By Sydney Dixon Howell & Nathan Proudlove | The University of Manchester

Abstract

In a large retail chain we used statistical regression to relate shrinkage to explanatory variables such as staffing, security, store layout and catchment-area demographics. There were large measurement errors in the retailer’s estimates of shrinkage, together with high correlations among the potential causal variables. These effects caused our models to give poor predictions of shrinkage for individual stores, but the models were highly statistically significant, which means that they can accurately forecast the average (and hence the aggregate) effects of policy changes affecting hundreds of stores. Factors associated with lower shrinkage included high turnover of stock, and high densities on the sales floor of staff, pay-points and customers. Our results suggest that crowding among staff and customers may be a more effective inhibitor of shrinkage than many traditional formal security precautions, such as CCTV and store detectives. Keywords: shrinkage, theft, retail crime, regression modelling

INTRODUCTION

Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 2 Shrinkage is a major retailing cost, but it has received surprisingly little academic attention. A typical US or UK retailer loses approximately 1.5% to 2% of sales revenue from shrinkage, a sum comparable to its total pre-tax profit (see for example Ernst and Young, 1991; British Retail Consortium, 2004; Hollinger and Langton, 2005). Such losses invite regression or other econometric analysis, though there should be extreme caution in inferring causal mechanisms. Econometric analysis, and in particular regression analysis, is rarely applied to shrinkage, though it has been used in various societal, retail and accounting fields (e.g. Carter and Holmberg, 1992; University of Florida, 1994; Brumm and Cloninger, 1995; Persons, 1995). Regression modelling is well established in other aspects of retailing management, such as predicting the potential turnover of a site (e.g. Brown, 1992; Birkin et al., 2002; McGoldrick, 2002). The objective of this paper is to report the results of a regression study of shrinkage inside one major retailer. Our results suggest that, with the exception of security tagging, many expensive security precautions are ineffective, and therefore potentially not cost-justified. Our results also suggest that factors which increase the Return on Assets and Return on Sales, such as higher utilisation rates for stock, fixed assets and staff, are also associated with lower levels of shrinkage. These findings suggest a testable hypothesis, namely that a thief’s subjective sensation of close physical visibility to other people (customers or staff) may be a stronger inhibitor to shrinkage than the actual physical ability of staff or security systems to scan for potential thieves. Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 3 The company studied is a broad-spectrum dry-goods retailer with a multi-billion pound turnover. It operates over 600 stores in the UK, and within these stores it recognises some 48,000 individual departments (around 80 per store) specialised to various product groups. The company prefers to remain anonymous, and in order to keep confidential the exact parameter values of the shrinkage models we produced, we present here the qualitative structure of the models for 10 important product groups, the levels of statistical and substantive significance achieved (p values and R2 values), illustrative graphs, using rescaled axes for confidentiality, and the main economic and behavioural implications of our findings. The structure of the paper is that the next section reviews the limited literature on retail shrinkage. The methodology section describes the data set, together with the prior theories from Accounting, Marketing and Security that motivated the data collection. The main findings are then presented, and the paper concludes with a summary and discussion of these findings, suggestions for future research, and practical recommendations for retailers.

RETAIL SHRINKAGE

Retail shrinkage can be defined as the value of inventory (including cash) that is lost through theft or through fraud by staff, customers or suppliers, or through administrative error. Losses from major fraud, robbery and burglary are usually excluded from shrinkage (Bamfield, 2004b) as are known losses due to management action, such as markdowns and scrapping. Losses are usually calculated at retail prices (Beck, 2004) though our subject company first estimated shrinkage at retail Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 4 prices, and adjusted this to estimated shrinkage at purchase prices, by deducting a standard gross margin for each product group. Several recent surveys illustrate the size of the losses from shrinkage and its main sources (see Table 1). Such surveys vary in how they aggregate the figures (Bamfield, 2004b).

The UK has the highest shrinkage levels in Europe. In addition to the £1.7 billion in shrinkage cost shown in Table 1, there are costs to UK retailers of nearly £1 billion in security precautions (staff and anti-theft technology) plus associated demands on the criminal justice system, tax revenue lost, and violence against retail staff, which jointly cost society a further £0.7 billion (Bamfield, 2004a). Across Europe, in the fast-moving consumer goods sector, 93% of shrinkage is thought to occur in stores (Beck, 2004). The contrast between the UK/Europe and the US/Canada, in the shares of shrinkage by customers and employees, is probably real, but reducing (Bamfield, 2004b). In the UK theft by staff is increasing faster than theft by customers, and around 16% of customer theft involves collusion with staff (e.g. on markdowns) (Bamfield, 2004a). In the Netherlands, theft by staff has prompted the establishment of a register of dishonest shop staff (Platform Detailhandel, 2004). However European retailers remain over-focused on theft by customers (Beck, 2004). Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 5 Over the last few years shrinkage has been falling in higher-shrinkage countries, such as the UK and US, and stable in lower-shrinkage countries (British Retail Consortium, 2004; Hollinger and Langton, 2005). The suggested reasons include an increasing use of security tagging (visible to customers). Around 50% of retailers plan to trial radio frequency identification (usually not visible to customers) in coming years (Bamfield, 2004a). According to Elder (1989) few retailers understand how shrinkage affects them, or how best to target security efforts. Ernst and Young (2003) recommended retailers to build ‘cause-and-effect’ models, which can identify the best and worst performing stores, evaluate various shrinkage-reduction efforts and guide the construction and testing of ‘optimal’ strategies. This paper describes the first stage of such a model, namely regression analysis to identify the potential candidates for causal variables.

