To do in main/casual models

• Add concurrency parameters
• Add absolute difference in square root of ages
• Add nodemix terms
• Truncate casual degree to 4 max ongoing partnerships

Current Main Formation Specification

main_formation <- ~edges +
                  nodefactor("race.eth", levels = -1) +
                  nodefactor("age5", levels = -1) +
                  nodefactor("deg.casl", levels = -1) +
                  concurrent +
                  nodematch("race.eth") +
                  nodematch("age5") +
                  nodematch("ai.role", levels = c("R", "I"))

Ribbon: middle 95% of simulations

Issues/Questions

• Removed degrange term altogether: since main degree truncated at 2, redundant to have both a concurrency term and a term specifying the number of nodes with degree of ≥ 2?

• Tried an agemix term, but many coefficient estimates returned as NA, similar issues when using absdiff(sqrt_age): went with a nodematch term for age

Current Casual Formation Specification

casl_formation <- ~edges +
                  nodefactor("race.eth", levels = -1) +
                  nodefactor("age5", levels = -1) +
                  nodefactor("deg.main", levels = -1) +
                  concurrent +
                  degrange(from = 4) +
                  nodematch("race.eth")

Ribbon: middle 95% of simulations

Issues/Questions

• If I add either nodematch("age5") or nodematch("ai.role", levels = c("R", "I")), ergm kicks back an "excessive correlation between model terms" warning.

To do in one-time partnerships:

• Estimate one-time contact rate
• Specify formation model

Document created variables and write up relevant methods from DataCleaning.Long.R and the network stats/estimation files.

Adjusting Target Stats

• Unfortunately, I wasn't able to finish adjusting the target stats today to get the models running with age- and race/ethnicity nodematches.

• For instance, I think my original target stats were such that the calculations for the nodematch terms implied different numbers of edges than the edges term, and I suspect the same goes for the other terms in the formation models.

• I'm still in the process of recalculating the target stats using a standardization scheme that I think should fix these inconsistencies and hopefully allow for inclusion of the terms that are currently causing problems. Will discuss on our call.

Node factor calculation for causal degree by main degree implies incorrect number of total edges in casual formation model (about 1,000 edges off).

Need to fix a kludge for correcting imputed partnership start dates that were filled in as occurring after sub_date.

Doesn't affect very many records. Low overhead and low priority for the moment.

Main/casual model estimations

• The model diagnostics look really dicey, particularly the main, but I still have to add a couple of parameters (estimated, just not added to the models yet). SEE UPDATE IN COMMENT THREAD

• Could not accounting for heterogeneity in dissolution rates contributes to this? I saw in the ARTNet paper that the dissolution rate varies by race/ethnicity and age

• Currently focusing on the demographic inputs to the model, but will eventually add HIV diagnosis status too.

MCMLE procedure

• Each time I run the models, the number of steps to convergence (when achieved) differs each time

  • Natural, or am I missing where to set a random seed?

• Despite the procedure's reporting convergence, many of the trace plots don't look great to me.

Estimates of ongoing partnerships

• For the most part I'm getting the same or very similar answers, comparing them to the Epidemics paper, but I can't seem to get the ongoing/inactive partnership breakdowns within main and casual partnerships, respectively, to match.

• A couple of small differences in how I analyzed that I don't think affect the estimates:

  • Imputed age slightly differently
  • Assigned partnership start date differently, but ongoing partnerships flagged based on PARTXONGOING and PARTXMAINONG.

• Perhaps I've assigned ongoingness incorrectly (see: 02b-DataCleaning-Long.R)?