Household- and Area-Level Differences in Burglary Risk and Security Availability over Time

Abstract

This chapter is based upon findings from a project which sought to establish which burglary security devices work for whom and in what context. A large body of previous research suggests that crime risk and vulnerability vary across individuals, households and areas. From this, we can assume that anti-burglary security devices may not exert the same protective effect for all households in all areas. Certain households may also be less likely to have anti-burglary devices installed at all. This chapter investigates how household and area differences may explain unequal burglary risks and security availability. It examines the relationship between burglary risk and the availability of the most effective security device combination ‘on a budget’ – window locks, internal lights on a timer, double door locks and external lights on a sensor (WIDE) – across population groups of different ethnicities, household composition, tenure, income, number of cars, and type of area of residence from 1993 to 2011/2012. It thus provides context to the security hypothesis for the crime drop using burglary in England and Wales as a case study.

Appendices

Appendix B

B.1 Data and Methodology

5.2.1 B.1.1 Variables

Population group-specific findings of burglary risk and effective security availability will only be discussed for the most effective of the combinations: WIDE. WIDE security availability compares the number of households with window locks, internal lights on a timer, double locks/deadlocks and external lights on sensor with those that have no security. Other results are available upon request.

Burglary risk is defined here as the likelihood of experiencing at least one attempted burglary or burglary with entry. So, unlike Chap. 4, the two burglary types are examined together to ensure there is an adequate number of respondents for analysis from each population group. This has theoretical disadvantages as attempts are (often) a product of the presence of security devices that were effective in thwarting a burglar. However, they did not deter burglars in the first place, and therefore the two burglary types together relate to household characteristics and security effective in deterring or thwarting.

Fourteen sets of explanatory variables that may affect both burglary risk and security availability entered the models, which were (reference category in italics):

• Number of adults (16 years old or older) in the household (one adult, two adults, three or more adults)

• Presence of children (under 16 years old) in the household (children, no children)

• Lone parent (lone parent, not a lone parent)

• Ethnicity of the respondent and in 2008/2009–2011/2012 ethnicity of the HRP (Black, Asian, Mixed/Chinese/Other, White)

• Annual household income before tax (£4999 and under, £5000–£9999, £10,000–£19,999, £20,000–£29,999, £30,000–£49,999 or £30,000 or more in 1994–1996, £50,000 or more which exists since 1998, no income information)

• Tenure (social rented sector, private rented sector, owner)

• Number of cars owned/used by the household in the last year (no car, 1 car, 2 cars, 3 + cars)

• Area type (rural, inner city, urban). The following are not discussed in the main text of Chap. 5:

• Social class, classifying households in three groups according to whether the Household Reference Person (HRP) is in routine (formerly manual or blue collar) or intermediate occupations, or never worked/not classified social class in comparison to professionals

• Accommodation type, classifying households in three groups according to whether they live in a semi-detached or terraced (row in the USA) house or flat/maisonette/other as opposed to a detached house

• Length of residence at current address, where the base of ‘10 years or more’ is compared to residence for ‘1 to 2 years’, ‘2 to 5 years’ or ‘5 to 10 years’

• Hours home left unoccupied, a variable attempting to gauge household’s guardianship if the home is left unoccupied ‘less than 3 hours’ or ‘3 to 7 hours’ compared to ‘7 or more hours’ on a typical weekday

• Region, contrasting North East (North in 1994–1996), Yorkshire and Humberside, North West, East Midlands, West Midlands, East (East Anglia in 1994–1996), London, South West and Wales to the South East

• Age of the Household Reference Person formerly termed ‘Head of Household’ (HRP) who is the individual in the household who owns or rents the accommodation

As seen in Chap. 4, the study employed all CSEW sweeps since 1992. A large number of household characteristics however are not available in the 1992 CSEW data set that currently exists in the UK Data Service – social class, number of hours home left unoccupied, area type and region. In addition, the routine activity variable – number of hours home left unoccupied – is not available in the 1994 and 1996 CSEW. To overcome the large number of missing household characteristics, two sets of models were estimated for this early period:

• 1. Based on the 1992–1996 CSEW data sets whilst including only those limited number of variables which were available

• 2. Based on the 1994–1996 CSEW data sets and including all variables consistently available over time (as shown in the above bullet points with the exception of hours home left unoccupied)

In this work the estimated model from the aggregate 1994–1996 (excluding 1992) CSEW data with the wider set of household characteristics and area type is reported in order to ensure that the results are consistent over time. The 1992–1996 model results with fewer explanatory variables are available upon request.

