Reflektioner kring trygghetsmätning i Malmö

Jag har haft lite fullt upp med annat och därför inte haft chansen att blogga på ett tag. Men ett och annat som kan vara intressant att kommentera har trots allt hänt, inte minst har Malmö stad gått ut med ett pressmeddelande om Malmö OMrådesundersSökning (MOMS). Pressmeddelandet fokuserade på att 7 av 10 Malmöbor uppger sig vara trygga i sitt bostadsområde, vilket i sammanhanget betyder att de inte känner sig otrygga ute i bostadsområdet på kvällen samt inte anger att det finns specifika personer eller platser i bostadsområdet som de är rädda för eller undviker. När jag twittrade om det fick jag mothugg från somliga som var skeptiska till om det kan stämma eller till och med menade att undersökningen är friserad.

 Jag tänkte i sammanhanget att det kan vara intressant att säga några ord om undersökningen och vad den faktiskt kan visa. Jag vill dock poängtera att jag inte är involverad i skrivandet av den rapport om undersökningen som snart kommer, det är framför allt lektor Anna-Karin Ivert samt adjunkt Alberto Chrysoulakis som arbetar med den, och jag har bara sammanställt lite kartor till rapporten. Det hindrar dock inte att jag kan föra en del mer allmänna resonemang om undersökningen.


MOMS är en stor trygghetsmätning, som framför allt skiljer sig från de trygghetsmätningar som polisen gör i Skåne och stora delar av Sverige genom att den 1) innehåller betydligt fler frågor vilket medför bättre möjligheter för analys och 2) gör ett försök att mäta resultat inom betydligt mindre geografiska enheter. Enkäten skickades ut till 7965 slumpmässigt utvalda personer i åldrarna 18-85 och fick till slut 4195 svar vilket ger en svarsfrekvens på ca 53%. Svarsfrekvensen är generellt lägre i områden med högre nivåer av fattigdom och liknande, vilket var väntat och har hanterats genom att dessa områden översamplats (fler personer har alltså fått enkäten där). Dessutom är det lite fler kvinnor än män som besvarat enkäten samt för många äldre och för få yngre. Trots dessa problem utgör den statistiska basen med 4195 respondenter ett synnerligen starkt underlag för att säga något om vad Malmöbor generellt upplever. Att det finns folk på twitter som menar att dessa 4195 Malmöbor har fel är alltså inget att ta på allvar.

Däremot är osäkerheten betydligt större när det gäller enskilda delområden, och det kan vara värt att ta upp för diskussion. Särskilt med tanke på att det i media blivit ett stort fokus på enskilda områden med hög otrygghet (Sydsvenskan) (SR). De kartor jag producerat för presskonferensen och andra presentationer är generellt färgkodade från rött till grönt, där mörkgrönt signalerar bra värden, rött signalerar sämre värden, och gult är någonstans i mitten. Dessa intervall är baserade på standardavvikelser , där varje färgkod motsvarar en standardavvikelse, och gult alltså motsvarar mellan minus 0.5 och plus 0.5 standardavvikelser från genomsnittet. Jag tar här en karta över "oro att utsättas för brott" som exempel, för fler exempel se presentationsmaterial från presskonferensen.

 Här är det viktigt att notera att skillnaden mellan gul och orange alltså i teorin kan vara väldigt nära 0, vilket i princip alltid är fallet när det gäller en hypotetisk skillnad mellan två angränsande kategorier. Ett hypotetiskt område som ligger på 0.50 blir gult medan ett hypotetiskt område som ligger på 0.51 blir orange. Skillnaden mellan gul och orange är alltså rent teoretiskt som minst (nästan) 0, som mest 2 (från -0.49 till +1.50) och i genomsnitt 1 standardavvikelse.Även om vi beaktar den genomsnittliga skillnaden på 1 standardavvikelse så är det i sammanhanget inte jättemycket. Det beror på att värdet för respektive område är baserat på de boende i det området, i genomsnitt ca 40 respondenter, vilket medför en stor statistisk osäkerhet. Jag har inte tillgång till rådatan för att beräkna sannolikheter, men vi kan konstatera att det är troligt att det i vissa fall är slumpen snarare än faktiska skillnader som ger utslag på färgkodningen av kartan.

