Human resource management in US subsidiaries in Europe and Australia: centralisation or autonomy?

Abstract

We explore determinants of subsidiary autonomy in setting human resource management (HRM) practices within US-parented multinational enterprises (MNEs), in Europe and Australia. We examine both the effect of strategic context and the effect of the institutional location of the subsidiary. We find that US MNEs show greater centralisation of control over HRM where the subsidiary faces global markets, in coordinated market economies vs liberal market economies, and where union density is low.

Appendix: Treatment of missing values

This appendix describes the procedure we adopted to account for the significant proportion of missing values in the data set. The CRANET survey has a substantial proportion of missing values (34.7% listwise across our analysed variables and sample). Fifteen per cent are missing on the union density variable, a key variable for our analysis. There are also significant numbers of missing values for two control variables: organisation size (17.7%) and organisation age (5.0%).

Appropriate strategies for handling missing values in data depend on the nature of the “missingness”. If missing values are randomly distributed, with no relationship to the value of the missing variable or to other variables in the data set (known as missing completely at random or MCAR), then both listwise and pairwise deletion of missing values will yield unbiased estimates when analysing the data, although, in the case of pairwise deletion, standard errors will be biased towards zero and listwise deletion will be inefficient, because it discards observations. Multiple imputation likewise yields unbiased estimates of parameters, but is more efficient than listwise deletion, and gives unbiased estimates of standard errors.

A less stringent assumption concerning missing data is that they are missing at random (MAR). That is, if x is a variable with missing (i.e., unobserved) values and x_mis is a variable coded 1 when x is missing and 0 otherwise, then, after controlling for other observed variables in the analysis, x_mis is independent of the value of x.

Where data are MAR but not MCAR then both pairwise and listwise deletion will yield biased estimates for parameters. In contrast, multiple imputation yields unbiased, fully efficient estimates for parameters and unbiased estimates of standard errors.

Where x_mis is related to the underlying value of x, after controlling for observed values, this is known as non-ignorable (NI) missing data. Where data are NI missing, pairwise deletion, listwise deletion and multiple imputation all yield biased estimates of parameters. Further, although it is possible to distinguish empirically between MCAR and MAR using observed data, it is not possible to determine from the observed data whether data are NI missing. However, in many cases it is possible to convert NI missing data to MAR data by controlling for further variables. There may also be a priori reasons to believe data to be MAR as opposed to NI (Allison, 2001; King, Honaker, Joseph, & Scheve, 2001).

In the case of our missing union membership data, there is good a priori reason to believe missingness will be related significantly to country, since in some European countries legislation constrains management from collecting data on union membership.

Analysis of our data set shows that country does indeed predict missingness accounting for 27% of the variance. Thus our data are certainly not MCAR. We cannot rule out that the data are NI. However, there is no a priori reason to believe respondents will be more or less likely to respond to questions about union membership as a consequence of high or low levels of union membership. Further, we are in a position to draw on a wide range of relevant variables in the Euronet-Cranfield data set in order to impute missing values for union membership.

In addition to the variables in our reported analyses, we included the following variables in the imputation procedure for missing values: Home country, sector, whether national level pay bargaining, presence of works council, percentage of workforce under 25, percentage of workforce over 45, percentage of manual workers in workforce, and whether a union is used as a channel of communication with the workforce. Using all variables to predict missingness on the union membership variable with the general linear model accounts for 50% of variance (Adj. R²=0.50). The largest proportion of variance is accounted for by country (partial η²=0.27). Thus, having controlled for a significant proportion of variance in missingness, and there being no a priori reason to suspect missingness to be related to level of union membership, we may have some confidence that the data are MAR, not NI missing.

Thus the most appropriate approach to dealing with the missing data is to use multiple imputation. We used the approach described by King et al. (2001). The missing values are imputed multiple times (usually three to five is sufficient) to generate multiple data sets. Analysis is then carried out on each imputed data set separately. The missing value imputation was carried out using the King et al. (2001) software Amelia, with five imputed data sets. Model parameters are calculated as the average of the parameters from each separate analysis. This process gives unbiased estimators for model parameters, and allows an adjustment to standard errors to allow for the uncertainty of imputed values. Standard errors are calculated by

where m is the number of imputed data sets, SE(q _j ) is the standard error of the jth estimate of parameter q, and