Selected Research Papers in Social Change, Education, Labour Market, and Criminology Volume II

partner at the time of the crime

Yes 134 44 % No 172 56 % Sex of offender Male 204 75 % Female 67 25 % Victim–offender relationship Stalking by an ex-partner 102 38 % Stalking by a non-ex-partner 164 62 %

2.2.3 Impact of stalking profile (criterion variables)

The effects of stalking on victims can be diverse. The spectrum of negative consequences of stalking range from physical damage over impairment of one’s mental health to suffering social and economic disadvantages (Hoffmann 2006: 149 ff). In order to place some structure on the myriad of potential consequences of stalking, we will differentiate between external (physical, social, and economic) and internal (psychological) effects of stalking.

Figure 4 depicts the external consequences of stalking surveyed, and their frequencies of occurrence. For each of the possible consequences of obsessive pursuit listed, the victims were to specify whether they had been affected by it or not.

The next step determined, for each of the respondents, how many of the external consequences of stalking named in Figure 4 had occurred. This resulted in a sum score with a range between 0 and 11. On average 2.2 negative consequences were reported (standard deviation: 1.89).

Figure 5 provides information concerning the internal consequences of stalking assessed and their frequencies of occurrence. For each of the psychological impairments listed, the respondents were to indicate whether or not they had experienced them as a consequence of being stalked.

In order to quantify the number of consequences stalking has on the victims’ psyche, a further sum index was created. Within a possible range from 0 to 11, the respondents took note of, on average, 4.6 negative psychological reactions (standard deviation: 2.63).

Patterns of Stalking Victimization: A Behavioral Typology 119

Figure 4: External consequences of stalking for the victim (n = 311, multiple answers possible)

0% 20% 40% 60% 80% 100

%Avoidance of specific places

Withdrawal from social circles

Change in recreational activities

Avoidance of new contacts

Repair costs

Physical injury

Loss of job

Cost of protective measures

Change of residence

Payment for orders placed

Other consequences

47%

43%

36%

18%

17%

13%

12%

11%

10%

4%

10%

Figure 5: Internal consequences of stalking for the victim (n = 311, multiple answers possible)

0% 20% 40% 60% 80% 100

%Stress

Anxiety and panic attacks

Depressed moods

Fatigue and loss of sleep

Loss of confidence

Lack of concentration

Apathy

Reduced productivity and reasoning skills

Loss of self-esteem

Self reproach

Other consequences

85%

64%

56%

51%

41%

35%

34%

33%

32%

23%

2%

Patterns of Stalking Victimization: A Behavioral Typology 120

2.3 Strategy of analysis

The segmentation of the victim population was accomplished with the assistance of a latent class analysis (LCA) (Hagenaars/McCutcheon 2002; Magidson/Vermunt 2004; McCutcheon 1987). The LCA has applied in its function as probabilistic cluster analysis for categorical variables (Bacher 1996: 353 ff). “The primary aim of cluster-analytic statistical methods is to condense a set of classification objects into homogeneous groups (= classes, clusters, types) or – to put it simply – to discover an empirical classification (taxonomy, typology)” (Bacher 1996: 1, translated from original German). The amalgamation of the objects to be classified (here: persons) is conducted in accordance with two criteria: (1) Persons who are to be grouped together should be as similar to one another as possible. (2) Persons who belong to different clusters should be as different from one another as possible. The results of such an analysis are groups of persons which can be interpreted as types.