Om vi istället fokuserar på skillnaden mellan ett område kodat som rött (eller mörkgrönt) och ett område kodat som gult blir det hela dock lite tydligare. Här är (den teoretiska) skillnaden som minst 1 standardavvikelse (0.5 till 1.51), i genomsnitt ca 2 standardavvikelser (0 till 2), och i extremfall kan den nå högre än 3 standardavvikelser från genomsnittet (-0.49 till mer än +2.5). Det är fortfarande teoretiskt möjligt att ett sådant samband är orsakat av slumpen, men det är betydligt mindre sannolikt. Det är därför bra att media generellt fokuserat på områden där det statistiskt sett inte är så troligt att det är en slumpmässig variation som orsakat resultatet. I sammanhanget kan det däremot ses som olyckligt att Sydsenskan publicerat en (väldigt snygg) karta , eftersom en sådan karta lätt kan leda folk till att dra slutsatser som är statistiskt osäkra. Samtidigt är dock artikeln, där Anna-Karin Ivert citeras, ganska nyanserad och fokuserar framför allt på de skillnader som troligen får betraktas som faktiska skillnader. Det artikeln också gör bra är att lyfta fram det mer generella sambandet - att stora delar av Malmö i princip inte har någon otrygghet alls medan det i ett fåtal områden istället finns en väldigt utbredd otrygghet.

Utan att föregripa den kommande rapporten kan det konstateras att de områden där otryggheten är hög karaktäriseras av att de som besvarat enkäten anger en relativt låg tillit och informell social kontroll samt en relativt hög problemnivå, något som Sveriges radio på ett förtjänstfullt sätt tog upp i sin nyhet om undersökningen. Själva rapporten kommer ut i december någon gång, och då räknar jag med att blogga mer om det hela.

Tillägg: Det slog mig att det kanske är en bra idé att säga något om hur kartorna bör tolkas om man nödvändigtvis vill tolka enskilda områden. Ett förslag är då att titta på skillnaden inte mot närmsta kategorin, men mot två eller tre kategorier bort. Det är sannolikt en skillnad mot något som är två kategorier bort (i genomsnitt 2 standardavvikelser), och mycket sannolikt en skillnad mot något som är tre kategorier bort (i genomsnitt tre standardavvikelser).
Exempel 1: Ett orange område på kartan är sannolikt inte ett tryggt område (ljusgrönt) och det är mycket sannolikt att det inte är ett väldigt tryggt område (mörkgrönt).
Exempel 2: Ett gult område är sannolikt varken ett väldigt otryggt (rött) eller ett väldigt tryggt (mörkgrönt) område.

Avoidance behaviour - some results from a case study of 4 neighborhoods

Between 2010 and 2012 I worked on a case study of the multi-family housing part of four neighborhoods in the city of Malmö. The main focus of the study was on the connection of social capital/collective efficacy and physical disorder in the form of vandalism, graffiti and arson, but fear of crime was also discussed extensively. The study included a fairly wide array of methods, including key informant- and focus groups interviews, systematic social observation (SSO) of disorder and a community survey. The final report of the project (Gerell 2013) which I submitted in august the 2012 did not (due to time constraints) include all survey-responses that were mailed in (see below), and did not include multi-level analysis of responses. I have since re-visited the full data-set with some additional analysis, and in this blog post I will present some findings related to fear of crime and avoidance behavior. This is data that I don’t believe I will ever try to get published in an academic journal, but I did write some up of it up for a presentation I held in a research discussion group with members from the municipality, police, rescue services and academia and so I might as well make it more publicly available as well. 