5.2.2 B.1.2 Data and Sample Sizes

The data are taken from the 1994 to 2011/2012 CSEW merged into five aggregate data sets and descriptive statistics of all variables used – household characteristics, area type, region, burglary risk and WIDE security availability – including sample sizes across the five aggregate CSEW data sets (1994–1996, 1998–2000, 2001/2002–2004/2005, 2005/2006–2007/2008 and 2008/2009–2011/2012) are given in Table 5.3. Prior to 2001, the full recall period was from 1 January of the year preceding interview until the date of interview – a period of about 14 months. For example, interviews for the 1996 BCS were conducted from January 1996 to June 1996, with incidents therefore reported from January 1995 to June 1996. After 2001 and a move to continuous interviewing, the ‘moving reference period’ includes the current month plus the 12 months prior to the date of the interview.

Appendix Table 5.3 Descriptive statistics of household and area characteristics, burglary risk and WIDE security availability from the WIDE sample of the CSEW aggregate data sets (1994–2011/2012)

The analysis reported in Chap. 5 requires information about security devices in both the ‘general population’ and burgled households. Some security devices are not strictly comparable over time due to changes in question wording. In addition, on a small number of occasions, device information was only collected for burglary victims and not the general population. For example, in the 1992–1996 sweeps, burglary victims were asked whether they had security chains, window bars/grills or dogs at the time of the event, but this information was not collected consistently from the general population sample. As a result, in the 1992–1996 sweeps individual devices for some victims may in fact represent these in combination with security chains, window bars or grills and/or dogs. The same devices may also exist unacknowledged within security combinations in the 1992–1996 data.

It should also be noted that the term ‘no security’ should be taken to mean ‘none of the CSEW listed devices’. Therefore, strictly speaking ‘no security’ is not comparable over time except between the 1998 and the 2007/2008 CSEW sweeps. ‘No security’ in the 1992–1996 sweeps means no burglar alarm, no double locks, no window locks and no lights. From the 1998 sweep onwards, more categories were included so in addition to the previous list, ‘no security’ means no security chains, no indoor lights on a timer and no external lights on a sensor. Therefore ‘no security’ in 1992–1996 means something different to ‘no security’ in the following sweeps. This may, to some extent, explain the higher frequency of ‘no security’ in the earlier 1992–1996 CSEW sweeps in Appendix Table 5.3. The same argument in theory applies for the pre- and post-2008/2009 sweeps of the CSEW due to the introduction of questions about the availability of CCTV in the respondent’s home, but the very small proportion of households with this device makes this issue negligible.

The samples for the statistical analyses reported in the current Chap. 5 are subsets of the data used in Chap. 4. In particular the samples of the analyses on the effectiveness of security against burglary nationally across England and Wales in Chap. 4 consist of all burglary victims and non-victims who (after being randomly selected) completed the Crime Prevention module (see Appendix A – Appendix Table 4.6, Chap. 4). To reiterate, information about household security availability of Crime Prevention module respondents who were also victims of burglary have been taken from the Victim Form. It therefore refers to the time of the first burglary rather than the time of the interview. For details please see Tseloni et al. (2014) and Appendix A in Chap. 4.

The data for the statistical modelling reported in the current Chap. 5 consist of all households (burglary victims and Crime Prevention module non-victim respondents) with effective (WD, WDE and WIDE) security availability and households with no security. For this reason, the sample sizes depend on the specific combination investigated. Indeed, the sample sizes examined in this work are dependent on the availability of effective security. The likelihood of having the respective security reduces from the most common effective combination of WD, which was present at roughly 14 percent of households, to the least common one, WIDE, which was found in 4 percent of the households in England and Wales between 2001/2002 and 2011/2012 (Tseloni et al. 2017). The later Appendix Table 5.4 gives the sample sizes of all data sets across the three (WD, WDE and WIDE) effective security combinations over the five sets of CSEW aggregate data. Comparing with the previous chapter’s Table 4.5 – the last row of which gives the sample sizes of households with no security, WD and WIDE – it can be seen that the sample sizes for the 2008/2009–2011/2012 statistical models of WD (5973) and WIDE (2766), shown in the third and last rows, respectively, of the last column of Appendix Table 5.4, have been obtained by adding the first figure in the last row of Table 4.5 to the second and the third, respectively.

Appendix Table 5.4 Correlation (standard error of covariance) between effective security availability and burglary risk during the crime drop

5.2.3 B.1.3 Statistical Model and Modelling Strategy

Models were run for all the effective security combinations: WD, WDE and WIDE from 1998 to 2000 onwards. The 1992–1996 group of sweeps used the following combinations, WD and LWD, due to the fact that the distinction between internal and external lights was not made during this period. As mentioned in the main text of Chap. 5, the results of the estimated models of burglary risk and WIDE security availability are discussed here, whereas results for WD and WDE are available upon request.

