Get More Value From Your Survey Data: 4 Techniques for More Effective Analysis

White Paper

Like many survey researchers, your day-to-day data analysis tools and techniques are likely to include cross-tabulations, bar charts, and finding mean differences between groups. However, these methods, while valuable, may be too simplistic to enable you to derive the most value from your survey data.

This white paper introduces you to four types of advanced analysis - cluster, factor, reliability and regression - that can help you gain important insights that you might miss using more basic methods. By expanding your survey analysis toolkit, you can delve deeper into your data to increase your understanding of survey responses and respondents, create better measures of important concepts and make more accurate predictions about behaviors and attitudes.

Get the download

Below is an excerpt of "Get More Value From Your Survey Data: 4 Techniques for More Effective Analysis". To get your free download, and unlimited access to the whole of bizibl.com, simply log in or join free.

download

Descriptive survey research

All survey researchers use descriptive statistical methods to summarize data and get a description of the responses to questions. These methods include frequency tables, cross-tabulations (stub and banner tables) and finding mean differences between groups or correlations between questions. For example, Figure 1 is a typical frequency table showing overall customer satisfaction for the hypothetical BSI hotel chain.

Though this type of analysis is necessary in any survey project, a series of such tables, or even cross-tabulations, only provides limited information about the attitudes and behaviors of the respondents. That is because the real world is actually multivariate, meaning that many factors play a part in an individual response. Bivariate approaches, such as analyzing data one question at a time, or using a cross-tabulation table to determine whether two questions relate, create an oversimplified view of the customer.

The table in Figure 1 doesn’t provide information about which customers give similar answers to several questions. Analyzing only one or two questions makes it impossible to see which set of questions measures similar concepts. In addition, using a table to see how responses to one question help predict responses to a second question ignores the factors that influence the second question. Thus, using a rating of overall service to predict satisfaction at the hotel chain ignores other factors, such as frequency of stay, rating of restaurants and rating of room quality.

The advanced survey methods covered in this paper enable you to analyze many survey questions simultaneously, in order to cluster respondents, group questions and make predictions with greater accuracy.

Segmenting respondents with cluster analysis

Cluster analysis enables you to group respondents with similar behaviors, preferences or characteristics into clusters, or segments. Through segmentation, you gain a greater understanding of important similarities and differences between your respondents. You can use this information to develop targeted marketing strategies, or to provide subgroups for analysis. In the case of survey data, clustering enables researchers to group respondents who provide similar responses on several questions.

Clustering, or segmentation, is a multivariate technique that analyzes responses to several questions in order to find similar respondents. Clustering is based on the concept of creating groups based on their proximity to, or distance from, each other. Respondents within a cluster, therefore, are relatively homogenous.

There are two types of cluster analysis:

  • Hierarchical: Observations are joined in a cluster and remain so throughout the clustering
  • Non-hierarchical: Cases can switch clusters as the clustering proceeds. The most common non-hierarchical method is K-means.

Cluster analysis requires you to:

  • Check the number of respondents in each cluster, as clusters of only a few respondents are not very usefu
  • Assess whether the clusters make sense, and whether their characteristics are easy to understand and describe
  • Validate the clusters by analyzing how they relate to other variables

The following cluster analysis example uses the Two-Step Cluster procedure in IBM SPSS Statistics, which incorporates statistical criteria to determine the optimal number of clusters. In this example, we apply the Two-Step Cluster method to customer survey data from the hypothetical BSI hotel chain. Though we measured the survey questions on various scales, this does not present a problem, as the Two-Step Cluster method is able to use data on nominal, ordinal or interval scales. We intend to create clusters of customers based on the following criteria:

  • Frequency of stay at BSI hotels
  • Length of customer relationship
  • Usefulness of Internet access
  • Customer company type
  • Degree of involvement in company’s decision to use BSI hotels
  • Importance of travel to job
  • The first output from the Two-Step method tells us how many clusters were found, and how many respondents are in each cluster. From the table in Figure 2, we see that the Two-Step method found three natural groups or clusters in the data, based on the responses to the six questions above.

