analyse-NEED-EULF-2014-models-v2.0.do

* Script to analyse DECC's 2014 EULF NEED data to:
* - investigate % variance of energy consumption due to dwelling type variables as a way to infer the % of variance due to people

* NB this script uses 2 data files derived from the original data using the 'process' script

* Original data available from: UK DATA ARCHIVE: Study Number 7518 - National Energy Efficiency Data-Framework, 2014
* http://discover.ukdataservice.ac.uk/catalogue/?sn=7518

/*   

Copyright (C) 2014  University of Southampton

Author: Ben Anderson (b.anderson@soton.ac.uk, @dataknut, https://github.com/dataknut) 
	[Energy & Climate Change, Faculty of Engineering & Environment, University of Southampton]

This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License 
(http://choosealicense.com/licenses/gpl-2.0/), or (at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

#YMMV - http://en.wiktionary.org/wiki/YMMV

*/

clear all

capture noisily log close

* written for Mac OSX - remember to change filesystem delimiter for other platforms
local home "/Users/ben/Documents"
local proot "`home'/Work/Data/Social Science Datatsets/DECC"
* for clam
* local proot "`home'/Work/NEED"
local dpath "`proot'/NEED/End User Licence File 2014/processed"
local rpath "`proot'/results/NEED"

*local version "1.0"
* initial models - all households for electricity models

*local version "1.1"
* restrict to gas only households to avoid complications of:
* - primary electric heating (presumably)
* - oil heating

*local version "v2_1pc"
*local sample 1
*local sampleby "EE_BAND PROP_TYPE"
* changed from using log consumption to consumption decile to avoid complications due to variable rounding ranges in original data (see readme)
* restricted analysis to households where gas is main heat source as it is better predicted by variables included & is more relevant to EPC (heat)
* uses 1% sample (c 30k) making sure keep proportions of property type and EE_Band to see if linktest fails with smaller n

local version "v2_10pc"
local sample 10
local sampleby "EE_BAND PROP_TYPE"
* uses 10% sample (c 300k) making sure keep proportions of property type and EE_Band to see if margin plots and co-efficients are the same 
* (linktest etc will probably now fail due to larger n)

*local version "v2_100pc"
*local sample 100
*local sampleby "EE_BAND PROP_TYPE"
* uses full sample (c 3m) to see if margin plots and co-efficients are the same 
* (linktest etc will probably now fail due to larger n)

set more off

log using "`rpath'/analyse-NEED-EULF-2014-models-`version'-$S_DATE.smcl", replace

* use the pre-processed wide form file which contains all years of consumption data but not the constant values which are in the xwave file
use "`dpath'/need_eul_may2014_consumptionfile_wide_`sample'pc.dta", clear

* we're going to use 2012 data only

keep HH_ID *2012*

* merge in the pre-processed cross-year fixed values file
merge 1:1 HH_ID using "`dpath'/need_eul_may2014_xwavefile_`sample'pc.dta"

* check what's valid
tab Gcons2012Valid Econs2012Valid, mi // O = off gas, V = valid, L = too low, G = too big, M = missing
tabstat Gcons2012, by(Gcons2012Valid) s(mean min max n)
* do off-gas use a lot more electricty (heating)?
tabstat Econs2012, by(Gcons2012Valid) s(mean min max n)

histogram Gcons2012, by(MAIN_HEAT_FUEL, total) name(histo_Gcons2012)
graph export "`rpath'/graphs/histo_Gcons2012_by_main_heating_fuel_`version'.png", replace

tabstat Gcons2012, by(MAIN_HEAT_FUEL) s(n mean min max)

* keep if valid gas & gas = main heat fuel
keep if Gcons2012Valid == "V" & MAIN_HEAT_FUEL == 1

***** random sample ****
* select a random sample but ensure proportions of sampleby are kept
di "* Keeping `sample'% sample by `sampleby'"
sample `sample', by(`sampleby')

tab `sampleby', mi

* log the consumption as it's very skewed -> becomes semi-normal & OK for linear regression
* Gcons = gas
* Econs = Electricity

* create log & deciles
* log - creates a normal distribution
* deciles - avoids the consumption rounding range differences (hopefully)
gen log_Gcons2012 = log(Gcons2012)
egen Gcons2012_dec = cut(Gcons2012), group(10)
gen log_Econs2012 = log(Econs2012)
egen Econs2012_dec = cut(Econs2012), group(10)

* combine consumption
* treat missing (gas) as 0
egen Allcons2012 = rowtotal(Gcons2012 Econs2012)

*gen log_Allcons2012 = log(Allcons2012)
egen Allcons2012_dec = cut(Allcons2012), group(10)

* create log consumption quintiles
*egen quinlog_Allcons2012 = cut(log_Allcons2012), group(5)
*egen quinlog_Gcons2012 = cut(log_Gcons2012), group(5)
*egen quinlog_Econs2012 = cut(log_Econs2012), group(5)

* fix some of the variables

* combine IMD: this is a bit dodgy as they are not strictly comparable
gen ba_imd = IMD_ENG
replace ba_imd = IMD_WALES if ba_imd == .

