the system. Others asked generally about being able to use the system at work or for patient care. One study used a question about the dimension of PBC referred to as controllability by psychological theorists , which in the case of IT refers to the degree to which using the system is up to the clinician.2
Table 5 summarizes which relationships specified by TAM and related models (Fig. 1) were significant and non-significant across the studies reviewed here. Only two studies tested the relationship between behavioral intentions and actual use [102,110]. In one study, the relationship was significant . In another, the relationship was only significant for one of two types of use behavior . The effect of ATT on intention to use, as specified by the original TAM, was significant in five out of six tests. The original TAM also specified that PEOU and PU are related, and this was the case in 10 out of 12 tests. In the studies testing the TAM postulate that PEOU and PU are both related to ATT, PEOU-ATT was significant in one of two tests , and PU-ATT was significant in three of three [32,101,105]. Every one of the 16 tests of the relationship between PU and intention to use was significant. The PEOU-intention relationship was much more inconsistent, significant in only seven of 13 tests. The effect of subjective norm on intention was similarly questionable, significant in only four of eight tests. In contrast, all five tests of the PBC-intention relationship were significant.
Some studies, though not all [102,109], attempted to explain the lacking predictive power of PEOU and SN. Some suggested that the lack of effect of PEOU was due to clinicians experience with the IT. One such explanation was that having no actual or little hands-on experience with the IT led to a poorer estimate of PEOU [96,104,108]; a second was that with time and practice, the PEOU of the system became no longer important to acceptance and use [97,100]. Looking across the reviewed studies, there appears to be some evidence for the first explanation, that lack of exposure appears to be associated with non-significant PEOU-ATT and PEOU-BI relationships. Another common explanation was that there was something special about physicians, such as their greater intellect and ability to learn to use technology [49,95,99], their relative disinterest in usability as long as the system is useful [96,97,108], and the availability of support staff to deal with the system for the physician [49,99]. Indeed, of the seven studies with non-significant PEOU-ATT or PEOU-BI relationships, six involved physician users. Of the eight studies with significant relationships, one had an all-physician sample , one had a sample of occupational therapists , two had all-nurse samples [102,103] and the remaining five had mixed (mostly-non-physician) samples. Note that sample differences in profession were sometimes confounded with sex and age differences (e.g., compare one physician sample of 70% males and 80% in their 40s and 50s  to one allfemale nurse sample with 98% less than 35 years old ). Physician-specific characteristics were also named as reasons for the non-significant effects of SN, namely their independence and general immunity to peer influence [49,97,108]. Indeed, only one of the four studies with a significant SN effect had an all-physician sample . Two studies suggested that the lack of SN effect was due to the system use being mandatory [97,108]. In opposition to that hypothesis, in three of four studies with non-significant SN effects, IT use was voluntary [49,97] or was perceived to be voluntary , and IT use was mandatory in two of four studies with significant SN effects [102,110].
Finally, Table 2 gives the percentage of variance explained in the dependent variable (R2 or an equivalent statistic) by all of the variables tested in each study or set of studies. We caution the reader about too strongly interpreting these values, because of study heterogeneity and other factors that limit comparability. For example, some studies tested as few as two predictors of behavioral intentions , resembling the original TAM, whereas others used as many as five predictors [103,105]. Generally speaking, however, the percentage of variance explained in behavioral intentions (or actual use) is reasonably high.
From the preceding review, it is evident that TAM has had widespread application in explaining health care providers’ reactions to health IT. The recent increase in the use of TAM appears justified, with many of the relationships specified by TAM repeatedly validated in health care settings and fairly large proportions of variance explained in the dependent variable (be it intention to use or actual use). Perhaps most impressive is that the relationship between PU and intention to use or actual use of health IT is significant in every test, suggesting that to promote use and acceptance, the health IT must be perceived as useful. But in addition to the apparent strength of TAM in health care studies, there are remaining challenges. Both are addressed next, followed by a discussion of future directions of TAM research in health care.