The main task of an LCA is to enable the formation of such clusters for categorical variables. Since conventional methods of cluster analysis require grouping variables measured at interval or ratio level, the LCA was developed to allow for the use of dichotomous, and later also nominal-polytomous and ordinal variables. The central idea behind such an LCA is simple to describe. A population of classification objects (here: respondents) is decomposed, based on their answer profiles for a series of dichotomous items, into subgroups with identical response patterns. These subgroups form the categories of a latent categorical variable (= latent classes). A background assumption is that the individual item answers are not independent from one another, and that the associations between the item responses can be explained by class affiliation. Within the individual classes, no systematic correlation may exist between the item answers (= local independence). It is now up to the LCA to determine the smallest number classes necessary to absorb the relationships between the variables observed. Model parameters of such an LCA are on the one hand class sizes (the proportions of the latent classes) and on the other hand the conditional item probabilities (the probability that the expression i of the nominal variable j will occur in latent class k).

In addition to the identification of population segments with identical response profiles, a further objective of the LCA is the classification of the units of analysis. Supposed class membership is determined by the maximum of the posterior probability of belonging to a particular class: A person is assigned to that class for which the membership probability is highest.

Our LCA was calculated with the statistics program Mplus 5.2. (Muthén/Muthén 2007). The robust maximum likelihood procedure (MLR) was applied to estimate the models. In order to facilitate the interpretation of the results, the segmentation of the victim population was not based on the individual stalking behaviors depicted in Figure 2, but rather rests upon the general stalking tactics proposed by Spitzberg and Cupach (2007) that were presented in Figure 3.

In the identification of the determinants of class membership, multinomial logistic regression analyses provided a major recourse (Orme/Combs-Orme 2009: 91 ff).

Patterns of Stalking Victimization: A Behavioral Typology 121

In order to examine the influence that the victimization profile has on the consequences of stalking for the victims, one-factorial analyses of variance (Turner/Thayer 2001) were implemented. Logit models and variance analyses were calculated with SPSS 16.0 (Bühl 2008).

3 Results

3.1 Segmentation of the victim population

When processing the results of an LCA, the first step is always to decide on the number of latent classes. This decision is usually made based on statistical parameters of model fit as well as on substantive interpretability of the classes identified. By default, the analysis starts with a one class solution. Then the number of classes is increased, each time by one, until one arrives at a model which provides an adequate fit to the data. As an additional constraint, it is required that the classes identified allow for a convincing characterization with regard to their contents.

13Table 4 supplies information on the goodness of fit of the one to five class models. The determination of the optimal number of latent classes requires a comprehensive consideration of a whole string of fit indices. Of central importance are, on the one hand, information criteria of model fit, and on the other hand statistics pertaining to the significance of model improvement resulting from the introduction of additional classes. Information criteria are based on the deviation between observed and predicted cell frequencies. They indicate how well a particular model fits the empirical data. Various information criteria differ with regard to the weight given to the number of estimated parameters and sample size. Here Akaike’s Information Criterion (AIC) (Akaike 1987), the Bayesian Information Criterion (BIC) (Schwartz 1978), and the Sample Size Adjusted BIC (SSABIC) (Sclove 1987) were applied. Lower values denote a ‘better’ fit. In addition to these information criteria, likelihood ratio tests can be conducted which examine whether a model with more parameters (classes) fits the data significantly better than a model with fewer parameters (classes). Specifically speaking, the LoMendell-Rubin Adjusted Likelihood Ratio test (LMR-LRT) (Lo et al., 2001) and the Bootstrap Likelihood Ratio Test (BLRT) (McLachlan/Peel, 2000) were applied. Both tests compare a K class model with a K-1 class model. If the corresponding p-value falls below .05, this indicates a significant gain in goodness of fit and implies that the K class model is preferable to the model with one class fewer.

On balance, the fit statistics assessed point towards a decision for the four class model. Although the BIC attains its lowest value with the two class solution, all other measures of model fit favor the four class model. The AIC and the SSABIC indicate that the data can be best reproduced by a model with four latent classes. The two likelihood ratio tests recognize a steady improvement in goodness of fit up to the four class solution. However, an increase in the number of classes from four

13 From the AIC to the BIC to the SSABIC, an increasing degree of importance is accorded to parameter number and sample size.