The neighborhoods

In the study we wanted to get a deeper understanding of community level processes by focusing on four neighborhoods. Four adjacent neighborhoods in the southern part of the city of Malmö were selected. Two of the neighborhoods are included in the municipality “area programs” which are targeted at especially troubled neighborhoods. Of the other two neighborhoods one is made up of tenant ownership associations with a fairly old population and the other is made up of rental apartments, about half of which are public housing with fairly low income levels. The project was focused on multi-family housing, and therefore excluded single-family housing present in two of the neighborhoods and a fairly large retirement home section in one neighborhood. The total population of the studied parts of the neighborhoods is about 12 000, with the tenant ownership association neighborhood significantly smaller than the other three. A downside to the multi-family housing focus is that official registry data is based on the neighborhoods as a whole, and thus direct comparison to registry data is not possible. In my report I did attempt to make some comparisons to official data to assess whether respondents of the survey were comparable to the population as a whole. It can be noted that among respondents there are far too many in the oldest age group (age over 60), too many females and too few young persons. Compared to a large survey of Malmö recently completed by my department the differences in this study are similar, but larger. 

An early result in the project was that neighborhoods were percieved by respondents in key informant and focus groups interviews to be too large to capture social processes of interest. The four neighborhoods were therefore divided into twelve micro-neighborhoods, with an average of about 1000 residents each. Each micro-neighborhood has its own access road, often giving the micro-neighborhood its name, and parking lot, contributing to the routine activities of the location. In this text micro-neighborhoods are used as level 2 units.

The commmunity survey

The sample for the community survey was constructed by mapping all addresses (each address consist of one stairwell with apartments) in the multi-family section of the four neighborhoods (N=416), then randomly sampling one address from each yard (N=59), and an additional 28 addresses sampled from the remaining 357. Out of the total 87 sampled addresses entrance was achieved to 86 which make up the final sample in the study. In each of the 86 addresses every household was included for a total of 1255 sampled households, with a total of 689 responses (54.9%). Within households the sample was non-random, whoever opened the door when the staff came knocking. The community survey was performed with the aid of the city part municipality between April and June 2012, with six persons from the studied neighborhoods recruited to knock doors for the survey. They were recruited based on attachment to the local community and knowledge of appropriate languages, covering the three main language groups of residents with foreign background living in the four neighborhoods (two interviewers each fluent in Albanian, Arabic and Afghan languages (Pashto/Dari)). Sampled households that couldn’t be reached received a response-envelope and instructions on how to mail their survey in.

Research design

In order to analyze outcomes of avoidance behavior related to fear of crime a 2-level logit regression have been performed using HLM 6 where the second level is made up of 12 micro-neighborhoods which were constructed based on key informant interviews while taking physical and social boundaries into account. Since the outcome variable is dichotomous the error variance is always set to pi^2/3 and ICC-values are computed based on tau^2/(tau^2+3.289868). In the empty model ICC-values are calculated for the outcome variable. In model 1 we add individual level variables of demographic character that can be assumed to have an impact. In model 2 concentrated disadvantage is added on level 2. With concentrated disadvantage, collective efficacy and physical disorder all being strongly correlated which may impact on the results in such a small data set we then try collective efficacy and physical disorder independently of each other before testing a full model. Model three thus adds collective efficacy to concentrated disadvantage as level 2 predictor while model 4 instead of collective efficacy introduces physical disorder. In model 5 finally we test both concentrated disadvantage, collective efficacy and physical disorder simultaneously. It should of course be noted that the inclusion of more than one level 2 predictor for such a small number of level 2 units is questionable (Raudenbusch & Bryk 2002).

Data

The dependent variable is based on one question in a residents’ survey in four neighborhoods of the Swedish city Malmö performed in 2012: “Have you abstained from participating in an activity due to concerns of safety in the neighborhood where you live?” with responses graded in frequency, “Very often”, “Often”, “Sometimes”, “No, rarely” or “Never”. The responses were recoded so that “No, rarely” and “Never” registered as a 0 while the three other responses were coded as a 1 (=avoidance behavior).It should be noted that "concerns of safety" is a somewhat unprecise translation to english. The actual word used, "otrygghet", literally means "not feeling safe", and it is related to the english language "fear of crime", but has a much broader meaning.

Individual level independent variables based on the survey (for descriptives see Table below) include dummies for sex, foreign background, age over 60, single household, length of residence over 5 years and subjective poverty. 

