• No results found

In Section 4.7 we saw that the configuration of contraints—the selection and ordering—may affect the simulation. In practice my choice of con- straint was limited to variables that were available in both the census and

Understanding Society, but this still offered a rich pool of information

to choose from. To ensure the constraints I chose created an optimal model I selected them based on theory (see Section 2.7) and empirically tested these to ensure they correlated well with my target variable. This section describes how I made my initial selection of constraints, how I

matched these in both the census and Understanding Society, and how I ensured the populations in the census matched across variables. In the next section I describe how I empirically tested their correlation with my target variable, then how I ordered these for entry into the simulation.

5.5.1

Initial choice

My choice of constraints included essential demographic information, such as age, ethnicity, and sex. I also chose to include highest education qualification as this has a known relationship with health outcomes. I also wanted to include an indicator for poverty or deprivation, as this is known to be a powerful predictor of health outcomes and inequality (see Section 2.7). There is no income variable in the census, so instead I used indirect measures as indicators of poverty, including car availabilty, housing quality, and home ownership (Senior, 2002). These are asked in both the census and in Understanding Society so can be used as constraint variables.

I am not suggesting that the lack of these amenities or non–ownership housing tenure should be enough to consider a household in poverty. Instead they may indicate reduced or limited financial resources which is linked to poorer health outcomes. Similar measures have been used as markers of deprivation, for example by Townsend et al. (1988).

My final choice of constraints for this pilot were: age; ethnicity; sex; highest educational qualification; car or van availability; and housing tenure.

5.5.2

Matching variable levels

In order to be used as constraints the variable levels must match precisely between the survey and the census data (Tanton and Edwards, 2013b). The definitions used in many demographic variables matched exactly, so

no further recoding was required. In some cases it was necessary to recode some variables before performing the logistic regression tests. These are performed by the scripts data-raw/0-prep-understanding-society.R for survey variables and data-raw/2-simulate-pilot.R for census vari- ables, and are described in brief here.

Ages were given as individual years in Understanding Society, so were recoded or ‘cut’ into bands that matched those already available in the census.

Ethnicity was only available with a reduced classification in the census with age information, which was necessary to remove children and young people aged under 16 from the constraints. To match the census, I therefore had to recode ethnicity in Understanding Society to match the reduced levels available in the census with age information by combining some ethnic groups.

Tenure was grouped differently in the census and in the survey. Those who owned their home with a mortgage or loan and those with shared owner- ship were grouped together in the census but separate in Understanding

Society so were recoded in the survey to merge these two groups. In the

census private renters and those living rent–free were grouped together, while social renters were a separate classification. In Understanding So-

ciety both private renters and social renters were grouped together. To

address this it was necessary to amalgomate private renters, social renters, and those living rent–free in to one group in both the census and the survey. This was less than ideal, but the only classification the two data sources allowed.

Education was perhaps the most challenging variable to prepare for comparison between Understanding Society and the census. The census provides the highest qualification by level, from no qualifications, through Level 1 qualifications, Level 2 qualifications, apprenticeships, Level 3 qual- ifications, Level 4 qualifications and above, and finally other qualifications not captured with these categories. This allows regulated qualifications in

England and Wales from different frameworks to be compared (Gov.uk, 2016a).

Understanding Society, on the other hand, provides the highest education

qualification the respondent reported, based largely on academic quali- fications types, such as degree, higher degree, ‘A’–level, GCSEs, other qualifications, and no qualifications. To compare the two classifications, I converted the qualifications given in Understanding Society to levels. GCSEs were problematic because GCSEs grades D–G are considered level 1, while five GCSEs grades A*–C are considered level 2 (Gov.uk, 2016b). To solve this I coded all GCSEs as a ‘Level 2 qualifications’, in effect treating this as ‘up to and including Level 2’. For the census data to match I simply added the number of individuals with Level 1 qualifications and Level 2 qualifications together, to create a comparable ‘up to and including Level 2 qualifications’ classification in the census. AS and A level qualifications were coded as level 3 qualifications. Degree and other higher degree qualifications were coded as level 4. Appren- ticeships in Understanding Society were coded as ‘other qualification’, so apprenticeships in the census were added to ‘other qualifications’ to match. Figure 5.6 shows the final coding and their respective proportions in Understanding Society.

5.5.3

Matching census populations

If the populations of the individual census constraints do not match the model will fail to constrain and not produce results. The number of respondents in each zone for each constraint must match exactly for the simulation to work. For example, the population of each zone in the age census table must match the population of the same zone in the ethnicity census table.

For most census tables the populations already matched so no preparation was necessary. In some cases the populations did not match and it was

0 10000 20000

qual_0 qual_2 qual_3 qual_4_plus qual_other

Highest qualification level

Count

Figure 5.6: Highest qualification

necessary to re–weight these to match populations in other variables. It is possible that some form of statistical disclosure control, for example record swapping or small cell adjustment, has been applied to maintain the anonymity of indvididuals, particularly for the small areas geographies I am using, which could explain why the populations did not match exactly (Office for National Statistics, n.d.: 6–7). Some variables are collected about households or a household reference person, rather than individuals within the household. This too could cause the population of the zones to be different. In either case, the solution to infer populations is acceptable as many characteristics of the household, such as socio–economic position, are determined by the resources—such as cars—available to household overall.

There are three possible values for car ownership in the census, and these were recoded in the survey to match. These are the household has no cars, one car, or two or more cars. To impute ‘correct’ zone populations for the car ownership table I calculated the proportion of respondents that had no cars, one car, or two or more cars in each zone. This was then multiplied against the known population of the zone. The result was a fraction, which can be fatal for individuals, so this number was rounded to a whole number. Finally, where the imputed population now had one

person too many or too few due to rounding, this was taken away from or added to the modal category to minimise the change in the proportions. This procedure is documented in data-raw/2-simulate-pilot.R.