The second visit was used for this cross-sectional analysis because questions about past 48-hour drug use were added to the questionnaire starting with that visit; thus, it was the first available time point for comparison and validation of self-reported use and urine testing. Women were asked about type of ATS use ; frequency and route of use since their last visit; number of days of use in the past month; and use in the past five days, with specific questions including “today,” “yesterday,” and “two,” “three,” “four,” and “five” days ago. Urine toxicology testing was conducted to qualitatively screen for recent ATS, opiate, and cannabis use. Women were asked to void into prelabeled sterile collection cups in a private lavatory; the specimens were passed through a private window to the on site laboratory for testing. The test included four strips, which yielded positive results for amphetamine and/or methamphetamine if either exceeded 1000 ng/mL; for opi ates if morphine in urine exceeded 2000 ng/mL; and for cannabis if the concentration 11-nor-Δ9-tetrahydrocanna binol-9-carboxylic acid exceeded 50 ng/mL. A positive amphetamine or methamphetamine screen was considered indicative of ATS use in the past 48 hours.Overall, results suggest high validity of self-reported ATS use among FSW when compared with urine toxicology screening. In almost all cases where women reported no ATS use in the past two days, negative urinalysis corroborated self-report. The majority of participants with positive urine tests reported ATS use during the same detection period. However, only 81% of participants who reported ATS use had positive urine tests.

One possible explanation of the low positive predictive value is that women in the study actually used ATS but in such a small quantity that the urine tests failed to detect it. Since ATS is illegal and its purity is unknown,cannabis drying trays some women could have used the less pure forms of ATS, which may not have been potent enough to be detected by urine testing. The NACD has reported that, among 151 pill samples of ATS tested, 25% of the samples had purities below 10%. Al though the proportion of women self-reporting ATS use was slightly higher than the urine test results , these rates are not inconsistent and are near perfect. Other studies have documented higher self reported use compared with urinalysis results, leading to recommendations that multiple methods be used to assess drug use exposures. The high concordance between self-report and test results are suggestive of high internal validity of self-report of ATS in our study population. Some differences were seen in the performance of self report compared with urinalysis when examined by age, HIV status, and sex-work setting. Most notably, there was lower precision between positive self-report and urinalysis tests among younger women and among women working in entertainment or service settings. The lower PPV may relate to lower prevalence of ATS use among these subgroups. We have previously shown that women working in entertainment and service sec tors in Cambodia are less likely to use ATS than women working in brothels. Prevalence of ATS among younger women is slightly lower but not significantly so. Importantly, specificity was high overall, with subgroup analyses showing valid self-report of no ATS use in our sample. This is import ant for further studies of ATS exposure in this population, for public health surveillance, and potentially for intervention and implementation of drug prevention programs. The high validity of self-report may be associated with several factors. The women in this study were not reluctant to answer the survey questions or to take the test, as indicated by the high participation rate. This could be due, at least in part, to the fact that the participants were recruited by a known and trusted community-based agent, our collaborating partner , and were comfortable with the staff involved in data collection.

Moreover, the women in the study knew that providing truthful responses about their drug use would not result in negative consequences or punitive action. This study had several limitations. Due to the small sample size and non-systematic sampling, our estimates lack precision and results may not be representative of all young women engaged in sex work in Phnom Penh or Cambodia. This is particularly true for the stratified analyses, where cell sizes were very small in some cases and prevalence of ATS was lower. Poor recall may have contributed to some discordance, including the relatively low PPV found overall. Approximately one in five women incorrectly reported recent ATS use. Recall of ATS use could be affected by recent ATS use and its side effects, including sleep deprivation and confusion. It is unknown if this would result in over- or underestimating of self-report. Since women were all informed about the testing as part of the informed consent process and ongoing study-procedure education, some women may have over reported use for the periods about which they were queried. Moreover, urine toxicology tests are not perfectly accurate. Although the urinalysis test is widely accepted as a “gold standard” for substance use validation, exclusive reliance on such results does not necessarily improve valid ity because of problems with false negatives. Many studies comparing self-report, urine, and hair testing results suggest that hair analyses provide higher rates of recent drug use than can be detected by either urine tests or self reports. Various authors suggest multi-modal testing for the most accurate results. Despite these limitations, our results suggest a high level of concordance between self-reported ATS use and urine toxicology results in this group of women. Results indicate high prevalence of ATS use among FSW, who are also at elevated risk of HIV and other sexually transmit ted infections. There are few, if any, community-based options for ATS users in Cambodia. The finding that self report, especially specificity, is valid among young FSW is important because of potential utility in surveillance as well as drug prevention and intervention programs in this population.