On the micro-neighborhood level concentrated disadvantage is a dummy extrapolated from a grid of income levels in quartiles where a micro-neighborhood have been assigned  1 (=concentrated disadvantage) if a majority of the surface in the micro-neighborhood consist of grids with over 50% of the households registered as low income according to statistics Sweden 2008 and the remaining micro-neighborhoods coded as 0 (=no concentrated disadvantage) (See Map below). 

 

The level 2 variable for physical disorder is based on observed physical disorder from a Systematic Social Observation (SSO, Sampson & Raundebusch 1999) performed in two rounds during 2011, once during the summer and once during the winter (See Map below for observations plotted). Each observation is based on an object (ie, waste bin, park bench, door etc) with at least one marker of physical disorder (ie, grafitti, vandalism) which was documented with a photograph and tagged with a GPS-location. The log value of the combined number of observations from both SSOs have been aggregated to the micro-neighborhood in this paper.   

The third level 2 variable finally is collective efficacy which is based on two items each for trust (Cronbach alpha = .710) and informal social control (Cronbach alpha = .782) departing from the same questions used in Sampson & Wikström (2008). The questions are graded in likert scales from ”Always” to ”Never” , with somewhat low but acceptable reliability (Cronbach alpha for collective efficacy = .735). The data presented here is the micro-neighborhood aggregate of the mean of the four survey items included.

Results

Empty models show that about 7% of the variance lies at the micro-neighborhood level, which after controlling for individual level differences drops to 2.6%. In model 1 length of residence is significant at the 95% level, and subjective poverty at the 90% level. Being single, being old, being female or having a foreign born parent are all insignificant. In models including micro-neighborhood predictors subjective poverty is insignificant, while length of residence remains a stable predictor. It is interesting to note that length of residence has a positive correlation with avoidance behaviour. Someone who has lived over 5 years in the same place is 70% more likely to report avoidance behaviour due to not feeling safe ("otrygghet"). Micro-neighborhood disadvantage is significant in model 2 and 3, but not after introducing observed micro-neighborhood disorder in model 4 and 5. Before controlling for observed disorder a respondent living in a disadvantaged micro-neighborhood is roughly two times more likely to have expressed avoidance behaviour than a respondent living in a non-disadvantaged micro-neighborhood.

Collective efficacy and observed disorder are both just significant on the 90%-level, and in the final model none of the level 2 variables are significant. This is not very surprising however, considering the high correlation of the three level 2 variables and the low number of level 2 units. But it does appear that the disadvantage variable is more important than collective efficacy. This is in line with some European research (ie, Bruinsma et al 2013) that tends to show that collective efficacy is not as important in Europe as it has been shown to be in the US (ie, Sampson et al 1997). 

The results also point to the need for small-scale geographical units of analysis. The effect of micro-neighborhood disadvantage would not have been (as) visible if the whole neighborhoods had been studied. The two neighborhoods where the disadvantaged micro-neighborhoods can be found each have a micro-neighborhood with tenant ownership associations as well where income levels are significantly higher, and in addition one of those neighborhoods include a large single-family housing area not iucluded in this study but with even lower levels of disadvantage. 

Regression table, for avoidance behaviour in HLM 6.

