Psychometric Properties of the Neuropsychiatric Inventory: A Review

INTRODUCTION

More than 40 million individuals worldwide are estimated to have Alzheimer’s disease (AD) or other dementias [1], driving research efforts to understand these conditions. Neuropsychiatric symptoms (NPS), sometimes referred to as non-cognitive symptoms or behavioral and psychological symptoms of dementia [2], are disturbances in behavior, thought, and emotion related to neurocognitive disorders. These symptoms are common [3], but their measurement remains a challenge.

NPS are often assessed using informant questionnaires. The information obtained with them is not always easy to interpret, as the theories of NPS research either await formulation or their relationship to NPS measures is unclear [4]. Furthermore, NPS constructs overlap with one another [5, 6], and the associations between constructs, or the behavioral and psychological phenomena of interest, and measures (i.e., scores on an informant questionnaire) [7] remain incompletely described [8, 9]. The lack of clarity about measures arises from the relatively limited psychometric data available regarding most NPS scales.

The most commonly used scale in NPS research is the Neuropsychiatric Inventory (NPI) [10]. This scale is a flexible and a relatively brief informant questionnaire that screens for the most common NPS in AD and other neurocognitive disorders. It has been utilized in both research and the clinic, with applications in clinical trials, neuroimaging, descriptive psychopathology, and as a validity benchmark for developing subsequent NPS scales [11–13]. Modifications for more specific contexts also have been made, yielding the Neuropsychiatric Inventory-Questionnaire (NPI-Q) [14] for very brief screening, the Neuropsychiatric Inventory-Nursing Home for nursing home contexts [15], and the detailed Neuropsychiatric Inventory-Clinician version (NPI-C) [16], integrating information from the patient and informant. Variants of the NPI have been applied outside the context of dementia, including in stroke [17], delirium [18], and cognitively healthy elderly [19]. Some researchers predict that the NPI could be applied to central nervous system (CNS) disorder clinical trials in general, covering “anticonvulsants, Parkinson’s disease therapies, hyperactivity/attention deficit disorder, pain therapies, CNS cancer therapies, and multiple sclerosis treatments” [11].

The NPI probes into 10 or 12 symptom domains depending on the version used. The 10 original domains are: delusions, hallucinations, agitation, depression, anxiety, euphoria, apathy, disinhibition, irritability, and aberrant motor behavior [10]. Sleep disturbances and appetite disturbances were added later [12]. Two important features distinguish the NPI from several conventional scales in psychiatry and neurology that rate the severity of individual symptoms on a Likert scale, e.g., from 1 to 5, and sum these values for the total score.

The first distinguishing feature of the NPI is the skip-question format (Fig. 1). In the NPI, informants are first asked screening questions related to the 10 or 12 symptom domains. Following a positive answer, the informant is asked six to nine subquestions per domain. Following a negative answer, the interviewer proceeds to the next domain without asking the subquestions in the screen-negative domain. This procedure saves time but is not without psychometric problems because of the assumption that the subquestions are only relevant if there is a positive answer on the screening question [20].

Fig. 1

The skip-question and scoring procedure of the Neuropsychiatric Inventory. This procedure is repeated until all domains are covered.

The second unique property is the scoring system. A screen-positive domain is given a domain score based on the product of the frequency (1 = occasionally, 2 = often, 3 = frequently, 4 = very frequently) and severity (1 = mild, 2 = moderate, 3 = severe) of the symptoms in the behavioral domain. The range is then 1 to 12 per subdomain, and scores of 5, 7, 10, and 11 are not possible. A total score for the NPI can be calculated as a measure of general level of psychopathology (maximum of 120 and 144 for 10 and 12-domain versions, respectively).

This scoring method was chosen with clinical trials in mind: it was considered that the effect of a treatment could manifest separately in reduced symptom frequency or severity [12]. This rationale has been criticized [8], as has the ability of informants to judge the severity of symptoms on behalf of patients [21]. In the NPI-Q, the frequency score was omitted as redundant, because the severity score correlates highly with caregiver distress and the frequency scores in the original NPI [14]. Acknowledging the variability in scoring symptoms of differing frequency and severity under a domain, Connor et al. [22] recommend that the interviewer repeat the positive subquestion symptoms to the informant, asking the informant to consider these symptoms as a whole in determining the severity and frequency scores. This approach addresses the challenge of weighing symptoms of differing frequency and severity by relying on the holistic judgment of the informant [22]. In sum, the subquestions are considered to reflect the same unitary construct as the screening question, the informant condenses the subquestion symptoms into single frequency and severity scores, and these scores are then combined to arrive at the multiplication score of 1 to 12.

