A concern reported in the literature is the burden on patients of answering all survey questions at one time [6-9]. More than 3.6 million patients completed mail-in surveys from the UK NHS (600,000 patients). The US survey (3.0 million respondents) required approximately 6 hours of time per month and cost approximately US $70 per month [3] to examine people’s health service experiences [10]. However, the UK NHS 70-item questionnaire is significantly longer than the US HCAHPS 25-item survey [11,12]. To reduce patient burden, it is first necessary to shorten the item length of the UK NHS inpatient questionnaire to increase response rates without compromising its assessment reliability [2,13].

Many studies [6-9] have reported that item response theory (IRT)-based computer adaptive testing (CAT) has the advantages of both long-form and short-form questionnaires [14-16] in precision and efficiency. Since many patients (or their guardians) already own mobile phones, which they are comfortable using, it makes sense to use them in hospitals and for hospital surveys. At this time, no studies have been published reporting online CAT via mobile phones in medical fields.

However, many skip items (see Multimedia Appendices 1 and 2) exist in the UK NHS 70-item questionnaire, which can be confusing and may perplex researchers on CAT implementation. Thus, our second aim was to tackle the problem of skip items in the UK NHS questionnaire and to implement the online CAT.

Many researchers have contributed to the dichotomous [6] and polytomous [7-9] formats used by CAT. The UK NHS questionnaire comprises items with different categories (eg, 3 and 6 categories for Items 40 and 41; see Multimedia Appendix 1). It is suited for applying the Rasch partial credit model (PCM), that is, items with a different number of responses and with an equal discrimination parameter [17], or the generalized partial credit model, that is, items with a different number of responses and with unequal discrimination parameters [18], if those items form a unidimensional construct. None was jointly available for a comparison of precision and efficiency differences of CAT estimation with the aforementioned methods commonly used in literature, such as PCM, answering all items (AAI), and the randomized selection method (RSM).

Further, as mobile phones have become ubiquitous in the health care setting [19], it is important to offer an alternative online Rasch PCM-CAT assessment to gather hospitalization experience feedback from patients. We propose access to the questionnaire using a quick response (QR) code via mobile phone.

The aims of the current study were to investigate whether CAT can (1) be efficient with item reduction and (2) be used with QR codes used for mobile phones.

The UK NHS 70-item questionnaire regarding patient experience was downloaded from the NHS official website [11]. The item and its threshold difficulties (lower summation scores for an item imply that it was more difficult for examinees to respond) were roughly determined by hand computation according to the key findings report for the 2008 inpatient survey [20]. We simulated an interactive metric of 1000 persons (following a normal distribution [~N(0,1)], called true scores) and 70 items (estimated with aforementioned item difficulties) using the Rasch PCM model [17,21]. Nine items originally designed to automatically select different paths were set with different probabilities by the authors. The remaining 61 items were allocated different weighted scores (see Multimedia Appendix 1). A set of 24 items (ie, regarding sections of the ward, doctors, nurses, patient care, and treatment) was extracted from the UK NHS 70-item questionnaire to be the CAT item pool (see Multimedia Appendix 2). We assumed that the set of 24 items is unidimensional based on the report from the previous study paper [13]. Because these 9 conditional selection path items make CAT difficult to design for a computer, they were excluded from the CAT item pool. Multimedia Appendix 3 shows the file layout and fields we designed for use with the datasets.

The Rasch model, named after Georg Rasch [22], is a psychometric model for analyzing categorical data as a mathematical function of the trade-off interaction between (1) the respondent’s latent trait (eg, hospitalization perception level in this study) and (2) the item difficulties. The dichotomous Rasch model and its extensions (eg, family models: rating scale model [23], PCM [17]) are successfully used in other areas, including the health profession [24] and market research [25], because of their general applicability [26].

