10.3: Sample quality

Learning Objectives

Learners will be able to…

• Assess whether your sampling plan is likely to produce a sample that is representative of the population you want to draw conclusions about
• Identify the considerations that go into producing a representative sample and determining sample size
• Distinguish between error and bias in a sample and explain the factors that lead to each

Okay, so you’ve chosen where you’re going to get your data (setting), what characteristics you want and don’t want in your sample (inclusion/exclusion criteria), and how you will select and recruit participants (sampling approach and recruitment). That means you are done, right? (I mean, there’s an entire section here, so probably not.) Even if you make good choices and do everything the way you’re supposed to, you can still draw a poor sample. If you are investigating a research question using quantitative methods, the best choice is some kind of probability sampling, but aside from that, how do you know a good sample from a bad sample? As an example, we’ll use a bad sample I collected as part of a research project that didn’t go so well. Hopefully, your sampling will go much better than mine did, but we can always learn from what didn’t work.

 

How representative of the coastline is the image in the frame compared to the whole photograph?

 

Representativeness

A representative sample is, “a sample that looks like the population from which it was selected in all respects that are potentially relevant to the study” (Engel & Schutt, 2011).\(^4\) For my study on how much it costs to get an LCSW in each state, I did not get a sample that looked like the overall population to which I wanted to generalize. My sample had a few states with more than ten responses and most states with no responses. That does not look like the true distribution of social workers across the country. I could compare the number of social workers in each state, based on data from the National Association of Social Workers, or the number of recent clinical MSW graduates from the Council on Social Work Education. More than that, I could see whether my sample matched the overall population of clinical social workers in gender, race, age, or any other important characteristics. Sadly, it wasn’t even close. So, I wasn’t able to use the data to publish a report.

 

Exercises

Critique the representativeness of the sample you are planning to gather.

• Will the sample of people (or documents) look like the population to which you want to generalize?
• Specifically, what characteristics are important in determining whether a sample is representative of the population? How do these characteristics relate to your research question?

Consider returning to this question once you have completed the sampling process and evaluate whether the sample in your study was similar to what you designed in this section.

 

Many of my students erroneously assume that using a probability sampling technique will guarantee a representative sample. This is not true. Engel and Schutt (2011) identify that probability sampling increases the chance of representativeness; however, it does not guarantee that the sample will be representative. If a representative sample is important to your study, it would be best to use a sampling approach that allows you to control the proportion of specific characteristics in your sample. For instance, stratified random sampling allows you to control the distribution of specific variables of interest within your sample. However, that requires knowing information about your participants before you hand them surveys or expose them to an experiment.

In my study, if I wanted to make sure I had a certain number of people from each state (state being the strata), making the proportion of social workers from each state in my sample similar to the overall population, I would need to know which email addresses were from which states. That was not information I had. So, instead I conducted simple random sampling and randomly selected 5,000 of 100,000 email addresses on the NASW list. There was less of a guarantee of representativeness, but whatever variation existed between my sample and the population would be due to random chance. This would not be true for an availability or convenience sample. While these sampling approaches are common for student projects, they come with significant limitations in that variation between the sample and population is due to factors other than chance. We will discuss these non-random differences later in the chapter when we talk about bias. For now, just remember that the representativeness of a sample is helped by using random sampling, though it is not a guarantee.

 

Exercises

• Before you start sampling, do you know enough about your sampling frame to use stratified random sampling, which increases the potential of getting a representative sample?
• Do you have enough information about your sampling frame to use another probability sampling approach like simple random sampling or cluster sampling?
• If little information is available on which to select people, are you using availability sampling? Remember that availability sampling is okay if it is the only approach that is feasible for the researcher, but it comes with significant limitations when drawing conclusions about a larger population.

 

Assessing representativeness should start prior to data collection. I mentioned that I drew my sample from the NASW email list, which (like most organizations) they sell to advertisers when companies or researchers need to advertise to social workers. How representative of my population is my sampling frame? Well, the first question to ask is what proportion of my sampling frame would actually meet my exclusion and inclusion criteria. Since my study focused specifically on clinical social workers, my sampling frame likely included social workers who were not clinical social workers, like macro social workers or social work managers. However, I knew, based on the information from NASW marketers, that many people who received my recruitment email would be clinical social workers or those working towards licensure, so I was satisfied with that. Anyone who didn’t meet my inclusion criteria and opened the survey would be greeted with clear instructions that this survey did not apply to them.

At the same time, I should have assessed whether the demographics of the NASW email list and the demographics of clinical social workers more broadly were similar. Unfortunately, this was not information I could gather. I had to trust that this was likely to going to be the best sample I could draw and the most representative of all social workers.

 

Exercises

• Before you start, what do you know about your setting and potential participants?
• Are there likely to be enough people in the setting of your study who meet the inclusion criteria?

You want to avoid throwing out half of the surveys you get back because the respondents aren’t a part of your target population. This is a common error I see in student proposals.