METHODOLOGY

We had meetings with the security, merchandising and accounting staffs of the subject chain in order to identify variables which they expected to be correlated with shrinkage (and potentially causal). Data were provided in cross-sectional form for a single year, for 10 departments in each of over 600 stores. Some of the variables were defined at department level (for example security tagging was used only one department), but other variables applied to the whole store, such as the rate of violent incidents in store, and the demographics of the store’s catchment area. Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 6 In total 421 variables were collected. We do not list or discuss all 421, since from a noisy sample of only 600 stores we expected to be able to estimate at most 20 robustly significant coefficients. We first looked for variables which had a statistically significant pairwise correlation with shrinkage. However any single variable (e.g. stock) might be acting as a proxy for some other, more relevant variable (e.g. size of store, which might itself be a proxy for speed of stock-turn), and such a proxy might lose its predictive power when used alongside the more relevant variable. We therefore only retained variables which remained robustly significant when used alongside other variables, in several models. This left only eight “robust” variables, most of which were statistically significant in all the models reported later. The sample size of 600 was ample for estimating their coefficients. Hypotheses Since this was exploratory research, we formed only very general hypotheses before or during it. Most came from the company or the literature, but we believe two are original.

The company’s accounting and security communities held the intuitive hypothesis that the main process which inhibits shrinkage is the deliberate detection and prevention of crime by staff and technology. Factors such as CCTV, tagging, strong staff sightlines, and separation of the power to move an asset from the power to record its movement, were all expected to reduce shrinkage. Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 7 Security staff also expected a “busy-ness effect”, by which there is less theft in the largest and busiest stores, and at the busiest times of the week and the year. Our cross-sectional annual total data could only compare busy with less busy stores. On demographics, security staff expected that older customers, older staff, longer-serving staff, more full-time staff and more socio-economically privileged customers might all be associated with less shrinkage. Merchandising and marketing staff suggested a possible tradeoff between displays meant to maximise purchase, and displays meant to minimise theft. We ourselves suggested what we call a “marketing” (or “merchandising”) model of security. This uses the contrast between a “sales” person (whose role is to make a sale face to face) and a “marketer” (who may never make a sale face to face, but tries to create the conditions to assist many sales).

Applying this contrast to security, the “sales” or face-to-face security approach searches for individual crimes, and tries to prevent or to punish every one. We can call this a rational model of prevention by staff. In contrast “marketing” security staff might prefer to watch crime passively, and interview suspects without prosecution, in order gain insights to inhibit future crime. Their observations might include irrational behaviours or perceptions, which are best studied by marketing-type methodologies. In the “marketing” model of shrinkage, a thief is a client for a ‘service’ offered by the store – the opportunity to steal. The population of these ‘clients’ has demographics, needs, preferences and consumption patterns for the ‘service’ of theft, which need to be analysed, so that the store can market the service negatively. Our study collected no data on individual thieves or thefts, but its results suggest that decisions not to steal Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 8 have more to do with irrational subjective fears on the part of the thief (a “marketing” factor) than with the objective probability of being detected (a rational staff preventive factor). We also hypothesised what we call a “time constant risk” model, namely that when stock is exposed to a fixed population of predators (staff and customers) their activity per unit of time is constant. On this “rational” model a faster stock-turnover should reduce the stock’s dwell time in the danger zone, reducing the shrinkage percentage, but this model makes no unconditional prediction for the effect of increasing the total population of staff and customers, with unchanged proportions of honest and dishonest. Method of analysis We expected to find outliers, collinearity between the variables, and heteroscedasticity, since all are seen in large business data sets. A few large outliers were easily detected and explained. Collinearity, especially with size, is an unavoidable effect of many corporate policies (e.g. “no cafés in stores with turnover below £X”) and we discuss it later. An unexpectedly serious problem was very large error in the measurement of shrinkage itself, and this restricted the type of analysis we could do. On average, shrinkage over the whole company was positive (a loss of stock) as expected, but the reported shrinkage for any one department was almost as often negative (an unexplained gain of stock) as positive, despite being aggregated over a year. The high level of measurement error made it impossible to “predict” the reported shrinkage of an individual department (low R2). However our models could Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and