On average, the reviewed studies yield (1) reasonably high R2 values, (2) frequently large effect sizes (not reported) for relationships between study variables, and (3) some consistently significant relationships. There is strong evidence to conclude that the perceived usefulness of an IT will have some impact on whether clinicians accept and subsequently use a health IT. Whether the IT is perceived to be easy to use may not be as likely to affect acceptance, but it does appear to correlate with usefulness, perhaps reflecting the notion that IT that is difficult to use cannot possibly be perceived as useful . The implication is that design, training, and informational sessions must focus on ensuring that health IT is (or, at least, is perceived to be) capable of improving important outcomes and is not difficult to use. The consistently significant relationship between PBC and acceptance suggests that, regardless of how useful and easy to use the health IT is, steps will need to be taken to ensure that end users feel confident in their ability to use it (self-efficacy), that using the IT is under their volitional control (controllability), and that barriers are removed and sufficient support provided (facilitating conditions).
There were a number of other strengths. Almost every study used well validated questionnaire items from previous studies, something that has been discussed as a major benefit of the penetration of TAM in IT research , and of the use of theory in health IT research, generally . Studies also tended to report the psychometric properties of scale items, which were usually reasonable and did not differ substantially between studies. There were, however, cases of measurement misspecification, for example the use of a single-item measure in one study , or the measure of the attitude construct using a “heightened enjoyment” scale in another , reflecting a common misunderstanding of attitude theory . Other cases of misspecification have to do with the lack of correspondence between construct definitions and operationalizations, discussed in Section 3.5. Many studies included a variety of health care professionals, and some compared the groups . Finally, almost every studied added variables to TAM in an attempt to better understand the antecedents of acceptance or health IT use behavior. This added variables approach, discussed below, represents the growth of the discipline in its incorporation of multiple theories and multiple sources of causality to account for the complexity of health care’s socio-technical systems [20, 105, 123-125].
There were also a number of challenges faced by the body of reviewed studies. As mentioned above, tests of expanded models allowed researchers to test potentially important relationships not specified in the original TAM , such as the effect of subjective norm [49, 97, 99, 102, 105, 108]. At the same time, the variety in model specification—so great that not a single study tested the same model—greatly limits quantitative or qualitative comparison across studies. Perhaps the lack of standardization is due to the variation in TAM research, generally , seen in the variations between TAM, TAM2, and UTAUT, or in early influential work testing direct relationships between PU and PEOU on one hand, and actual use on the other . Indeed, some scholars have criticized TAM research for adding variables haphazardly, leading to a progressively less coherent theory . Even the language used in the reviewed studies is not always standardized, with the terms adoption and acceptance used interchangeably and incorrectly. Standardization of the models tested need not be in conflict with the addition of new variables and relationships: researchers can test a standard TAM for comparison, and can separately test expanded models as well.
Few studies assessed moderators [97, 100, 105], despite their importance in recent writings on TAM [36, 38, 51, 127]. Among moderating or contextual factors, one worth exploring may be the voluntariness of health IT use [46, 50, 102, 110]. Another is the stage of the health IT; some studies reviewed here studied prototypes , trial systems , or to-be-implemented systems , whereas others studied implemented systems that had been around for different lengths of times. Studies of TAM have shown that over the life course of an IT, the relationships in the model may change; for example, ease of use may be critical at first and less important as time goes by [51, 100, 128]. Type of health care professional is another moderator deserving of further inquiry [97, 103, 111]. Although individual or profession-specific differences may not have been possible to study in homogeneous samples, such as those comprised of only young physicians [99, 109], all-females , or all-males , there were studies in which comparisons were possible but not carried out, such as the study aggregating data from physicians, nurses, and information systems department staff , one aggregating data from pharmacists, physicians, nurses, top and middle managers, technicians, and more , or one aggregating across dental hygiene, physician assistant, and radiology professions .
