Assessing Spillover Effects of HIV Testing on HIV Incidence in Rural KwaZulu-Natal, South Africa

Transcript

• 00:00Geared to that. So,
• 00:02Faith, you are welcome to
• 00:04come and the floor is
• 00:06yours.
• 00:10Hello, everyone. Thank you for
• 00:11being here today and,
• 00:14thank you for having me
• 00:15today for this presentation
• 00:16of my work, basically.
• 00:18And,
• 00:20today's this part of this
• 00:21talk is about assessing the,
• 00:24cause effects,
• 00:25particularly
• 00:26the spillover and the direct
• 00:27effects of HIV testing on
• 00:30the HIV instance,
• 00:31in the in the
• 00:36rural South Africa. Sorry. I'm
• 00:38a little bit nervous.
• 00:47So first of all, I'd
• 00:48like to send all our
• 00:49collaborators
• 00:50and all the investigators,
• 00:53who collected this data. And
• 00:55the the project was founded
• 00:56by NIMH, and the other
• 00:58results I presented today is
• 01:00solely on our our officers'
• 01:02responsibilities,
• 01:04not related
• 01:06represent any official
• 01:08opinion.
• 01:12So first is a background.
• 01:13We know that HIV is
• 01:15a virus such,
• 01:17that that will affect
• 01:19the human immune system and
• 01:20the without the treatment, it
• 01:22were developed to the eyes.
• 01:25And, we know that HIV
• 01:26we now have the HIV
• 01:28care test data, which is
• 01:29ninety five, ninety five, ninety
• 01:31five target. The first ninety
• 01:32five is,
• 01:34for people who who are
• 01:35living with HIV,
• 01:37we hope to get the
• 01:38ninety five percent of them
• 01:39getting the HIV
• 01:40rapid test. So getting a
• 01:42HIV test
• 01:43so they can know their
• 01:44HIV status.
• 01:46The second ninety five is
• 01:48we hope that, with the,
• 01:50the people who are living
• 01:51with HIV and also getting
• 01:53the HIV test and knowing
• 01:55their status positive status,
• 01:57they could be,
• 01:58on treatment, the ART treatment.
• 02:01And the the third one
• 02:02is we hope for the
• 02:04all the people who are
• 02:05on ART treatment,
• 02:07we can achieve that ninety
• 02:09five percent of the people
• 02:10can get the barrier load
• 02:11suppression.
• 02:12So this is our, HIV
• 02:14care
• 02:15test kit, target.
• 02:17And, we have, some treatment
• 02:19guidelines
• 02:21by the WHO for the
• 02:22HIV.
• 02:23And the, the guidelines has
• 02:25already changed over the last,
• 02:28fifteen years.
• 02:29So at the beginning,
• 02:31first, they have,
• 02:33delay delaying the ART and
• 02:35treatment,
• 02:37until the on CD4 cell
• 02:39counts, it's less than two
• 02:41hundred
• 02:42cell count two two hundred
• 02:43cells per,
• 02:45milliliter.
• 02:46Then they change the HR
• 02:47in twenty thirteen to,
• 02:50delaying the ART when the
• 02:52CD6 cell
• 02:54count is less than three
• 02:56hundred five fifty,
• 02:58cells per milliliter
• 03:00blood. And,
• 03:01on the fifteen,
• 03:03twenty fifteen, they changed that
• 03:05again to the five
• 03:07five hundred
• 03:09five
• 03:11I'm so sorry.
• 03:13So, yeah, I missed the
• 03:15one. So the before the
• 03:16twenty eight twenty,
• 03:17ten, the,
• 03:19the guideline is
• 03:21a significant count less than
• 03:22the two hundred cells per
• 03:24milliliter. And, from the twenty
• 03:27ten to twenty three, they
• 03:28has,
• 03:31to twenty thirteen,
• 03:33they have the, delay
• 03:35between the twenty ten to
• 03:37twenty thirteen,
• 03:38the they're delaying the, ART
• 03:40until the CD4 is less
• 03:41than three hundred fifty,
• 03:43cells.
• 03:44And, from the thirteen to
• 03:46fifteen,
• 03:47they has the,
• 03:48threshold is less than the,
• 03:52three hundred fifty
• 03:54thirds
• 03:55per milliliter.
• 03:56And after
• 03:57the, twenty fifteen, they has
• 04:00the,
• 04:01they have the new
• 04:03policy just to start with
• 04:04ART immediately
• 04:06after the
• 04:07diagnosis of the HIV.
• 04:11So we know that HIV
• 04:12testing is the first step
• 04:13in the h r HIV
• 04:15care test data, and it
• 04:17is an important tool for
• 04:18reducing the HIV
• 04:20infection.
• 04:21And,
• 04:22spillover effect is another
• 04:26is another,
• 04:28keyword in the in this
• 04:29talk, and, it it can
• 04:31also be called as the
• 04:32indirect effect. So I
• 04:34in fact, this is the
• 04:36effect of one individual's
• 04:39treatment will affect
• 04:40analysis outcome,
• 04:42and, this may occur in
• 04:45the human
• 04:47human network because people interact
• 04:49with each other, so they
• 04:50may affect,
• 04:51they may have some,
• 04:53connections.