The study data need to meet the following criteria to fit the Rasch model: the infit and outfit mean square errors (MNSQ) of all items are ˂1.5 for unidimensionality and ˃0.5 for local independence [27]. Simulation data were generated fitting to the Rasch PCM model [21].

Three scenarios were designed to compare their efficiency and accuracy on the UK NHS 70-item questionnaire: (1) the AAI (answering all items on those 24 items), (2) the RSM (randomized selection method to draw 12 items), and (3) the CAT (at least 5 items and stop at person reliability of 0.80) responding to the 24-item pool.

We applied CAT stop rules, such as when person reliability reaches 0.80 (=[1 − SEM_pi]×[1 − SEM_pi], where SEM_pi=person standard error of measurement on item i=1/variance_pi=1/information_pi), and when the last 5 average consecutive person estimation change is less than 0.05 after the minimum necessary completed number of items is ≥5.

In addition, we ran an author-made VBA (Visual Basic for Applications) module in Microsoft Excel to conduct a simulation study (see Multimedia Appendix 4). Another Web-based CAT was programmed for use on mobile phones. The maximum likelihood estimation algorithm [28] (see Multimedia Appendix 4) was used to (1) estimate person measures on the three scenarios, (2) compute correlation coefficients between estimated person measures among the three scenarios and the original true scores to verify CAT accuracy, and (3) analyze CAT efficiency of item length shortened by CAT compared with the other two scenarios (ie, AAI and RSM) (see Figure 1).

The set of 24 items (see Table 1) was taken as unidimensional due to simulation data fitting the Rasch model’s requirement with values of infit and outfit MNSQ between 0.5 and 1.5 [21]. Each item has its own threshold difficulties (see Multimedia Appendix 4).

Table 2 indicates that the CAT relates the true scores (r=.97 in column 2) and the AAI (r=.97 in columns 3 and 4) to a high association, indicating that the CAT earns an equivalent accuracy compared to the AAI and a higher accuracy than the RSM (in column 5 of the estimation section). The summation scores have a higher correlation (r=.98) to the within counterparts (eg, summation RSM scores vs estimated RSM logit scores) and a slightly lower correlation (r=.92-.97) to the between counterparts (eg, summation RSM scores vs estimated AAI or CAT logit scores), implying that the raw summation scores have a high correlation (r=.98) with the estimated logit scores shown in the last 4 columns of Table 2. The bottom row of Table 2 shows that the CAT earns the shortest item length, indicating the CAT has advantages in efficiency over AAI and RSM.

By scanning the QR code (see Figure 2), the CAT icon appears on the patient’s mobile phone. The mobile CAT survey procedure was demonstrated item-by-item in action (see Figure 3). Person fit (ie, infit and outfit MNSQ) statistics showed the respondent behaviors. Person theta is the provisional ability estimated by the CAT module.

The standard error in Figure 3 was generated by the following formula (see Multimedia Appendix 5): 1/√(Σ variance(i)), where i refers to the CAT finished items responded to by a person [30]. In addition, the residual (resi) in Figure 3 was the average of the last five change differences between the pre-and post-estimated abilities on each CAT step. CAT will stop if the residual value is less than 0.05. “Corr” refers to the correlation coefficient between the CAT estimated measures and its step series numbers using the last five estimated theta (=person measure) values. The flatter the theta trend, the higher the probability that the person measure is convergent with a final estimation.

We verified that computer adaptive testing can be (1) efficient with questionnaire item reduction and (2) used with QR codes on mobile phones.

The CAT item pool was designed using Items 15-38 of the UK NHS 70-item questionnaire. We found that CAT can be more efficient for answering questions than both AAI and RSM without compromising its measurement accuracy, which is consistent with previous studies [6-9]. Our online CAT inpatient survey for gathering satisfaction responses from patients was accessed by scanning a QR code on a mobile phone, which has never been demonstrated previously.