Additional information from the Two-Step method enables us to understand what type of customer each cluster represents. In Figure 3 on the next page we see, for example, that customers in Cluster 1 travel more for work than customers in the other two clusters, and that they frequently stay at BSI hotels. Measured by length of customer relationship, however, customers in Cluster 1 are not the most loyal.

Once you are satisfied that you understand the clusters and their characteristics, you can create descriptive labels. In this case, we label Cluster 1 customers “Frequent BSI Road Warriors.” Cluster 2 customers are “Less Frequent Travelers,” and Cluster 3 customers are “Loyal BSI Road Warriors.”

Using the clusters in analysis

Now that we have identified the clusters of respondents, we can use them in analyses and reports. For example, to see how overall customer satisfaction relates to cluster membership, we use a clustered bar chart (Figure 4). The chart shows that customers in Cluster 3 are more satisfied with BSI hotels than are customers in Cluster 1. Since Cluster 1 customers stay at BSI hotels more often than customers in the other clusters, these results may be cause for action with that segment. Knowing the characteristics of the most and least satisfied customers can help you make important business decisions.

Presenting your results

The Custom Tables procedure in IBM SPSS Statistics enables you to create attractive and complex stub- and banner-type tables. The table in Figure 5 shows the mean of customer satisfaction by cluster membership. For each cluster, we nest the customers’ willingness to stay again at a BSI hotel. Each row includes questions about use of hotel spas and billing problem resolution. The Custom Tables procedure enables you to condense many types of information into an attractive and compact table.

Grouping questions with factor analysis

Most questionnaires include several questions about each key topic. A questionnaire used to measure patient satisfaction with an HMO, for example, might include several questions about physician care, as well as several questions about the performance of support staff, such as x-ray technicians. While it is helpful to look at each question individually, it is often possible to create more reliable and valid measures by using the responses to several questions simultaneously. Compound measures of critical concepts can make your analyses more powerful.

Factor analysis enables you to discover clusters or groups of questions about similar concepts, based on correlations or covariances between questions. You can use the factors to:

  • Create scales or compound measures composed of several questions
  • Reduce the number of questions on a questionnaire by, for example, identifying questions that measure the same concept
  • Understand the relationships between several questions simultaneously

Factor analysis is a general linear model (GLM) technique, which means that it assumes data are measured on an interval scale. As is common in survey research, we can easily use variables measured on five-, six- or seven-point scales.

Want more like this?

Want more like this?

Insight delivered to your inbox

Keep up to date with our free email. Hand picked whitepapers and posts from our blog, as well as exclusive videos and webinar invitations keep our Users one step ahead.

By clicking 'SIGN UP', you agree to our Terms of Use and Privacy Policy

side image splash

By clicking 'SIGN UP', you agree to our Terms of Use and Privacy Policy

How IBM SPSS Statistics helps you find the appropriate analysis technique

The Statistics Coach makes it easy to select the most appropriate analysis method for your data. In this case, we want to group questions, but we’re not sure which method to use. From the “Help” menu, we select “Statistics Coach” to display the first screen shown in Figure 6. Then, we choose the “Identify groups of similar variables” option, click “Next,” and choose the data type. The program immediately opens the Factor Analysis dialog box, shown in the second screen.

Factor analysis includes two distinct steps. The first step involves extracting a small number of factors from the data. Think of the factors as underlying attitudes reflected in answers to specific questions. There are several extraction methods; principal components extraction and principal axis factoring are used most frequently. In the second step, the factors are rotated to ease interpretation. Varimax is the rotation method used most frequently with survey data.

In this example, we apply the principal components extraction and varimax rotation methods to the following statements taken from the survey of BSI hotel customers. These statements had a five-point response scale, ranging from “Strongly Disagree” to “Strongly Agree:”

  • BSI services are a good value
  • BSI offers high-quality services
  • BSI makes it easy to make reservations
  • BSI makes my job easier
  • I am satisfied with BSI dining facilities
  • BSI facilities are up to date
  • BSI rooms are appropriately priced

The resulting table (Figure 7) contains the two factors extracted from the survey statements. The procedure automatically identifies factors that explain more variance than individual statements. The two factors in the table account for approximately 59 percent of the total variance among the statements, which is quite satisfactory

The most critical output of our factor analysis, however, is the rotated components table (Figure 8), which shows the loading, or correlation, between each question and the two extracted factors. Questions with high loadings on one factor and low loadings on other factors are associated with the high-loading factor. For example, the statement “BSI makes it easy to make reservations” is associated with Factor 1 because it has a high correlation (.716) with that factor and a low correlation (.260) with Factor 2.