In the development of the NPI, a host of psychometric properties were examined [10, 12]. Although these qualities are important, they are only a subset of the features that are usually desired in scale validation [23, 24]. Previous reviews of the NPI [8, 11, 13, 25, 26] have summarized these data, taking them at face value, and solidified the notion that the psychometric data are sufficient and indicative of good reliability and validity. However, drawing inferences based only on the reported metrics may have led to biased conclusions for at least two reasons: the conditions for stricter psychometric analyses have not been met, or the useful properties come at the expense of unexamined properties through overoptimization [24]. Indeed, recent reports suggest that common scales in behavioral sciences either report insufficient psychometric data to evaluate their utility [27], or when examined more closely, have suboptimal validity evidence [24]. All scales face these challenges, but the unconventional structure and scoring method of the NPI may cause additional difficulties in validation.

Despite wide use of this scale, several knowledge gaps persist regarding the NPI. For instance, aspects of construct validity, such as correlations with other tests, are rarely mentioned. Similarly, differences between translated versions and scoring procedures should be considered. Given these cautionary observations and the aspiration that the NPI could be used as an outcome measure in CNS disorders more broadly [11], a thorough review of its psychometric basis is timely and needed. The aim of this narrative review is to examine the psychometric properties of the NPI. Instead of focusing only on what has been reported, we highlight which psychometric properties are seldom included in the translation and validation studies of the NPI. The identified gaps are subsequently discussed as opportunities for refining the NPI.

VALIDITY AND RELIABILITY OF THE NPI

FA of the NPI

Only one of the reviewed studies included an analysis of the structural validity of the subquestions. Here, it is useful to follow the terminology that these authors used, in which they distinguished between an “item score” and a “domain score” [21]. The researchers modified administration of the NPI, asking all subquestions without the screening questions, and the item score referred to the individual frequency rating of a subquestion. It is important to note that here, the score refers to a single item, not to the domain. As mentioned earlier, the domain score is typically used in NPI research and is based on the screening procedure, subquestion probing, informant judgment of the holistic frequency and severity of the symptom domain, and ultimately the product of the frequency and severity scores. This painstaking description is crucial for unraveling FA studies of the NPI, in which domain scores are used similarly to simpler scores, such as a Likert rating on a single symptom.

The NPI domain scores have been subjected to numerous FAs, in which the objective has been to derive syndromes of related symptom domains. For example, a psychotic syndrome could include delusion, hallucination, and aberrant motor behavior domain scores of the NPI. This post hoc line of research was not specified in the development of the NPI. It also has a different aim from FA, using item scores acquired via alternative scoring methods, as in the Gallo et al. study, to demonstrate whether the subquestions refer to a single, independent construct in the first instance (Fig. 2). Structural validity based on subquestion data would substantiate the inferences concerning any domain score (including zero) and further combinations of domain scores in syndrome models.

Fig. 2

Two levels of factor analyses of the Neuropsychiatric Inventory. 1) The syndrome level is where the majority of the NPI factor analytic studies has taken place. The aim of this research is to explore whether latent variables (syndrome ABC), often called a syndrome, would underlie the correlations between different domain scores (A, B, C) of the NPI. These studies implicitly assume that the domain scores can be used as useful indicators for a latent variable. This is the level of factor analytic research reviewed by Canevelli et al. [48]. 2) The subquestion level is critical for establishing structural validity, but it has not been extensively studied, indicated by the dashed lines. The aim of this research is to show that the subquestions (e.g.., A₁, A₂,  ...   A_n) address a unidimensional construct (the one suggested by the screening question), justifying their use in scoring the domain. Structural validity studies can reveal, for example, that the relationships between subquestions and the latent construct are not strong enough, that the subquestions under a single domain address more than one construct like in the study by Gallo et al., or that subquestions from a domain could reflect some other construct instead, or in addition to, the one it is intended to. To explore potential cross-loadings (e.g., C₁ to domain C and A), asking all subquestions from the informant without screening questions is required.