Consumer Research 17:2, 101-120. 9 identify trends shared by hundreds of similar departments, with high precision (high F)1. The high level of measurement noise restricted the range of possible analysis methods. It was not appropriate to use artificial neural nets (ANNs) since they tend to “over fit” data, especially noisy data as here – Ripley (1994). Also ANNs do not provide coefficients for the effects of individual variables, and so give little guidance on policy. Simultaneous equation models (SEM) are also known to amplify noise, relative to Ordinary Least Squares (OLS), even when mutual causality is present, if the data are noisy (e.g. Wonnacott and Wonnacott (1970)). It seems that in such cases OLS can reduce the mean squared error by more, through not amplifying sample noise, than it inflates the MSE, through the bias it creates by ignoring two-way causality. A further drawback of SEMs (for this study) is that mutual causality seems unlikely here.

For example the level of theft in a store’s catchment-area may plausibly predict (if not affect) the level of theft in the store, but it is not likely that the level of theft in a single store (or department) can detectably affect the total level of crime in its catchment area. In the special case of CCTV some mutual causation seems hypothetically possible, on the assumption that if theft drives CCTV usage up, the increased CCTV usage somewhat reduces theft. However we found no evidence for this: there was positive correlation of CCTV with theft, but no detectable negative partial correlation (details omitted for reasons of space). This suggested that a high level of theft drove up the level of CCTV, but CCTV itself had no effect. The positive correlation seemed due to company policy (“install extra CCTV if shrinkage exceeds £Y”). We avoided Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 10 making a statistical estimate, based on noisy data, of the CCTV policy which the company had itself set 2. For all these reasons, and for the sake of robustness and ease of interpretation, we built reduced form models using stepwise OLS regression. As and when the company can reduce the noise level in its measurements of shrinkage, it would be interesting to apply two-stage least squares (2SLS), but the low level of mutual causality may lead to little change from the OLS models. The data set We summarise the variables in Table 2. The subject group measured shrinkage as the “error” found in the stock accounting equation, upon taking physical stock at a notional cost price (Howell and Lehocky, 1997). On this definition shrinkage includes process losses and administrative errors, but excludes known losses such as scrapping and markdowns3. In the case of tagging, which had been introduced by different stores at different stages of the year, one variable measured the fraction of the year for which tagging was in use.

We have noted the excess of variables (421) relative to data points (600) and the problems of collinearity and measurement error further limited our ability to estimate coeffcients robustly. Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 11 Collinearity among the independent variables We expected useful collinearity among the catchment-area demographic variables, such as local levels of income and of crime, and we expected these to correlate with the rate of public order offences within the store. We also expected correlation between the average age of the staff and the age of the customers, and some correlations (of varied signs) between both these ages and the level of cash losses from tills, and the level of compliance with company audit procedures. Any of these variables might also be correlated with shrinkage. More seriously, as we have already noted, there was large management-induced collinearity with sales. Performance targets tend to be set as multiples of sales, and revenue and investment spending policies are often sales-driven. Further policy-induced collinearity arises because chains of stores choose to build in standard types of locations, leading to standardised demographics for both staff and customers, and to standardised sets of local competitors. The classic response to collinearity is either to delete variables, or, where feasible, to set up experiments to break the collinearity. The latter was not possible. We deleted all variables which had no significant pairwise correlation with shrinkage, and we also deleted variables which did not remain significant when used alongside other variables, on the assumption that that the other variables might be more efficient proxies for (potentially) even deeper factors.

This very conservative policy was unavoidable, given the large ratio of variables to observations, but the fact of deleting a variable by this policy does not mean that that the variable truly has no effect. A less Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 12 noisy measurement of the dependent variable might have allowed us to identify additional robustly significant independent variables. Heteroscedasticity There was extreme heteroscedasticity (see Figure 1). Shrinkage ratios for departments with small levels of sales were impossibly noisy, although their aggregate money value was small. Figure 1 has already been trimmed of a few extreme outliers in the y dimension (ratio of shrinkage to sales), but adjustment is needed for the remaining, orderly heteroscedasticity. However, adjustments for heteroscedasticity can interact with the causal and statistical specifications which are being used, either to adjust the data or to estimate the resulting model, so we used them sparingly.

We tried specifying shrinkage in absolute money value (with or without a heteroscedasticity correction) in the hope that money values would prove more orderly than ratios to sales. Unfortunately this was thwarted by multicollinearity with sales, since all of the monetary values were correlated with sales, as were many of the operating ratios.