There is a significant need for evidence-based prevention and drug treatment resources in Cambodia, including potentially cognitive behavioral therapy, contingency management, and possibly new pharmacotherapies to reduce ATS use. The forthright self-reporting of drug use by women participating in this study shows that, in a safe and non-punitive setting, disclosure of accurate drug use is possible. These findings, which are consistent with other studies showing high validity of self-reported drug use, may also be relevant to other vulnerable populations in Cambodia reported to have high rates of ATS use and who may also be in need of interventions, including children, young adults,heavy duty propagation trays and men who have sex with men. Indeed, with escalating manufacture and use of ATS throughout Southeast and East Asia, and in consideration of the need for expanded surveillance of drug use to more accurately inform public health and policy responses, self reported use may be a reliable data collection method. For surveillance, research, and health-care settings, it is import ant that providers and others address drug-related health issues in a nondiscriminatory manner and without punitive consequences in order to accurately assess and effectively address health and safety issues in high-risk populations.The COVID-19 pandemic has created a pressing need for tools to combat the spread of misinformation. Since the pandemic affects the global community, there is a wide audience seeking information about the topic, whose safety is threatened by adversarial agents invested in spreading misinformation for political and economic reasons. Furthermore, due to the complexity of medical and public health issues, it is also difficult to be completely accurate and factual, leading to disagreements that get exacerbated with misinformation. This difficulty is compounded by the rapid evolution of knowledge regarding the disease. As researchers learn more about the virus, statements that seemed true may turn out to be false, and vice versa. Detecting this spread of pandemic-related misinformation, thus, has become a critical problem, receiving significant attention from government and public health organizations , social media platforms , and news agencies . In this paper, we introduce the COVIDLIES dataset for misconception detection on Twitter. COVIDLIES comprises of 62 common misconceptions about COVID-19 along with 6591 related tweets, identified and annotated by researchers from the UCI School of Medicine. Given a tweet, we annotate whether any of the known misconceptions, curated by the researchers, are expressed by the tweet. If they are not, then they are considered No Stance. If they are, we further identify whether the tweet propagates the misconception or is informative by contradicting it . Example misconception-tweet pairs for each label are illustrated in Figure 1.1. We provide benchmark results for each of these sub-tasks. First, we evaluate text similarity models on their ability to detect whether a tweet is relevant to a given misconception . Next we evaluate zero-shot and few-shot models for the ability to detect the stance of each towards retrieved misconceptions For the zero-shot setting we train on the pre-existing tasks of natural language inference and fact verification. For the few-shot setting we train on COVID-19 Health Risk Assessment task combined with a dataset of COVID-19 tweet-misconception pairs annotated for stance by researchers from the UCI School of Medicine.

Our results show that existing models struggle at both tasks , however improve considerably after domain adaptation ; 74.3 Hits@1 for retrieval and 46.3 macro F1 on zero-shot stance detection. We see some further improvement on stance detection when using the few-shot setting . While our initial results using domain adaptation and few-shot learning are encouraging, they leave much room for improvement. There is still much work that needs to be done before NLP systems can be seriously considered for combating COVID-19-related misinformation, and we hope COVIDLIES will be useful to help researchers understand when such systems are ready to be deployed. Due to limited availability of labeled data specific to this problem, we expect that models will need to be supervised on other, related tasks. For misconception retrieval, for example, relevant misconceptions can be ranked by measuring the semantic similarity between the tweet and each misconception, e.g., using cosine similarity between average word embeddings or more recent transformer-based methods such as BERTSCORE . For the stance detection sub-task we perform zero-shot learning by training on the pre-existing tasks of natural language inference and fact verification. We also perform few-shot learning by training on COVID-19 Health Risk Assessment task combined with a dataset of COVID-19 tweet-misconception pairs annotated for stance by researchers from the UCI School of Medicine. The current dataset contains 62 misconceptions, along with 6591 annotated tweet-misconception pairs. Statistics about the distribution of labels are provided in Table 3.2. The distribution is heavily skewed, containing mostly No Stance tweets, and a higher proportion of Agree tweets than Disagree. The heavy skew towards No Stance tweets could be a due to the dataset construction methodology, specifically using BERTSCORE without fine-tuning to retrieve tweets per misconception. As we show in 4.0.2, domain adaptation significantly improves misconception matching. Further, presence of more Agree than Disagree tweets could be due to a bias in BERTSCORE towards scoring agreement higher. Top misconceptions for each class are shown in Table 3.1. We only consider misconceptions with more than 80 annotated tweets, and rank the misconceptions for each class by the proportion of tweets that are annotated as that class. We present the top three misconceptions for each class with their corresponding percentage. There are misconceptions for which 100% of the paired annotated tweets express No Stance, which we do not see for the other two classes. We notice there is a misconception with nearly 50% of paired tweets labeled as Agree; and the highest proportion of Disagree labeled tweets found for any misconception in the Disagree class was 51%. COVIDLIES, however, is an evolving dataset; annotation is not yet complete for all 62 Wikipedia misconceptions matched to 100 tweets using BERTSCORE, and we are continually identifying additional misconceptions, as well as collecting more recent tweets for annotation. Further, we will gather more relevant tweets by using domain-adapted retrieval models, which, as we will see in the next section, considerably outperform the current approach to retrieval, BERTSCORE. We obtain vectorized representations of tweets and misconceptions using word embeddings.