The statistical model used is the bivariate logit regression model (Snijders and Bosker 1999) of two associated outcomes – the likelihood of burglary and WIDE security availability – over a set of explanatory variables, here the household characteristics, area type and region outlined earlier. The choice of the statistical model is justified by the fact that both outcomes are in theory affected by the same household and area characteristics, whilst they are also interrelated. The model allows estimating the correlation of the likelihood of burglary and WIDE security availability in addition to the effects of the explanatory variables on each outcome. Further explanation of this statistical model within criminology (with respect to multiple fear of crime measurements) and a discussion on how to interpret its fixed and random parameters, including the correlation between outcomes, can be found in Tseloni and Zarafonitou (2008). The model across the five sets of data was estimated via the computer software MLwiN version 2.10 (Rasbash et al. 2009). For a complete guide go to http://www.bristol.ac.uk/cmm/software/mlwin/).

The modelling strategy included four stages as follows: for each of the five sets of aggregate data from 1994–1996 to 2008/2009–2011/2012, first a baseline model, whereby respective pairs of random constant terms are only estimated, to establish that burglary risk and WIDE security availability are overall correlated. The ‘national average’ figures in Appendix Table 5.4 give the estimated baseline correlations. The last row of figures in Appendix Table 5.5 show the baseline constant estimates. Secondly all variables in the next Appendix Table 5.3 were added to the baseline models to give over time estimates of all theoretically relevant effects on burglary risk and security availability. These models also provide a consistent profile of the RH over time. Not all estimated parameters however were statistically significant. In the third stage of the model fitting, characteristics with no statistically significant (at p-value greater than 0.10) parameter in both burglary and security regressions were omitted from the model. To put it differently, household characteristics were retained on the basis of having at least one parameter at p-value lower than 0.10 in at least one outcome, burglary or security. Further in the fourth and final stage, individual categories with p-value greater than 0.10 were removed from the models only if they could be merged with the base category of their encompassing characteristic. The following example may clarify the third and fourth stages of the modelling strategy. Number of cars the household used in the previous year is a single household characteristic which in the regression is represented via three dummy variables, no car, one car and three or more cars, whilst two-car households are the comparison base. All the categories of this variable were included in the third stage because at least one parameter out of six (three for burglary and three for security) had a p-value less than 0.10. However, since three or more cars had p-value greater than 0.10 in both burglary and security regressions and it is adjacent to the base of two cars, it was omitted. Thus in effect two or more cars become the base of comparison for the effects of number of cars in the household on burglary risk and effective security availability across most periods.

Appendix Table 5.5 Estimated odds ratios, Exp(b), of fixed effects of household characteristics, area type and region of England and Wales on joint burglary risk and WIDE security availability over time, 1993–2011/2012 – based on bivariate logit regression models of CSEW aggregate data sets (1994–2011/2012)

B.2 The Correlation of Burglary Risk and Effective Security Availability Nationally, 1993–2011/2012

Appendix B.2 offers statistical details of the findings discussed in Sect. 5.4.2 with regards to the overall and conditional correlation between burglary and effective security. Unlike the remainder of Chap. 5, it shows the results for all three effective security combinations highlighted in Chap. 4: WD, WDE and WIDE. Appendix Table 5.4 shows the estimated residual correlation between burglary risk and presence of security across the three effective combinations identified in Chap. 4 (WD, WDE and WIDE), over five aggregated CSEW data sets (1994–1996, 1998–2000, 2001/2002–2004/2005, 2005/2006–2007/2008 and 2008/2009–2011/2012) and for two sets of models. In this instance, the correlation coefficient coincides with the estimated covariance between burglary and security, and the respective standard error is given in brackets. Indeed, in joint or bivariate logit models, the respective variances of the two binary dependent variables are one by construct; therefore their covariance is also their correlation. The two sets of models refer to the baseline or empty model, whereby only the random intercepts of the outcomes, burglary risk and presence of effective security, have been estimated in the joint models, and the final two-equation model including all statistically significant covariates for each outcome. The empty model essentially gives the national average correlation of burglary risk and presence of effective security. The final model’s correlation gives the residual value after the mediating effects of household and area characteristics and region have been incorporated. In this light it refers to the correlation specific to the RH.³⁹

The (conditional) correlation of burglary risk and effective security presence between two randomly selected households with an identical profile is lower than the national average (see Appendix Table 5.4). However, the household profile can only be delineated with respect to all measurable characteristics in the CSEW and is portrayed in the RH, of which more will be said in the next section. The reduction in the correlation is modest and does not exceed 21 percent (calculated as (50–63)/63 from Appendix Table 5.4). Since context matters (as seen already and in the following sections) we may conclude that it is less than perfectly measured in this work which relies on the CSEW. The relationship between burglary risk and effective security presence depends on household characteristics which are not measured in the CSEW.⁴⁰ These may include detailed area socio-economic characteristics, neighbour social networks and house and area of residence physical layout and access to other places (see also Chaps. 2 and 3).