If we used sales as a single independent variable, there seemed to be a favourable effect of scale on shrinkage, in the form Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 13 y = axb where y is money shrinkage, x is money sales, and a and b are parameters with a, b > 0 and b < 1. Figure 2 shows one example. If we added further variables alongside sales, the results became insignificant unless we constrained b to be unity for all the variables. This forced us to choose between a univariate model, containing sales alone and containing an unexplained economy of scale, and a multivariate model, containing more of the potentially “causal” variables (e.g. stock-turn), but with no ability to detect whether these variables had any scale effects in their own right, or were acting as proxies for other scale effects (for example because larger stores had faster stock turns). We chose the multivariate type of model, because both types made similar errors, and a “sales only” model can produce no useful recommendation on policy, other than the empty one of making uniformly large sales in all stores. At worst, the robustly significant variables in a multivariate model can be empirically valid proxies for the true effects of larger sales, and they also give detailed suggestions for future study. We accordingly specified most of the variables as ratios to sales, even though this caused problems of heteroscedasticity, and required care in dimensional analysis for these variables.

We eliminated a few “outlier” data points, and we corrected for the remaining heteroscedasticity where this was large. Procedure to select a subset of candidate variables The 421 variables fell into the categories listed in Table 2. We first looked for high bivariate correlations with either shrinkage in money value, or shrinkage as a Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 14 percentage of sales, both before and after partialling out the correlation of shrinkage with sales. Only 74 variables survived this test. The variables rejected up to this point included all data on competitors, pay-points and in-store facilities, although several of these had plausible causal links with shrinkage if considered in isolation. We then built multivariate models using more than one of the 74 variables. Unfortunately only 8 variables continued to have robustly significant coefficients when used alongside other variables, plus a ninth variable, tagging, whose effect was robust but partly artefactual. This does not mean that the rejected 65 variables had no effects on shrinkage (or even that the retained 9 variables had more than proxy relationships with the causes of shrinkage). Cleaner data, and/or experiments designed to break collinearity, might reinstate any of the 65 omitted variables, and we discuss hypotheses on them in our conclusions. To avoid data mining we used all 9 of the “robust” variables in models for all 10 product groups, even though not every variable was significant in all models, since Wonnacott and Wonnacott (1990) among others recommend including all the plausible variables, as a precaution against over-fitting (in the form of falsely rejecting a relevant variable, on the basis of insignificant results in a single sample).

FINDINGS

Pairwise correlations between individual causal variables and shrinkage Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 15 Table 3 lists the variables significantly correlated with shrinkage, after partialling out the effects of sales. These pairwise correlations were mostly small (r2 < 0.22). The low r2 is expected from the high measurement error for shrinkage (and perhaps for other variables) but the large sample size of 600 allows the correlations to have extreme levels of significance, ranging from 0.01 to less than 0.0005. This means that the reality of the correlations is well established, and the aggregate effects of the correlation in large samples of departments can be predicted with some precision. Figure 3 plots an example of the pairwise correlations. Although the plot is noisy, it is free of gross outliers, and normality is acceptable as an approximation for regression slope testing. The low r2 causes large errors in “predicting” shrinkage for a single department by regression, as seen in the wide outer confidence interval plotted. However, because of the large sample size the confidence interval around the estimated regression line itself, also plotted, is narrow. Hence the regression line can predict with great accuracy the average or aggregate effect of the correlation upon a large sample of departments. The narrower, inner confidence interval plotted is for the true regression line (which is theoretically the mean effect on an infinite sample of departments). For clarity we omit the confidence interval around the regression predictions for the company’s actual finite sample size of 600 departments, but this is little wider than the interval as plotted.

Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 16 This chain recognised 34 frequent competitors, and the pairwise correlations showed that shrinkage was significantly higher when these competitors were present (detail omitted). This may mean that the competitors had better security than this group, and so were channelling thieves towards it. On this interpretation, in 2 product groups this chain was significantly and sharply inferior in security to its competitors; and in 7 product groups it was less sharply, but still significantly, inferior. The remaining product group showed no significant difference. For some competitors an explanation in terms of rational preventive security quality is plausible: two of the “significantly superior” competitors restrict the access of their customers to stock much more severely than our subject group does. However many of the competitor correlations might be proxies for effects not related to security performance as such. For example if one competitor favours demographically “tough” locations, the presence of this competitor – if used as the only independent variable – may be a proxy for a more theft-prone location. Conversely if one competitor preferentially attracts a more law-abiding staff and customer base, its presence nearby may bias the customer base of our subject chain towards potential thieves. This might justify collaborative research between the security and marketing staffs within our subject chain, and even some collaboration between competing chains. The results shown in Table 3 may seem disappointing to security staff, but encouraging to general managers, because they seem to suggest that expensive security precautions, such as store detectives and CCTV, have no independent effect on shrinkage.

