Predictors of Disguised and Simulated Handwritten Text

Introduction

Disguised and simulated writings are challenging for forensic handwriting experts (FHEs) when assessing authorship. Previous research has reported that authorship opinions on these writing types attract higher inconclusive and misleading rates than natural writing does (Found & Rogers, 2005, 2008). While research by the authors has found that FHEs have a skill over laypeople in correctly determining the writing process of pairs of naturally written and disguised writings (Bird, Found & Rogers, 2010a), there does not appear to be a skill (in either FHEs or laypersons) associated with determining the process of production of individual samples that are either naturally written or disguised (Bird, Found & Rogers, 2012a). Unpublished research by the authors has explored the skill of FHEs in discriminating between disguised and simulated writing samples and found that while the strength of the evidence of an FHE was greater when they gave an opinion that a sample was simulated rather than disguised, the probative value of the evidence for both disguised and simulated writings is low (Bird, Found & Rogers, 2010b).

The task of discriminating between disguised and naturally written samples has been found, in further unpublished research by the authors, to be affected by the success of the disguise strategy employed. Those disguised writings that display many of the writer’s natural writing characteristics were often mistaken for being naturally written (Bird, Found & Rogers, 2012b). This means that when examining questioned natural and disguised samples, FHEs had a tendency to call disguised samples natural. Furthermore, when examining questioned disguised and simulated samples there appears to be a propensity for calling simulated samples disguised (Bird, Found & Rogers, 2010b). These results suggest that FHEs’ expectations of the predictor features of these unnatural writing types are not accurate.

This paper investigates the relationship between FHEs’ responses on the process of production of questioned disguised and simulated handwriting samples and their verbal statements relating to the features they observed as indicative of the particular unnatural writing behaviour. A clear relationship between FHEs’ verbal statements and correct responses will enable elucidation of predictor features of disguised versus simulated writings.

Methods

Analysis

The observations recorded by FHEs regarding the features they observed as indicative of either disguise or simulation behaviour were free text. The groups’ observations were reviewed so that suitable identifiers could be developed. This led to the collation of a list, aligned to features commonly reported in the literature (Alford, 1970; Harrison, 1958, Keckler, 1997; Konstantinidis, 1987; Regent, 1977; Totty, 1991) and used in other research by the authors (Bird, Found & Rogers, 2010a, 2012a, 2012b). This list was further refined by grouping features together which relate to a common theme, to facilitate statistical analysis. The groupings (factors, in the statistical analysis) are given in Table 1. The identified factors, as shown in Table 1, are SPEED, CONSTRUCTION, SIZE, SPACING, NEATNESS, SLOPE, PRESSURE, FORMAT and TREMOR.

Table 1

Groupings of observed features into factors.

Factor	Observed features
Speed	Writing speed
Construction	Altered letter construction/style Disconnected letter forms Altered ascenders or descenders Altered initial or terminal strokes Angularity Altered stroke order Unusual/awkward letter forms/embellished Proportions No attempt to simulate Connectivity Inconspicuous significant dissimilarities Diacritic Similarities in letter forms/details Imitation of obvious features
Size	Letter size
Spacing	Spacing
Neatness	Neatness Retracing/overwriting Corrections Lack of internal consistency
Slope	Altered slope
Pressure	Apparent writing pressure
Format	Altered baseline Altered punctuation Spelling Margin habits
Tremor	Complexity/skill level Fluency

To ascertain how the identified factors are associated with FHEs’ responses, Population-Average Generalised Estimating Equations (PA-GEE) with binomial distribution and LogitLink was applied (Liang & Zeger, 1986). Also taken into account was how FHEs’ responses compared to the actual behaviour employed (disguised or simulated) for each questioned handwriting sample. There were 29 subjects in the sample set, but each subject contributed 100 repeated observations (one ‘observation’ for each of the 100 questioned samples). As the repeated observations are nested within the separate subjects, the data set exhibits a typical two-level structure (see Figure 1). The PA-GEE was applied because of the multi-level data structure. However, there was no real advantage to apply other hierarchical or multi-level models such as the Generalised Linear Mixed Model (GLMM), where the fixed and random effects must be specified. The repeated observations are assumed to be correlated in an exchangeable pattern in PA-GEE. Theoretically speaking, this specification is most appropriate and it has been proved that the results would remain consistent if the working correlation matrix is mis-specified. Analysed with Stata 11.0 (Stata Corp, Texas, USA), all statistical tests were conducted at 5% level of significance. Tests were undertaken using the unforced responses of the 29 participants. The model was analysed to determine if there were significant differences in the factors associated with correct or incorrect answers when taking into account the handwriting behaviour, i.e. if different factors are associated with obtaining a correct (or incorrect) answer for disguised versus simulated samples.

Figure 1

Results

Table 2 gives the results for the model, where examiners were willing to offer an opinion as to whether a questioned sample was disguised or simulated. It should be noted that where a significant Adjusted Odds Ratio (AOR) is above unity (i.e. greater than 1) for a feature, there was a higher chance of obtaining a correct answer for samples of that handwriting type. Conversely, where a significant AOR is below unity (i.e. less than 1), there is a higher chance of obtaining an incorrect answer for samples of the relevant handwriting type.

For disguised samples, the only factor significantly related to a FHEs’ responses is SLOPE, which has a positive relationship. Conversely, for simulated samples, there is a higher chance of having a correct answer without SLOPE. SIZE is also significantly, negatively associated with FHEs’ responses for simulated samples, while CONSTRUCTION and TREMOR are significantly, positively associated with responses.

Discussion

These results suggest that the identification of altered slope in a questioned sample when compared to a naturally written sample may be a predictor of disguise behavior (Figure 2). Harrison (1958) states that deliberate changes to slope are rarely constant in disguised writings. This internal inconsistency in slope may also assist in detecting the disguise strategy in individual handwriting samples. If altered slope is not observed as a feature in the comparison of questioned and naturally written samples, this may be a predictor of simulation behaviour.

Figure 2

Slope and size are both relatively obvious features in handwriting as opposed to stroke order, stroke direction or writing pressure, for example. It may thus be supposed that these obvious features are copied reasonably well by simulators; hence the greater chance of a correct answer in the absence of their observation by FHEs in questioned simulated samples.

The positive association of letter construction and tremor features with a correct answer for simulated samples, along with these features not being significantly associated with answer for disguised samples, means that they may be used as predictors of simulation behaviour in questioned samples (Figure 3).

Figure 3

A limitation of the analysis undertaken revolves around the diversity in disguise strategies. Although the literature reports some disguise strategies as commonly used (altered slope, letter design, size, see Huber & Headrick, 1999), the success of disguise, even when the same strategy was employed, varied widely from writer to writer (Bird, Found & Rogers, 2012b). Any single predictor feature will only be relevant in cases where the particular strategy (or by-product) has been employed (or is observable), so will not apply to all disguised samples.

Conclusion

The results of this study suggest that observations of altered slope, differences in features of construction and size and the presence of tremor may be used as predictors to discriminate between disguise and simulation behaviour. Observing an alteration in slope was found to be significantly associated with correct answers for disguised samples while observations relating to construction and tremor were significantly associated with correct answers for simulated samples. The absence of observations relating to size is also significantly associated with correct answers for simulated samples.