Calibration to KAIS 2012 Data

Issue

We lack data on PLHIV Diagnosed for Kenya for all years except 2012, when the 2012 KAIS Survey was published.

"Among all HIV-infected persons in the KAIS 2012, 46.9% self-reported positive based on their last HIV test." - Chapter 8, page 71.

I have been assuming that this equates to 46.9% of PLHIV being diagnosed as of 2012. To convert this to a number I just multiplied the proportion by the estimate of PLHIV I have from Spectrum. The result was 622,733 PLHIV Diagnosed in 2012.

Calibration

We proceeded with model calibration to this data point: screen shot 2016-06-09 at 10 59 14

However, it became apparent that the majority of simulations deviated with this datum and found the number of persons aware of their serostatus appeared to be higher.

Reconciliation

My thoughts on this problem fall into two categories:

1. Issue with KAIS data

According to the 2012 KAIS report, "46.9% self-reported positive". This self-reporting opens up the possibility of considerable bias in the result, both up and down.
Additionally, this survey only reached 18,000 individuals from 8,000 households of ~43M persons living in Kenya. The results are presented as being "nationally-representative" but questions over whether results can be extrapolated to the entire nation remain.
Finally, a proportion of PLHIV diagnosed was reported. I then manipulated this figure into a numeric value for use in the model. If the estimate of PLHIV from Spectrum is low, then the resulting value of PLHIV Diagnosed will also be underestimated.
2. Issue with the structure of the model
The model does not replicate the entire HIV epidemic, nor does it attempt to mimic up-ticks in testing scale up.
The model attempts to calibrate to available data between 2010 and 2015 using a simple structure that captures HIV progression, associated mortality and progression through care. The model uses fixed rates of transition between various state compartments to capture the dynamics of care. These rates are adjusted during calibration to fit to available data.
As rates are fixed over time, the model has difficulty fitting to data representing a significant change in an element of care; for example, testing scale up. In 2015, there were estimated to be ~1M people in care, but in 2012 only 500k people were aware of their infection. The model therefore has difficulty reconciling this substantial change with a single fixed testing rate.
Finally, the model continuously tracks all individuals, including those lost from care. Enumerating "PLHIV Diagnosed" involves counting all individuals who have been diagnosed and are still alive, this includes all patients in care, on ART, lost from pre-ART and lost from ART care. The KAIS study may not have been able to identify and consider persons lost from care, and therefore may mean that directly comparing the model to the KAIS data compares two different measures, and that the model potentially overestimates the values reported by the KAIS.
Conclusion

The model has been designed to be a simple tool to describe the HIV cascade for multiple countries and provide estimates of where gaps in care exist. The model takes into consideration available data on the cascade and imposes a strict structure of the dynamics of care to make sense of data from various and potentially conflicting sources. Additionally, the model must be able to easily adapt to countries with drastically different epidemics. As a result, model calibration may result in a compromise between irreconcilable data sources, but aims at identifying parameter sets that fit the majority of data.

jackolney / WhoCascade