Helen Chester University of Manchester. Brief overview of study and findings Focus on issues and recommendations for: Researchers wishing to do similar.

Helen ChesterUniversity of Manchester

Brief overview of study and findings

Focus on issues and recommendations for:

Researchers wishing to do similar research with BCS or datasets

Policy and practice

Research background Study objectives Definitions How the BCS is used for this study Approach Summary of findings Issues and recommendations

Research has shown that the chance of a household becoming a victim of property crime may depend on both the characteristics of the household and the property (e.g. household composition, income, type of housing, presence of security measures) and the type of area in which the household lives (e.g. population size, socioeconomic and demographic structure of the local population)

Uncertainty as to whether household or area level variables are more important (Hope, 2007)

To put ‘property crime in context’ bycarrying out a multilevel analysis of theBritish Crime Survey (BCS) to investigatethe relationship between the

characteristicsof households and those of the areas inwhich they are located on householdchances of being a victim of property

crime.

Property crime includes theft (theft from the household by someone who has been

granted entry), burglary (theft by someone who has not been granted

entry) and criminal damage.

This definition includes attempted and actual crime and excludes minor crimes

such as milk bottle theft.

Takes into account the ‘clustering’ in data and so better estimation of coefficients and standard errors

Unlike ordinary regression where there is only one random variable which is the error term, in multilevel analysis there are two random variables at level 1 (household) and level 2 (area) (Brown and Rasbash, 2004).

Allows assessment of proportion of variation in property crime victimisation attributable to household and area characteristics so the relative importance of these in explaining an outcome.

Area level residuals (that is the remaining variation in property crime victimisation attributable to particular area types) can be used to rank areas from highest to lowest.

This information can be used to compare areas to one another and to the average for England and Wales.

Can also examine how these rankings change when household and area variables are controlled for.

Differences between households in terms of their characteristics are expected to lead to differences in the chance of becoming a victim of property crime according to lifestyle/routine activities theory

Examples of potentially important variables include: Composition of households (i.e. number of adults,

children) Characteristics of head of household (e.g. age,

marital status) Type of dwelling (e.g. detached, flat) Income level (actual and proxy measures of this)

If households live in different areas they are expected to have different chances of property crime victimisation due to expectations from social disorganisation theory

The following variables are expected to lead to higher exposure within these areas of households to potential offenders:

Economic status of neighbourhood (i.e. Deprivation)

Residential mobility Family disruption Urbanisation Neighbourhood heterogeneity

Property crime victim (1)

Non victim (0)

Statistical models to examine the influenceof a number of different variables on ahousehold’s probability of victimisation –

whatincreases or decreases the chance of

victimisation

Datasets from three waves (1998, 2000, 2001/02)

ONS ward type to define ‘area of residence’

Census data to measure characteristics of areas

Level 2 →Area level → Social disorganisation theory/routine activities theory

(from the 1991 census)

Level 1 → Household level → Routine activities/lifestyle theory

(from the BCS)

Linked using ONS ward type

Single level logistic regression models and multilevel models estimated for 1998, 2000, 2001/02

Examined how the inclusion of area level characteristics affected the relationship between household characteristics and property crime victimisation

Relative importance of household and area characteristics in predicting property crime

Benefits of multilevel modelling compared to single level logistic regression modelling

Both household and area level characteristics important in predicting property crime

Evidence of between and within area variation in property crime

The inclusion of area level variables into models containing only household level variables affected the conclusions drawn from this models: some household level variables became insignificant (e.g. social renting); effect of some household level variables changed (e.g. dwelling type)

Findings (2)Findings (2)

Significant (but small amount of) variation in property crime victimisation between ONS ward types (area of residence) in each year of the survey. Some area types were found to be significantly different from one another and the average for England and Wales .

High degree of correlation between the rankings over area types across the three years of the survey with some area types consistently having the highest and lowest property crime victimisation.

The ranking of area types ‘made sense’ when it was examined how these compared on ‘social disorganisation’ indicators, in particular deprivation was thought to play a key role.

In fact the area level measures of social disorganisation were found to explain most if not all of the remaining variation in property crime victimisation between ONS ward types

Findings (4)Findings (4)

Overall household characteristics explained a much greater proportion of variation in property crime victimisation suggesting that these were more important than the area level characteristics (associated with residential location) in explaining differences in victimisation.

This study provided support for findings of Trickett et al. (1995) that the standard errors and coefficients in multilevel modelling are virtually the same as in single logistic regression models.

However multilevel modelling did appear to have a number of benefits including:

the ability to rank areas according to property crime victimisation whilst controlling for household and/or area level variables;

the possibility of examining change in rankings of these areas over time;

a clear distinction between composition and context by attributing clearly the effects of variables at different levels; and

the estimation of within and between area variation in property crime.

Feel it is not possible to definitively conclude that household characteristics are more important than area characteristics in predicting property crime victimisation

This is because the assessment of the relative importance of household and area characteristics is dependent on a number factors: the variables included in the models; the choice of ‘geographical indicator’ used to define ‘area of residence’; and the sampling strategy of the survey.