We in fact found positive correlations between shrinkage and the use of CCTV and likewise for the numbers of detectives and arrests (detail omitted). This suggests that this chain’s large spend on store detectives, prosecutions etc. was in Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 17 effect an expensive way (and statistically not a very precise one) to confirm where the rates of shrinkage were highest. It seems generally plausible that store detectives have little deterrent effect on shrinkage, for many retailers, because store detectives make arrests discreetly. One possible topic for future “marketing” research on shrinkage is to investigate, by interview, the acquaintanceship networks among shop-lifters (on the hypothesis that there is a network of thieves, among whom information about arrests is preferentially transmitted). This would allow security staff to predict how many arrests would be needed to alarm the whole of the local network. If this number were large but feasible, it might prove useful to concentrate detectives, on a rotating basis, to achieve saturation fear messages in the local thief population. Conversely if saturation proves infeasible or unhelpful, it might be efficient to reduce the spend on detectives. Longer use of security tagging by the only department which used tagging during the subject year was correlated with lower shrinkage in that department, as expected. However, longer use of tagging in the store was correlated with higher rates of shrinkage in some other department, which did not use tagging. This seems an artefact. Tagging was applied first in the “worst” stores, where theft was high in many departments. Hence tagging (applied to one department only) could be an empirical proxy for a higher risk of shrinkage in every other department, when used in bivariate correlations. The tagging variable remained significant even at the later stage of multivariate modelling, when it was competing with local demographic variables, and this suggests that there are site-specific risk factors, which standard local demographics do not capture. Howell SD and Proudlove NC (2007).

“A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 18 Multiple regression models to predict shrinkage As already noted, we found only nine variables (including tagging) which were robustly significant for most product groups. Table 4 shows the nine variables, together with the usual signs (directions) of their relationships with the shrinkage ratio in the models for the 10 different product groups. The potential implications for management are interesting.

In Table 4 the demographic environment appears most often (three variables), but demographics are outside management’s influence, short of changing the portfolio of sites. The remaining variables reflect internal store performance, and hence offer opportunities for management. Among these, higher money stock turnover and higher turnover per stockroom square metre are associated with lower shrinkage. Among staff demographics, only staff age is robustly correlated with shrinkage in all the product groups. Only one efficiency variable from the internal audit is retained (till cash losses). Several of these variables imply that some of theft is by staff – as suggested by e.g. Platform Detailhandel, 2004; Bamfield, 2004a; Beck, 2004. More details of the regression models for the 10 departments are in Table 5. In every product group but one the regression was highly significant (p < 0.00005). For the exceptional product group the coefficients were non-significant, but they resembled those of the other product groups in sign and approximate size. This product group Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 19 had unusually small total sales, so its data may contain greater noise, and there may also have been occasional misallocation of accounting entries between this group and a larger product group, with which it was commercially related. The adjusted R2 values for the nine “significant” product group models were low, ranging from 0.06 to 0.16. This is because the dependent, measured shrinkage, was itself mostly measurement noise. However in large samples measurement noise cancels out, and overall the models are highly significant, as measured by high F ratios. Therefore these models have high predictive accuracy for the effects of any policy decisions or experiments, which affect a sufficiently large number of stores.

Even our surviving set of 9 independent variables (including tagging) was collinear. To test this we calculated the principal components of the correlation matrix of all 9 independent variables, for each product group. On the null hypothesis that all 9 variables are fully independent, their 9 eigenvalues (which must sum to 9) should all be close to 1, and deviations above and below 1 are equally probable. In every product group only one eigenvalue was above unity, being always at least 7.2, and the first four eigenvalues explained around 93% of the total standardised variance. This suggests that even our 9 robustly explanatory variables contain little more than 4 fully independent sources of variation. We suspect that the single very large eigenvalue reflects (in linearised form) the shared effect of sales volume on all the independent variables, and the remaining three factors may be in part an approximation to some of the non-linear and/or variable-specific effects of sales. Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 20 Setting aside the variables of tagging (used in only one product group) and percentage of full time staff (whose effects were weak) the remaining 7 variables tended to have the same sign in most of the 10 departmental models. This agreement of sign, between the coefficients of the same variable in different departments, can be verified non-parametrically, using a binomial test. The null hypothesis is that a variable’s coefficient takes either sign in any given model with probability p = 50%, the number of trials is 10 (models) per variable, and the number of successes is the total number of positive signs. All 7 variables violate the null hypothesis (that the coefficient’s sign is random from model to model) at 5% significance. Uses of the regression models Having built robust models, we then used them to investigate particular issues of concern to the company’s management.

It is interesting that although our models give wide confidence intervals around the predicted shrinkage for an individual department, in extreme cases even these wide confidence intervals can be useful. For example our models identified one store as having ‘significantly’ low shrinkage. We heard on enquiry that the local manager had recently been sacked for “fiddling the books” in order to conceal theft. Another site was identified by the model as having shrinkage significantly higher than predicted, but on re-audit the shrinkage proved to have been grossly mis-measured. We have noted that the confidence intervals shrink sharply as the sample size rises, and therefore any shared excesses (or deficits) of shrinkage become quickly detectable Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 21 when aggregated across larger numbers of stores. In this way we found that a new format for displaying goods had unusually high shrinkage. We suggested a design change, and we calculated a suitable sample of stores on which to experiment, in order to verify quickly that this had the desired effect. The models could also test some intuitive beliefs held by the group’s managers.