Many studies have discussed patient perceptions about hospitals and the benefit of listening to other patient experiences when choosing a hospital [31,32]. There has been a rapid increase in the number of websites that allow patients to rate their hospitals [33,34]. Almost all health care providers have been explicitly required to conduct surveys of their patients’ health care experiences. However, those surveys often use an individual item-by-item approach to disclose patient views on hospital service quality, which does not provide hospital staff with information to make further improvements without considering the overall hospital performance [13].

We demonstrated that an NHS inpatient experience questionnaire with shortened items can be used with an IRT-based CAT technique without compromising its measurement accuracy. Using a CAT approach with such complex question structure jointly with item pools and conditional path skip items is rare. Our online CAT module used by scanning a QR code on a mobile phone can be extended to many dimensions simultaneously in a survey. For example, the Clinical Dementia Rating scale [35] used in patients with dementia consists of six domains. We could design a module using CAT through several procedures in a common questionnaire in the future.

Hospital staff must consider both the efficiency and utility of assessment for the selection of the CAT items [36]. The traditional survey collects all feedback from patients through particular sets of questions to assess what causes patient difficulty or dissatisfaction. The assessment results help hospital managers determine where improvements can be made [36]. We can use the Rasch simulation technique to overcome the problem in questionnaires of unanswered items (ie, which do not provide hospital staff with information to make further improvement). This Rasch simulation technique [21] can be used to fill in the expected responses to those unanswered CAT items according to the final person theta (ability) and the specified item difficulties. Thus, the CAT can provide efficient assessments and the full information needed to make improvements.

Furthermore, the person outfit mean square in CAT is also saved in our database (see Multimedia Appendix 3). An outfit mean square of 2.0 or greater for a patient indicates a possibly aberrant response pattern [37], such as cheating, careless responding, lucky guessing, creative responding, or random responding [38], which makes it hard to reveal valuable information using the traditional survey method.

Six limitations of this study are addressed. First, the study was based on the assumption of unidimensionality across those 24 CAT items. Although several articles have supported the notion that the UK NHS 70-item questionnaire can construct a one-dimension domain [13,31], those items cannot be generalized to the 24 CAT items used in different countries or by different groups. Future studies should further verify those 24 items to make the CAT module valid and feasible in health care practice.

Second, the first CAT item was selected from a randomized item pool. The CAT selection rule for the first item can be redesigned referring to the previously completed items and inferring a provisional theta (ie, person measure) to select the first item with the maximum information (ie, variance) in the item pool so that the questionnaire length could be shorter (see Multimedia Appendix 6).

Third, only one CAT module was designed in the NHS inpatient questionnaire due to the conditional selection path skip items that existed in non-CAT items. Future studies are recommended to overcome this barrier and to design a CAT-by-CAT approach in the questionnaire so as to reduce more item length in a questionnaire.

Fourth, we have not discussed the issue of participation options using traditional postal mail or email. Because not all patients possess a mobile phone, specifically a smartphone, and 3G/4G WiFi communication, all options (mail or email) must be offered to patients when invited to participate in the survey. Readers have found the email option useful to answer questions either by connecting to the Web, or by scanning a QR code on a mobile phone (eg, Figure 2) if applying the CAT demonstrated in this study. Future studies are needed to further explore and improve the processes of the CAT survey.

Fifth, we conducted a simulation of 1000 patients’ true scores followed by a standard normal distribution. This might contradict the general experience of satisfaction surveys with ceiling effects that impede standard normal distribution. Future studies are needed to sample from a negatively skewed population to further verify whether the CAT can be efficient on item reduction over AAI and RSM.

Sixth, the unidimensionality of the 24 CAT items may be questioned given the different realms of hospital ward, doctors, nurses, patient care, and treatment. It might be implausible to assume that these 24 items are unidimensional. Future studies are required to further investigate the issue.

With advances in technology, we can now offer patients alternative ways via mobile phones to gather their feedback on hospitalization satisfaction. The online CAT can reduce the number of survey items for patients to respond to, as well as be accessed via mobile phone using a QR code.