Data was collected at two points. At Time 1 (August-October 2020), 1,660 participants were recruited from Amazon Mechanical Turk (MTurk), an international platform accessible to anyone with internet and English proficiency. No geographical restrictions were applied, resulting in a naturalistic, international sample primarily from the United States (n = 1,136, 68.4%), India (n = 209, 12.6%), Brazil (n = 117, 7.0%), Canada (n = 66, 4.0%), Italy (n = 38, 2.3%), and 94 participants (5.6%) from 29 other countries. Participants were compensated $0.90 USD. The study was approved by the Peking University IRB (No.: 2020-04-18).

Eligibility criteria included informed consent, correct responses to an embedded validity check (judging the size of a common animal relative to a human; 25). Individuals with a confirmed COVID-19 diagnosis were excluded. This was necessary to ensure the scale specifically measured the fear of contracting the virus, rather than capturing the distinct psychological responses, such as illness-related distress or anxiety about recovery, that follow a confirmed infection, which would confound the construct of interest. The final Time 1 sample comprised 821 men, 830 women, and 9 individuals of other genders, with a mean age of 36.94 years (SD = 11.97).

Time 2 data were collected one month after Time 1, in November 2020. Following the same inclusion criteria, 355 participants were retained (181 men and 174 women), with a mean age of 39.21 years (SD = 12.69). Due to the high attrition rate (21.4%), a logistic regression analysis was conducted to examine potential attrition bias between participants who completed both time points (n = 355) and those who completed only the baseline assessment (n = 1,305). The analysis revealed that younger age (OR = 0.97, p < 0.001), higher baseline health anxiety (OR = 1.06, p < 0.001), and US residency (OR = 1.69, p < 0.001) were significant predictors of attrition. While this indicates some systematic patterns in dropout, the follow-up sample remains adequate for test-retest reliability analysis, though estimates should be interpreted with the understanding that they may be slightly conservative due to the selective attrition of more anxious participants.

The participant’s characteristics are shown in Supplementary Table S1 in the supplementary material.

All analyses were conducted using SPSS 25, Mplus 8.3, and R. Parallel analysis, exploratory factor analysis (EFA), assessments of convergent and discriminant validity, internal consistency, test-retest reliability, Receiver Operating Characteristic (ROC) analysis, and Pearson correlations were performed in SPSS. Confirmatory factor analysis (CFA) was conducted in Mplus. Item Response Theory (IRT) analysis was performed in R using the ltm package.

To determine the factor structure of the PHAID, the full Time 1 sample (n = 1,660) was randomly split into two equal subsamples (n = 830 each). The first subsample was used for EFA. Prior to EFA, a parallel analysis was conducted to determine the number of factors to retain, which involved comparing actual data eigenvalues with the 95th percentile of eigenvalues from 1,000 simulated random datasets (29, 30). Factors were retained if the i-th eigenvalue from the data exceeded the corresponding random value (31, 32). EFA was performed using principal axis factoring with Promax (oblique) rotation (33). Items were iteratively removed if they had a factor loading below 0.35 or a cross-loading difference less than 0.2 between primary and secondary factors.

A series of CFAs were conducted in the second subsample to evaluate and compare three competing models: (1) the original SHAI factor structure, (2) the two-factor structure derived from EFA, and (3) a modified version of the EFA-based model. For the EFA-derived model, items with loadings below 0.50 were excluded. In line with methodological recommendations to avoid overfitting (34), model modifications were made only for a limited number of item pairs that demonstrated both high modification indices (MI > 10, 35) and a strong, a priori substantive rationale based on close semantic content. The model fit was evaluated against conventional standards (36). An adequate model fit was indicated by 1 < χ²/df < 3, comparative fit index (CFI) and Tucker Lewis index (TLI) values ≥ 0.90, root-mean-square-error of approximation (RMSEA) ≤ 0.08 (more liberal) or ≤ 0.05 (stricter), and standardized root-mean-square residual (SRMR) value ≤ 0.05.

Scale validity and reliability were assessed via multiple methods. Convergent validity was evaluated by correlating the PHAID total score with the HADS anxiety subscale; discriminant validity was assessed via correlation with the HADS depression subscale. Internal consistency was estimated using Cronbach’s alpha for the total scale and subscales. Test-retest reliability was assessed by correlating PHAID scores at Time 1 and at the one-month follow-up (Time 2).

To establish a clinical cutoff for the PHAID, ROC analysis was conducted using the HADS anxiety subscale and its established cutoff (28) as the reference standard. Regression analyses were performed to examine the relationship between PHAID total score and hand washing total score. Multiple models (linear, quadratic, cubic, power, logarithmic, and exponential) were tested, and model fit was compared using adjusted R². Notably, the regression analysis examining the relationship between PHAID scores and handwashing was conducted for exploratory purposes to investigate the form of this relationship, rather than to build a predictive model.