For example they believed that shrinkage was excessive in Greater London: our models confirmed this, and quantified a confidence interval for the effect. We also used the regression models to simulate the probable company-wide benefit of implementing changes to the more controllable of the explanatory variables. We predicted a payoff of some millions of pounds per annum pre-tax from extending tagging and making changes to stockrooms and staffing. We suggested a suitably large sample of stores and time periods, in order to test the proposed changes of policy with acceptably small error. The models also implied that the company could gain benefit at department level, if it could reduce the error in its measurements of shrinkage. This would allow more detailed models of local effects, and narrower confidence intervals to detect good and bad shrinkage management. However pending such improvements, our models could only detect extremely gross effects at the individual departmental level, as already noted, and the company could only accurately measure shrinkage for large aggregates of stores. The largest such aggregate is the total company, whose shrinkage is needed for financial accounting. Our models showed that for this purpose the company was “over-measuring”.

Despite the large measurement error for any one department, the company could in fact estimate its total shrinkage rate accurately from a relatively Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 22 small sample of all departments. This fact could be exploited either by doing far less measurement annually, or by detecting changes in total shrinkage by sample checks at monthly intervals.

DISCUSSION AND SUMMARY

Discussion In this section we discuss wider implications for future work. We have seen that the sample size of 600, combined with high measurement noise and collinearity, made it impossible to estimate independent effects for all the 74 variables (including tagging) which had significant pairwise correlations with shrinkage. It is possible that any of these variables may prove to have an independent effect on shrinkage in later study, so we include hypotheses about them. We have seen that many costly security precautions (other than tagging) such as overt and covert security guards, CCTV etc., may have little value in reducing shrinkage for this retail chain. This implies that shrinkage risk is mainly driven by the dynamics of the total selling operation, which also drive millions of honest, accurately recorded transactions. Shrinkage crime seems to correlate with other crime types, both in the catchment area and in-store, but shrinkage does not correlate clearly with local poverty, since theft seems lower among the most disadvantaged and housebound Acorn group. Older Howell SD and Proudlove NC (2007).

“A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 23 customers and staff are associated with lower shrinkage, and this may mean that they are less likely to steal in their own right, and/or that (testably) they may be more diligent, and confident, in intervening against crime by younger customers or by younger staff. A link between age and propensity to commit retail theft is found nationally (e.g. Bamfield, 2004a). Whatever the demographic risk factors, the internal dynamics of the store operation matter. The bivariate association between higher stock-turn and lower shrinkage supports our “time constant exposure model” of shrinkage risk, i.e. stock is less at risk if it moves through the system fast. An alternative explanation is that insiders and outsiders prefer to steal from smaller, less busy stores, located outside city centres, perhaps also closer to home, and it is only coincidental that these stores have slower stock-turn. “Marketing” type research on the psychological preferences of thief segments should clarify this. The finding that bigger and more productive stores suffer less theft is also indirect support for a hypothesis of the subject company’s security staff that “busyness” inhibits theft – but why?

This partly conforms to our “time constant” model, but other models are possible. One obvious hypothesis is that larger stores can monitor potential theft more easily – for example larger stores have more pay-points per square metre, and more staff per square metre, so that all stock is on average closer to some member of staff. However this fact does not improve the efficiency with which staff can scan the customers, because there are more customers per staff member in a large store, given that large stores tend to have higher sales per staff member. Hence each staff member has less time free to watch customers, and a larger number of customers to Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 24 watch. Further, when the large store is busy, staff may be able to scan only a small radius into the crowd, which causes a quadratic loss of floor area scanned. This suggests that in a larger and busier store a thief is objectively safer from staff observation, so the “sales” or rational staff preventive model of security is less effective. Yet these are the conditions where least theft occurs, so some other, less rational explanation, based on the psychology of the thief may be needed (the “marketing” approach to security).

We suggest a hypothesis of “customer mutual inhibition” from theft. This suggests that in busier stores, although the thief is less visible to staff, he or she is objectively more visible to fellow customers. Perhaps more importantly, the thief sees more fellow customers (and more of the absolutely more numerous staff) and this may give the thief an erroneous sensation of being watched (as when speeding drivers slow down when or even after they see a police car). Customer mutual inhibition should also be strong at peak times of the week or year (as reported by the company, but not directly observed by us). However stock-turn is then also high. It would seem interesting for future research to explore the relative roles of rational staff preventive explanations such as “time constant risk”, and the (potentially) less rational world of “customer mutual inhibition”. A similar effect may apply to staff thieves, and this may explain the robust finding that shrinkage is lower when stockrooms are smaller, relative to the sales rate (so that there is presumably more crowding of both stock and staff). Stock crowding ought to increase the efficiency with which a would-be staff thief can scan the stocks, thus increasing risk according to the “time constant risk” model, but crowding may also increase the scanning efficiency of the honest staff (a “rational staff preventive” factor). However staff crowding may also increase the thief’s sense of unease at the Howell SD and Proudlove NC (2007).