Overall, there was great variability in measures used within and between studies, despite the constructs being similarly defined and the existence of established and validated questionnaires. Evidence of variability certainly challenges the degree to which the different tests of TAM were actually similar enough to compare. It also indicates to us deeper problems with TAM, at least as it is used in health care. The generic definitions of TAM were differently interpreted, a potential sign of lacking theoretical explication of the constructs. At the very least, even if there are theoretic positions on and explanations of constructs, there may be lacking agreement or exposure, because TAM investigators in health care differently interpreted constructs defined the same way. For example, there is work looking at the meaning and makeup of PU , and other work suggesting that useful outcomes can be seen at multiple level—e.g., IT that is useful for the individual versus for the organization —but these theoretical details did not systematically transfer to the studies reviewed here
Sometimes study investigators used measures that did not fit with the definitions and measures that are not typically included in studies of TAM. This may be because investigators felt that the established definitions and measures did not match their particular study context. TAM was developed outside of health care, and therefore some of its core concepts and measures may not appear relevant to health care investigators . For example, the focus on personal productivity as a measure of usefulness may not be meaningful, and certainly not sufficient, in a health care context. Thus, some studies incorporate measures of increased quality of care [97, 106] and similar health care-specific measures of usefulness [32, 104]. Nevertheless, most measures were generic and not health care-specific. Why was this? One possibility is that the health care researchers were not aware of the benefit of contextualizing TAM constructs for health care and the costs of not doing so. Another is that they did not have much basis upon which to determine a priori what would constitute contextualized TAM measures. What does it mean for health IT to be useful to clinicians? What does it mean for health IT to be easy to use? Who are the referents exerting social influence on clinician end users? What kind of barriers and facilitators are important for health IT use? We believe that those kinds of questions are not trivial and require a rigorous, data-driven approach, which we recommend and describe in Section 4.4.
Finally, looking at Table 5, a few other limitations are obvious. First, certain variables such as actual use have been too infrequently measured, precluding tests of several important relationships, such as the one between intention and actual use [102, 110]. When the target IT is not yet implemented at the time of the study, measuring variables such as actual use may require longitudinal research, something that is greatly lacking in the reviewed studies. Some relationships have been inconsistent (or consistently weak), raising the possibility of moderating effects or other theoretically important differences between studies. Are the effects of PEOU spurious, given their inconsistent predictive power? Perhaps so, but in some studies, the effect of PEOU is as strong as that of PU , if not much stronger . There is also evidence that how strongly PEOU, or other variables for that matter, is related to attitudes, behavioral intentions, or actual use depends on factors such as subject type and technology type . Indeed, the two studies with relatively strong PEOU effects both had young, mostly non-physician samples in Taiwanese medical centers who had hands-on experience with the IT [103, 109]. It may also be that the measures of PEOU in the reviewed studies were not sensitive to the ease of use dimensions that are important in health care settings, something that is discussed below.
Given the nascence of TAM research in health care, several future general directions and specific research questions can be suggested. Table 6 provides a non-inclusive list of suggestions, many of which apply to TAM research in general. The first two general directions are next described in more depth.
Above we praised what one might call the “added variable” approach to testing TAM in health care settings. Such an approach was commonly taken in the reviewed studies, in order to either better understand the factors that might cause behavioral intentions or actual use of health IT, or to understand the causes of other factors in TAM, such as attitudes or PEOU and PU. As certain variables reappear in models tested in health care, they help researchers, designers, and others to better understand health IT use and acceptance, and there may be benefits to developing general IT theory as well [58, 130]. For example, the common addition of variables related to compatibility between the health IT and clinical work processes in the reviewed studies [48, 49, 98, 103, 105, 106, 109] is in line with growing interest in the concept of “fit” as a critical need for successful IT design [131, 132], especially in health care [21, 28, 57, 62, 72, 100, 133-135].
The tradition of adding variables is common in research on TAM [51, 136, 137] and on more general theories of human behavior [138, 139]. Theory-based additions to the prediction and explanation of health IT use and acceptance is a welcomed approach [28, 56], and it appears to be the approach advocated by some researchers for furthering the use of TAM in health care [30, 58, 87, 99, 100, 109]. But where will these added variables come from? And what modifications can be proposed for the variables that are already included in TAM? In the following section, we argue for an approach that is complementary to the added variables approach, but that makes explicit how researchers and practitioners can better expand and modify TAM for the health care context.