In a review of the FA literature related to the NPI, Canevelli et al. [48] found that none of the studies produced a unidimensional factor structure (as sometimes approximated in the validation studies reviewed in the present paper), and no two studies identified the same factor structure. The authors also reported that an exhaustive factor model was not established in almost 30% of NPI FAs (i.e., some domains often failed to associate with any factor, commonly aberrant motor behavior, appetite disturbances, and sleep disturbances). The NPI domain syndromes are also temporally unstable [49]. These findings are important for two reasons. First, they call into question the utility of estimating the internal consistency of the entire NPI, as was done in some studies reviewed here. It is theoretically unlikely and not even desired that the combination of all domains reflects a unidimensional structure. Second, in addition to other factors, such as sample size issues and study population variability, the heterogeneity of the syndrome structures may stem from domain scores being suboptimal indicators for a latent variable, given the assumption made with little empirical support that they reflect only one construct.

DISCUSSION

We examined the psychometric properties of the NPI in a framework including those properties pertinent for construct validation and health-related outcome measurement in general. In agreement with previous reviews, we found that aspects of reliability are major strengths of the NPI, perhaps explaining the measure’s “gold standard” status [26]. The inter-rater and test-retest reliabilities are particularly key in clinical trial settings. Although flexibility and relatively quick administration were not properties explicitly included in the review, they also are major strengths of the NPI. Thus, it is not surprising that the NPI has been applied in several settings where a rating scale of similar scope has been lacking.

The main issue, however, is that the reported properties cover only some of the generally used psychometric properties, representing what might be necessary but insufficient reliability and validity evidence for the NPI. The psychometric data appear to have significant gaps, at least in part because the small sample sizes in the studies precluded more comprehensive analyses. Regarding construct validity, we find it concerning that only one study [21] examined structural validity with the NPI subquestions. Those authors suggested that more than half of the subquestions currently considered to reflect a single domain could represent two or three constructs. These findings should be considered preliminary, however, because this group used exploratory methods and a small sample size relative to the number of items. If some domains are indeed multidimensional, this factor could limit their utility as validity benchmarks for other scales (and vice versa), given the lack of clarity about the contributions these different constructs make to a correlation [50, 51]. Discriminant validity, or the extent to which theoretically distinct constructs do not correlate with one another, was mainly investigated regarding global cognition. This assessment was important for demonstrating that the NPI cannot be reduced to a measure of dementia severity. However, it would also be pertinent to show that theoretically distinct neuropsychiatric symptom constructs do not correlate in undesired ways. The existing FA literature on NPI domain scores addresses a different aim altogether, namely one of exploring possible higher-order syndromes, and should not be confused with construct validation in the traditional sense.

Coefficient alpha was used in the studies as an estimate of internal consistency of either the entire NPI or domains of it. The interpretability of coefficient alpha hinges on several assumptions, the most pertinent being that the items form a unidimensional latent variable model [23, 52, 53]. Most studies did not include examination of this property, which would have required larger sample sizes and alternative scoring methods [20, 23]. As the authors of one of the reviewed studies noted [42], using coefficient alpha for the entire NPI seems counterintuitive because the domains should not form an internally consistent aggregate; rather, the aim is to assess 10 to 12 distinct symptom domains.

Data for a subset of patients enrolled in the studies were analyzed in the test-retest and inter-rater reliability analyses. Although the reporting was transparent regarding how many patients were included in these analyses, the number of patients who were screen-positive for the assessed domains in these subsets was less obvious. For example, the original NPI study [10] had no patients with euphoria in the test-retest analysis, and a correlation of 1.0 between hallucination severity scores at the first and second evaluations suggests that few patients were assessed in this analysis.

None of the reviewed studies included assessment of measurement error. In our experience, this finding is typical in the scale validation literature, and its significance should be interpreted in the context of the intended use of the scale. Because the NPI scoring method was chosen with clinical trials in mind [12], information about measurement error could aid in interpreting trial results where the clinical significance of changes in outcome measures is of interest. Sample sizes in trials of NPS with dementia are increasingly large—some authors even suggest that they are too large [54]—and thus, sufficiently powered to detect even minute changes in the primary outcome, often some variant of the NPI. Thus, it could be of practical importance to assess the lower limit of changes that the NPI can detect and the magnitude of change that can be considered clinically important [55]. Additional challenges may arise from the non-parametric distribution of domain scores.

Taken together, the reviewed studies suggest that the validation and translation of the NPI seem to have followed homogenous procedures in the last two decades. Although this consistency has the benefit of allowing estimation of stability of some properties across studies and cultures, it has led to systematic neglect of some aspects of validity. Below, we outline some recommendations for future research.