• 04:54And,
• 04:55existing studies have suggested that
• 04:57HIV testing has a significant
• 05:00effect in reducing the HIV
• 05:02infection.
• 05:03However, these studies have not
• 05:05fully be, assessed by the
• 05:07spill spillover effect or indirect
• 05:10effect.
• 05:12So we want to take
• 05:13the presence of the sphere
• 05:14of effect factor into account
• 05:18into account the,
• 05:19our understanding of
• 05:21the HIV drug test. It
• 05:23will help us to conduct
• 05:25more robust ablutions to force
• 05:26the effects of HIV testing
• 05:28instead of just to know
• 05:30the overall.
• 05:31And,
• 05:32we want to consider
• 05:33one test
• 05:37component at one time and
• 05:38the checking the others
• 05:40as the confounders.
• 05:48So so,
• 05:50the overall project goal is
• 05:51assessing the direct and all
• 05:53effects
• 05:54of HIV testing on the
• 05:56HIV instance.
• 05:57And we will have two
• 05:59type of analysis. One is
• 06:01the Ivanovo
• 06:02analysis, we call that. This
• 06:03is just a one time
• 06:05expiry. The other one is
• 06:07longitudinal
• 06:08analysis. We will have the,
• 06:10repeated exporters,
• 06:12the longitudinal data. And the
• 06:14expiry for the Ivanovo analysis
• 06:16is if this person ever
• 06:18received HIV testing
• 06:19in the type, in the
• 06:21test study in our data.
• 06:22And outcome
• 06:23of each of these HIV
• 06:25instance, the cover words are
• 06:27all baseline my measurements, and
• 06:29the inference side was defined
• 06:31by the homestead homestead at
• 06:33Graceland.
• 06:37So, let's start with the
• 06:39abnormal analysis.
• 06:40This is a overview of
• 06:42all the variables that we
• 06:43use in the analysis.
• 06:45The explorer,
• 06:47is I will receive or
• 06:48not receiving the HIV test
• 06:50and the outcome HIV instance.
• 06:53And the homestead level cohort
• 06:54here is HIV prevalence
• 06:56in the homestead. And,
• 06:58individual level cohorts, we have
• 07:00five cohorts
• 07:01considered in this analysis.
• 07:03First is the gender and
• 07:05the the mere circumcision status,
• 07:07then is age, medical status,
• 07:10also the number of sector
• 07:11partners in the last twelve
• 07:13months
• 07:14and the, if
• 07:16experienced any sexual violence in
• 07:18the last
• 07:19twelve months.
• 07:21So as we can see
• 07:22here,
• 07:23in total, we the data
• 07:25was,
• 07:28in total, we have the
• 07:29twelve thousand and seven hundred
• 07:31and zero three people involved
• 07:33in this analysis.
• 07:35Most of them receiving the
• 07:36HIV,
• 07:38HIV test. So only one
• 07:40thousand and sixty three
• 07:42did not tested.
• 07:44And the for the outcome,
• 07:45we can say we have
• 07:46a few outcome.
• 07:47It's four hundred and sixty
• 07:49five for all the other
• 07:50types of and twelve thousand,
• 07:53they are still take keep
• 07:55HIV negative at the end
• 07:56of the study.
• 07:59So this is,
• 08:01is,
• 08:02descriptive
• 08:03analysis of the baseline characteristics
• 08:05we mentioned.
• 08:06So the final, HIV status,
• 08:09we only have less than
• 08:10four percent of people had
• 08:12the HIV survival version.
• 08:14And, if we look at
• 08:15the, HIV premiums in the
• 08:17homestead level, it's around the
• 08:20fifth and seventeen, eighteen percent.
• 08:23And these are some distribution
• 08:26of other,
• 08:27baseline coverage.
• 08:30So,
• 08:31that's the background, but then
• 08:33we started with our method
• 08:35part. The notation is specifically
• 08:38for the, annual number analysis.
• 08:40So we have the k
• 08:41is the index of the
• 08:42homesteads.
• 08:43The k value from the
• 08:45one to capital k, and
• 08:46that's the this,
• 08:48n k is,
• 08:50subsample in the homestead k.
• 08:52It's just a a collection
• 08:54of all members in the
• 08:56homestead k. And it's a
• 08:59is the number of the
• 09:00individuals in the homestead k.
• 09:02I is the index of
• 09:04individuals
• 09:04in homestead homestead k. And
• 09:07this I, right, can be
• 09:09from one to the n
• 09:10k.
• 09:11So then x is
• 09:13overall, x is also,
• 09:16overall, x is the covariance.
• 09:18The a is expiry, and
• 09:20the y is the outcome.
• 09:22Then the index just,
• 09:24represent the different type of
• 09:26covariance
• 09:27expiry.
• 09:28And we also have the
• 09:29outcome.
• 09:30So the y k I
• 09:32is the observed outcome and
• 09:34the y k I given
• 09:35the a k is
• 09:37the potential outcome.
• 09:44So,
• 09:45we use, we have some
• 09:47assumptions here. First is the
• 09:49extendibility.
• 09:49We assume that is the
• 09:51outcome is independent from the
• 09:53treatment assignment given the,
• 09:56cover
• 09:57condition on the cover.
