00:36:08 MK: Wasn’t it said that RTPCR would not be done if RAT was negative? 00:37:05 Nishant Chandgotia: Wasn’t that for people with moderate and high risk? Also I am guessing that this is only being done as an example. 00:38:23 Rajesh Sundaresan: "Wasn’t it said that RTPCR would not be done if RAT was negative?" Yes. Siva is just giving an example. However, in some districts, the staff didn't realise that they needn't have administered the RTPCR, and went ahead and did it. 00:39:00 Rajesh Sundaresan: Correction: RAT positive --> RTPCR was not done. 00:41:13 MK: The test pattern is not fixed but adaptive? Is this likelihood calculation right? 00:42:37 Rajesh Sundaresan: Good point. Test pattern does depend on risk category. We ignored it, for simplicity. 00:43:49 MK: Does it depend on category or on the results of the tests themselves? I mean second test depends on result of first test? 00:44:33 Rajesh Sundaresan: It depends on category (low risk didn't get RAT), and also on test outcome on-site (RAT positive). 00:45:26 Rajesh Sundaresan: So a more precise likelihood calculation will account for likelihood of the test pattern as well. 00:47:18 Rajesh Sundaresan: There are also some test patterns that arose and were beyond our control, lost samples, unidentified samples, etc. 00:47:47 Frank den Hollander: The percentages are rather high. Why? 00:48:32 Rajesh Sundaresan: Siva is explaining ... 00:49:23 Mriganka Basu Roy Chowdhury: What is the difference between Acute Infection and COVID 19 Prevalance? 00:50:32 Rajesh Sundaresan: Acute infection fraction is the fraction of individuals who see active infection. They test positive for viral RNA. COVID-19 prevalence stands for fraction of people infected in the population (both past infection and current infection). 00:51:32 Deepayan Sarkar: Re high active rate, how are you making sure there is no selection bias? How are you sampling people who are staying home? Maybe this was explained in the beginning. 00:51:57 Samyak Rajanala: What are the confidence bands in the estimated prevalence counts? 00:52:10 Abhay: Can the sample values be extrapolated to the whole population? 00:52:12 MK: My question too. How was randomization on site ensured? 00:52:44 Abhay: The design is missing an entire group who is not part of the three cohorts 00:53:11 Rajesh Sundaresan: Sampling bias, Deepayan's question. The market area sampling was systematic. The elderly were supposed to come from the census list. Those with comorbidities were to come from a vulnerable individuals list compiled in April 2020. Both lists were systematically sampled. But there were violations of this protocol. 00:53:39 Rajesh Sundaresan: Samyak, confidence intervals will be discussed in the next talk as well. 00:55:21 Aditya Gopalan: the 0.05% fatality rate estimate seems like great news! aren't rates reported generally in the 1-2% ballpark? 00:55:33 Rajesh Sundaresan: Abhay and MK: Agreed. This is a sentinel survey rather than a completely random survey. The hope was that the design effect will account for the correlation. A truly random survey would have required quite a bit of travel for the surveyors. 00:55:57 Rajesh Sundaresan: The time and cost considerations didn't permit such an approach. It was a tradeoff that was made. 00:56:52 MK: Systematic sampling is usually better than random sampling. My only question was whether it was ensured on site (following Deepayan’s question) 00:57:25 Rajesh Sundaresan: 0.05% IFR: This is an underestimate. Not every COVID-19 death would have been counted. The IFR in more urban areas like Bengaluru would be a better indicator. 00:57:47 Mriganka Basu Roy Chowdhury: 1:40 CIR means you're reporting 1 for every 40 true cases? 00:58:45 Deepayan Sarkar: I wouldn't be surprised if many low-risk individuals were just staying home, especially among high / middle income groups, particularly in urban areas like Bangalore. Not something you can do anything about, but could be a confounding factor. 