• No results found

Chapter 9 discusses the findings and conclusions o f the study and in the light of these makes some recommendations for best practice The chapter also

7. Keeping the respondent informed of how far through the interview they are

5.5 Analysis of indicator variables

5.5.2 Indicator development

Analysis of interviews and the development of indicators were completed blind to the pupil data. Enforcement-level characteristics were each summarised into 2 and 3 level variables describing between-school variation in these characteristics. These are described in Sections 6.3-6.8. Once this was done, a 2 -level indicator was created to describe whether policy-level characteristics tended to support or undermine consistent no-smoking messages in each school. Construction of this variable is described in Section 6.9 and involved the re-classification of each policy-level indicator on the grounds of which levels tended to support and which tended to undermine the production of consistent no-smoking messages. These were then amalgamated into the final supportiveness of policy-level characteristics variable.

An indicator was then devised to describe the extent to which the WSE supported the smoking policy. Variation in enforcement-level characteristics could be related to variation in the extent to which each characteristic supported or undermined the policy. Ideally, these would have been combined in order to give an indication as to the extent that the WSE supported or undermined the policy. However, as many of these enforcement-level characteristics emerged from the analysis, data on them were limited across the whole sample. Instead, the two enforcement- level characteristics that there were data on across most schools were used to create a 3-level variable describing the extent to which the WSE in each school supported or

undermined the policy (Section 7.5). As before, this was then re-classified into a 2-level indicator describing the extent to which each WSE could be said to support or undermine the production of consistent no-smoking messages (Section 7.5).

Finally, the indicators describing the extent to which policy-level and the WSE/enforcement- level supported or undermined the production of consistent no-smoking messages were amalgamated to create a policy context indicator describing the extent to which the school tended to support or undermine no- smoking messages (Section 7.6).

5.5.3 Analysis

Once the indicators had been created, analysis of them in association with HBSC data achieved Objective 4.

It is not the intention or place of this thesis to defend the mathematics behind the statistical techniques undertaken for the collaborative analysis of indicator variables. It should also be noted that, as described above, during this stage of analysis the author worked closely with Dr Wiium as she ran the analyses, interpreting the findings in light of his qualitative data. The author is making no claim to having devised this analysis. Instead, it should be noted that the contents of this section are based largely upon discussions with Dr Wiium and the draft paper on which they have been collaborating as well as the author’s own understanding of and reading around the subject.

HBSC pupil data were provided as an SPSS file by WAG. These data had already been cleaned at a national level, however for this analysis, all schools that participated in HBSC but did not take part in interviews were removed from the data set. One HBSC question asked whether pupils smoked every day (daily smokers); at least once a week but not every day (weekly smokers); less than once a week or never. As daily and weekly smoking prevalence was very low in younger pupils (Table 5.7), only year groups 10 and 11 were used in the

final analysis. Having made this decision, School 08 had to be removed from the sample as there were no data for these year groups in this school3. This gave a final sample for analysis of 1941 pupils across 45 schools.

Table 5.7 Frequency o f smoking pupils by year group

Year Group % weekly smoking1 % smoking daily

Year 7 1.9 0.9

Year 8 6.1 4.0

Year 9 11.7 9.0

Year 10 20 14.8

Year 11 21 16.7

Weekly smoking is created by collapsing daily and weekly smoking in order to capture all those respondents smoking at least weekly

Social research commonly involves people in social contexts and groups which they both influence and are influenced by (Hox, 2002). These are often seen as hierarchies, with people ‘nested’ in their social contexts and as such, observations within social contexts cannot be assumed to be independent and such data, due to the presence o f these hierarchical levels, is termed multilevel (Hox, 2002; Rasbash et a l, 2000). Pupils nested in schools is a classic example of this and the employment of cluster sampling to account for this has already been discussed (Section 4.4.2.1). As traditional statistical techniques assume that observations are independent (Hox, 2002), multilevel techniques have more recently been developed to account for hierarchical data structures (Rasbash et al, 2000). Not to account for this can lead to false positive findings. As this study used a hierarchical data set, it employed multilevel techniques.

3 WAG had no record o f why there was no data for these year groups in this school, but it was most likely because these year groups were involved in examinations at the time of HBSC data collection.

In the first instance, a simple cross-tabulation of each indicator against weekly, daily and daily smoking on the school premises was conducted. The latter of these was included to test the hypothesis that smoking bans may merely displace smoking behaviour from the school site (Gordon & Turner, 2003a; Northrup et a l’s, 1998; Pentz et al, 1998; Turner & Gordon, 2004a). This descriptive analysis allowed investigation into the proportion of pupils in each level and the assessment of any patterns in the data.

Any indicators that demonstrated a pattern in the expected direction, were subjected to a multilevel logistic regression analysis against weekly smoking, daily smoking and daily smoking on the school site. Logistical regression was used since smoking is a binary variable (i.e. yes or no). For each policy indicator variable, the reference category was taken as the level of the variable assumed to be the strongest policy characteristics with odds ratios showing the comparative likelihood of being a weekly, daily or daily on the school site smoker across the other levels of the indicator.

A third analysis was then conducted on findings that were significant at the second stage. This multilevel logistic regression again tested these indicators against weekly smoking, daily smoking and daily smoking on the school site, this time controlling for pupil-level variables theoretically linked to smoking behaviour (e.g. parental smoking). In this way, the association of these factors with smoking can be assessed. If any given factor is strongly linked to adolescent smoking (e.g. parents smoke) and many of the pupils in the school exhibit this characteristic (i.e. the parents of many pupils smoke) then any apparent relationship between the school-level variable (i.e. policy) and smoking prevalence may be due to this compositional characteristic rather than the contextual characteristic of the school (policy). These confounding relationships may lead to false results. By including these pupil-level variables in the model, the amount of variation in smoking prevalence that they explain may be controlled for and the effect of any confounding relationships accounted for. At this stage, five models were tested:

Model A:

a random intercept model adjusting for no covariate (i.e. the null model)

Model B:

a random intercept model including all significant pupil-level variables in the present study

Model C:

a random intercept model including significant school policy indicators and controlling for all pupil- level predictors

Model D:

a random intercept model including significant school policy indicators and controlling for all pupil- level predictors as well as best friend smoking

Model E:

a random intercept model including significant school policy indicators and pupil-level predictors and possible interaction between school and pupil variables

The random intercept model refers to the multilevel model which allows for between school variation in smoking prevalence. The extent to which this between-school variation is the result of policy characteristics is the subject of this analysis. These are discussed further in Chapter 8.

Finally, due to the findings of the analysis, a descriptive cross-tabulation of pupil perceptions of policy compared to staff reporting of policy was also conducted.

The remainder of this thesis presents the results of this analysis and discussion of the findings.