Statistical Analysis Sample: Multiple Regression: Predictors of Employee Job Satisfaction

Model Answer

Introduction and Research Question

Employee job satisfaction is widely recognised as a predictor of retention, productivity and organisational commitment, which has made it a central outcome variable in human resource management (HRM) research (Locke, 1976). Despite decades of study, organisations continue to struggle to identify which levers most reliably improve how satisfied employees feel in their roles. This report presents a multiple linear regression analysis undertaken to establish which of four commonly cited workplace factors — perceived organisational support, work-life balance, pay satisfaction and job autonomy — best predict overall job satisfaction among employees of a UK-based logistics and distribution company.

The practical stakes of this question are significant. The Chartered Institute of Personnel and Development has repeatedly highlighted the direct and indirect costs of voluntary employee turnover, including recruitment, induction and lost-productivity costs, reinforcing the value to employers of identifying concrete, actionable predictors of satisfaction rather than treating it as an abstract attitudinal construct of interest only to researchers (CIPD, 2023). A statistically grounded account of which levers matter most therefore has direct implications for where a resource-constrained HR function should focus its improvement efforts.

The analysis was commissioned following an internal engagement survey that returned a below-benchmark satisfaction score, prompting the HR department to request a statistically grounded account of which factors were driving the result, rather than relying on anecdotal exit-interview evidence. Multiple regression was selected because the outcome variable, job satisfaction, is continuous and was expected to be jointly influenced by several correlated predictors; a series of separate bivariate tests would not have allowed the unique contribution of each predictor to be isolated while holding the others constant (Field, 2018).

The research question guiding the analysis was: to what extent do perceived organisational support, work-life balance, pay satisfaction and job autonomy predict employee job satisfaction, and which of these factors makes the strongest unique contribution once the others are controlled for? A standard (simultaneous) multiple regression was chosen over a hierarchical or stepwise approach because there was no strong a priori theoretical basis for entering the predictors in a particular order, and all four variables were considered of broadly comparable conceptual importance within the job characteristics and social exchange literatures (Eisenberger et al., 1986).

Two hypotheses were tested. The null hypothesis (H0) stated that the four predictors, taken together, do not significantly explain variance in job satisfaction. The alternative hypothesis (H1) stated that at least one predictor makes a statistically significant unique contribution to job satisfaction once the model is fitted. The sections that follow describe the sample and variables, the assumption checks undertaken prior to interpretation, the regression output itself, and a plain-English interpretation of what the findings mean for the client organisation.

Data and Variables

Data were collected via an anonymous online survey distributed to all non-managerial staff across three UK distribution depots, yielding 220 usable responses after 14 incomplete or straight-lined returns were removed (N = 220, response rate 61%). All five variables were measured on 5-point Likert-type scales (1 = strongly disagree, 5 = strongly agree) and were treated as continuous for the purposes of the analysis, which is consistent with common practice in organisational psychology when composite scale scores are used (Howitt and Cramer, 2020).

Job satisfaction, the dependent variable, was measured using a six-item short-form adapted from established job satisfaction instruments, with items averaged to produce a single composite score for each respondent. Perceived organisational support (POS) was measured with a five-item adaptation of the Eisenberger et al. (1986) scale. Work-life balance was captured with a four-item scale addressing perceived control over working hours and the ability to meet non-work commitments. Pay satisfaction used a three-item scale addressing satisfaction with base pay, benefits and pay relative to comparable roles in the sector. Job autonomy used a four-item scale addressing perceived discretion over how, when and in what order tasks are completed during a shift.

Table 1 summarises the descriptive statistics for all five variables prior to the regression being run.

All variables showed means comfortably within the scale range and standard deviations indicating reasonable spread rather than floor or ceiling effects. Job autonomy returned the highest mean (M = 3.68, SD = 0.66), suggesting staff generally felt they had reasonable discretion in their roles, while pay satisfaction returned the lowest mean and the largest spread (M = 3.05, SD = 0.85), indicating more varied views on reward across the sample.

The internal consistency reliability of each composite scale was assessed using Cronbach’s alpha prior to computing the composite scores used in the regression, since a scale with poor reliability would undermine confidence in any resulting coefficients. Table 2 reports the reliability coefficients obtained.