00:59:38 Rajesh Sundaresan: Yes, systematic sampling was followed as much as possible. If not enough were available, the staff had to sample the next day. The protocol required that every 3rd pregnant woman to be recruited, every fifth OPD visitor's attendee, every fifth vegetable vendor, etc. 01:00:08 Rajesh Sundaresan: "1:40 CIR means you're reporting 1 for every 40 true cases?" Yes, that's correct. 01:01:39 Rajesh Sundaresan: Yes, Deepayan, we may be missing those who have been extremely cautious. We didn't know how to correct for this. Suggestions most welcome. We will have two more rounds in Dec/Jan and one in March (if needed). 01:02:41 Rajesh Sundaresan: (Correction: One round in Dec/Jan and one round in March). 01:05:24 Gautam Menon: The symptoms were self-reported? This is sometimes inaccurate, especially because people might not remember (or misattribute) past symptoms 01:39:05 Gautam Menon: Siva: What fraction of RAT negatives received/did not receive a followup PCR test? 01:40:51 Siva Athreya: As per protocol: from each participant two swabs were taken. If RAT was positive then the second swab was not processed. RAT negative was sent to RTPCR lab for processing. We are awaiting 1600 odd results when we decided to give up the chase. 01:41:20 MK: Is it usually a reasonable assumption that the errors of different tests on the same individual are independent? 01:43:16 Siva Athreya: As far as my limited knowledge the sensitivity and specificity of a test are based on past performance of them. I am not sure of the biology of it. Perhaps Gautam might have more information. 01:45:38 Gautam Menon: Yes, they are calculated based on tests on samples that are known to be definitely negative (taken long before the pandemic) or definitely positive (taken from patients showing active symptoms). The PCR test uses RNA while the RATs use other protein parts of the virus. They’d be expected to be independent. However, different types of RAT tests may not be completely independent, but that’s not a problem here. 01:46:51 Aditya Gopalan: slightly technical question: if your estimate pHat is in some parts close to 0 (e..g, for some category p_i), then isn't the Fisher confidence interval unnecessarily large? (variance should be much lesser for a coin toss with prob. close to 0) 01:48:19 Siva Athreya: Aditya, your comment is correct. The interval size will depend on phat. 01:48:46 Deepayan Sarkar: Yes, ideally CIs should be computed by inverting tests, but probably the extra programming and computational burden is not worth it. 01:49:03 MK: Isn’t the Fisher information 1/p(1-p)? I am confused about Aditya’s question 01:49:04 Aditya Gopalan: Ok thanks. 01:49:54 Deepayan Sarkar: Siva: I don't remember you saying anything about correlations in \hat{p}, was there anything interesting there? 01:50:54 Deepayan Sarkar: Manjunath: The asymptotic result assumes \hat{p} is Normal, but it cannot be because it's bounded to be in (0,1). This matters close to the boundaries. 01:51:40 MK: Yes, indeed, the normality goes out, thanks. I was wondering if I was confusing Inf with 1/Inf. 01:52:07 Siva Athreya: Deepayan: the covariances were non-trivial but I don't think anything stood out to show anything significant. 02:00:12 MK: This analysis is for a fixed u? 02:00:34 Siva Athreya: for total burden. So yes u = (1,1,1) 02:03:22 MK: Given the sensitivity and all that, could it not be possible that a better thing is to reduce sample size and administer only RT-PCR? 02:06:08 Siva Athreya: Manjunath: It is easier to say in words. I will try to write it out. Rajesh can answer. Given budget and optimisation criteria (reduction of variance) the sample size is a given and the solution seems to be based on test patterns. 02:06:22 Aditya Gopalan: Question: Did this design + budget calculation the driver for the N=432 sample number in the survey? Or was it something else? 02:06:32 Siva Athreya: Aditya: no. 02:07:35 Siva Athreya: Aditya: this is what we worked on post survey vexed by the idea that RAT was excluded for one risk group. If Budget was a criteria and reduction of variance was a goal then is there a better way of doing test pattern. 02:14:57 Aditya Gopalan: Thanks for the great talk and discussion! Best wishes for the future stages.