Concurrent validity of an Estimator of Weekly Alcohol Consumption (EWAC) based on the Extended AUDIT

Background and Aims: The 3-question Alcohol Use Disorders Identification Test (AUDIT-C) is commonly employed in healthcare to screen for levels of alcohol consumption. AUDIT-C scores (0-12) have no direct interpretation and do not provide information on alcohol intake, an important variable for behaviour change. The study aimed to (a) develop a continuous metric from the Extended AUDIT-C, offering equivalent accuracy, and providing a direct estimator of weekly alcohol consumption (EWAC); (b) evaluate the EWAC's bias and error using the Graduated-Frequency (GF) questionnaire as a reference standard of alcohol consumption. Design: Cross-sectional diagnostic study based on a nationally-representative survey. Settings: Community-dwelling households in England. Participants: 22,404 household residents aged [≥]16 years reporting drinking alcohol at least occasionally. Measurements: Computer-assisted personal interviews consisting of (a) Extended AUDIT and (b) GF. The primary outcomes were: mean deviation <1 UK unit (metric of bias); root mean squared deviation <2 UK units (metric of total error) between EWAC and GF. The secondary outcome was an equivalent receiver operating characteristic area under the curve for predicting alcohol consumption in excess of 14 and 35 UK units compared to AUDIT-C and AUDIT. Findings: EWAC had a positive bias of 0.2 UK units [95% confidence interval: 0.04, 0.3] compared to GF. Deviations were skewed: while the mean error was +/-11 UK units/week [9.8, 12.0], in half of participants the deviation between EWAC and GF was between 0 and +/-2.1 UK units/week. EWAC predicted consumption in excess of 14 UK units/week with a significantly greater area under the curve (0.921 [0.917, 0.925]) than AUDIT-C (0.871 [0.866, 0.876]) or the full AUDIT (0.854 [0.849, 0.860]). Conclusions: The EWAC https://ewac.netlify.app is designed to estimate weekly alcohol consumption using answers to the Extended AUDIT-C questionnaire. Using the detailed GF as a reference standard, the EWAC met the targeted bias tolerance. Its accuracy was superior to that of both AUDIT-C and the full AUDIT in relation to consumption thresholds, making it a reliable complement to the Extended AUDIT-C for health promotion interventions.


Introduction
consumption. Instead, we employ methods developed for quantity-frequency-variability 119 instruments [20]. For every individual ݅ , the EWAC is computed as the product of were not asked any further AUDIT or GF questions. They were thus excluded. 139 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint 2.3 Measures which participants described how many times they consumed given quantities of alcohol 145 during the last 4 weeks [21]. A longstanding obstacle in alcohol research and treatment lies 146 in the absence of undisputed 'gold standard' or biomarker for objectively determining alcohol 147 consumption. Instead, a number of instruments measure self-reported consumption with 148 varying validity and reliability over different durations. Comprehensive reviews [22-27] 149 indicate that yesterday recall and prospective diaries tend to record higher (and more 150 accurate) alcohol consumption by minimising recall bias, followed by GF measures. The GF 151 schedule's main advantage lies in measuring occasional heavy consumption, which can 152 constitute an important proportion of total consumption. Although widely employed in 153 population surveys, the GF schedule is uncommon in clinical practice. 154 We examined the EWAC's coverage of per-capita alcohol retail sales for England in 2014 155 [28], which capture all alcohol produced/processed in or imported to England for sale or 156 consumption. 157 We also cross-examined the EWAC's distribution against other reference estimates obtained 158   Table S6) were fitted solely in respondents with an AUDIT-C score 197 5 or an AUDIT score 8, for whom additional characteristics were recorded during 198 interview: favourite drink (beer; wine; spirits alone; mixed spirits; cider; other); and whether 199 the respondent had attempted to restrict alcohol intake in the last 12 months. 200

Receiver Operating Characteristics 201
We tested the EWAC's superiority to the traditional AUDIT and AUDIT-C scores in predicting 202 consumption exceeding 14 or 35 UK units/week. These correspond to UK thresholds for 203 characterising alcohol use as 'increasing risk' (predicted by an AUDIT-C score of 5-7), and 204 'higher risk' (AUDIT-C8) which is above 35 units for women and 50 units for men [36]. We 205 tested the hypothesis that the EWAC has an identical receiver operating characteristic full 206 area under the curve (AUC) to the AUDIT-C and the full AUDIT scores using nonparametric 207 paired AUC tests [37]. AUDIT-C and AUDIT scores were calculated from the Extended 208 AUDIT by capping the contribution of each question to 4. 209

