update plans

Author: huisaddison

Diff for: google_symptoms_exploration/plans.md

@@ -0,0 +1,33 @@
+## Plans
+- Overview: Old GHT, new Google Symptoms.  Go through previous notebook (copy
+  that notebook into this directory)
+- Demo correlations notebook: very crude time analysis where we aggregate over 
+  all counties and examine the correlation with respect to time
+- Open problems; each one can be its own notebook:
+    - Amount of missingness by county, symptom.  Should we just discard some 
+      of the symptoms because it's available so rarely?  Many of the symptoms
+      seem _a priori_ totally unrelated to COVID-19, e.g., excessive hoarding,
+      should we toss those out to make our lives easier?
+    - Overwhelming number of symptoms reported: 422.  How to efficiently 
+      perform a correlations analysis for all of them at once?  (We would
+      want to examine over time; across counties; at various time lags
+      against cases)
+    - Correlations analysis can give us a sense of what symptoms to include 
+      in a new indicator; then we would need to weigh them appropriately 
+      into the new indicator.  Unexpected correlations (that are not
+      spurious) may provide new insight.
+- My view on two approaches to a new indicator (we could do both, assuming 
+  we have the people):
+  - "Scientifically guided approach": consult medical / public health 
+    literature & experts to obtain a subset of symptoms, weigh them
+    appropriately, this is the new indicator.  Google considered such an
+    approach, but decided to defer construction to users because of
+    lack of uniform advice across different agencies with which they
+    consulted.  PRO: May give us signal beyond (often unreliable) case
+    reporting.
+  - "Brute force approach": set up supervised problem(s) against cases 
+    (at various lags), through in all of the symptoms, see what allows us
+    to best mimic cases.  PRO: Should be well-correlated with cases; if
+    trained against appropriately backfilled cases, it can give a "nowcast"
+    of cases.
+