Those in health care who advocate the use of TAM commonly point out how different physicians and other health care providers are from the users of IT in other studies of TAM, and how the general health care context differs from that of the industries where TAM originated [30, 32, 58, 61, 96, 98, 100]. For example, one writes, “Generally, the essential characteristics of users and technologies in professional healthcare differ greatly from the customary commercial context … thus, any model developed for the general public may not apply to a healthcare environment” . The importance of context cannot be overstated . In a different context, not only may there be differences in the variables needed to understand IT use and acceptance, but also the meaning of variables such as PU and PEOU may be drastically different. The great variability in how reviewed studies interpreted and operationalized TAM constructs was discussed above. We suggested that the variability could be attributed to the lack of systematic contextualization of TAM to health care settings, health care users, and health care tasks. Different health IT researchers may differently interpret what is “useful” about health IT, in the absence of health IT-specific definitions of and measures for usefulness (see Table 4). Similarly, the sources and maybe even the route of social influence may be different in health care compared to in the other industries where TAM was developed and refined. For a physician, social influence might come from patients, patient families, professional groups, regulatory agencies, employing organizations, insurance providers, fellow physicians, referring physicians, nurses, bosses, and others. A contextualized version of TAM would allow researchers and practitioners to become aware of the health care-relevant dimensions of PU, PEOU, SN, and PBC. Further, we suspect that part of the reason for the inconsistent predictive power of variables such as PEOU and SN in health care may have to do with a poor match between construct operationalization and the context in which it is measured. For example, the study of Yi et al  is one of few that find an effect of SN, and the authors used three survey items that referred to specific sources of social influence, colleagues, superiors, and subordinates rather than asking general questions (see Table 4). It is difficult to formally assess our suspicion given the limited number of studies and missing information about measures in several studies. Thus, we only tentatively suggest that if contextualized TAM measures replace the practice of using generic measures or specific non-standardized measures based on a particular research group’s “common sense” , perhaps we will see better predictive power.
There is still the matter of how one can arrive at contextualized study of TAM. We propose a specific process for doing so. The process is called beliefs elicitation [42, 142-145], and it is the preferred method for contextualizing theories of behavior to a specific setting (health care), with a new population (clinicians), and a new behavior of interest (health IT use) . In an elicitation study, researchers find out about the beliefs that participants have about a certain behavior, based on participant responses to open-ended questions. Those beliefs become the basis for subsequent theory. As a matter of fact, beliefs elicitation was the process that Davis and colleagues used in order to fit general behavioral theory to the context of IT use, to form TAM . They found out, based on this preliminary work, that the two key beliefs were PEOU and PU, and they defined those based on what was important to their test population; thus, for example, PU was defined in reference to improvements to personal productivity, as opposed to improvements in patient outcomes, because the participants in Davis’ study were college students, not clinicians.
We believe that conducting elicitation studies with clinicians, focusing on the salient aspects of their health IT use in the health care context will lead to better refined, contextualized theories of health IT use and acceptance, just as early studies in other industries led to refined and contextualized theories of IT use and acceptance by students, bank tellers, programmers, and others. Few have done this kind of beliefs elicitation work with respect to health IT [70, 72, 76, 96, 102], and none to our knowledge have revised TAM based on such work.3 Yet guides on how to carry out such work are readily available [42, 142, 146]. Those guides contain directions for asking about, for example, important others who approve or disapprove of health IT use behavior, in order to determine who specifically is believed to exert social pressure on clinicians.
In the process of carrying out a beliefs elicitation study, the researcher is able to learn not only about what are the salient beliefs held by clinicians, but also the reasons for those beliefs. This has practical implications because knowing the reasons for health IT use beliefs permits designers, policy makers, and others to take action and make changes, a benefit recognized by the authors of two studies reviewed here:
The major advantages of eliciting beliefs from a sample of the population of interest are that there is a greater guarantee that the beliefs will be relevant to the population … and that intervention strategies may be properly targeted at the key issues. 
In general, beliefs are important not only because they influence subsequent behavior, but also because they are amenable to strategic managerial manipulation through appropriate interventions such as user involvement and user participation … In the health care context, an understanding of what causes physicians to hold certain beliefs about the target CIS [Clinical Information System] would be of value not only to individuals responsible for overseeing implementation of these systems, but also to researchers interested in explicating the paths through which technology use behavior is manifested. 