Aggregate concurrent validity 210
We compared the empirical cumulative distributions of (1) the EWAC computed in the ATS; 211 (2) the GF estimator in the ATS; Poststratification survey weights adjusted for nonresponse bias in sources (1-3), and self-215 selection into prospective diary data collection in source (4). 216 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint and error and were used for the remainder of the analysis. With those, the EWAC's 220 Pearson's correlation with GF was estimated at The root mean squared deviation (RMSD) at 10.9 units/week [95% CI: 9.8, 12.0] was 225 significantly greater than the pre-registered 2-unit total error tolerance ‫,)100.0<(‬ suggesting 226 that the EWAC falls on average 11 units away from the GF reference standard. 227 The RMSD masked a dispersed and skewed distribution of error (Table 2). In 50% of 228 participants, the EWAC fell within

Subgroup accuracy 237
MDs and RMSDs were regressed against respondent characteristics to identify potential 238 subgroup differences in bias or precision from the reference category (White British females 239 aged 25-34 years without educational qualifications, who never smoked). Model predictors 240 accounted for <2% of the variance, indicating stability in MD and RMSD across subgroups 241 (Table S5). 242 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint

Receiver Operating Characteristics 262
We examined the EWAC's ability to predict consumption exceeding 14 or 35 UK units/week. 263 The full areas under the receiver operating characteristic curves (AUC, Figure S7 is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint AUDIT-C score ‫)100.0<(‬ and the full AUDIT score ‫.)100.0<(‬ The best cut-off for detecting 273 consumption in excess of 35 units/week using the EWAC is 17 units/week (Table 4). 274

Strengths and limitations 288
This paper is the first to (a) develop an EWAC using a well-accepted and validated alcohol 289 screening tool such as the AUDIT; and (b) quantify its bias and precision with respect to a 290 continuous measure of alcohol consumption. One study [38] previously reported mean 291 consumption by AUDIT-C score, but without quantifying bias or precision of such a measure. 292 Other studies have evaluated the AUDIT-C's accuracy in estimating alcohol consumption, 293 but only in relation to predicting consumption in excess of predefined thresholds [9]. Such 294 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review) preprint
The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint studies achieved AUCs ranging 0.83-0.96. Our study found the EWAC to be superior to both 295 the AUDIT-C and the full AUDIT in predicting GF exceeding 14 and 35 units/week. At the 14-296 unit threshold, the specificity gain from 0.684 (AUDIT-C) to 0.876 places the EWAC with the 297 other best-performing diagnostic tools. 298 Our study provides strong confidence in the internal and external validity of findings in 299 England on account of the large sample size and extensive range of subgroup analyses 300 reported. Bias was mostly consistent across subgroups examined (age/sex, education, 301 smoking status, religion), with one exception. EWAC overestimated alcohol consumption by 302

2-3 UK units/week in Black/Other ethnic groups; though overall precision was better in 303
Black/Asian/Other ethnic groups than in White British respondents. Variation in the 304 sensitivity of AUDIT-C across ethnic groups has previously been noted in the US [39]. 305 Repurposing a well-known tool such as the AUDIT-C has several advantages. It is already 306 translated in many languages and adapted to the varying standard drink sizes adopted 307 internationally [7]. The Extended AUDIT scores can be converted into traditional AUDIT 308 scores by capping items to 4, thereby offering a point of comparison with existing evidence.  The EWAC, by limiting itself to an estimation of alcohol consumption is transparent across a 321 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review) preprint
The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint populations excluded from most sampling frames (e.g. communal/carceral institutions, 327 homeless/no fixed abode populations). 328 We also note that the reference standard, as all self-reported measures, is imperfect. This 329 may introduce bias into estimates of precision: by definition, the reference standard's own 330 error will inflate the RMSD. In other words, it is likely that a proportion of the RMSD is 331 attributable to error in the GF measures rather than the EWAC. 332 Previous research offers reassurance that the EWAC's error is comparable to that of other 333 (even the more time-consuming) questionnaires. One telephone interview study [43]  and suggests that EWAC's agreement with GF is similar to that of diaries. 341

Potential applications 342
Being equivalent or superior to the AUDIT in speed, accuracy, and international 343 standardisation, the EWAC appears suitable for use in clinical practice to support brief 344 interventions and to feed back a reliable 4-unit wide interval estimate of alcohol consumption 345 (eg: '9-13 units/week' or '70-100 g/week'). One such calculator is available online: 346 https://ewac.netlify.app 347 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint (observable entity) [45] is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted December 11, 2020. ; https://doi.org/10.1101/2020.12.11.20247106 doi: medRxiv preprint