Usuário:Abdo/MMMNS/Ampliação da Rede/Who knows who

This is about what do we need "knows" to mean in the scale up survey, given we want to use the Jogo dos Contatos estimates for transmission error.

Summary

The definition of know in the scale-up survey needs to include knowing by name or nickname and adding a question to the scale-up survey.

Matt's notes: Names and nicknames in the definition of know

The headline is that I think that we all think that we should make the following changes

1) use the following definition of "to know" in both the jogo de contatatos and the scale-up survey: "to know someone means that they live in Curitiba; you know them and they know you by sight, name, or nickname; you can contact them in person, by telephone or by mail; and that you have been in contact in the last two years." The change here is that we've added nicknames. The "standard" definition from Killworth et al. is: ". . . you know the person and they know you by sight or by name; you can contact them in person, by telephone or by mail; and you have had contact with the person in the past two years”. (from Killworth et al, 2003 footnote 1)

2) Add a question to the scale-up survey. First we will ask, "How many heavy drug users do you know?" If the response is 1 or more, then we will add the question "How many of them know you by name?"

Now I will explain how this all started

This issue of including nicknames was raised by Maeve our final afternoon after testing the jugo de contatos and seeing that if we use a definition of know that does not include nicknames we will get many fewer contacts. For example, the person she interviewed knew 5 people named Carlos that also knew him, but only 2 of them knew him by name; the rest knew him by nickname. Since many drug users go by the nickname there are fewer people that will know them by name. Since we are already having fewer contacts than we expected this could be a problem.

Then the puzzle become: Why did we remove nicknames when we met in the morning? I think this is because we figured out that if we allow nicknames then the transmission error gets more complicated.

Before we arrived in Curitiba, we thought there were two types of transmission error. If the respondent in the scale-up survey is a heavy drug user and if the respondent in the scale-up survey is not a heavy drug user. We can write this as:

p(a | u) = probability that the respondent is aware that one of his contacts is a heavy drug user if the respondent is a heavy drug users

p(a | ~u) = probability that the respondent is aware that one of his contacts is a heavy drug user if the respondent is not a heavy drug users

The hypothesis is that p(a | u) > p(a | ~u).

Then we realized that in the jogo de contatos that we are only sampling alters that the RDS heavy drug user knows by name; our sampling method will never sample alters that the RDS heavy drug user only knows by nickname. This creates a problem if we think that the flow of information is different between people that know each other by name and people that know each other by nickname. In other words, this creates a problem if the probability of transmission depends not just on the characteristics of the respondent, but also on the characteristics of the relationship. This creates more complication and I think that is what was confusing us. Imagine that there are 4 transmission errors:

p(a | u, kbn) = probability that the respondent is aware that one of his contacts is a heavy drug user if the respondent is a heavy drug users and they know each other by name

p(a | u, ~kbn) = probability that the respondent is aware that one of his contacts is a heavy drug user if the respondent is a heavy drug users and they don't know each other by name

p(a | ~u, kbn) = probability that the respondent is aware that one of his contacts is a heavy drug user if the respondent is not a heavy drug users and they know each other by name

p(a | ~u, ~kbn) = probability that the respondent is aware that one of his contacts is a heavy drug user if the respondent is not a heavy drug users and they don't know each other by name

Our jogo de contatos only allows us to estimates p(a | u, kbn) and p(a | ~u, kbn). However, in the scale-up survey, when the respondent is asked "How many heavy drug users they know?" this includes both people they know by name and by nickname. Imagine for example that the respondent is not a heavy drug user and they say they know 3 heavy drug users. We would apply the correction p(a | ~u, kbn) to each of these 3 alters even if maybe we should apply p(a | ~u, ~kbn) to some of them. I think this is the reason that we decided that we did not want to allow people in scale-up to report on people that they know by nickname.

However, this introduces problems in the jogo de contatos. I think we can switch the scale-up definition to include nicknames if p(a | ~u, kbn) = p (a | ~u, ~kbn). Is it reasonable to assume that in a relationship between a heavy drug user and non heavy drug user that the chance that the non heavy drug user knows is the same whether they know each other by name or by nickname? I'm not Brazilian so I'm not sure. This doesn't seem obviously wrong to me. Also, we should think about if it names sense that p(a | u, kbn) = p (a | u, ~kbn), but this does not matter much because there will be very few heavy drug users in the scale-up survey.

I think it gets even more complicated if A knows B by name, but B knows A by nickname.

What should we do going forward?

As I wrote at the top, I think our consensus is that going forward I think we should:

1) use the following definition of "to know" in both the jogo de contatatos and the scale-up survey: "to know someone means that you know them and they know you by sight, name, or nickname; you can contact them in person, by telephone or by mail; and that you have been in contact in the last two years." The change here is that we've added nicknames.

2) Add a question to the scale-up survey. First we will ask, "How many heavy drug users do you know?" If the response is more than 0, then we will add the question "How many of them know you by name?"

One reason for this second question is that we could potentially apply different transmission errors. For example, I know 3 heavy drug users and 2 know me by name. Therefore, we would apply p(a | ~u, kbn) to 2 and p(a | ~u, ~kbn) to the other 1. The problem is that we don't know p(a | ~u, ~kbn) because we can't estimate it from jogo de contatos. Therefore, I think the only thing this second question is good for is just to see how much of a problem we might have; the more that the heavy drug users that the scale-up respondent knows that know the scale-up respondent by name, the less potential problem we have if (a | ~u, kbn) ~= p(a | ~u, ~kbn).

In the end I think adding nicknames makes the extra assumption p(a | ~u, kbn) = p (a | ~u, ~kbn) and p(a | u, kbn) = p (a | u, ~kbn) in order to get more contacts in jugo de contatos. I think this is a good trade.

Errata

Sorry I was a bit confusing. I wrote:

"Is it reasonable to assume that in a relationship between a heavy drug user and non heavy drug user that the chance that the non heavy drug user knows is the same whether they know each other by name or by nickname?"

What I should have written was:

Imagine a relationship between a heavy drug user and a non-heavy drug user. Do we think that the probability that the non-heavy user is aware that the other is a heavy user depends on whether they know each other by name or if they know each other by nickname?

In other words:, Let

p(a | ~u, kbn) = probability that the respondent is aware that one of his contacts is a heavy drug user if the respondent is not a heavy drug users and they know each other by name

p(a | ~u, ~kbn) = probability that the respondent is aware that one of his contacts is a heavy drug user if the respondent is not a heavy drug users and they don't know each other by name

If we are willing to assume that p(a | ~u, kbn) = p(a | ~u, ~kbn) then I think we are OK. If we are not willing to assume this, then we have a problem that I don't see a solution to. In general one issue with the jogo de contatos is that it only samples relationships where the drug user knows the alter by name. If those relationships are different from other relationships, then our estimate will be biased. I don't see a way around this and I don't think it is a fatal problem. I think we can just note this as a limitation in the paper.