Patterns of Stalking Victimization: A Behavioral Typology 122

to five no longer resulted in a significant gain in goodness of fit. Therefore, the five class solution was the first solution to be rejected.

Table 4: Goodness of fit of the 1 to 5 class models

Model fit statistic Model LL AIC BIC SSABIC LMR-LRT BLRT One class model -1228.34 2470.69 2496.87 2474.67 ---- ---- Two class model -1115.11 2260.23 2316.32 2268.75 .000 .000 Three class model -1099.92 2245.84 2331.85 2258.90 .028 .000 Four class model -1086.53 2235.06 2350.99 2252.67 .011 .000 Five class model -1079.95 2237.90 2383.75 2260.06 .442 .429

LL = Log-Likelihood; AIC = Aikake’s Information Criterion; BIC = Bayesian Information Criterion; SSABIC = Sample Size Adjusted BIC; LMR-LRT = Lo-Mendell-Rubin Likelihood Ratio Test; BLRT = Bootstrap Likelihood Ratio Test

After determining the appropriate number of classes, the next step is to characterize the subgroups identified. Table 5 displays the proportional share of each of the four classes in the victim population and the conditional item probabilities, i.e. the probabilities of the various stalking tactics surfacing, given a specific class membership. These conditional item probabilities were applied as cluster profiles in the description of the latent classes.

Table 5: Conditional item probabilities in the four class solution Class Stalking tactic 1 (Polytropic -intensive stalking) 2 (Polytropic -moderate stalking) 3 (Distant stalking) 4 (Surveillance)

Mediated contact .97 .97 .90 .00 Intimidation and harassment 1.00 .87 .52 .44 Surveillance .99 .66 .07 1.00 Interactional contact .92 .61 .00 .10 Invasion .93 .28 .05 .40 Physical aggression and violence .69 .19 .08 .14

Hyper-intimacy .33 .21 .00 .15 Class size (persons) 96 127 68 20 Proportion of entire category 31 % 41 % 22 % 6 %

Class one comprises close to one-third (31 %) of the population and is characterized by high prevalences of all seven stalking tactics. Its members were de facto all victims of attempts to establish unwanted interactional and mediated contact, surveillance, intimidation and harassment, as well as the invasion of private space. Also, the level of physical aggression and violence is highest here: More than two-thirds of the victims considered members of this class were confronted with physical attacks. One-third of the members of this group reported hyper-intimacy behaviors on the part of the perpetrator; even this proportion is higher than in the other classes. Due to the wide range of offensive behaviors the victim has to deal with, which is expressed in high likelihoods of appearance for

Patterns of Stalking Victimization: A Behavioral Typology 123

all, or at least most, of the stalking strategies, the victimization pattern underlying class one can be labeled polytropic-intensive stalking.

14The second, and largest, class encompasses 41 % of the stalking victims surveyed. The members of this group reported, almost without exception, being the targets of continued attempts at contact via communication media, and were very often subjected to repeated intimidation and harassment. Personal advances and surveillance activities were also frequently mentioned. In sharp contrast to the first class, however, physical aggression, violence, and behaviors which infringe the privacy or personal space of the victim play a minor role. In this class, stalking behaviors that are extremely offensive and highly intrusive, which in and of themselves interfere with the victim’s physical safety or cause damage to their personal property, are largely absent. Therefore, one can refer to the corresponding victimization pattern as polytropic-moderate. The third latent class is comprised of 22 % of the respondents. Their victimization profile is dominated by persistent attempts to establish indirect contact performed through means of communication. In just over half of the cases this was connected with intimidation and harassment of the victim. From other stalking tactics, these persons remain to a large extent spared. An image emerges of a perpetrator who physically remains at a distance, one that avoids personal contact and attempts to get in contact with the victim by communication media. In demarcation from the other classes, this group can be labeled victims of distant stalking.