A final benefit of beliefs elicitation is its ability to develop what we call an evolved model of health IT use and acceptance. By evolved, we mean a model that is unconstrained by the behavioral theories of the 1960’s and 70’s on which the original TAM was based [41, 147]. Rather, a researcher who asks questions based on more recent theoretical advances in the prediction and explanation of behavior [33, 120, 122, 139, 148] can begin to build a theory of health IT use that better parallels current knowledge about why people behave in certain ways, and why they use or do not use certain technologies. An evolved theory may thus incorporate recent distinctions between cognitive and affective causes of behavior, the effect of facilitating conditions versus self-efficacy versus controllability, both external normative and personal normative (e.g., moral/ethical) influences on behavior, and more.
Note that a beliefs elicitation approach is compatible with the added variable approach because elicitation studies can suggest and support additional model variables. The main difference—and the point on which we may disagree with Yarbrough and Smith—is whether new variables or variable modifications should be made solely on the basis of existing theory or on a combination of theory and empirically elicited beliefs. Although a purely theory-based approach—which might for instance add the compatibility variable from Rogers’ Innovation Diffusion Theory —can be useful, empirical work may still be necessary to (a) determine whether those variables are relevant to health care, and (b) contextualize these new variables in order to determine, for example, what “compatibility” means in health care [57, 135].
TAM is a well regarded theory of technology acceptance and use that has been widely researched outside of health care and has lately become an important theoretical tool for health IT research. Designers, purchasers, and others involved with IT projects are routinely advised to use the TAM to aid the design or purchasing process, training and informational sessions, implementation, and other activities. To the extent that the factors predicting acceptance are controllable, they can be strong levers for acceptance and use. TAM has a strong track record in numerous industries; will it be a useful theoretical tool in health care? In this paper, we assessed TAM’s future in health care by examining its past. Although TAM did a fair job predicting, and perhaps explaining, clinician end-user acceptance and use of health IT, there is much room for improvement. Aside from a need for standardization, more tests of certain relationships, and better reporting of data, there is also a need to continue exploring new theoretically motivated variables and relationships that can be added to TAM.
More important, researchers should conduct studies for the purpose of identifying salient beliefs that clinicians have about using health IT for at least two reasons. First, this will allow researchers to probe about a wide range of theoretically interesting clinician beliefs, which could make the theory more robust and relevant in health care. Second, and equally important, by contextualizing TAM to health care, there is the opportunity to develop the “left side of the model.” That is, contextualization uncovers the specific and, importantly, actionable meaning and causes of generic variables such as usefulness and ease of use, helps determine who are the “important others” of subjective norms and what are the actual barriers and facilitators to IT use. With the US National Research Council concluding in 2009 that current health IT is not designed to adequately support the cognitive work of clinicians , the time to begin uncovering specific, contextualized, and actionable constructs is now.
Overall, there is great interest in TAM in health care4 and ample opportunity for its success, but whether TAM evolves into a theory of health IT, as opposed to a theory for health IT, is still to be seen.
Author R.J.H. is supported by a training grant from the US National Institutes of Health (TL1 RR025013-01). The work was also funded in part by grants by the US National Library of Medicine (R01 LM008923) and the US Agency for Healthcare Research and Quality (R01 HS013610) to author B.K. The funding agencies had no involvement with this paper. We appreciate the helpful comments of several anonymous reviewers.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
1Although the authors reviewed eight studies that tested TAM, five of those analyzed the same sample of 408 Hong Kong physicians, and the authors noted this fact.
2Because of certain theoretical considerations, some of the studies also separately measured voluntariness or mandatoriness of use, which is like controllability. Further, some studies measured self-efficacy, which is a perception of one’s own ability to do something .
3One study reviewed here  found that a contextualized model created using beliefs elicitation slightly outperformed the generic TAM in a health care setting, but did not suggest modifications to TAM. Unfortunately, the study did not provide much detail about differences between the two models or the process and results of the elicitation process, making it difficult to draw independent conclusions.
4In fact, in the one year period following this review, six additional health IT studies of TAM were published [150-155], as were several qualitative/descriptive studies using TAM or a related framework [156-158].
Conflict of interest None.