B.3 Estimated Bivariate Logit Regression Models of Burglary Risk and WIDE Security Availability During the Crime Drop

The odds of burglary victimisation are the ratio of the likelihood of being burgled over the complement probability of not being burgled (Long 1997, p. 51). With the exception of age of HRP, the estimated odds of burglary for each household characteristic, area type and region in Appendix Table 5.5 are in comparison and therefore as a ratio to their respective base category (given in brackets in Tables 5.1 and Appendix Tables 5.3 and 5.5 and as a horizontal line at value y = 1 in Figs. 5.2, 5.3, 5.4, 5.5, 5.6, 5.7, 5.8 and 5.9). Examples of how to interpret the odds ratios (Long 1997, p. 81) will follow in the last paragraph.

The odds of WIDE security availability is the ratio of having this combination over no security at all. Given the sample employed in this study, the two outcomes are exhaustive and complement each other, although in principle and in reality households have a number of security combinations (see Chap. 4 for a full list). As such WIDE security availability is in principle part of a multinomial logit (at least) three outcome response (contrasting WIDE or any other security combination to no security). However, this statistical specification would have prohibited investigating its correlation with burglary risk in context: ‘[M]ultinomial response variables cannot be included in multivariate response models, but can be used in univariate response models’ (Rasbash et al. 2017, p.228). Therefore the study focused on examining the population group-specific relationship between burglary risk and each effective (WD, WDE and WIDE) security combination in isolation. The results of the estimated bivariate logit regression models of burglary victimisation and WIDE security availability across the five aggregate CSEW data sets from 1994 to 2011/2012 are presented in Appendix Table 5.5.

The last two rows of Appendix Table 5.5 give the estimated constant terms from the final and baseline model obtained from the fourth and first model fitting stages, respectively (see Appendix B.1 discussion), for each period examined. With the exception of age all variables in the estimated models are qualitative and therefore denoted by dummy variables contrasting each category of the nominal variable to a respective category (see Appendix Table 5.3). For example, ethnicity of the respondent or HRP contrasts each ethnic minority group to White. The effect of all base categories together on the regression outcome, here the likelihood of burglary and effective security availability, is given in the respective intercepts or constant terms (see Appendix B.1). The constant terms are therefore here the log odds of experiencing a burglary (attempted or with entry) and WIDE security availability for the RH (penultimate row) and nationally (last row) in each set of CSEW data. For example, in the 1994–1996 CSEW data, the national average log odds of burglary and of WIDE security were −1.55 and −0.09, respectively. For a household with the RH profile generally⁴¹ as given in Sect. 5.4.1, the respective figures were −0.92 and −0.79. Comparing across rows it is worth noting that the RH’s burglary risk was higher than the national average except in 2001/2002–2004/2005 and 2008/2009–2011/2012, whilst its WIDE security availability was unequivocally lower.

Burglary risks in this study should only be viewed in relative terms, such as risk of the RH compared to the national average or of population groups under question compared to the RH’s risk. Due to the limited sample of the Crime Prevention module and employing all burglary victims’ data in this study, our estimates of national or RH-specific burglary risk entailed in the constant terms of the respective regressions in the last two rows of Appendix Table 5.5 are overestimated. A rough calculation⁴² of the extent of this overestimation is given in Appendix Table 5.6 which compares burglary risks between the entire CSEW sample and the WIDE samples used here across the five periods. Based on the discrepancy between risks in the two samples over time, the last column of Appendix Table 5.6 gives a deflation factor which can be multiplied with the absolute burglary risk estimates from our models. Interested readers may apply the figures provided in the last column of Appendix Table 5.6 to the estimated risks entailed in the baseline constant terms of Appendix Table 5.5 in order to obtain more realistic estimates of national burglary risks in absolute terms.⁴³ Simple (without compositional effects) burglary risks of population groups from the entire CSEW samples are provided in Table 5.2 of the main text in Chap. 5.