• 10:00And the the also the
• 10:01positive pay will assume that
• 10:03the probability of receiving the
• 10:05treatment
• 10:06is larger than zero for
• 10:08all possible,
• 10:09treatment,
• 10:10assignment and, possible cover.
• 10:14Also, we have the partial
• 10:16inference assumptions. This
• 10:18is assumed that the individual's,
• 10:21outcome
• 10:22depends on not only his
• 10:24own treatment, but also the
• 10:25labor's treatment. And labor's is
• 10:27defined by the homestead
• 10:29in this
• 10:32study. So,
• 10:33we also have the treatment
• 10:35of variation I mean, irrelevant
• 10:37irrelevance,
• 10:38assumptions. These assumptions assume that
• 10:41we only have one version
• 10:42of, intervention and one version
• 10:44of not, not given intervention.
• 10:46That's if there is multiple
• 10:48versions of interventions,
• 10:50we assume they are not
• 10:51related to the,
• 10:53our parameter of being fixed.
• 10:58Sub question?
• 11:00So in the assumption page,
• 11:02on exchangeability,
• 11:05so here,
• 11:07is it a so so,
• 11:09does this hold by design?
• 11:11Is this a randomized study?
• 11:16The step Right? So it's
• 11:17just assuming that at the
• 11:19homestead level, the entire treatment
• 11:21assignment value is independent of
• 11:23all the potentials, even the
• 11:25collection of the entire
• 11:28co covariance set within a
• 11:29homestead. Right? Yes. Okay.
• 11:32As it's it's observational.
• 11:35Well, it's cluster randomized.
• 11:36Yeah. But this particular package
• 11:38component is self selected, not
• 11:40randomized.
• 11:41So and then a k
• 11:43here is which we can
• 11:44make is that is that
• 11:45is the receipt of the
• 11:46component. Right. So Yes. Simon
• 11:48is random. Right? Right. But
• 11:50this is not really about
• 11:51the assignment. Right. Because it
• 11:53is it's different. It's it's
• 11:54on the receipt of the
• 11:55Yeah. Yeah. That's right.
• 11:57Okay.
• 11:58And in the notation system,
• 11:59we actually haven't discussed the
• 12:01conversation at all. Okay. I
• 12:02just wanna make sure.
• 12:05Yeah. I think I think
• 12:06there there shouldn't be any
• 12:07bars here either. It should
• 12:08just be
• 12:11Yeah. It should just be
• 12:11yeah. With the Oh, really?
• 12:14The The selection of the
• 12:15assignments in Yeah. The assignment.
• 12:18Yeah. That's a typo. Sorry.
• 12:19Bye bye. Oh, it's okay.
• 12:20It's okay. Yeah.
• 12:21That's great.
• 12:28So this is our estimate.
• 12:31We use,
• 12:33the it's the marginal mean
• 12:34of the plantar outcomes. So
• 12:36the expectation of the plant
• 12:37out of plants
• 12:39given the, individual treatment, AKI
• 12:42and the neighbor's treatment, the,
• 12:44the
• 12:45other members in the, same
• 12:47homestead.
• 12:48It's,
• 12:49equal to the all user
• 12:51logistic regression, so it's equal
• 12:53to the exponential,
• 12:54of the,
• 12:56intercept and the then the
• 12:57individual,
• 12:59the coefficient side
• 13:01one times the individual,
• 13:02filament
• 13:03and the
• 13:05coefficient side two times the,
• 13:07neighbors'
• 13:08treatment
• 13:09to summarize the treatment.
• 13:11And, here we summarize the
• 13:12treatment by just, some of
• 13:14the plus treatment is binary.
• 13:17So we just sum out
• 13:18the neighbors'
• 13:19treatment then divided by the
• 13:20number of neighbors. So this
• 13:22that will be some summarize
• 13:24the neighbors' exposure.
• 13:26And we define the, direct
• 13:28effect and spill over over
• 13:29effect in the risk ratio
• 13:31in terms of risk risk,
• 13:33risk ratio. And this
• 13:35is, the direct effect and
• 13:37the spillover effect, will look
• 13:39like in the,
• 13:41our model setting.
• 13:47And, the estimator we use
• 13:49is an existing,
• 13:52published estimator. It's a MS,
• 13:54marginal structure model,
• 13:57with the inverse probability
• 13:58weights.
• 14:00So the IPW
• 14:01inverse probability weights is used
• 14:02to adjust the further confoundings,
• 14:04and MSM is used to,
• 14:07model the financial outcomes.
• 14:09And, the estimate here, where
• 14:12is a solution for the,
• 14:13for
• 14:14weighted generalized estimating
• 14:18equation.
• 14:19So I think
• 14:23people should be familiar with
• 14:25this, so I wouldn't be
• 14:26spend a lot a lot
• 14:28of time here. Not every.
• 14:30Oh, sorry. Yeah. I'd say
• 14:31especially especially the modeling of
• 14:33the the weights.
• 14:35Oh, yeah. So the SW,
• 14:37represents the,
• 14:39stabilized weights. SWK is stabilized
• 14:41with for the,
• 14:44homestead,
• 14:45And the the numerator is,
• 14:47the
• 14:49the probability of receiving the
• 14:51treatment,
• 14:52only given the,
• 14:56yeah, receiving the treatment,
• 15:01and the the, denominator
• 15:04is a probability of receiving
• 15:06the treatment
• 15:07is equal receiving the treatment
• 15:09is
• 15:10depends on the cover words.