“A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 25 presence of others (staff “mutual inhibition”). Research seems justified to explore these competing mechanisms, and the components of their effects. The fact that staff or customer thieves may use non-rational ways to assess the likelihood of being caught presents many topics for “marketing” types of study. Such psychological effects can also explain several known results, such as the fact that it is useful to put public warning labels on the displays most in danger of theft (Carter and Holmberg, 1992), and why mirrors placed at eye-level are said to be more effective (Sanderson, 2004). These measures might increase the reality and/or the subjective sense that the thief is being surveyed by other customers, as well as by nearby staff. We conjecture (testably) that a policy of greeting customers may also inhibit theft, just as it is known to encourage sales (Underhill, 2000). We expect our findings to be highly pleasing to store profit managers, since they imply a “win-win” opportunity, to improve profit while reducing shrinkage, and without incurring large costs for security precautions. It seems that many factors known to boost profitability (such as faster turnover of the fixed and stock assets, higher staff productivity and greeting customers) may also reduce shrinkage. Conversely, costly security precautions, which reduce profitability without increasing sales, such as uniformed guards, detectives and CCTV, appear not be productive, at least for this chain (although it would be prudent to experiment before reducing them). Prospects for future research Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”.

The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 26 Our cross-sectional study needs supplementing with longitudinal studies, and with controlled experiments, in order to break the strong collinearities in the data. They need replication in other types of retailer. Where shrinkage measurements are noisy, the sample sizes required for experiments (and for efficient and fair management control actions) are computably very large. As Ernst and Young (2003) point out, EPOS environments are now permitting more accurate and cheaper measurement of shrinkage itself. This should permit higher R2 values in models, finer discrimination of the effects of all the causal variables, and more sensitive detection of good or bad shrinkage management. Our two hypotheses of “constant time exposure” and “customer and staff mutual inhibition” need comparative study.

The first is a rational staff preventive modelling approach according to which one should formally analyse, and experimentally test, the potential and actual rates at which customers and staff scan each other, at various levels of store occupation density, and in various layouts. From the potentially less rational world of customer and staff mutual inhibition, it may be possible to devise experiments comparable to those used in merchandising, to identify the determinants of a thief’s sense of unease at being watched. Potential ways to increase the psychological proximity of staff and other customers might include transmitting live images of working staff to many places in the store, or elevating the positions at which some staff work (with the goal of making the staff visible to the customers, rather than the converse, which is the goal of a rational staff preventive model). We have anecdotal reports (only) of success in the U.S. from placing life-size cardboard images of police in the store. If so, it should be testable whether life-sized cut-outs of Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”.

The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 27 staff might have an equal effect on potential thieves, and /or a better effect on purchasing by honest customers. Objective increases in customer density (desirable under the customer and staff mutual inhibition model, but less obviously desirable under the time constant model) could be achieved by shrinking the sales and stock floors, either permanently or, in large stores and at quiet periods, by closing some areas off temporarily (as restaurants do). Increases in the perceived or actual customer density may also stimulate buying, as happens in restaurants. To sum up our recommendations to retailers: use EPOS and RFID to measure shrinkage; use older staff; maximise the flow through stockrooms and the sales floor; minimise idle space in store and stockroom; use marketing and merchandising research tools to study the segments of your thief base, as closely as you would study the segments of your customer base (the two overlap). Include thieves’ choices between stores, their movements within store, their preferences, needs, subjective beliefs and their reference social networks. Finally, design experiments to find the best ways of negatively merchandising the opportunity for theft, such as making staff more visible to customers, and customers more visible to each other, and of optimising the balance between the security and the attractiveness of displays.

Summary

We obtained a year’s total shrinkage from 10 product groups in over 600 stores form a major retail chain, together with 421 potentially explanatory variables, and looked for Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 28 correlates of shrinkage in the ten product groups. The potential explanatory variables included internal financial efficiency measures such as stock-turn and sales per employee, and various internal and external variables describing the layout, staffing, customers and competitive environment of the store in total, and/or of each product group. The total of 600 stores provided on average only 1.5 data points per variable for each product group, which is too little to estimate robust coefficients for all 421 independents. We found huge errors in the measurement of shrinkage. The consequences of noisy shrinkage data, and the danger of using them as a basis for rewarding or penalising staff, were described in earlier work (Howell and Lehocky, 1997).

