diff --git a/docs/investigation_into_drivers_of_heterogeneity.md b/docs/investigation_into_drivers_of_heterogeneity.md
new file mode 100644
index 00000000..5aebc5b8
--- /dev/null
+++ b/docs/investigation_into_drivers_of_heterogeneity.md
@@ -0,0 +1,109 @@
+---
+editor_options:
+  markdown:
+    wrap: 72
+---
+
+## Wastewater forecasting evaluation: Investigation into drivers of heterogeneity in the impact of wastewater on forecast performance
+
+This document is intended to outline concretely the proposed analyses
+requested as a follow-up to the previously decided on [evaluation
+plan](https://github.com/cdcgov/wastewater-informed-covid-forecasting/tree/prod/docs).
+This evaluation focused on two major components: - retrospective
+comparison of forecast performance with and without wastewater using
+vintaged datasets of COVID-19 hospital admissions and wastewater at the
+jurisdiction and wastewater treatment plant level, across the 2023-24
+epidemic season - comparison of the performance of the
+wastewater-informed model to other models submitting to the COVID-19
+forecast Hub, both in real-time (when submitted from Feb - March of
+2024) and retrospectively (using the real-time COVID Hub submissions and
+our retrospectively produce submission), to be presented in the
+manuscript: "Bayesian generative modeling for heterogeneous wastewater
+data applied to COVID-19 forecasting". In the interest of maintaining
+scientific integrity, we specified the plan for this evaluation prior to
+performing the analysis.
+
+The goal of the proposed analyses here is to investigate, using the
+empirical wastewater data, hospital admissions data, model performance
+metrics, and the model parameters, the potential drivers of the
+heterogeneity in the relative performance of wastewater. This analysis
+is intended for hypothesis generation, rather than hypothesis testing or
+assigning any causal relationships, as we believe a full-fledged
+analysis of the impact of different characteristics of the wastewater
+and hospital admissions data on forecast performance is out of scope for
+this paper and should be performed as a separate independent analysis.
+
+We are writing this analysis plan with the intention of coming to a
+consensus on the scope of the required additional analysis and the form
+of the presentation of the results, prior to running the analysis, again
+with the intention of promoting scientific integrity and holding
+ourselves accountable to presenting the results in an unbiased manner.
+
+We plan to address the following questions via the proposed analyses:
+
+All of the planned analysis will focus on the retrospective model
+performance with and without wastewater data.
+
+1.  How strong is the correlation in the recent trend in wastewater
+    signals (with one another), and is that consistent with the eventual
+    trend in hospital admissions data? How does it impact performance?
+
+-   We will quantify correlation in recent trend in wastewater signals
+    using only data from the 2 weeks prior to the forecast date. We will
+    pool data from all sites in those two weeks, calculate a correlation
+    coefficient, and estimate the instantaneous exponential growth rate
+    in the observed data. We will compare this to an estimate of the
+    exponential growth rate in the hospital admissions using the
+    evaluation data from 1 week prior to the forecast to 2 weeks beyond
+    the forecast date, in an attempt to characterize the trend of
+    hospital admissions into the forecast period. Next, we will bin
+    forecasts into the following categories:
+
+    -   high correlation in wastewater signal, trend in same direction
+
+    -   high correlation in wastewater signal, trend in opposite
+        direction
+
+    -   low correlation in wastewater signal, trend in same direction
+
+    -   low correlation in wastewater signal, trend in opposite
+        direction
+
+We will then display and summarize the distributions of forecast
+performances (average and relative CRPS of the full 38 day horizon for
+an individual forecast date and location) in each of the bins.
+
+2.  How did variability in the wastewater data impact forecast
+    performance?
+    -   We will quantify the variability in wastewater data in two ways
+        1 by computing the coefficient of variation across the time
+        series for each wastewater treatment plant (returning a
+        distribution of CVs) and 2 by looking at the posterior estimate
+        of the mean obsrvation error across sites.
+
+    -   We will summarize the "average variability in the wastewater
+        signal" for each location and forecast date by taking the mean
+        of each of the empirical and model-based distributions, and we
+        will plot the mean variability from both methods compared to the
+        forecast performance (average and relative CRPS of the full 38
+        day horizon for an individual forecast date and location)
+3.   How much does latency impact forecast performance?
+    -   We will bin forecasts by those containing more than 5 sites with
+        wastewater concentration data within the last 20-15, 14-11 days,
+        10-8 days, and 7-0 days of the forecast date.
+
+    -   We will present violin plots demonstrating the distribution of
+        forecast performance (average and relative CRPS of the full 38
+        day horizon for an individual forecast date and location)
+4.  How does relative vs absolute forecast performance compare, e.g.
+    does wastewater improve forecasts when the performance would have
+    otherwise been poor, or vice versa?
+    -   scatterplot of the relative CRPS of the wastewater informed
+        model versus the absolute CRPS of the hospital admissions only
+        model
+
+    -   scatterplot of the absolute CRPS of the wastewater model versus
+        the absolute CRPS of the hospital admissions only model
+
+We plan to add the proposed plots and analysis to the supplement of the
+manuscript.