• 15:12So the difference between the
• 15:13cover words and the, denominator
• 15:15no.
• 15:17Not numerator and denominator
• 15:20is,
• 15:21if the cover words exists
• 15:23in the,
• 15:25in the function or in
• 15:26the in the formula.
• 15:28And
• 15:29the
• 15:31gamma k is a homestyle
• 15:33homestyle level random effect,
• 15:35we were used under the
• 15:37v k is working
• 15:38on coherence matrix.
• 15:41And the h k is
• 15:42just a random function of
• 15:44x and treatment a, and
• 15:46the epsilon k is the
• 15:48random error for the,
• 15:50homestead k.
• 16:00The next question again.
• 16:03So
• 16:03the
• 16:04on this page, I guess,
• 16:06when we when we estimate
• 16:08the stabilized weights, right,
• 16:10how did we,
• 16:12numerical like, how is the
• 16:14random effects integrated out in
• 16:16the calculation? Is it just
• 16:17by Gauss Herman quadrature? Or
• 16:20Yes. How we okay. We
• 16:21have to assume they're now,
• 16:22like, distributed. So Right. That's
• 16:24the annoying thing I'm not
• 16:25everyone new is actually
• 16:28using Chet Chet Chet Chet
• 16:29Chet Chet and Cool's test
• 16:30Brent Cool's test to make
• 16:32sure that
• 16:33the random effects are actually
• 16:34normally distributed because normally that
• 16:35assumption is not as important.
• 16:37But here, we're actually integrating
• 16:38You have to over, like,
• 16:39assuming it's a normal distribution.
• 16:41Yeah. So that's actually a
• 16:42good reminder, because we didn't
• 16:43check that explicitly here. We
• 16:45definitely need to check it.
• 16:46Right. I So you make
• 16:47everyone, but not. Well, I
• 16:49don't know. I just forgot.
• 16:50I feel like I was
• 16:51on a real kick with
• 16:51it, like, a couple years
• 16:52ago Yeah. Where everyone I
• 16:54was like, you have to
• 16:54run this test. Test. They're
• 16:55like, what if it fails?
• 16:56And I'm like, we'll figure
• 16:57it out. The other thing
• 16:59I wanna the other thing
• 17:00I wanna mention is that
• 17:01I think this is actually
• 17:02where I thought a potential
• 17:05approach to simplify the calculation
• 17:07Mhmm. Was, I think, in
• 17:08the
• 17:09classical binary regression type of
• 17:11modeling
• 17:13class, we always learn about
• 17:14when we want to marginalize
• 17:17out the random effects. There
• 17:18is actually a convergent formula
• 17:20on the logistic model, and
• 17:21the approximation was actually pretty
• 17:23accurate.
• 17:24I think it's, like, square
• 17:25root,
• 17:27high square something. Mhmm. And,
• 17:30that that that's actually something
• 17:32that might be, like, might
• 17:33be interesting to be tested.
• 17:35Yeah. Because then it actually
• 17:37saves a lot of combinations.
• 17:38It's an integration
• 17:40of all the.
• 17:42Yeah. It's a product and
• 17:43product. So you're not really
• 17:44integrating each component separately. Right.
• 17:47Right. It's integrating
• 17:49a product of,
• 17:50logistic from the that he
• 17:51is. Right. I see. Okay.
• 17:52That's the So I don't
• 17:53think it would work necessarily.
• 17:55That's inter it's interesting to
• 17:56to to see if that
• 17:57actually would work because I
• 17:58know the quadrature takes a
• 18:00long time to run. Just,
• 18:02a binary treatment, maybe we'll
• 18:04have that. But Yeah. The
• 18:05progress. I don't know. Right.
• 18:07And then the other thing
• 18:08was that the working covariance
• 18:09matrix, is it in
• 18:11the working
• 18:13covariance matrix? Because I think
• 18:15Eric has a paper showing
• 18:16that if you have non
• 18:18independent, then you could bias.
• 18:20It gets yeah. If you
• 18:21have non independent in the
• 18:22longitudinal
• 18:23setting, I think it probably
• 18:24also applies to the network
• 18:26setting.
• 18:27I think it does. Like,
• 18:28cluster setting in general. Although,
• 18:30actually, right, one of our
• 18:31former students tried to argue
• 18:32that it didn't, but I
• 18:33wasn't totally convinced.
• 18:35Oh, and it did anything
• 18:36about that. In the network,
• 18:37it didn't. But Yeah. Interesting.
• 18:39We can go back to
• 18:40one of the other thing,
• 18:41actually, Ben, that would be
• 18:42a good seed to plant.
• 18:43Is there a relationship between
• 18:44informative cluster size and that
• 18:46bias that comes up
• 18:47without using the independent working
• 18:49correlation measures, or is that
• 18:50just a coincidence?
• 18:52I think this is already
• 18:53assuming a way, though. But
• 18:54once we focus our
• 18:56interest on the estimate parameter
• 18:58wise by the marginal structural
• 19:00model, we assume that's correct.