The table in Figure 8 shows that Factor 1 is associated with the first five statements. Factor 2 is clearly associated with the two statements about pricing and value.

At this point in your analysis, you can create descriptive factor labels to use in reports and subsequent analyses. In this case, we label Factor 1 “Quality,” since it measures various aspects of the quality of hotel services and facilities. Factor 2 is labeled “Value,” since it measures satisfaction with room price and the perceived value of services.

Though the individual responses to each statement may be useful, we now have two factors that are more valid measures of the quality and value of the hotel chain than any single statement. That is the essential benefit that factor analysis provides.

To use factors in future analyses, create a combined measure of the questions or statements associated with each factor. There are two ways to accomplish this:

  • Automatically create factor scores with the factor procedure using standardized (z) scores
  • Compute a new variable by adding the raw responses for the statements associated with each factor, and dividing by the number of questions, to create a mean score

Determining the reliability of factors

It’s important to verify that survey responses are valid and reliable. The same is true for the factors that you discover. Fortunately, factors are by definition valid, because all of the associated questions or statements measure aspects of the same concept. It is still important, however, to establish the reliability of the factor.

The reliability analysis procedure in Statistics enables you to determine whether a set of survey questions, items, or statements forms a reliable scale. This means that the items measure a single concept with reasonably high intercorrelations. To perform reliability analysis, you don’t need more assumptions about the data than you do for factor analysis. As an added benefit, the output is usually easy to interpret.

To demonstrate the reliability analysis capabilities of Statistics, we apply the technique to the “Quality” factor that we discovered using factor analysis. The screenshot in Figure 9 (on the next page) shows the main output. The key value in the output is Cronbach’s alpha, which in this case is .6946. This statistic varies from zero to one, and though alpha has several interpretations, the cutoff value is more useful in determining whether a scale is reliable. The standard rule of thumb is that alpha must be greater than approximately .70 to conclude that the scale is reliable.

Since the output shows that alpha for the “Quality” factor is just below .70, we could decide that it’s close enough for use in subsequent analyses. The output, however, also suggests which items can be removed from the reliability scale to increase alpha. For example, if we omit the “Update” item (which represents the statement, “BSI facilities are kept up to date”), alpha increases to more than .70.

Using a factor in analysis

Once you determine that a factor is valid and reliable, what can you do with it? The answer is simple. Treat the factor as you would treat a new variable, and use it as you would use any question in which you have confidence. Relationships are typically more clear and distinct with factors than with individual questions or statements.

To illustrate, we examine the relationship between perceived quality and frequency of stay at BSI hotels. First, we use the single statement about quality (Figure 10). We then perform the same analysis using the “Quality” factor (Figure 11).

Both charts show a positive relationship between perceived quality and frequency of stay. Notice, however, that the relationship is much stronger for the “Quality” factor than for the single statement about quality. Hotel quality is comprised of several aspects of service, and single questions by definition don’t do well at capturing all aspects of a general concept. This example is a direct illustration of the benefits of using factor and reliability analysis together.

Making predictions with regression

Multiple regression, a general linear model technique, is the most popular method for studying the relationship between an outcome variable and several predictor, or independent, variables. It is often used with survey data, because it enables you to combine many variables into one predictive equation. In addition, multiple regression helps to determine the unique role of each variable in predicting the outcome, provides a measure of the total explanatory power of the model (R2), and provides an estimate of whether a variable is a statistically significant predictor or not.

Multiple regression is often called linear regression, because a linear, or straight-line, relationship between predictors and outcome is assumed. Relationships between variables may not always be linear, but it is best to assume that they are in order to create a useful model. As with factor analysis, multiple regression works best with variables measured on an interval scale, but you can also use typical survey response scales.

