This is a Medium Rare template. Explore our complete range of designs here.
top
The most common approach to making sense of the huge amount of data generated by employee surveys is to identify the lowest scores and build action plans to address them. It is an intuitive approach and a perfectly valid one. Key driver analysis offers a different way of looking at the data.
Rather than focusing on the lowest scores, it establishes which factors are most strongly associated with key organisational outcomes — and where the evidence suggests effort and investment will make the most difference.
It seems obvious that improving a low score should be a priority — but only if there is a meaningful relationship between that score and the outcomes that matter. Key driver analysis tells you whether that relationship exists.
Throughout this article employee engagement is used as the primary organisational outcome — referring to the attitudinal outcomes that employee surveys most commonly measure: how positively employees feel about their work and their organisation, their intention to stay, their sense of pride, and their willingness to recommend the organisation as a place to work.
What follows sets out how to run a key driver analysis from preparing the raw data to understanding the outputs and the decisions that need to be made to ensure the results stand up to scrutiny. Done well, the process produces findings that can be trusted. Done carelessly, it produces unreliable findings.
The analysis involves four steps. The first three — data preparation, selecting items for analysis, and item screening — lay the groundwork. The fourth is the key driver analysis itself.
Survey platforms usually export data as a spreadsheet — one row per respondent, one column per survey question. Responses are typically stored as text labels exactly as they appeared to the respondent: "Strongly agree", "Disagree", and so on. Before any analysis can begin, these labels must be converted to numbers.
Strongly Agree, Agree, Disagree, and Strongly Disagree would be assigned values 4, 3, 2, and 1 respectively — higher numbers representing more positive responses — an important convention to establish before analysis begins.
"No Opinion" and "Not Applicable to Me" responses are treated as missing data rather than assigned a numeric value. "No Opinion" represents an absence of a view and "Not Applicable to Me" signals that the item falls outside a respondent's experience entirely. Both are analytically uninformative as scored responses, but tracking them separately matters. A high "No Opinion" rate on a particular item may indicate the question is poorly worded or that employees feel unable to assess it. A high "Not Applicable to Me" rate suggests the item may be too role-specific to function as a driver for the workforce as a whole.
Survey platform exports occasionally contain formatting issues that need resolving before analysis. Column headers may contain special characters and formatting inconsistencies — curly apostrophes rather than straight ones, for instance — that cause mismatches when the data is processed. A good practice is to produce a clean version of the raw file with these issues resolved at the outset, leaving the original untouched as a permanent record.
Not all survey questions belong in the key driver analysis. Before any statistics are run, each item must be classified — and those that fall outside the scope of the analysis excluded.
Driver items are the attitudinal statements that form the input to the key driver analysis — capturing how employees experience their work across themes such as management effectiveness, leadership, culture, development, and reward. These are the items that will be examined for their association with engagement outcomes.
Outcome items are the measures of employee engagement that the analysis is trying to explain. Rather than being used as inputs to the driver model, they are kept separate and used as the dependent variable — the thing the driver items are correlated against. Typical outcome items cover overall feeling about working for the organisation, intention to stay, pride, and willingness to recommend.
Demographic items capture respondent characteristics — department, tenure, grade. These are retained for subgroup analysis but excluded from the key driver analysis.
Other items may be excluded on statistical or content grounds. These are covered in Step 3.
In this step the driver items are screened and those that do not meet the criteria for inclusion removed. There are three potential reasons for removal — statistical, content-based, or redundancy.
Each driver item is examined against a set of descriptive statistics. An item where almost everyone responds the same way carries no analytical signal — it cannot correlate meaningfully with other items or with the engagement outcomes, and correlation is the foundation of everything that follows.
Standard deviation is the primary indicator. On a four-point scale, a standard deviation below 0.6 is a flag for review — the decision to remove or retain requires judgement.
Mean and skewness together describe the shape of the response distribution. A high mean combined with a low standard deviation suggests responses are bunching toward the positive end of the scale — moderate negative skewness of this kind is expected in engagement data and not a concern in itself. A skewness statistic beyond ±2 is a flag for review, and where 80% or more of respondents choose the highest response option the item is no longer distinguishing usefully between respondents and becomes a stronger candidate for removal.
No Opinion and Not Applicable rates are reviewed for each item. Where more than 20% of respondents selected either option the item is flagged. A high No Opinion rate may indicate the question is poorly worded or that employees feel unable to assess it. A high Not Applicable rate suggests the item may be too role-specific to function as a driver for the workforce as a whole.
Two examples illustrate how different statistical issues can lead to exclusion. An item such as "I have the skills I need to do my job effectively" may show a mean of 3.47 on a four-point scale with a standard deviation of 0.55 and nearly half of respondents at the ceiling — almost everyone agrees, leaving too little variance to be analytically useful.
Conversely, an item about the organisation's handling of a sensitive workplace issue might show reasonable variance but a No Opinion rate above 20%, indicating that a significant proportion of employees cannot reliably assess it — perhaps because they have no direct experience of the situation the item describes.