• 19:01And I think those concerns
• 19:02basically go over. Yeah. If
• 19:03you formalize the estimates nonparametrically,
• 19:06then it becomes an issue.
• 19:07That's I guess, so it
• 19:08was like it was like
• 19:09a significant incident that independent
• 19:11working correlation is supposed to
• 19:12possible solution or Yeah. Cluster
• 19:14size. And Right. It is
• 19:16possible. Okay. Right. But then
• 19:18that requires us to map
• 19:20the marginal structural parameter to
• 19:22a nonparametric call to estimates
• 19:23using line one by zero.
• 19:24We can know this is
• 19:26not explicit here. Yep. But,
• 19:28yes, I think that's also
• 19:29protection using independent scoring correlation.
• 19:32Yeah. So we did that
• 19:33here. That's good. But the
• 19:35terms that's Great to know.
• 19:35Yeah. For both reasons, actually
• 19:37That's great to know. They're
• 19:37not entirely convinced that that
• 19:39problem goes away in the
• 19:40network setting and then also
• 19:41to kind of skirt around
• 19:42the informative cluster size.
• 19:46Lots of things about Yeah.
• 19:49So here is the results.
• 19:51I show the two kind
• 19:53of two types of results.
• 19:55One is using the unadjusted
• 19:56model, not just for any
• 19:58coherence.
• 19:59The other one,
• 20:00is using the MSM adjusted
• 20:03by, using the IPW
• 20:04adjusted for the compounding.
• 20:06And,
• 20:08to interpret the
• 20:10results for the adjusted model,
• 20:12so the direct effect
• 20:15we got the estimation of
• 20:16the risk ratio is zero
• 20:17point fifty seven seven, which
• 20:20means that the risk of
• 20:21HIV infection were decreased
• 20:24by forty
• 20:25forty two point three percent
• 20:27under the HIV testing versus
• 20:29no testing,
• 20:30with the HIV testing for
• 20:32others in your, in the
• 20:33same homestead keep cons constant.
• 20:36And the spillover effect, the
• 20:38estimates of risk ratio is
• 20:40zero point seventy six two.
• 20:42And,
• 20:43we can it it can
• 20:45be
• 20:47it can be integrated as
• 20:49the HIV
• 20:51infection risk were decreased by
• 20:54twenty three
• 20:55point eight percent
• 20:57compared to the when the
• 20:59individual themselves,
• 21:01doesn't receive any, HIV test,
• 21:04but their neighbors
• 21:05all of their neighbors receiving
• 21:07the HIV test versus all
• 21:09of them neighbors not receiving
• 21:11the HIV test.
• 21:12The
• 21:13HIV risk,
• 21:15risk, HIV infection risk will
• 21:18decrease by
• 21:19the percentage of component eight
• 21:21percent.
• 21:22But, we can say that
• 21:23the active factor, the ninety
• 21:24five percent confidence interval span,
• 21:26didn't span zero. So
• 21:28it didn't span one, so
• 21:30it mean that this is
• 21:32is equally significant.
• 21:33But for the spillover factor,
• 21:36the ninety five percent confidence
• 21:38interval actually span one. So
• 21:40this is not a statistically
• 21:42significant results.
• 21:48And, we also do the
• 21:50this is, alpha one minus
• 21:51alpha zero. It's just a
• 21:53different,
• 21:54allocation strategy, different proportion of
• 21:57neighbors
• 21:58back in the treatment.
• 21:59And I tested I tested
• 22:01the different,
• 22:04comparison.
• 22:05So when the difference between
• 22:06r r one r one
• 22:08alpha one and alpha zero
• 22:10is zero, is zero point
• 22:11two, zero point four
• 22:13until the zero point eight
• 22:14and one. So what work
• 22:16is the risk ratio?
• 22:17And all the results
• 22:19are pretty similar. They have
• 22:21the,
• 22:23risk ratio is less than
• 22:24one, but the ninety five
• 22:26percent confidence interval
• 22:28always, unspun one, which is
• 22:30suggested
• 22:31the results is not statistically
• 22:34significant.
• 22:34And this is the results
• 22:36for the,
• 22:37just the one time
• 22:39exporter, not the time variant
• 22:41one.
• 22:41So next step, we will
• 22:43move to the time variant
• 22:45one.
• 22:46First of all, I need
• 22:47to mention that this part
• 22:48is there,
• 22:49in progress. So,
• 22:51there will not get the
• 22:52results to present today. But,
• 22:55I will show some,
• 22:58some plan and some
• 23:02steps we we are we
• 23:04already,
• 23:05went through in the process.
• 23:08So first of all, to
• 23:09prepare for the, the human
• 23:11data for the for this
• 23:12analysis,
• 23:14we need to do some,
• 23:16exploration.
• 23:17So for the, observations,
• 23:20the present methods,
• 23:21which both have valid HIV
• 23:23status.
• 23:24HIV status is amusing. We
• 23:26need to remove them. And
• 23:28if the person,
• 23:29has don't have the HIV
• 23:31inactive neighbors
• 23:32at the baseline visit, We
• 23:34also need to remove all
• 23:36the visits related to this
• 23:37person because,
• 23:39if there's no HIV negative
• 23:42neighbors, there's
• 23:43it is impossible to have
• 23:45any spillover effect.