* = p<0.05, ** = p<0.01 *** = p< 0.001 + = p<.1

Variabel	Model 0	Model 1:  Individual level variables	Model 2:  M1+Lvl 2 concentrated disadvantage	Model 3: M2+CE	Model 4: M2+SSO	Model 5:  M2+CE+SSO
Fixed effects	Odds ratio (C.I)	Odds ratio (C.I)	Odds ratio (C.I)	Odds ratio (C.I)	Odds ratio (C.I)	Odds ratio (C.I)
Individual level variables
Female		0.88 (0.57, 1.34)	0.86 (0.56, 1.31)	0.88 (0.57, 1.34)	0.87 (0.57, 1.33)	0.88 (0.57, 1.34)
Foreign born parent		1.38 (0.86, 2.21)	1.29 (0.08, 2.08)	1.23 (0.76, 1.99)	1.21 (0.75,  1.96)	1.21 (0.74, 1.95)
Age over 60		0.84 (0.52, 1.37)	0.83 (0.51, 1.36)	0.85 (0.52, 1.39)	0.87 (0.53, 1.42)	0.87 (0.57, 1.42)
Single		0.93 (0.60, 1.44)	0.95 (0.62, 1.48)	0.95 (0.61, 1.46)	0.93 (0.60, 1.44)	0.93 (0.60, 1.44)
Length of residence >5		1.73 (1.05, 2.87)*	1.71 (1.03, 2.83) *	1.73 (1.04, 2.87) *	1.70 (1.02, 2.82) *	1.71 (1.03, 2.84) *
Subjective poverty		1.50 (0.94, 2.39)+	1.45 (0.91, 2.33)	1.44 (0.90, 2.30)	1.44 (0.90, 2.30)	1.43 (0.89, 2.23)
Micro-neighborhood level variables
Concentrated disadvantage (majority of space in grids registered for >50% of households low income)			2.43 (1.27, 4.65) *	1.97 (1.05, 3.68)*	1.31 (0.49, 3.51)	1.42 (0.49, 4.11)
Collective efficacy				2.92 (0.80, 10.71)+		1.65 (0.23, 12.08)
Observed disorder (log of SSO results)					3.33 (0.77, 14.44) +	2.38 (0.32, 17.98)
Intercept	0.27 (0.16, 0.44) ***	0.14 (0.06, 0.32) ***	0.11 (0.05, 0.24) ***	0.12 (0.06, 0.26) ***	0.14 (0.06, 0.31) ***	0.13 (0.06, 0.31) ***
Between neighborhood variance
ICC	7.23%	2.63%	0.20%	0.04%	0.20%	0.22%
Reliability	0,787	0.671	0.365	0.204	0.359	0.372

 Descriptives

Variable	N	Min	Max	Mean	Std dev
Sex (Female = 1)	667	0	1	.61	.487
Foreign background (At least one parent foreign born = 1)	676	0	1	.48	.500
Age (Over 60 = 1)	681	0	1	.47	.499
Household composition (Single = 1)	677	0	1	.55	.498
Length of residence (Over 5 years = 1	677	0	1	.70	.457
Subjective poverty (Insufficient household income = 1)	658	0	1	.25	.434
Avoidance behaviour (Always, often or sometimes = 1	656	0	1	.20	.400
Level 2 (Micro-neighborhood)
Concentrated disadvantage (majority of micro-neighborhood surface registered as > 50% low income = 1)	12	0	1	0.33	0.492
Collective efficacy (Aggregated from survey, individual level descriptives in parentheses. Reverse coded, 1-5, with 1 being highest possible)	12 (596)	2.32 (1.0)	3.31 (5.0)	2.7944 (2.7802)	.282 (.80670)
Observed physical disorder (log of total number of observations)	12	1.26	2.44	1.8858	.384

 References

Bruinsma, G. J., Pauwels, L. J., Weerman, F. M., & Bernasco, W. (2013). Social Disorganization, Social Capital, Collective Efficacy and the Spatial Distribution of Crime and Offenders An Empirical Test of Six Neighbourhood Models for a Dutch City. British Journal of Criminology (Online advance access, published May 20, 2013)

Gerell, M. 2013. Skadegörelse, bränder, grannskap och socialt kapital. Malmö University Publications in Urban Studies (MAPIUS), Malmö

Raudenbusch, SW. & Bryk, AS. (2002). Hierarchical linear models: Application and data analysis methods. Sage publications inc

Sampson, RJ., Raudenbusch, SW, Earls, F. (1997) Neighborhoods and Violent Crime: A Multilevel Study of Collective Efficacy, Science Vol. 277, No. 5328, pp 918-924 

Sampson, RJ. & Raudenbusch, SW. (1999) Systematic Social Observation of Public Spaces: A New Look at Disorder in Urban Neighborhoods. American Journal of Sociology, vol. 105., No. 3, pp 603-651 

Sampson, R. J., & Wikström, P. O. (2008). The social order of violence in Chicago and Stockholm neighborhoods: A comparative inquiry. Order, conflict, and violence, 97-119.