In general, regression analysis should include only variables that may be good predictors, or variables that you want to include for reasons that are practical (the customer type variable is important for your business, for example) or theoretical (previous work suggests that customer gender is a key predictor). Though it’s possible to include dozens of predictor variables in a regression equation, it’s best to be more selective.

The first steps in regression analysis are to identify the predictors and the variable or question you want to predict. In this example, we apply regression techniques to the hypothetical BSI data to predict overall customer satisfaction. We use nine predictors, including the “Quality” factor identified earlier, to demonstrate how to link regression and factor analysis.

The first results produced by Statistics are in the two tables shown in Figure 12. The R2 value is .583, so more than half of the variation in customer satisfaction responses is explained by these nine predictors. The second table shows that the set of predictors is statistically significant at predicting customer satisfaction at the .01 level of significance (because the “Sig.” value is below this number). This means that there is a less than one in one hundred chance that our regression results occurred by chance.

The next step is to determine which of the questions are significant predictors in the multivariate model (Figure 13). For this, we check the “Sig.” column for values below approximately .05. Seven of the nine questions meet this standard (one just barely). This is useful information, since it tells us which variables are significant, and which are not. We see that satisfaction with room price and facilities are not important when predicting customer satisfaction.

We then use the nonstandardized coefficients in the B column to assess the effect of each predictor. For example, in the B column for “BSI services are a good value,” the value .468 means that for every one unit increase in the response to this question, satisfaction increases .468 units. To put it another way, if we compare customers who agree (score=4) that BSI services are a good value to customers who strongly agree (score=5), we then predict that overall satisfaction will be about .47 units higher for the latter group. This relationship controls for the other variables in the model. In other words, we can make this absolute statement because multiple regression automatically controls for all variables in the model. By using the B coefficients for all significant predictors, we can create a prediction equation to use for overall customer satisfaction.

Note that the B coefficient for frequency of stay is negative. This means that customers who stay more often at BSI hotels are less satisfied than infrequent customers, after we take other factors into account. This could be a key finding for BSI analysts and management to explore.

The next step is to determine the relative importance of the variables using the Beta coefficient from the table in Figure 14. Beta values, which vary from -1 to 1, are important to use when questions have different response scales. The higher the absolute value of Beta, the more important the variable is in predicting customer satisfaction. In this case, we see that the most important predictors are the perceived value of services and the “Quality” factor from our factor analysis example.

These results could enable the BSI hotel chain to develop an efficient and effective strategy for improving customer satisfaction, with a focus on only the most relevant features. The chain could also use this information to make predictions about dissatisfied customers, and target future marketing efforts appropriately.

In the BSI example, regression enables us to understand that customer satisfaction is related more to the perceived value of services than to the price of rooms. Although “BSI services are a good value” is the best predictor, “BSI rooms are priced right” is not even a significant predictor. This is surprising, since we might assume that price has an impact on satisfaction. The discovery suggests that price as compared to perceived value is more important to customers than price alone. Using regression analysis, we gain critical insight into the relationship between these two factors and the real-world attitudes of BSI customers.

Summary

As demonstrated in the examples given in this paper, advanced survey analysis methods enable you to extract detailed, valuable information about the attitudes and behavior of survey respondents.

IBM SPSS Statistics offers powerful techniques – such as the cluster, factor, reliability, and regression methods discussed here – that enable you to discover relationships you might miss using more basic software and methods. When it’s time to present your results, Statistics provides the professional graphics and reporting capabilities you need to effectively communicate complex ideas.

Any survey researcher can use the advanced methods covered in this paper, because Statistics makes analysis intuitive and straightforward. The more you use the methods presented here, the more comfortable you will become, and the more insight you will gain from your survey data. Further, these advanced methods can give your organization a competitive advantage, and enable you to make more informed decisions in a timely and effective manner.

Want more like this?

Want more like this?

Insight delivered to your inbox

Keep up to date with our free email. Hand picked whitepapers and posts from our blog, as well as exclusive videos and webinar invitations keep our Users one step ahead.

By clicking 'SIGN UP', you agree to our Terms of Use and Privacy Policy

side image splash

By clicking 'SIGN UP', you agree to our Terms of Use and Privacy Policy