• 23:47And, third, we also need
• 23:48to remove all the people
• 23:50who were HIV positive at
• 23:52the baseline because our outcome
• 23:54is HIV instance. If you
• 23:56are already HIV positive at
• 23:57the baseline, then there is
• 23:59no point outcome is already
• 24:01there.
• 24:02So we need to remove
• 24:03all the visits
• 24:04for the individuals who were
• 24:06HIV positive at the baseline
• 24:08at the beginning.
• 24:09And we also need to
• 24:11remove the,
• 24:12person also,
• 24:14the person who only has
• 24:16one visit.
• 24:17So in that way, we
• 24:18won't we won't have any
• 24:20outcome. We only have one
• 24:21visit. We need at least
• 24:22two visit to cancel baseline
• 24:25and outcome.
• 24:26So,
• 24:27we will also remove also
• 24:29the last visit and last
• 24:31observation because for the last
• 24:32observation,
• 24:33there won't be next next
• 24:35visit. So there's no next
• 24:37HIV status
• 24:39confirmation, so we don't have
• 24:40the outcome anymore.
• 24:42And, also,
• 24:43once the,
• 24:44if once the information
• 24:46here as this person has
• 24:48convert has got got the
• 24:50HIV serial conversion, we will
• 24:52censor the this person's,
• 24:54and
• 24:55remove all the follow-up phrases
• 24:57after this confirmed the positive.
• 25:00So this is, what we,
• 25:02we did with the when
• 25:04building the longitudinal data site.
• 25:06There's some other things that
• 25:08we went through and that
• 25:09we need to make the
• 25:10decisions for the longitudinal data
• 25:12site.
• 25:13For example, how to determine
• 25:15the time, this creative time,
• 25:17and
• 25:18how to make the,
• 25:19how to summarize the neighbors'
• 25:21progress
• 25:22when they get the, visits
• 25:24at a different date.
• 25:26Something like this.
• 25:27And,
• 25:29so for the locations,
• 25:31this is,
• 25:33similar, but we have more
• 25:35locations
• 25:36for this longitudinal analysis. Although
• 25:38we have a new factor
• 25:39time. So the k and
• 25:41the I are still the
• 25:43same as before. K is
• 25:44a homestead and I is
• 25:46a individual.
• 25:47And, we have a new
• 25:48factor, j is the index
• 25:50of ten pounds.
• 25:52And the the alpha is
• 25:53the still Bernoulli probability of
• 25:55the individual explorer in the
• 25:57homestead.
• 25:58We refer to allocation strategy.
• 26:00And pure is number of
• 26:01the cover rates,
• 26:03and the x is, again,
• 26:05the cover rates,
• 26:07the cover words and the
• 26:08display from the index and,
• 26:10with always r two bar.
• 26:12It's just the different type
• 26:13of cover words. For example,
• 26:15this is,
• 26:16x k I j. It's
• 26:18just the original,
• 26:21it's just as a, covariance
• 26:23vector for the individual I
• 26:25in the homestead k at
• 26:27the time j, and the
• 26:28x x bar k I
• 26:30j
• 26:31is the,
• 26:33the coverage
• 26:34the matrix
• 26:35of j times t that
• 26:37denotes the coverage of individual
• 26:39I I in
• 26:41the,
• 26:42homestead k,
• 26:44all the coverage history from
• 26:46time one up to time
• 26:48j, including up to and
• 26:50including the time j. And
• 26:52the the expert page
• 26:54is the, history of the
• 26:55covers for all the individuals
• 26:58in the homestead k,
• 27:00up to and, including the
• 27:02time j.
• 27:04So similar things apply to
• 27:06the expo
• 27:07exposure and the outcome,
• 27:09and, we have more on
• 27:10different index.
• 27:12The a k minus I
• 27:14j means the vector of
• 27:16treatment
• 27:17at, at for the all
• 27:19the individuals in the homestead
• 27:21k excluding the individual I,
• 27:24up to and,
• 27:26including the time j. And
• 27:28the
• 27:29a bar k I j
• 27:30is the history,
• 27:33individual treatment history for
• 27:35the individual eye in the
• 27:36homestead
• 27:37pay,
• 27:39actual and including the time
• 27:41j. So all the, cover
• 27:43all the locations with bar
• 27:45means the, history rate, and
• 27:48the without bar means just
• 27:49about current
• 27:50time point.
• 27:52And the y k I
• 27:52j is the upload of
• 27:54time point. The individual I
• 27:55is the homestyle
• 27:57k at the time j.
• 27:59And the a k j
• 28:01where AKG bar will be
• 28:02the,
• 28:03matrix of j times n
• 28:05k that denotes all the
• 28:07exposed explorer phase query,
• 28:09of the all in all
• 28:11individuals in the,
• 28:13homestead pay at time.
• 28:17All the pays period from
• 28:18the temp one to and
• 28:21including.
• 28:22So this is how we
• 28:24do the annotations.
• 28:27So we will use the,
• 28:29this is just
• 28:32like, to show how the,
• 28:34relationship
• 28:35how the relations,
• 28:36between the different,
• 28:38explorer outcome is. So here
• 28:40we have the two person.