Item Response Theory (IRT) analysis was conducted to inform the development of a short form (PHAID-S). A Graded Response Model (GRM, 37) was fitted to the data. For each item, two parameters were estimated: a discrimination parameter (a) and three difficulty parameters (b1, b2, b3), corresponding to the thresholds between the four ordered response categories (0-3). Test Information Curves (TIC) and Item Information Curves (IIC) were also calculated to evaluate the measurement precision of the scale and its individual items across the latent trait of health anxiety. A ROC analysis was also conducted to establish a cutoff of PHAID-S.

The data were suitable for EFA, as indicated by a significant Bartlett’s test of sphericity (p <.001) and a high Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy (0.945). Parallel analysis supported the extraction of two factors, with the actual eigenvalues for the first two factors exceeding the 95th percentile eigenvalues from simulated random data (see Supplementary Table S2 in supplementary material). Item 4 was removed due to a factor loading below 0.30, and Items 17, 15, and 3 were subsequently removed due to excessive cross-loadings. The final solution, presented in Table 1, consisted of two factors that collectively accounted for 61.93% of the total variance. All retained items demonstrated strong loadings, ranging from 0.51 to 0.89. Negative cross-loadings are a known artifact of oblique rotation and should be interpreted in the context of the strong primary loadings and the inter-factor correlation. Based on the content of the items, Factor 1 was labeled Catastrophic Thinking, and Factor 2 was labeled Infection Worries.

The original SHAI factor structure showed all standardized factor loadings were statistically significant (p <.01 for item 4; p <.001 for all others); however, model fit was poor: χ²/df = 10.04, p <.001, CFI = .843, TLI = .821, RMSEA = .104, SRMR = .061. The two-factor EFA-derived structure required the removal of Items 8 and 16 due to loadings below 0.50. All remaining items loaded significantly (p <.001), and model fit improved substantially: χ²/df = 7.94, p <.001, CFI = .941, TLI = .927, RMSEA = .091, SRMR = .046. Moreover, error covariances were added for three pairs: item 12 and 9 (self-diagnosis), item 13 and 10 (misinterpretation of bodily sensations), and item 5 and 2 (core fear of infection). The modified model (see Figure 1) demonstrated excellent fit: χ²/df = 6.53, p <.001, CFI = .956, TLI = .942, RMSEA = .082, SRMR = .040.

As presented in Table 2, Pearson correlation analyses revealed strong and statistically significant positive associations (all p <.001). The correlations between the total score and the two subscales were very high, ranging from 0.619 to 0.975. Similarly, all items showed strong correlations with the total score (r = 0.634 to 0.821), and with their respective subscales (r = 0.756 to 0.835 for Catastrophic Thinking; r = 0.874 to 0.888 for Infection Worries). The Cronbach’s alpha and McDonald’s omega were excellent: α = 0.931 and ω = .947 for the total scale, α = 0.928 and ω = .939 for the Catastrophic Thinking subscale, and α = 0.851 and ω = .856 for the Infection Worries subscale.

One-month test–retest correlations were high for the total score (r = 0.826, p <.001), as well as for the Catastrophic Thinking (r = 0.780, p <.001) and Infection Worries (r = 0.821, p <.001) subscales.

The PHAID total score and its subscales showed significant, moderate-to-strong positive correlations with the HADS anxiety subscale, supporting convergent validity (PHAID total: r = 0.622, Catastrophic Thinking: r = 0.617, Infection Worries: r = 0.454; all ps <.01). In contrast, correlations with the HADS depression subscale were weaker (PHAID total: r = 0.263, Catastrophic Thinking: r = 0.267, Infection Worries: r = 0.175; all ps <.01). The association between the PHAID total score and anxiety was significantly greater than with depression (Z_H = 15.45, p <.001), supporting the convergent and discriminant validity of the PHAID.

The ROC curve for the PHAID, using the HADS anxiety subscale (cutoff ≥ 8) as the criterion, yielded an area under the curve (AUC) of 0.83 (p <.001, 95% CI [0.81, 0.85]; see Figure 2). The optimal cutoff score was 24 on the PHAID total, corresponding to the maximum Youden’s index of 0.550. At this threshold, sensitivity was 0.760 and specificity was 0.790. Sensitivity, specificity, and Youden’s index values for additional cutoff scores are presented in Supplementary Table S3 in supplementary material.

The relationship between PHAID total scores and hand washing total score was best described by a cubic regression model (adjusted R² = .02), as shown in Figure 3. The cubic model demonstrated a significant overall fit, F(3, 1656) = 13.81, p <.001. All predictors were statistically significant: linear term (b = 1.39, t = 5.91, p <.001), quadratic term (b = −0.05, t = −5.60, p <.001), and cubic term (b = 0.0006, t = 5.30, p <.001). Compared to the cubic model, the quadratic, power, logarithmic, exponential, and linear models accounted for less variance (adjusted R² = .007,.005,.005,.002, and.002, respectively). The cubic regression curve indicated that hand washing total score increased with PHAID scores up to a total score of 24, plateaued between scores of 24 and 36, and increased again at higher scores. It is important to note that this regression analysis was exploratory in nature. The low adjusted R² confirms that health anxiety is just one of many factors influencing handwashing, and the PHAID is not intended to predict this behavior. The key finding is the significant non-linear pattern itself, which provides preliminary construct validity by aligning with the theoretical model of adaptive versus maladaptive anxiety.