Appendix Table 5.6 Burglary risks in absolute values in the entire and employed (WIDE security focused) CSEW samples over time

Appendix Table 5.7 Household Reference Person (HRP) ethnicity as a percentage of respondents’ ethnicity, 2006/2007 CSEW

Our focus is on burglary and security inequality and how they changed during the crime drop, and therefore absolute estimates are outside the scope of this study and indeed, with regard to security, unattainable due to lack of data . The following paragraphs interpret the relative to the RH estimates of burglary risk and WIDE security across all population groups in the CSEW data.

Apart from the constant terms, all figures in Appendix Table 5.5 present the exponentials of the estimated parameters because, as will be seen shortly, they have a more intuitive interpretation than the parameters themselves (Long 1997, pp. 80–81). The exponentials of the combined estimated parameters for age and age squared of HRP give the (non-linear) change in the odds of burglary (as given in even number columns) and WIDE security availability (shown in odd number columns bar the first) for each year the HRP grows older. The remaining set of figures in Appendix Table 5.5 give the odds ratios of experiencing a burglary (attempted or with entry) and WIDE security availability against no security for a household that belongs to the respective population group in comparison to the base category for the same characteristic with all other household and area attributes and region being equal. They also provide an indication of their statistical significance. For example, in the 1994–1996 CSEW data, the odds ratio of burglary and of WIDE security of single adult households did not significantly differ from those of a two-adult household. The same can be said for three or more adult households with respect to burglary. However three or more adult households had 46 percent (calculated as 100 × (0.54–1)) lower odds of WIDE security availability than two-adult households. Moving along the top rows of figures, the 2001/2002–2004/2005 estimates can be interpreted similarly to the ones from 1994 to 1996 with one exception. Three- or more adult households had, compared to two-adult households, 50 percent (calculated as 100 × (1.50–1)) higher odds of burglary (albeit with weak statistical significance, p-value between 0.05 and 0.10). The remaining figures can be interpreted in a similar way bearing in mind the respective indications for their statistical significance.

Prior to our work discussed in Chap. 5, the question of who is at highest risk of burglary and who is least likely to have the most effective security had not been investigated. One of the contributions of this work is using a methodology that accounts for the composition of each population group. Ignoring compositional effects for a moment, the highest burglary risk was faced by the following household types: HRP of Mixed, Chinese or Other ethnicity, single adult households, those with children, lone parents, social renters , households earning under £5000 per year, without a car and living in inner cities of England and Wales in 2008/2009–2011/2012.

With one exception, the above profile remains largely the same when the household profile of maximum burglary risk is examined across all possible contributing factors simultaneously. The exception refers to income; the statistical modelling analysis found that households at £5000–£9999 per year or without income information are at 73 and 32 percent higher burglary risk, respectively, than others during the same period. Although the profile of households mostly at risk of burglary is roughly the same, the estimated effects (odds ratios) of each population socio-economic characteristic on burglary risk disagree between the methodologically rigorous estimates discussed in the previous section (Figs. 5.2, 5.3, 5.4, 5.5, 5.6, 5.7, 5.8 and 5.9 and Appendix Table 5.5) and the bivariate cross-tabulations (Table 5.2).

A few examples of the most startling differences refer to lone parents, inner-city residents and social renters demonstrating in a clear manner the implications of not considering each group’s composition in the case of bivariate associations . According to Table 5.2, lone parents have 3.2 times (or 220 percent higher than) the burglary risk of others, whereas from the statistical models, this stands at a much lower 1.8 times (or elevated by 80 percent) the baseline risk (see Sect. 5.5.3 and Appendix Table 5.5). This is likely because the (bivariate analysis calculated) odds ratio entails the two individual effects of ‘single adult household’ and ‘living with children’ in the overall lone parent estimate. Indeed, the statistical model produced a similar overall lone parent odds ratio (at 3.45, see penultimate paragraph of Sect. 5.5.3). Similarly, inner-city households’ burglary odds ratio is overestimated in the bivariate analysis of Table 5.2. This is because it entails the elevated burglary risk of other (than area type) predominant characteristics of inner-city households, such as single adult ones or HMOs and without car households which are most likely to reside in inner cities. By contrast the odds ratio for social renters based on bivariate analysis underestimated their burglary risk to 2.2 times (or 120 percent more than) that of owner-occupiers – from the statistical models this was 3.6 times (or 260 percent higher than) the RH. The reader would recall that social renters are households on low income but in recent years not worse off (in terms of fuel poverty at least) than private renters and nearly a quarter are lone parents (Wilson and Barton 2017). Therefore, the underestimation possibly confounds the individual effects of income and household composition within tenure. To conclude, the population groups of highest burglary risk in this study are by and large the same regardless of methodology, but the estimate of the effects’ magnitude – and therefore prioritisation of preventive resources to those of greatest need – is compromised when using bivariate associations.