This large measurement error further reduced the efficiency of our estimation. We additionally found strong collinearity of all the variables with sales, which was partly caused by management policy, and which further reduced the efficiency of estimation. As a result we could identify only 9 variables which had robustly significant associations with shrinkage, and in consistent directions, across most product groups. Better measurement of shrinkage might permit better models, and therefore tighter management control action. Our resulting regression models had low R2 but high F values. The low R2 reflects a low ability to predict the shrinkage of a single department, as expected from the huge error in measuring shrinkage itself; but the high F shows that in samples of many departments, where measurement error cancels out, the equations can predict aggregate outcomes with high precision, and can therefore guide policy. Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”.

The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 29 We found some support for two competing models of theft inhibition which, so far as we know, are both original, namely the “time constant risk” model, which is based on rational prevention of shrinkage by staff, and the “customer (and staff) mutual inhibition” model, which suggests that thieves may form perceptions of being watched in non-rational ways. Retailers should research their thieves as carefully as they research their honest customers, and using similar marketing research methodologies.

  • There is a parallel with statistics on accidents health risks: the aggregate death rate from smoking is known, but it is impossible to predict an individual’s exact year of death from the fact of their smoking or not smoking.
  • One might choose to estimate statistically the impact of a policy that is already known, in order to check the level of enforcement of the policy, or (if that is known) to validate the statistical estimation process itself.
  • The company suspected that its managers were trying to conceal shrinkage by declaring it as known waste, but in a separate study not reported here, we found no evidence for this.

References

Bamfield, J. (2004a) European Retail Theft Barometer IV, (Nottingham: Retail Research Centre).

Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120.

Bamfield, J. (2004b) ‘Shrinkage, shoplifting and the cost of retail crime in Europe: a cross-sectional; analysis of major retailers in 16 countries’, International Journal of Retail & Distribution Management, 32(4/5), pp. 235-41.

Beck, A. (2004) Shrinkage in Europe: a survey of stock loss in the fast moving consumer goods sector, (Brussels: ECR Europe).

Birkin, M., Clarke, G. and Clarke M. (2002) Retail geography and intelligent network planning, (Chichester: Wiley). British Retail Consortium (2004) 11th Annual Retail Crime Survey 2003, (London: British Retail Consortium).

Brown, S. (1992) Retail location: A micro-scale approach, (Aldershot: Avebury).

Brumm, H.J. and Cloninger, D.O. (1995) ‘Violent crime and punishment: An application of the LISREL model’, Applied Economics, 27, pp. 719-25.

Carter, N. and Holmberg, B. (1992) ‘Theft reduction in a grocery store through product identification’, Journal of Organisational Behavior Management, 13(1), pp. 129-35.

Elder, A. (1989) ‘Retail theft: the hidden crime’, International Journal of Retail & Distribution Management, 17(3), pp. 39-41.

Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 31

Ernst and Young (1991) 13th Annual Survey of Loss Prevention in the Retail Industry, referenced in Shern, Verdisco and Forester (1992), ‘Loss prevention trends in retail: an executive overview’, Chain Store Executive, 68(1), pp. 6-13.

Ernst and Young (2003) Loss prevention survey, (Chicago: Ernst and Young).

Guthrie, J. (2003) New Zealand Survey of Retail Theft and Security 2003, (Dunedin: New Zealand Centre for Retail Research and Studies, University of Otago).

Hollinger, R. and Langton, L. (2005) National Retail Security Survey 2004, (Gainesville: University of Florida).

Howell, S.D. and Lehocky, M. (1997) Pattern analysis of data for control system diagnosis. European Management Journal, 15(2), pp. 167-173.

McGoldrick, P.J. (2002) Retail Marketing (2nd edition), Maidenhead:

McGraw-Hill. Persons, O.S. (1995) ‘Using financial statement data to identify factors associated with fraudulent financial reporting’, Journal of Applied Business Research, 11(3), pp. 38-46.

PlatformDetailhandel (2004) Report on Retail Crime in Europe, (Leidschendam NL: PlatformDetailhandel).

Howell SD and Proudlove NC (2007). “A statistical investigation of inventory shrinkage in a large retail chain”. The International Review of Retail, Distribution and Consumer Research 17:2, 101-120. 32 Retail Council of Canada (2004) Canadian Retail Security Report, (Toronto: Retail Council of Canada).

Ripley, BD (1994) Neural networks and related methods for classification, Journal of the Royal Statistical Society, B, 56 (3), pp. 409-456.

Sanderson, B. (2004) ‘Mirror, mirror on the wall, is it really the best security of all?’, Retail World, 57(1), p. 29.

Underhill, P. (2000), Why we buy – the science of retailing, (New York: Touchstone). University of Florida (1994) Security Research Project Report – referenced in Donnelly and Harrison: ‘Store security – the retail perspective’, Stores, 76(11), pp. 57-8.

Wonnacott T.H. and Wonnacott R.J. (1970) Econometrics, (New York: John Wiley). Wonnacott T.H. and Wonnacott R.J. (1990) Introductory statistics for business and economics, (New York: John Wiley).