All items showed discrimination parameters greater than 2.87. For Catastrophic Thinking, Items 9 (a = 3.21, b1 = -0.08, b2 = 0.74, b3 = 1.63), 13 (a = 2.89, b1 = -0.39, b2 = 0.66, b3 = 1.72), and 7 (a = 2.87, b1 = -0.43, b2 = 0.65, b3 = 1.63) had the highest discrimination and provided the most information according to the Item Information Curves. For Infection Worry, Items 5 (a = 3.68, b1 = -1.39, b2 = -0.03, b3 = 0.98) and 1 (a = 3.38, b1 = -1.67, b2 = -0.16, b3 = 0.96) had the highest discrimination and information values. These five items were selected for the PHAID-S. The 5 selected items adequately represent the core content of the original factors: the Catastrophic Thinking factor is represented by three high-discrimination items covering difficulty disengaging from thoughts, self-diagnosis, and persistent rumination, while the Infection Worries factor is represented by its two most central items (general worry and fear).

Confirmatory factor analysis was conducted to evaluate the two-factor structure of the 5-item PHAID-S. The model demonstrated an excellent fit to the data: χ²/df = 9.55, p <.001, CFI = .981, TLI = .953, RMSEA = .102, SRMR = .026. All standardized factor loadings were statistically significant (p <.001) and substantial, ranging from 0.79 to 0.81 for the Catastrophic Thinking factor (Items 7, 9, 13) and from 0.80 to 0.81 for the Infection Worries factor (Items 2, 5), as shown in Figure 4.

Test Information Curves showed maximum measurement precision for Catastrophic Thinking in the latent trait range of -0.5 to 1.5, and for Infection Worry in the range of -1.5 to 1.0 for both PHAID and PHAID-S, indicating that the PHAID-S covers a similar range of the latent trait as the full PHAID, although with a lower peak information. The short form demonstrated internal consistency of Cronbach’s α = 0.85. Detailed parameter estimates and information curves are presented in Supplementary Table S4 and Supplementary Figures S1–S4 in the supplementary materials.

For the PHAID-S, ROC analysis (criterion: HADS anxiety subscale ≥ 8) yielded an AUC of 0.81 (p <.001, 95% CI [0.81, 0.85]; see Supplementary Figure S3). The optimal cutoff score for the PHAID-S was 11, corresponding to the maximum Youden’s index of 0.470. However, considering the utility of a short form for efficient screening, a cutoff of 12 may also be considered, as it offers substantially higher specificity (0.845 vs. 0.744) for a moderate decrease in sensitivity (0.653 vs. 0.727), which could be preferable in settings where minimizing false positives is a priority. The cutoff of 11 is recommended for maximum sensitivity. Additional cutoff statistics are provided in Supplementary Table S5 in the supplementary material.

Several limitations of this study should be acknowledged. First, participants were recruited via Amazon Mechanical Turk (MTurk). While this platform enabled the collection of a geographically diverse sample, the participant pool may overrepresent. White, middle-class individuals with higher education levels and computer literacy. Moreover, while our sample included participants from multiple countries, its online nature and Western predominance warrant caution regarding cultural generalizability. Second, the data were collected during the early stages of the COVID-19 pandemic (August-November 2020). The psychological impact and associated health anxiety may have evolved as the pandemic progressed, and public health policies varied across regions. Consequently, the results might underestimate the potential severity of such anxiety in later phases of the pandemic or in regions experiencing severe outbreaks. Third, the test-retest reliability assessment was limited by the sample size, and attrition analysis revealed that the dropout was not completely random but was influenced by age, nationality, and baseline health anxiety scores. Future studies should collect more unbiased longitudinal data to further examine the scale’s temporal stability. Fourth, while the PHAID demonstrates good convergent validity with general anxiety (HADS-A) and discriminant validity from depression (HADS-D), its direct convergent validity with a measure of health anxiety itself (e.g., the full Short Health Anxiety Inventory) was not assessed in this study. Fifth, the interpretation of the significant cubic relationship with handwashing is exploratory and requires validation in studies specifically designed to test this behavioral pattern.

Future research can focus on several key areas to extend the utility of the PHAID. First, validating the PHAID and its short form across more diverse populations—including varied cultural, socioeconomic, and clinical groups—would help establish its broader applicability and stability across different contexts. Second, future studies could explore how factors such as media exposure, health literacy, and pre-existing vulnerabilities influence PHAID. Third, future studies could investigate how varying levels of PHAID scores correlate with a spectrum of health behaviors, from appropriate vigilance to excessive actions. Fourth, the clinical and public health utility of the PHAID cutoff scores could be evaluated in real-world settings to assess their effectiveness in identifying individuals who might benefit from targeted support.