Both items are excluded, but for different reasons. The first fails on variance; the second fails on relevance.
Some items pass the statistical checks but do not belong in the key driver analysis because of what they are measuring rather than how they are distributed. Three situations arise in practice.
First, awareness items — questions that measure what employees know rather than how they feel. Questions such as "I understand how annual pay increases are determined", "I understand the organisation's objectives and strategy", or "I am aware of the whistleblowing process" are knowledge checks, not attitudinal statements. They may correlate with outcomes — informed employees tend to be more engaged — but this does not make them drivers of engagement in a meaningful sense. Whether an employee understands the pay review process tells you nothing about how they feel about working for the organisation. Awareness items are better reported as standalone metrics alongside the main findings.
Second, items that are attitudinal in form but too distal from the day-to-day employee experience to function as meaningful engagement drivers. An item about the organisation's commitment to sustainability, for instance, may produce consistently positive scores but is unlikely to discriminate between engaged and disengaged employees in a way that leads to actionable interventions. These items are removed from the driver model and reported as standalone metrics.
Third, items that are time-bounded or context-specific rather than measuring a stable aspect of the employee experience. A question about sentiment toward organisational changes over the past year, or excitement about a particular strategic initiative, is capturing a moment rather than an enduring driver. If circumstances change — the initiative succeeds or fails, the change programme ends — the item loses its meaning. Items that measure anticipated organisational behaviour rather than current experience fall into the same category — "I believe the organisation will take action on the results of this survey" depends on what happens after the survey rather than on the quality of the employee experience at the time of measurement. These items are better reported as contextual indicators alongside the main findings.
In all cases the excluded items are not discarded — they are reported separately as contextual or thematic findings alongside the main analysis.
Where two items are so similar in wording that they are effectively measuring the same thing, one is removed. Including near-duplicate items distorts the analysis — the construct they both measure appears artificially more important than it actually is. The item retained is generally the one with slightly better statistical properties or the one whose wording is more precisely attitudinal.
The result is a final retained item list — a set of driver items that are statistically viable, conceptually appropriate, and non-redundant. Every exclusion is documented with a reason. That documentation matters: it makes the analytical choices transparent and the analysis reproducible if the survey is repeated in future years. The key driver analysis is built on a sounder foundation as a result.
With a clean set of driver items established, the analysis can now turn to the central question: which aspects of the employee experience matter most for engagement and related outcomes? Key driver analysis addresses this through a sequence of complementary approaches — beginning with a straightforward examination of how individual items correlate with the outcomes, grouping those items into reliable constructs through factor analysis, and then estimating the unique contribution of each construct. Each approach has a distinct contribution: item-level correlations provide granularity and transparency, factor analysis produces reliable and actionable constructs, and Relative Weight Analysis accounts for the relationships between those constructs when estimating their importance. In practice the three approaches tend to tell a consistent story — which is itself reassuring — with RWA adding statistical rigour and a principled basis for tiering the results rather than overturning the conclusions that simpler methods produce.
The starting point is to examine how each driver item correlates individually with the engagement outcome. This produces a ranked list of every survey question by the strength of its association with engagement — a granular and transparent picture that is straightforward to interpret and easy to explain to a non-specialist audience. A low-scoring item that correlates strongly with engagement is a different kind of finding to a low-scoring item that does not — and it is this distinction that item-level correlations begin to make visible.
The correlation statistic recommended here is Kendall's tau. Whilst Pearson correlation is more widely used in practice, tau is more appropriate for ordinal survey data — particularly where, as is common in engagement surveys, responses tend to cluster toward the positive end of the scale. Tau values range from -1 to +1. The higher the absolute value, the stronger the association between a driver item and the engagement outcome — positive values indicate that higher scores on the driver item are associated with higher engagement, negative values the reverse.
Rather than correlating each driver item against a single outcome question, the analysis uses a composite outcome score derived from all the outcome items. This composite produces a single score for each respondent that reflects their overall level of engagement, with items that are stronger indicators of the underlying construct contributing more to the score. Correlating driver items against this composite rather than a single outcome question produces a more reliable ranking — less susceptible to the particular wording of any one item.
Confidence intervals should be computed for each correlation using a bootstrap approach — resampling the data many times and examining the distribution of results to produce upper and lower bounds around each estimate. These intervals quantify the uncertainty in the ranking. A tau of 0.45 and a tau of 0.42 may look different as point estimates but if their confidence intervals overlap substantially the data does not support treating them as meaningfully different. This becomes important when presenting findings — items whose intervals overlap should be grouped into tiers rather than ranked precisely against each other.
The item-level ranking is a useful starting point and in some respects the most directly actionable output of the analysis — individual questions tell you specifically what to address rather than pointing to a broader theme. Its limitation is practical: a survey with fifty driver items produces fifty correlations, which is difficult to synthesise and communicate. Factor analysis, which follows, organises related items into a smaller number of reliable constructs — producing a smaller number of meaningful themes that are easier to interpret and more directly actionable.