Figure 6: Frequency of stalking profiles (n = 311)

60%

40%31% 41%

22

%20%Polytropic-intensive stalkingPolytropic-moderate stalking6% 0%

Distant stalking Surveillanc

e

The fourth class is also the smallest. It makes up 6 % of the sample. Based on the conditional probabilities of having experienced the various infringements surveyed, the stalking profile which crystallizes is a surveillance-focused one. Here the collection of information about the victim remains in the foreground. All members of this class report being victims of surveillance measures. In somewhat less than half of the cases, this was bound to intimidation and harassment of the victim. In four out of ten cases the monitoring was accompanied by invasion of the victim’s privacy or living space. No other stalking strategies were involved in the infractions

14 This characterization is also based on the fact that the probabilities of encountering the stalking scenarios “intimidation and harassment”, “surveillance” and “establishing direct contact” are lower in class two than they are in cluster one.

Patterns of Stalking Victimization: A Behavioral Typology 124

reported here. In short, the stalking behavior is characterized by systematic attempts to acquire knowledge and information about the victim, actions which of course do not always go unnoticed.

It should not go unnoted that we incorporate some ordinal information in the interpretation of a nominal class structure. But, from a substantive point of view, this makes sense here. Apart from differences in the number of heterogeneous stalking behaviors that can be found in the single classes, an interpretation of the established victimization patterns partly in terms of intensity is warranted by the legally protected interest. The aim of the anti-stalking legislation is to protect the freedom of independent living, which is restricted by the different phenotypes of stalking in various degrees.

17.59. This corresponds to a regression coefficient of 2.87 (p = 01)3.2 Determinants of the victimization profile In order to clarify the factors on which the concrete victimization profile is 15dependent, in a next step a multinomial logistic regression analysis (Orme/CombsOrme 2009: 91 ff) was calculated. A multinomial logistic regression examines the influence of one or several independent variables on a nominal polytomous dependent variable (here: class membership). Polytropic-intensive stalking was used as reference category with which the other categories of the target variable were contrasted. The independent variables included were age and sex of the victim, the victim’s relationship status at the time of the crime (whether the victim was living with a spouse or partner at the time of the stalking), sex of the offender, as well as victim-offender relationship. Table 6 shows the results. If the perpetrator and the victim engage in an intimate relationship beforehand, this drives the intensity of the stalking activities upwards. Pursuit by an ex-partner reduces the likelihood that the infringements will remain limited to mere surveillance or distant stalking. Also, the probability of the polytropic-moderate victimization pattern surfacing diminishes, but not to the same extent as for the aforementioned stalking types. Conversely this means that stalking by an expartner from a previous sexual relationship is linked to a higher risk of polytropicintensive victimization. In cases of persecution by a former partner, the probability that a victim will be a member of that class of victims which have to deal with the most comprehensive and serious perpetrator behaviors is raised by a factor of 16. This finding replicates the understanding gained through international research that cases in

15 The simultaneous integration of all five predictors into the regression model did not generate multicollinearity problems. All tolerances exceed .80; all variance inflation factors remain below 1.25.

16 The effect parameters for the most comprehensive and serious victimization patterns are not listed in Table 6 because class one was utilized as reference category here. They result from a modified logit model in which the reference group was not the class of “polytropic-intensive stalking” but the group of surveillance victims (without depiction).

Patterns of Stalking Victimization: A Behavioral Typology 125

17which the stalker is a former sexual partner embody the greatest potential for escalation (Hoffmann 2006: 8; Sheridan/Blaauw 2006: 74). In addition, there are indications that men fall victim to distant stalking somewhat 18more frequently than females. In that the gender of the victim, however, in total does not differentiate significantly among the different types of victimization under consideration, this relationship will not be further developed.