• 28:42The,
• 28:43a one one zero, it's
• 28:44means the individual one in
• 28:45the homestead one at the
• 28:47time zero, and the one
• 28:48two zero is individual two
• 28:50in the homestyle one at
• 28:51time zero. And then we
• 28:52have the y,
• 28:54one one one means that's
• 28:55the in, outcome of individual
• 28:58one in the homestyle one
• 29:00at time one. So you
• 29:02the as we can say,
• 29:03the,
• 29:04exporter at time zero will
• 29:06affect the exporter at the
• 29:08time one, and we have
• 29:09the cover as x one
• 29:10one zero here. It's the
• 29:12cover as of h f
• 29:14of individual I
• 29:16in, at homestead I,
• 29:18homestead one at time zero.
• 29:20And this work is,
• 29:22on converters. It will affect
• 29:24both the exporter exporter and
• 29:27the outcome.
• 29:28So the course is also
• 29:30at time zero. And, besides
• 29:32this so all the red
• 29:34arrows that I put here,
• 29:36it means the effect of
• 29:37the individual's own exporer history
• 29:40and the cover words.
• 29:42And, for the blue arrows
• 29:43like this, from the individual
• 29:46two, the explorer at time
• 29:47zero
• 29:48to the outcome of individual
• 29:50ones,
• 29:51at time one. This is
• 29:53the spillover effects.
• 29:55So one individual's x treatment
• 29:57ratio will affect another's outcome.
• 30:00And the black arrows is
• 30:01just the relationships
• 30:02influencing the ex exporters
• 30:05so the confounders can,
• 30:07include in the exporters.
• 30:09And the the cover rates
• 30:11from other there's other members
• 30:13in the same home side
• 30:14and also affect this person's
• 30:17export
• 30:18exporter.
• 30:21Yeah. The last type of
• 30:23the arrows I have here
• 30:24is the gray arrows. This
• 30:26is,
• 30:27yeah. I already mentioned this.
• 30:29I think that's all the,
• 30:31different types of possible
• 30:33connections
• 30:34between these,
• 30:36variables
• 30:37at fault for now. And
• 30:38we can see how complex
• 30:39it is already.
• 30:41Just the giving is the
• 30:42two person and three time
• 30:44points.
• 30:47And then there is no,
• 30:48connection between the outcome
• 30:51the explorer because,
• 30:53our outcome
• 30:54Sorry.
• 30:55Get rid of this one
• 30:56sec.
• 30:58Sorry,
• 31:00Okay. Okay. Thank you.
• 31:02Thank you.
• 31:03Okay. So No.
• 31:05No. It's not gonna play
• 31:06play.
• 31:17It says you are muted,
• 31:18so I don't.
• 31:20Oh, actually, you alright. I
• 31:21think we're good now.
• 31:23Go ahead. Oh, yeah.
• 31:25So we don't have the,
• 31:27in I we don't have
• 31:28the I rows from the,
• 31:31y one one one one
• 31:32one one one two eight
• 31:33one one one is because
• 31:35all our,
• 31:37our outcome is HIV. So
• 31:39it's it's server converter or
• 31:40not. Once it is HIV
• 31:42server converter, we will censor
• 31:44this individual. So all of,
• 31:46available queries outcome come will
• 31:48just be zero. So there's
• 31:50no difference.
• 31:51There's no
• 31:52error at all.
• 31:56And then we use a
• 31:57similar method. Just, this time,
• 31:59we need two assumptions. We
• 32:00need to get the,
• 32:03we we need to add
• 32:04add some factor the new
• 32:05factor time. And, now we
• 32:07had the that's just the
• 32:09exchangeability,
• 32:10and we have the sequential
• 32:11condition
• 32:12exchangeability.
• 32:13So why I k I
• 32:15j is the outcomes of
• 32:17individual I I from side
• 32:18of k,
• 32:20f I j.
• 32:21And, this is
• 32:22a similar idea as before.
• 32:25The outcome is independent from
• 32:27the treatment of assignment
• 32:29given the, condition on the
• 32:31cover rates and the the
• 32:32treatment history.
• 32:34So positive is similar,
• 32:36as before. It says that,
• 32:39the probability
• 32:40of receiving the treatment given
• 32:42any given the,
• 32:44cover rates
• 32:46history and the treatment history
• 32:49is larger than zero for
• 32:51any cohort's
• 32:52history and the any treatment
• 32:54history.
• 32:55And we have the temporal
• 32:57partial inference assumptions.
• 33:00This is the thing that
• 33:01is the outcome
• 33:02Only the outcome of individual
• 33:05only depends on the
• 33:07on their own treatment and
• 33:08the the treatment,
• 33:10of the members from the
• 33:12same homestead,
• 33:13not from other homesteads,
• 33:16And, also the treatment history
• 33:19involved of them. And the
• 33:20treatment evaluation,
• 33:22irrelevance, this is the same
• 33:23as before. We just assume
• 33:24there's one variant of treated
• 33:26and treat
• 33:28one variant of treatment and
• 33:29one variant of no treatment.
• 33:35So the estimate, similarly, we
• 33:37have the,
• 33:40log by new model. So
• 33:41we have the expectation
• 33:43of the outcome.