All five scales returned alpha coefficients above the conventionally accepted minimum of .70 (Field, 2018), with job autonomy returning the lowest value (α = .76) and perceived organisational support the highest (α = .88). No items were removed from any scale, as removing items would not have meaningfully improved reliability for any of the five measures.

Preliminary Checks and Assumptions

Before interpreting the regression coefficients, six standard assumptions for multiple linear regression were checked, following the sequence recommended by Field (2018), Pallant (2020) and Tabachnick and Fidell (2019). Checking assumptions before interpreting coefficients matters because a violated assumption, such as severe multicollinearity or non-normal residuals, can produce misleading standard errors and significance tests even when the underlying data are otherwise sound; skipping this stage is one of the most common errors seen in applied regression work of this kind.

First, the sample size was checked against Green’s (1991) rule of thumb (N ≥ 50 + 8k for testing the overall model, where k is the number of predictors), which for four predictors requires a minimum of 82 cases; the achieved sample of 220 comfortably exceeded this threshold. Second, each predictor was checked for a linear relationship with the outcome using scatterplots, and all four showed an approximately linear pattern with no evidence of pronounced curvature.

Third, multicollinearity was assessed using Variance Inflation Factor (VIF) and tolerance statistics, since two or more highly correlated predictors can distort coefficient estimates even when the overall model fits well (Hair et al., 2019). Table 3 reports the collinearity diagnostics for the four predictors.

All VIF values were well below the conservative threshold of 5 (and comfortably below the more cautious threshold of 2.5 recommended for social science data by Hair et al., 2019), and all tolerance values exceeded .70, indicating multicollinearity was not a material concern for this model.

Fourth, independence of residuals was checked using the Durbin-Watson statistic, which returned a value of 1.98. As this falls close to the ideal value of 2 and comfortably within the conventionally acceptable range of 1.5 to 2.5, the assumption of independent errors was considered satisfied (Field, 2018).

Fifth, homoscedasticity was assessed by plotting standardised residuals against standardised predicted values. The resulting scatterplot showed a roughly even, random spread of points around zero with no funnel or curved pattern, supporting the assumption of constant error variance across the range of predicted scores.

Sixth, the normality of the residuals was checked using a histogram with a superimposed normal curve and a normal P-P plot, supplemented by the Shapiro-Wilk test applied to the standardised residuals, W = 0.987, df = 220, p = .192. As this result was non-significant, the null hypothesis of normally distributed residuals could not be rejected, indicating the normality assumption was reasonably well met.

Finally, case-wise diagnostics were inspected for influential outliers using standardised residuals (none exceeded ±3.29) and Cook’s Distance (maximum value = .041, well below the commonly used cut-off of 1.0). On this basis, no cases were removed and all 220 responses were retained for the final analysis. With the assumptions satisfied, the standard multiple regression could proceed to interpretation with confidence.

Test Results

A standard multiple linear regression was conducted in SPSS (version 28) with job satisfaction as the dependent variable and perceived organisational support, work-life balance, pay satisfaction and job autonomy entered simultaneously as predictors.

The overall model was statistically significant and explained a substantial proportion of variance in job satisfaction. Table 4 presents the model summary.

Table 5 presents the accompanying ANOVA table testing the significance of the overall model.

Table 6 presents the standardised and unstandardised coefficients, standard errors, t-values and significance levels for each predictor.

As shown in Table 6, three of the four predictors made a statistically significant unique contribution to the model at the p < .05 level: perceived organisational support (β = .34, p < .001), work-life balance (β = .21, p = .002) and pay satisfaction (β = .18, p = .003). Job autonomy did not reach statistical significance once the other three predictors were controlled for (β = .10, p = .135).

Cohen’s f² was calculated to gauge the practical size of the overall model: f² = R² / (1 − R²) = .375 / .625 = 0.60, which exceeds Cohen’s (1988) threshold of 0.35 for a large effect. This indicates that, collectively, the four predictors accounted for a substantively as well as statistically significant share of the variation in employee job satisfaction across the sample.

As a robustness check on the non-significant autonomy predictor, a reduced hierarchical model (Model 2) was re-run with job autonomy removed, retaining only perceived organisational support, work-life balance and pay satisfaction. Table 7 compares the two models.

