How do you choose which constructs to include in a data analysis? This is an important question, as it guides the kind of data you need. The most desirable situation is to know already what you are interested in studying, and why it is important. It is true that, sometimes, you have data but no preconception of the real use or even meaning of the data. This sometimes happens in exercises such as data mining, which we discuss briefly later.

However, if you take a more systematic view, there are various options for the constructs and associated data that you choose.

You will always have one or more constructs on which you wish to focus, such as revenue (proxied by Sales data in the Accu-Phi case). Focal constructs are the key concepts of interest in your study, the things you care about and that stimulated you to start a research project in the first place. There are several considerations here.

Predictors are those constructs that you believe might explain or predict your focal constructs. You do not always need predictors. However, you do need them when you seek to explain or predict one or more focal constructs using their relationship to other variables.

For example, in the Chapter 1 Accu-Phi case, constructs such as trust, satisfaction, and enquiries might be thought of as predictors.

Often you also seek to include constructs that in themselves are not of primary interest to you, or even that you believe explain or predict the focal constructs, but that do express the relevant environment of your focal constructs.

Demographic descriptors and differentiators of your observations (such as individuals or firms) are often a case in point. Say you want to explain or predict a certain behavioral concept such as employee turnover. While you might have good reason to position some demographic constructs as direct predictors of turnover (e.g. older and longer-tenure staff are less likely to resign), you may not have strong reason to believe that others (perhaps those such as gender and race?) would actually cause turnover. Instead, these other constructs may be included as environmental factors in a way that is not quite cause-and-effect. Such control variables create a proper data context for the focal variables.

Understand that if you do not include measures of crucial control constructs, you are analyzing your core focal constructs out of their proper context and therefore you will not quite get the right statistical results in many types of analyses. When you do analyses that involve balancing multiple focus and predictor variables against each other, if you do not also balance these in the context of their environment (as represented by control constructs), then you may get the wrong picture.

As an example of the usefulness of controls, I was involved in helping a group of scientists study Fetal Alcohol Syndrome (damage to babies in the womb due to mothers drinking). Fetal Alcohol Syndrome damages many things, including growth (in this case, the focus variable). In this particular study they injected all the pregnant mice with alcohol and some of the mice with methanol. The idea was to test whether the growth in the babies, once born, was better among mice injected with or without methanol. The basic data showed no difference, however the data did not include the genders of the babies! Obviously male mice grow bigger and faster than female mice. So, when I included gender as a control variable, the data then showed an effect for the methanol. Gender was an important control variable in the context of babies’ growth; without it, the patterns in the data were obscured.

The basic lesson of the very first step is that you will always have constructs on which you want to focus (focal constructs), sometimes you wish to explain or predict these using other predictor constructs (not always), and sometimes you wish to analyze your focal constructs in their relevant context as represented by control constructs (not always).

Each time you gather a piece of data to capture into a database, in some way you are implicitly asking a question, even if it is not expressed verbally. There are three major considerations when deciding how to gather data:

With regard to question formats and who we are asking, there is a lot to think about.

First, there are various ways to gather data, that is, various ways to ask the questions. Some of these options include:

The main point to consider here is which data gathering method best suits your needs. GPS data on your logistics fleet may exist, but it may not be the best data for answering questions about why drivers stop.

Second, each question can have more or less accuracy and specificity about exactly what it is you are asking for. A specific question or data query can be expected to give more precise answers, and therefore better data. But beware: you need to be very sure of whether you are being as specific as you think you are. For example, how do you ascertain a customer’s age?

As stated previously, always try to think of a way to get continuous (ratio or interval) data, so long as this is feasible. Sometimes this is not possible: gender, for instance, will always be categorical. But at other times, asking things in ordinal categories, such as age bands, provides inferior data to asking for more finely tuned data (as in the example of age question formats above). In this regard, there is a trade-off sometimes in terms of quality of data vs. complexity, for which see the answer format material in the next section.

Third, some question formats involve broad, open-ended questions, such as ”where do you see yourself in five years’ time?”. Such query formats can obviously be expected to render very diverse answers dealing with lots of issues. You may need to do what is called content analysis on such data, where you break it down and do intermediate analysis on the content of the answers before it can be put into an analytics process. This relies on the analyst to have certain qualitative research skills.