Regardless of which variant of the NPI is going to be further refined, the sample sizes need to be larger for FA methods in structural validation. Some authors have recommended rules of thumb regarding minimum sample sizes in FA, such as 200 to 300 [23]. Others have refrained from such suggestions, however, because properties such as the communality of the items can lead to reliable FAs in smaller groups [56]. Nonetheless, this aspect can also be considered a question of research priorities. To advance the study of AD in large collaborative projects, neuroimaging can be performed on thousands of unaffected controls, at-risk individuals, and those with a diagnosis of a neurocognitive disorder, whereas collecting item-level data on NPS scales is possible for a fraction of the cost. Gathering data for further validation of the NPI can be achieved alongside many such pioneering research projects.

The skip-question format of the NPI poses specific problems for assessing construct validity at the level of individual items. In the end, the domain scores are a result of the informant ascribing some numbers to the whole of subquestions in the domain, so that analyzing these subquestions should be the basis for any future NPI validation projects. Ensuring this aim, however, requires an alternative scoring method in which informants are asked all subquestions, not only those tied to positive screening answers. The issue could be further explored in a test-retest setting assessing whether informants answer similarly to the standard NPI screening and related subquestions and the alternative version with subquestions only. This line of inquiry could delineate the limitations of having statistical contingencies between subquestions and screening questions. A perhaps more feasible approach would be further development of the NPI-C in which all items are queried. However, the extensive number of items on the NPI-C poses statistical and practical challenges.

Perhaps because of the relatively limited theoretical literature available, researchers often have not explicitly stated their hypotheses regarding convergent and discriminant validity. Statistical significance between two scales should not be a deciding factor per se, and the magnitude of correlations is what is important. The COSMIN guidelines suggest some generic hypotheses regarding systematic reviews of the convergent and discriminant validity of patient-reported outcomes: for example, scales measuring the same construct should correlate at 0.50 or higher, whereas those measuring distinct constructs should correlate under 0.30 [28]. With two decades of validating, developing, and translating the NPI, the literature is sufficient to allow formulation of a priori hypotheses regarding these properties.

This review has focused mostly on the traditional latent variable modeling approach to psychometrics. Rather recently, however, it has been suggested that modeling scale items as networks of conditional relationships could resonate with the nature of the constructs more accurately [57, 58]. The core idea of network models is that variables are connected to one another through conditional relationships, so that only relationships that have taken into account the shared variance of all symptoms are included. The result is a sparse network, depicting unique relationships among items. Networks correspond with the notion that symptoms of psychiatric disorders interact so that the presence of one symptom is likely to increase the possibility of another symptom. Regarding NPS, network methods have already been used in a few studies [59, 60]

Even more recent developments concern the so-called hybrid models, which combine both latent variable and network properties, such as using the residuals of latent variables to estimate networks [61]. A clinical interpretation of these models could be as follows: the latent variables represent a change in brain structure or metabolism associated with the disease, whereas the networks represent the remaining psychological and behavioral phenomena unaccounted for by the latent variable and their relationships with one another. This interpretation is compatible with etiological accounts of NPS [62].

The NPI literature can be confusing because the domain scores are often called “symptoms” or “items.” Given the lack of clarity about which and how many symptoms are referenced in each situation and that the scores result from a multiphasic process, we suggest using the term “domain scores.” The subquestions of the NPI in turn can be considered to reflect symptoms through item scores. These distinctions preserve the separation between constructs and measures [7], and attach the NPI to psychiatric and neurological scales more generally, in which symptoms often are more narrowly defined.

Finally, we have sought to include in this review some actionable suggestions for using and developing future NPS scales. Developers would benefit from proceeding according to the substantive, structural, and external phases outlined in Table 2. These phases build on one another, so that having data on a latter property is difficult to interpret in the absence of previous properties. Although this compilation of information is relatively recent [27], the literature itself is not. Indeed, the classical methods presented in the construct validity literature have been used successfully in NPS scale development [6].

CONCLUSIONS

The NPI is considered the gold standard for NPS scales and has been used to deepen understanding of NPS in both research and clinical contexts in the past two decades. Based on our review, the reliability of the NPI is its greatest psychometric strength. However, because the screening and scoring rely on information related to individual subquestions, we were concerned to find that only one study examined the structural validity of the subquestions in the NPI, whereas other investigations of construct validity were limited in scope. These findings point to uncertainty regarding whether the current NPI formulation reflects the intended constructs. Future research should target addressing these gaps and aiming for conceptual clarity.