Table 6: Multinomial logistic regression analysis to predict class membership (n = 244; reference category: polytropic-intensive stalking)

Regresscoeff. B B Effect-coeff. eWald statistic a error p

Surveillance Stalking by an ex-

partner -2.87 .06 6.86 .009

Male perpetrator .19 1.21 0.08 .773 Age of victim .00 1.00 0.00 .971 Male victim .12 1.13 0.02 .896 Residing jointly with .62 1.86 0.99 .321

partner at the time of the crime

Distant stalking Stalking by an ex-partner -2.28 .10 16.52 .000

Male perpetrator -.24 .79 0.27 .606 Age of victim -.02 .98 1.54 .215 Male victim 1.18 3.25 4.04 .044 Residing jointly with .60 1.83 1.90 .168

partnerat the time of the crime

Polytropic-moderate stalking Stalking by an ex-partner -1.14 .32 11.03 .001

Male perpetrator .38 1.46 0.83 .364 Age of victim -.01 .99 1.03 .309 Male victim .38 1.46 0.46 .499 Residing jointly with .22 1.25 0.39 .534

partner at the time of the crime

Model fit χ² = 55.08; p = .000; Nagelkerke’s R² = .22

17 The prior relationship between stalker and victim also significantly discriminates among victimization patterns on a global level. A likelihood ratio test which compares the goodness of fit of the full model encompassing all five predictors with that of the same model devoid of the effect of

the offender-victim relationship signals, with a χ² value of 29.66 and an associated a error below .001, a significant explanatory power of the above-mentioned relationship.

18 Global likelihood ratio test: χ ² = 4.81; p = .186.

Patterns of Stalking Victimization: A Behavioral Typology 126

None of the other potential predictors of class membership which had been taken into account showed significant explanatory power. The exact form of stalking victimization could not be determined by the age of the victim, nor by his/her relationship status at the time of the crime, nor by the sex of the offender.

3.3 Construct validity

In conclusion, an analysis should examine the effects of the phenomenology of 19stalking on the victim. An analysis of the significance of the victimization profile with regard to the consequent hardships suffered by the victim simultaneously contributes to construct validation. Within the framework of a construct validation, the task at hand is to formulate plausible hypotheses concerning the relationship between the concept under evaluation and other entities, and then to assess them empirically. Should the postulated hypotheses be confirmed, then this raises the confidence in the validity of the measurement of the critical concept (Bortz/Döring 1995: 186 f). At the center of further explanations is the assumption that with an increasing versatility and intensity of the stalking behaviors committed by the perpetrator, the resulting detrimental effects for the victim will increase in number. Following this logic, it can be expected that persons who were victims of polytropic-intensive stalking will report more negative consequences of the unwanted pursuit than respondents who experienced polytropic-moderate stalking. The lowest number of negative effects should be manifested among victims who were confronted with mere surveillance or distant stalking.

To test this hypothesis a multivariate analysis of variance (Turner/Thayer 2001) was calculated – multivariate because the two variables under investigation (the number of external and internal consequences of stalking) are systematically

correlated (Pearson’s r = .65; p = .001). Table 7 displays the results. As shown in Table 7, the theoretical expectations were confirmed in all respects. The global multivariate test of whether the four types of victims are affected to varying degrees by the negative consequences of stalking revealed significant differences. These differences can be found for both the external consequences of victimization as well as the internal repercussions. Substantially, the same pattern arises in both cases. The victims of polytropic-intensive stalking suffered the most disadvantages. Ranked second in this regard were the victims of polytropicmoderate stalking. The lowest numbers of negative consequences were reported by the victims of distant stalking and persons who had been subjected to no other

19 This decision becomes difficult when the expected relationships do not materialize. “Should the hypotheses not be confirmed, then it is not clear whether the validity of the measurement or the hypothesis itself should be placed in doubt” (Bortz/Döring 1995: 186 f, translated from original German). Therefore, in construct validation, reference should only be made to hypotheses for which their being true is not disputed.

Patterns of Stalking Victimization: A Behavioral Typology 127

20infractions but surveillance. These two victim groups do not systematically differ from one another in terms of the effects of victimization.