Removing job autonomy reduced R² by only .007, a change that was itself non-significant, F change(1, 216) = 2.25, p = .135, and which is numerically identical to squaring the autonomy predictor’s t-statistic from Model 1 (1.50² = 2.25), as would be expected for the removal of a single predictor. This confirms that autonomy was contributing negligible unique explanatory power beyond the other three predictors, and that the three-predictor model is a defensible, more parsimonious alternative for reporting purposes.

Figure 1 displays the standardised beta coefficients for the four predictors in the full model, illustrating their relative strength once the shared variance between predictors has been accounted for.

Interpretation

This section translates the statistical output above into plain-English conclusions relevant to the client organisation, reporting each result in APA style before explaining its practical meaning.

The overall regression model was statistically significant, F(4, 215) = 32.25, p < .001, R² = .375, adjusted R² = .363, indicating that the four predictors together explained approximately 37.5% of the variance in employee job satisfaction, with the adjusted figure suggesting minimal shrinkage were the model applied to a new sample. In applied HR terms, more than a third of the differences in how satisfied employees reported feeling in their roles can be statistically accounted for by how supported they feel by the organisation, how well they can balance work against other commitments, how satisfied they are with their pay, and — to a lesser, non-significant extent — how much autonomy they have.

Perceived organisational support emerged as the strongest unique predictor, b = 0.28, SE = 0.05, β = .34, t(215) = 5.60, p < .001. Holding the other three predictors constant, each one-point increase on the five-point POS scale was associated with a 0.28-point increase in job satisfaction. This is consistent with social exchange theory, which holds that employees who perceive their organisation as genuinely valuing their contribution and caring about their wellbeing tend to reciprocate with more positive work attitudes (Eisenberger et al., 1986). For the client organisation, this suggests that visible, day-to-day signals of support — fair treatment by line managers, recognition of effort, and responsiveness to employee concerns — may do more to move overall satisfaction than any single policy change.

Work-life balance was the second strongest significant predictor, b = 0.19, SE = 0.06, β = .21, t(215) = 3.17, p = .002. Each one-point increase in perceived work-life balance was associated with a 0.19-point increase in job satisfaction, controlling for the other predictors. Given that the sample was drawn from a logistics operation where shift patterns and unpredictable hours are common sources of strain, this finding is unsurprising and points to scheduling practice as a second lever worth examining in detail.

Pay satisfaction was also a significant, if slightly smaller, unique predictor, b = 0.15, SE = 0.05, β = .18, t(215) = 3.00, p = .003. This finding is worth flagging carefully to the client: while pay satisfaction does matter, its standardised contribution (β = .18) was smaller than that of organisational support (β = .34), suggesting that a pay review alone would be unlikely to close the satisfaction gap without also addressing perceived support and work-life balance.

Job autonomy did not make a statistically significant unique contribution once the other three predictors were controlled for, b = 0.09, SE = 0.06, β = .10, t(215) = 1.50, p = .135. This does not necessarily mean autonomy is unimportant to employees in general; rather, it indicates that within this sample, and once shared variance with the other three predictors was accounted for, autonomy did not add explanatory power beyond what support, balance and pay already captured. This pattern is consistent with prior arguments that autonomy effects can be partially absorbed by measures of organisational support when both are entered into the same model (Locke, 1976), since supportive managers often grant more discretion as a matter of course.

Taken together, the pattern of results suggests a clear priority ordering for HR intervention: initiatives that strengthen employees’ sense of being supported by the organisation are likely to yield the largest gains in job satisfaction, followed by improvements to work-life balance and, to a smaller degree, pay. The non-significant autonomy finding suggests that granting additional task discretion in isolation, without accompanying improvements in support or balance, may not by itself shift overall satisfaction scores within this particular workforce.

From a cost-benefit perspective, this ordering is also practically useful. Pay increases are typically the most expensive and least reversible lever available to an HR department, whereas manager-training programmes aimed at improving perceived support, and moderate scheduling reforms aimed at improving work-life balance, are comparatively low-cost and can often be piloted at a single depot before wider roll-out. The regression results therefore support a staged intervention: first targeting line-manager behaviour and communication (addressing perceived support), second reviewing shift-scheduling flexibility (addressing work-life balance), and only then, if satisfaction gains remain insufficient, considering a more costly pay review. This sequencing allows the organisation to test lower-cost interventions first and to re-survey staff before committing to the more expensive option.