* household level vars
* for models
local generic_hvars "i.BOILER_YEAR i.MAIN_HEAT_FUEL i.LI_YEAR i.LOFT_DEPTH i.FLOOR_AREA_BAND WALL_CONS i.CWI_YEAR i.PROP_TYPE i.PROP_AGE i.EE_BAND"
* for graphs
local generic_hvarsg "BOILER_YEAR MAIN_HEAT_FUEL LI_YEAR LOFT_DEPTH FLOOR_AREA_BAND WALL_CONS CWI_YEAR PROP_TYPE PROP_AGE EE_BAND"
* for models by type
local generic_hvarsnp "i.BOILER_YEAR i.MAIN_HEAT_FUEL i.LI_YEAR i.LOFT_DEPTH i.FLOOR_AREA_BAND WALL_CONS i.CWI_YEAR i.PROP_AGE i.EE_BAND"

* area level vars
local generic_rvars "i.ba_region i.ba_imd"

* define different property types
local ptypes "101 102 103 104 105 106"
local pt101 "detached"
local pt102 "semi"
local pt103 "end_terr"
local pt104 "mid_terr"
local pt105 "bungalow"
local pt106 "flat"

* check for weirdness
su `generic_hvarsg'

* now loop over the energy types & run linear regression models
* NB - the rounding of the consumption values may lead to modelling problems

* add Econs Allcons for electricity & sum of both
* rename so graph names don't break
rename log_Gcons2012 lg2012
rename Gcons2012_dec g2012dec
local vars "lg2012 g2012dec"
foreach v of local vars {
	* check distributions of original consumption values
	
	* all hhs model
	qui: regress `v' `generic_hvars' ///
		`generic_rvars' ///
		i.BOILER_YEAR
	
	est store `v'
	* test a variable
	predict `v'_r, resid
	foreach testv of local generic_hvarsg {
		di "Testing residuals against `v'"
		* based on http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm
		* can't use factor variables in acprplot...
		graph hbox `v'_r, over(`testv') name(rtest_`v'_`testv')
		graph export "`rpath'/graphs/rplot_`v'_`testv'-`version'.png", replace
		* possibly floor area - heteroskedasticity ?
	}
	

	di "* -> `v' estat to test for heteroskedasticity & omitted vars"
	estat ovtest
	estat hettest
	
	* we ought to be testing for linearity too
	di "* -> `v' linktest to test for model specification"
	di "* if p of _hatsq < 0.05 -> mis-spec"
	di "* http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm"
	linktest	
	
	di "* test EPC margins for `v'"
	margins EE_BAND
	marginsplot, name(mplot_`v'_EE_BAND) note("All dwellings")
	graph export "`rpath'/graphs/mplot_`v'_EE_BAND-`version'.png", replace
	
	* models by property type - to see if rsq & coefficients vary
	foreach p of local ptypes {
		di "* -> testing `v' for `pt`p''"
		qui: regress `v' `generic_hvarsnp' ///
			`generic_rvars'	///
			i.BOILER_YEAR ///
			if PROP_TYPE == `p'
		est store `v'_`pt`p''
		
		di "* -> `v' 2012 `pt`p'' - estat to test for heteroskedasticity & omitted vars"
		estat ovtest
		estat hettest
		
		* we ought to be testing for linearity too
		di "* -> `v' `pt`p'' linktest to test for model specification"
		di "* if p of _hatsq < 0.05 -> mis-spec"
		di "* http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm"
		linktest
		di "* test EPC margins for `v' (`pt`p'')"
		margins EE_BAND
		marginsplot, name(mplot_`v'_EE_BAND_`pt`p'') note("Type: `pt`p''")
		graph export "`rpath'/graphs/mplot_`v'_EE_BAND_`pt`p''-`version'.png", replace

	}
	* models for different consumption quintiles - to see if rsq & coefficients vary
	/* doesn't make much sense to do this if using deciles as dependent variable
	foreach q of numlist 0/4 {
		di "* -> testing log_`v'2012 for quintile: `q'"
		qui: regress log_`v'2012 `generic_hvars' ///
			`generic_rvars'	///
			i.BOILER_YEAR ///
			if quinlog_`v'2012 == `q'
		est store rlog_`v'2012q`q'
		
		di "* -> quintile: `q' - estat to test for heteroskedasticity & omitted vars"
		estat ovtest
		estat hettest
		
		* we ought to be testing for linearity too
		di "* -> quintile: `q' - linktest"
		di "* if p of _hatsq < 0.05 -> mis-spec"
		di "* http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm"
		linktest	
	}
	*/
}

* output all the results - that's a lot of t tests!
* we could put them all out in one file but it would be really hard to find the ones you want!
estout lg2012 using "`rpath'/models/NEED-EULF-2014-log-gas-model-`version'.txt", replace cells("b se p _star") stats(r2 r2_a N ll)
estout lg2012_* using "`rpath'/models/NEED-EULF-2014-log-gas-models-by-property-type-`version'.txt", replace cells("b se p _star") stats(r2 r2_a N ll)

estout g2012dec using "`rpath'/models/NEED-EULF-2014-gas-deciles-model-`version'.txt", replace cells("b se p _star") stats(r2 r2_a N ll)
estout g2012dec_* using "`rpath'/models/NEED-EULF-2014-gas-deciles-models-by-property-type-`version'.txt", replace cells("b se p _star") stats(r2 r2_a N ll)

di "* Done!"

log close