Factor analysis identifies groups of items that tend to be answered in a similar way — items that correlate strongly with each other are likely measuring the same underlying aspect of the employee experience. These groups are called factors. A factor made up of items about manager feedback, manager support, and manager recognition is measuring something coherent — what might be labelled manager effectiveness. Factor analysis formalises that intuition across all driver items simultaneously, producing a smaller set of constructs that capture the main themes in the data.
Each item is described by a factor loading — a number that indicates how strongly it is associated with each factor. Items with high loadings on a factor are strong indicators of that underlying construct. Items that do not load clearly onto any factor are excluded from the final model.
The number of factors is not fixed in advance. Statistical criteria are used to guide the decision, but the final solution also needs to make conceptual sense — the factors should be interpretable and tell a coherent story about the employee experience. In a typical engagement survey the analysis produces between six and eight factors covering themes such as manager effectiveness, leadership and communication, culture and inclusion, development and career progression, reward, and team working.
With the driver items organised into factors, the same correlation approach used at item level is repeated — but now working with factor scores rather than individual items. Each respondent has a score on each factor, and these scores are correlated with the composite engagement outcome using Kendall's tau. The result is a ranked list of factors by their association with engagement.
The factor-level ranking tends to tell a broadly similar story to the item-level ranking — the themes that emerged as important at item level will generally remain important at factor level. Where the two rankings differ, the factor-level results are the more reliable guide. A factor score based on multiple items is a more precise measure of the underlying construct than any single item, and the ranking is less susceptible to the particular wording of individual questions.
Confidence intervals are computed for each factor correlation in the same way as at item level — providing a basis for grouping factors into tiers rather than implying a precision in the ranking that the data does not support.
The final analytical step is Relative Weight Analysis (RWA). The factor-level correlations produced in the previous step measure the association between each factor and engagement independently — without accounting for the fact that the factors themselves are related to each other. In most engagement surveys, factors do correlate with each other — organisations where leadership is strong tend also to have stronger cultures, better manager effectiveness, and clearer communication. When factors are correlated in this way, their individual correlations with engagement can be difficult to interpret — it is hard to tell how much of a factor's apparent importance reflects its genuine relationship with engagement and how much reflects its association with other important factors.
RWA addresses this by estimating the unique contribution of each factor to engagement after accounting for the relationships between factors. The result is expressed as a percentage of the total variance in engagement explained by the model — if the model explains 57% of the variance in engagement, and organisational leadership accounts for 34% of that, its relative weight is 34%. The weights across all factors sum to 100% of the explained variance.
In practice RWA tends to confirm the ranking produced by the simpler factor-level correlations rather than overturn it. This is reassuring — it suggests the driver hierarchy is robust and not an artefact of the correlations between factors. Where the two rankings do diverge, the RWA ranking is the more defensible basis for prioritisation.
Confidence intervals are computed for each relative weight using the same bootstrap approach as for the item and factor level correlations. These intervals are used to group factors into tiers — factors whose confidence intervals overlap cannot be ranked precisely against each other and should be presented as a group rather than in a strict order.
The output of the analysis is a ranked list of factors by their relative importance for engagement, with confidence intervals that indicate the precision of each estimate. The first step in reporting is to translate this ranked list into tiers — grouping factors whose confidence intervals overlap rather than presenting a precise ranking that the data does not support.
Working down the ranked list, factors whose confidence intervals do not overlap with those immediately below them form a distinct tier. Where intervals overlap, the factors are grouped together. In practice this typically produces three or four tiers. The top tier represents the clearest priorities — the factors where the evidence for their importance is strongest and most distinct. The bottom tier typically contains factors that are either performing well already or less central to the engagement experience of this particular workforce.
The tiering principle matters because point estimates alone can be misleading. A model might show organisational leadership at 34% and culture and inclusion at 29% — a difference that looks meaningful but whose confidence intervals overlap substantially. Presenting these as jointly the top priority, with a combined weight of 63%, is more honest and more useful than implying a precision the data cannot support.
Key driver analysis provides the quantitative foundation for understanding what drives engagement. It identifies the factors most strongly associated with the outcomes that matter and provides a statistically defensible basis for prioritisation. But the numbers alone do not tell the whole story. Free-text comments illuminate the reasons behind the scores. Follow-up conversations with employees and managers add context that no survey can fully capture. And the judgement of people who know the organisation — its history, its culture, its current pressures — is irreplaceable in turning analytical findings into effective action.
The value of key driver analysis is not that it produces a definitive answer. It is that it provides a rigorous, transparent, and reproducible starting point — one that focuses the conversation on the right questions and ensures that decisions about where to invest effort are grounded in evidence rather than instinct alone.
Let’s start a conversation about how employee surveys can help you develop a workplace where people and performance grow together.
