To investigate our objective, we developed a survey aimed at measuring clinicians’ attitudes toward diagnosis and medication of ADHD as latent constructs. A team consisting of two clinical psychologists (AM, IL) and two psychiatrists with extensive clinical experience with ADHD patients (IB, AH) designed the questionnaire. To further improve validity and reliability of the items, five external clinicians with special interest in ADHD tested the survey during its development and were interviewed to collect feedback on points of improvement. In addition, the survey was piloted in one clinic.

The survey was developed as a one-time, self-administered, web-based questionnaire. Questions were closed-ended with four or six Likert-scaled, forced choice options (i.e., no neutral response possible). Items were developed assuming variation from restrictive to liberal attitudes toward ADHD diagnosis and medication.

Comment boxes for optional supplementary comments were provided for each item. All except for two background questions were possible to skip. To maximize response rate, we aimed for a survey that was quick to complete (< 10 min). The survey also included some items as part of a broader investigation of ADHD practice in the clinics that are not relevant to the objective in this article and thus not presented here. The complete, translated survey can be found in Appendix A.

To explore variation in clinician attitudes, eight items were used in analyses. To ease communication of results, all items have been given labels that are indicated in italics throughout the text. Three survey items covered background information about respondents (profession/educational background, work experience, and frequency of contact with ADHD patients, Table 1). Three items involved attitudes regarding diagnosis of ADHD (certainty when diagnosing, hypothetical prevalence in an ideal world, Table 2; await making a diagnostic decision, Table 3). One item considered treatment of ADHD (statements about medication, Table 4), and one item was originally intended to cover both diagnosis and medication (over/undertreatment in Norway today, Table 2). Two of these items (await, medication) consisted of a question followed by several statements (“sub-items”) to which respondents indicated to which extent they agreed. See Tables 1, 2, 3 and 4 for wording and options.

We aimed to include the whole population of clinicians, including all healthcare professionals of various educational backgrounds, currently involved in diagnosing and treating ADHD in the Norwegian CAMHS outpatient clinics. This represents most clinicians working with ADHD among children and adolescents in Norway, as the private health sector in child psychiatry is negligible.

As there is no central registry of employees in the 88 CAMHS outpatient clinics in Norway, we approached the heads of all the clinics and asked them to forward the survey to relevant employees in their unit. Each clinic received a unique link, enabling us to give feedback to the clinics on their average scores compared to national data (provided a response rate of > 50%).

The survey was accessible from September to December 2020. A reminder was sent to all heads of clinics two weeks after initial distribution. Clinics that still had not returned any responses a few weeks after the reminder received personalized follow-up emails or a call.

Answers to survey items concerning attitudes were scored along the hypothesized dimension with low scores representing restrictive and high scores liberal attitudes. To estimate response rate, we used the coverage of clinicians per 1,000 children in some catchment areas (made available to us by coauthor IB), which should be roughly generalizable to other regions.

Structural equation modeling (SEM) was applied to estimate and examine the relationships between our hypothesized attitude variables. A two-factor confirmatory factor analysis (CFA) was used to estimate the latent constructs of clinicians’ attitude toward diagnosis (DA) and attitude toward medication (MA).

Three survey items concerning attitude toward diagnosis were excluded from the final DA model, one due to high amount of missing (D8/other; Table 3) and two because of low factor loadings (certainty: \(\lambda\) = 0.12; ideal: \(\lambda\)= 0.11; Table 2). DA was thus estimated by seven sub-items of the item await (D1–D7; Table 3). Residuals were allowed to correlate between some indicators measuring highly related concepts: both D2/trauma and D3/health problems in the family could also be characterized as D1/psychosocial challenges; and D6 and D7 both refer to neurodevelopmental conditions (Table 3).

MA was estimated by seven sub-items of the item medication (M1–M9; Table4). Two sub-items were excluded due to low factor loadings (M2: \(\lambda\) = 0.18; M4: \(\lambda\) = 0.24). Residual correlations were allowed between M3 and M5 (both relating to concerns over side effects).

The item over/undertreatment (Table 2) was not included in the SEM analysis as its wording in terms of both question and options made it unclear where to fit in the model.

Standardized factor scores for DA and MA were extracted from the two-factor CFA model. The extent of variation in attitudes attributable to profession and clinic level was examined by intraclass correlation coefficients (ICC) in variance-components models, based on factor scores extracted from the factor models. Separate partially latent structural regression models were used to examine associations between experience and frequency (as single exogenous indicators) with DA and MA.

Models were estimated using maximum likelihood with missing values (MLMV) (missing 10% in final models). Model fit was assessed by \(^\) tests, comparative fit index (CFI), and root mean squared error of approximation (RMSEA). An insignificant \(^,\) CFI > 0.90, and RMSEA < 0.08 indicate a good model fit [28]. Focal strains in the solution were assessed with modification indices. As robustness checks, models were additionally estimated using MLMV with standard errors clustered by clinic; ML with ordinal family and logit link function to account for the Likert-scale of indicators; and mean- and variance-adjusted weighted least squares (WLSMV) with categorical indicators and standard errors clustered by clinic.

The online survey was constructed in Corporater Surveyor, an online survey tool pre-approved by the Data Protection Officer of Haukeland University Hospital. Analysis and data visualization was performed in Stata SE 17, except WLSMV estimated in Mplus v8.1.