There are concepts that : there is theory for but variables considered to

measure them are omitted from the analysis as they are not collected either in our own or secondary data

there is theory for and variables are available in datasets that could be used to measure them but they are excluded by the researcher

‘no one thought of’ - there may or may not be existing proxies or measures that could capture these but these are not included in the analysis as they are not viewed as relevant.

How area of residence is defined is very important as the level of aggregation at which the area characteristics are defined is likely to influence the results

Unfortunately theories do not always give guidance on what would be appropriate

Evidence that different relationships can be found between variables at different scales and aggregations.

This is known as the ‘modifiable areal unit problem’ (MAUP) and is considered to consist of two related problems:

The scale problem: “is the variation in results when the same areal units are

combined into sets of increasingly larger areal units of analysis” (Openshaw and Taylor, 1979 p.128).

Openshaw and Taylor (1979) argue that:“analysis of the same census data at scales ranging from

enumeration districts, wards, local authorities, up to standard regions will almost certainly provide

alternative results and possibly interpretations” (Openshaw and Taylor, 1979 p.128).

The aggregation problem is:“any variations in results due to alternative units of

analysis where n, the number of units is constant” (Openshaw and Taylor, 1979 p.128).

Selection of variables for the Selection of variables for the modelmodel

Omitting variables from either the household or area level will affect the estimate of the relative effect of household versus area level characteristics

Especially important if using evidence of variation at particular levels to provide support for a particular theory

Choice of ‘geographical Choice of ‘geographical identifier’ to define ‘Area of identifier’ to define ‘Area of

residence’residence’Use of ONS ward type problematic for a number

of reasons: large groups (43) that may contain quite different

wards ‘dampening’ the area effect as may contain areas with a mix of higher and lower victimisation levels;

do not know how variable wards are within a particular ward type and therefore how the ‘area effect’ for a particular ward differs from that of the ward type.

As variation between wards within ward types is omitted it is possible there is an over estimation of the variation in property crime victimisation attributable to household level characteristics (as this variation will be observed at the household level)

Sampling strategySampling strategy

Although the BCS is clustered, it is designed as a national survey and not as a survey to estimate between and within area effects.

Choice and selection of PSUs and households within them affect the mix of household and area types included in the survey (estimates of intra and inter area variation in property crime is likely to depend on how homogeneous/heterogeneous the population of areas are)

Because of the above, any changes in the sampling strategy between different sweeps is likely to affect comparisons between years of the survey

Recommend consider these Recommend consider these issues when doing similar issues when doing similar

research (1 of 2):research (1 of 2): Importance of how ‘area of residence’ is

defined (qualitative work, assess the utility of different definitions)

‘Omitted variable bias’ and its importance when assessing relative importance of area and household characteristics (data linkage, primary data collection both to avoid and inform)

Sampling strategy: how it affects estimation of household and area level effects and comparisons over time (investigate this)

Recommend consider these Recommend consider these issues when doing similar issues when doing similar

research (2 of 2):research (2 of 2): Importance of using measures of household and

area level characteristics in models (large source of omitted variable bias avoided if include both).

Importance of multilevel modelling Ensure standard errors and coefficients not biased Proportions of variance Ability to ‘rank’ areas

However ask yourself does multilevel analysis test your ‘theory’? Algorithm used simultaneous estimation of household and area level effects, none take precedence in the process - may not necessarily correspond with theory

Recognition in recent crime strategies that areas do not suffer equally from crime e.g. Violent Crime and Drug Strategies 2008; Youth Crime Action Plan 2008 all focus or recognise the value of targeting specific areas.

Move from national to locally set targets. Increasing emphasis on different agencies working in

partnership to tackle crime from the early days of CDRPs through to recent crime strategies.

Importance of information sharing between local agencies to tackle crime, and

Importance of local information in developing local strategies to target crime

Policy/practice Policy/practice recommendationsrecommendations

This type of research could point to areas that should be targeted as part of a strategy to tackle property crime and maybe those that lessons could be learned from.

Point to the importance of linking datasets from a number of agencies/departments in explaining variations in crime

BCS has several advantages over recorded crime data and this type of research demonstrates how the survey could potentially be used to explore variations within and between CDRPs

Recommendations for data Recommendations for data producers (1 of 2)producers (1 of 2)

Work to explore feasibility of redesigning theBCS to reconcile need to produce national estimates with need for data more representativeof smaller geographical areas (to enable relativeassessment of area and household level effects),taking note of existing studies which have begunto explore the extent to which this is possible(Vallée et al. 2007) . This would provide usefulinformation to allow investigation of trends which might not be possible using local surveys.

Recommendations for data Recommendations for data producers (2 of 2)producers (2 of 2)

Linkage of BCS data with other datasets e.g. health data, census data etc. or other datasets of relevance to a number of agencies

Consider inclusion of including modules of criminological research value e.g. ‘lifestyle’ module

Thank you for listening. Thank you for listening.

Any questions?Any questions?