Conclusion and Limitations

This analysis set out to establish which of four commonly cited workplace factors most strongly predicted employee job satisfaction in a UK logistics workforce, using a standard multiple regression. The model was statistically significant and explained a substantively large share of variance in job satisfaction (R² = .375), with perceived organisational support, work-life balance and pay satisfaction each making a significant unique contribution, while job autonomy did not.

For the commissioning HR department, the practical implication is that efforts to improve overall satisfaction scores are likely to be most effective if prioritised in the order suggested by the standardised coefficients: strengthening perceived organisational support first, followed by improvements to work-life balance, with pay-focused interventions treated as a secondary rather than primary lever. Given that job autonomy did not emerge as a significant unique predictor once the other variables were controlled for, resources aimed solely at increasing task discretion may be better redirected towards support- and balance-focused initiatives, at least in the short term. A practical first step recommended to the client is a short pilot programme at a single depot, combining manager-support training with a modest scheduling review, with satisfaction re-measured after three to six months to establish whether the predicted gains materialise before any wider, costlier roll-out is authorised.

Several limitations should be considered when interpreting these findings. First, the design was cross-sectional, meaning causal claims cannot be made with certainty; it remains possible, for example, that more satisfied employees perceive their organisation as more supportive, rather than the reverse. A longitudinal or experimental design would be required to establish direction of causality with greater confidence. Second, all variables were measured via self-report on the same survey instrument at a single time point, raising the possibility of common method variance inflating the observed relationships (Tabachnick and Fidell, 2019). Third, the sample was drawn from a single organisation in one sector, which limits the generalisability of the specific coefficient estimates to other industries or organisational contexts, even though the general pattern of support and balance outweighing pay is broadly consistent with the wider literature. Fourth, only four predictors were included; unmeasured variables such as line manager quality, perceived job security and organisational justice may also explain unique variance in job satisfaction and could usefully be incorporated in future work. Fifth, while the reliability of each composite scale was acceptable (all Cronbach’s alpha values ≥ .76), reliability does not guarantee validity, and it remains possible that the adapted short-form scales used here do not capture every facet of the constructs they are intended to measure as comprehensively as the longer, validated instruments from which they were derived.

Future research could usefully extend this analysis using a longitudinal design to strengthen causal inference, incorporate additional predictors identified in the wider HRM literature, and test whether the relative importance of support, balance and pay differs across departments, seniority levels or contract types within the organisation. Notwithstanding these limitations, the present analysis provides the client organisation with a statistically grounded, actionable starting point for prioritising its employee experience initiatives, rather than relying on anecdotal evidence alone.

References

CIPD (2023) Employee Turnover and Retention. London: Chartered Institute of Personnel and Development.

Cohen, J. (1988) Statistical Power Analysis for the Behavioral Sciences. 2nd edn. Hillsdale, NJ: Lawrence Erlbaum.

Eisenberger, R., Huntington, R., Hutchison, S. and Sowa, D. (1986) ‘Perceived organizational support’, Journal of Applied Psychology, 71(3), pp. 500–507.

Field, A. (2018) Discovering Statistics Using IBM SPSS Statistics. 5th edn. London: Sage.

Hair, J.F., Black, W.C., Babin, B.J. and Anderson, R.E. (2019) Multivariate Data Analysis. 8th edn. Andover: Cengage.

Howitt, D. and Cramer, D. (2020) Introduction to Statistics in Psychology. 7th edn. Harlow: Pearson.

Locke, E.A. (1976) ‘The nature and causes of job satisfaction’, in Dunnette, M.D. (ed.) Handbook of Industrial and Organizational Psychology. Chicago: Rand McNally, pp. 1297–1349.

Pallant, J. (2020) SPSS Survival Manual. 7th edn. Maidenhead: Open University Press.

Tabachnick, B.G. and Fidell, L.S. (2019) Using Multivariate Statistics. 7th edn. Boston: Pearson.