• 33:46Natural outcome is the
• 33:48a function of,
• 33:51treatment history individual treatment history
• 33:53and the neighbor's treatment history.
• 33:56And the the direct effect
• 33:57and the spillover effect is
• 33:59well defined in terms of
• 34:01risk ratio.
• 34:07So we have the similar
• 34:09thing for the estimator
• 34:11just adding the, new factor
• 34:13time. And,
• 34:15in this case,
• 34:16we need to have some
• 34:18history treatment, not just as
• 34:19one treatment.
• 34:21And the stabilized ways is
• 34:23similar as before. It's the
• 34:26now the numerator is the
• 34:28probability
• 34:29of,
• 34:30the treatment
• 34:32receiving treatment condition on the
• 34:34treatment history,
• 34:36not just the one time
• 34:37treatment.
• 34:38And the denominator,
• 34:40again, will depends on not
• 34:41just the the treatment history,
• 34:44but also depends on the
• 34:45cover rates.
• 34:47And
• 34:48the address are similar as
• 34:50the previous one.
• 34:54Okay. So that's all the
• 34:55results I have so far.
• 34:56And, based on the Ivanovo
• 34:58and lastly's results, it suggested
• 35:00that HIV testing
• 35:02is, has a statistically
• 35:04significant
• 35:05protective
• 35:06effect
• 35:07on the HIV,
• 35:08infection.
• 35:09But the,
• 35:10spillover effect results suggested that
• 35:12HIV testing
• 35:14may have a protective,
• 35:17effect on the HIV infection,
• 35:20but, we need more start
• 35:22need more study on this.
• 35:23Hopefully, the longitudinal
• 35:25study, analysis without
• 35:27can give us some new
• 35:28insights.
• 35:30And,
• 35:31the limitations
• 35:33first of all, the current
• 35:35analysis is for one time
• 35:37exporters. So the longitudinal,
• 35:39analysis results may differ from
• 35:41this.
• 35:42And the second way has
• 35:44that,
• 35:45it may have the model
• 35:46miss misspecified
• 35:48misspecified.
• 35:49And,
• 35:51the third
• 35:52plan possible limitation is that
• 35:54there may be some random
• 35:55error in the data because
• 35:56this is the data from
• 35:58observational study,
• 35:59not the random randomized trial.
• 36:01And,
• 36:03the last one is we
• 36:04may have some unmeasured components,
• 36:06which assumes there's no unmeasured
• 36:08components.
• 36:09There may be
• 36:12some. So for future directions,
• 36:15I will continue,
• 36:16working on the longitudinal analysis
• 36:18for the HIV testing to
• 36:20see which results I will
• 36:22got and I will get.
• 36:23And,
• 36:25next step, I will also
• 36:26investigate the effects of other
• 36:28commonly used inventions,
• 36:31being HIV state instance of
• 36:33the HIV based behavior,
• 36:35including the ART retention, the
• 36:37link here to pair, the
• 36:38mirror circumcisions,
• 36:39papers.
• 36:41And,
• 36:42after that, we're we're moved
• 36:44from the assessing one single
• 36:46intervention
• 36:47to the intervention package. So
• 36:49intervention package is a package
• 36:51including multiple intervention components.
• 36:54And we were developing method
• 36:56a new method for I'm
• 36:57looking for causal effects
• 36:59of the intervention package. We're
• 37:00tweaking all the intervention components
• 37:02as part of the package,
• 37:04not component components.
• 37:06So that's a future plan.
• 37:09And these are some selected
• 37:11references.
• 37:13Also, at the end of
• 37:14the,
• 37:15talk, I want to,
• 37:17show I want to just
• 37:19announce two job openings at
• 37:21the URI. One is the
• 37:23tenure track assistant professor in
• 37:26and,
• 37:27has a,
• 37:28web link here.
• 37:30And,
• 37:31it will start it's a
• 37:32nine months,
• 37:34position. We're beginning of fall
• 37:36twenty twenty six.
• 37:38Another one is a post
• 37:39doc,
• 37:42in the CIFARIS research team,
• 37:43and, this position requires a
• 37:46focus on the code inference
• 37:47in the networks for the
• 37:49HIV prevention.
• 37:50And,
• 37:51if you are interested in
• 37:53this, you can contact the
• 37:54doctor Ashley Pukanda,
• 37:56and I will put the
• 37:57email address there. Shameless plug.
• 38:00That must be the email
• 38:01address. And then you're not
• 38:03now looking for jobs. I
• 38:04know. Right? Let me see
• 38:05this.
• 38:06Yeah. I can also I'm
• 38:07on the search committee for
• 38:08the first one. So if
• 38:09anybody has any questions.
• 38:11Is it bias targets or
• 38:12biosphere?
• 38:13Biosphere. Biosphere.
• 38:16But it'll be the second
• 38:17biosphere station on campus, so
• 38:19I have a vested interest
• 38:20in this person. I'm saying,
• 38:21keep it with the pipe.
• 38:22I know. I saw that
• 38:23when I was up there,
• 38:24I was like, you know,
• 38:25this thing. That's it.
• 38:27Yeah.
• 38:28Wait. If you know what
• 38:28I do