In a similar vein, more subjective questions, such as people’s mental states or conditions (e.g. personality assessments), may need careful psychometric development before the questions are usable. In such cases using previously-developed question formats or sets is a good idea. Also, when assessing more subjective or hard-to-measure variables, it is often a good idea to ask more than one question about the variable. This would require a multi-item scale; see the next section for more on this.

Even with the same question, you can invite people to answer in different ways. For example, think of the question “Do you trust Accu-Phi’s cloud storage to keep your accounting data safe?” You could ask this in an informal interview or a focus group, and invite open-ended answers. Alternately, you could re-cast it as an invitation in a customer survey for people to respond to the statement “I trust Accu-Phi’s cloud storage to keep our accounting data safe,” and invite answers on a very specific answer format, which could then take various forms. The way you invite people to answer questions is important.

Several issues arise when thinking about answer formats. Again, specificity can be important. For example, you might ask in an employee form, “How long did you work in your last job ______________”. When thinking about this example, consider the following points:

When providing people with a specific answer scale you need to decide how to structure it. For example you could ask people to respond to the statement “I am satisfied with my career development options here.” You could offer them an answer scale as seen in Figure 4.1 Example of a question with a Likert scaled answer format (a Likert-type scale):

Or, you could use something like the format seen in Figure 4.2 A semantic differential answer format with an online slider scale (a semantic differential scale that is completed online using a slider that the respondent can slide anywhere between the two extremes):

Once again, see methodology books for various other response formats. There are many such formats. One mistake beginners make in research using surveys or similar instruments is to ask too many questions using the same response format, for example, fifty questions on five issues all of which are assigned the classic five-point Likert response format from Strongly Disagree to Strongly Agree. This is monotonous and often leads to poor responder engagement with the survey, whereas mixing up the answer format can help.

With regard to such formally-structured answer scales, several considerations arise. One is how to capture the data. You would often capture the answer as a number in the datasheet (e.g. if a respondent ticks the third option in the alternate semantic differential scale below, perhaps capture this as a ”3”, as in Figure 4.3 Correspondence between survey answer data entry below).

How many response options should you give? In the Likert example above there were five options; in the semantic differential scale there might be a wide range, depending on how the sliding scale works. This question is a hotly debated one in the world of research methodology:

As stated above, in some cases we need to ask more than one question about a variable to start measuring it accurately. This is dealt with next.

In data analysis, we often have only one measurement of a variable. In employee records, for example, we may only have one data entry for employee age, start date, etc. This is fine when the data is a measurement of something very objective and when there is little scope for measurement error.

Sometimes however we are measuring something that is very subjective, or might have high scope for measurement error. Psychometric measures, for example, when we are surveying people about subjective mental constructs such as satisfaction, are far trickier. In such cases, there is a greater risk of measurement error, since people understand survey questions differently. For instance, asking employees in a survey, “Are you satisfied with your job?” might be construed as:

These differences in interpretation or understanding make the answers less reliable. In addition, a given single-item measure might be biased by other things, such as an employee rushing to fill in the survey and not really thinking hard about a question. Even more objective assessments such as performance appraisals by supervisors are open to measurement error of various kinds.

In cases where the interpretation is in the control of the person organizing the measurements, it is generally better to gather data on a construct using more than one measurement. This is referred to as “multi-item assessment.” In our Chapter 1 Accu-Phi example, two of the constructs measured (trust and satisfaction) are measured with multiple questions all asking different aspects of the construct. For example, we might have constructed a survey using the questions seen in Figure 4.4 Possible multi-item measures of trust and satisfaction for each variable.

You will generally be able to find multi-item scales on most business, psychological, sociological and other constructs in academic journal articles, which you can often fairly safety adopt and adapt to your uses. These will have been previously validated through careful scientific methods. If you wish to use a previously designed scale – and you probably should, if possible, since these have been tested for validity – try to find scales tested and adapted for your local context (geography, industry, and the like). You often will.

There are two issues you may wish to consider when designing or choosing multi-item scales:

Finally, having designed your question and answer protocols for data collection, and having captured the data, you now have to deal with various issues in the raw data. I discuss these issues next.