 

Many of you will sample people from your agency, like clients or staff. Let’s say you work for a children’s mental health agency, and you wanted to study children who have experienced abuse. Walking through the steps here might proceed like this

1. Think about or ask your coworkers how many of the clients at your agency have experienced this issue. If it’s common, then clients at your agency would probably make a good sampling frame for your study. If not, then you may want to adjust your research question or consider a different agency to sample. You could also change your target population to be more representative with your sample. For example, while your agency’s clients may not be representative of all children who have survived abuse, they may be more representative of abuse survivors in your state, region, or county. In this way, you can draw conclusions about a smaller population, rather than everyone in the world who is a victim of child abuse.
2. Think about those characteristics that are important for individuals in your sample to have or not have. Obviously, the variables in your research question are important, but so are the variables related to it. Take a look at the empirical literature on your topic. Are there different demographic characteristics or covariates that are relevant to your topic?
3. All of this assumes that you can actually access information about your sampling frame prior to collecting data. This is a challenge in the real world. Even if you ask around your office about client characteristics, there is no way for you to know for sure until you complete your study whether it was the most representative sampling frame you could find. When in doubt, go with whatever is feasible and address any shortcomings in sampling within the limitations section of your research report. A good project is a done project.
4. While using a probability sampling approach helps with sample representativeness, it does not guarantee it. Due to random variation, samples may differ across important characteristics. If you can feasibly use a probability sampling approach, particularly stratified random sampling, it will help make your sample more representative of the population.
5. Even if you choose a sampling frame that is representative of your population and use a probability sampling approach, there is no guarantee that the sample you are able to collect will be representative. Sometimes, people don’t respond to your recruitment efforts. Other times, random chance will mean people differ on important characteristics from your target population. ¯\_(ツ)_/¯

In agency-based samples, the small size of the pool of potential participants makes it very likely that your sample will not be representative of a broader target population. Sometimes, researchers look for specific outcomes connected with sub-populations for that reason. Not all agency-based research is concerned with representativeness, and it is still worthwhile to pursue research that is relevant to only one location as its purpose is often to improve social work practice.

 

 

Sample size

Let’s assume you have found a representative sampling frame, and that you are using one of the probability sampling approaches we reviewed in section 10.2. That should help you recruit a representative sample, but how many people do you need to recruit into your sample? As with many questions about sample quality, students should keep feasibility in mind. The easiest answer I’ve given as a professor is, “as many as you can, without hurting yourself.” While your quantitative research question would likely benefit from hundreds or thousands of respondents, that is not likely to be feasible for a student who is working full-time, interning part-time, and in school full-time. Don’t feel like your study has to be perfect, but make sure you note any limitations in your final report.

To the extent possible, you should gather as many people as you can in your sample who meet your criteria. But why? Let’s think about an example you probably know well. Have you ever watched the TV show Family Feud? Each question the host reads off starts with, “we asked 100 people…” Believe it or not, Family Feud uses simple random sampling to conduct their surveys the American public. Part of the challenge on Family Feud is that people can usually guess the most popular answers, but those answers that only a few people chose are much harder. They seem bizarre, and are more difficult to guess. That’s because 100 people is not a lot of people to sample. Essentially, Family Feud is trying to measure what the answer is for all 327 million people in the United States by asking 100 of them. As a result, the weird and idiosyncratic responses of a few people are likely to remain on the board as answers, and contestants have to guess answers fewer and fewer people in the sample provided. In a larger sample, the oddball answers would likely fade away and only the most popular answers would be represented on the game show’s board.

In my ill-fated study of clinical social workers, I received 87 complete responses. That is far below the hundred thousand licensed or license-eligible clinical social workers. Moreover, since I wanted to conduct state-by-state estimates, there was no way I had enough people in each state to do so. For student projects, samples of 50-100 participants are more than enough to write a paper (or start a game show), but for projects in the real world with real-world consequences, it is important to recruit the appropriate number of participants. For example, if your agency conducts a community scan of people in your service area on what services they need, the results will inform the direction of your agency, which grants they apply for, who they hire, and its mission for the next several years. Being overly confident in your sample could result in wasted resources for clients.

So what is the right number? Theoretically, we could gradually increase the sample size so that the sample approaches closer and closer to the total size of the population (Bhattacherjeee, 2012).\(^5\) But as we’ve talked about, it is not feasible to sample everyone. How do we find that middle ground? To answer this, we need to understand the sampling distribution. Imagine in your agency’s survey of the community, you took three different probability samples from your community, and for each sample, you measured whether people experienced domestic violence. If each random sample was truly representative of the population, then your rate of domestic violence from the three random samples would be about the same and equal to the true value in the population.

But this is extremely unlikely, given that each random sample will likely constitute a different subset of the population, and hence, the rate of domestic violence you measure may be slightly different from sample to sample. Think about the sample you collect as existing on a distribution of infinite possible samples. Most samples you collect will be close to the population mean but many will not be. The degree to which they differ is associated with how much the subject you are sampling about varies in the population. In our example, samples will vary based on how varied the incidence of domestic violence is from person to person. The difference between the domestic violence rate we find and the rate for our overall population is called the sampling error.

An easy way to minimize sampling error is to increase the number of participants in your sample, but in actuality, minimizing sampling error relies on a number of factors outside of the scope of a basic student project. You can see this online textbook for more examples on sampling distributions or take an advanced methods course at your university, particularly if you are considering becoming a social work researcher. Increasing the number of people in your sample also increases your study’s power, or the odds you will detect a significant relationship between variables when one is truly present in your sample. If you intend on publishing the findings of your student project, it is worth using a power analysis to determine the appropriate sample size for your project. You can follow this excellent video series from the Center for Open Science on how to conduct power analyses using free statistics software. A faculty members who teaches research or statistics could check your work. You may be surprised to find out that there is a point at which you adding more people to your sample will not make your study any better.

Honestly, I did not do a power analysis for my study. Instead, I asked for 5,000 surveys with the hope that 1,000 would come back. Given that only 87 came back, a power analysis conducted after the survey was complete would likely to reveal that I did not have enough statistical power to answer my research questions. For your projects, try to get as many respondents as you feasibly can, but don’t worry too much about not reaching the optimal amount of people to maximize the power of your study unless you goal is to publish something that is generalizable to a large population.

A final consideration is which statistical test you plan to use to analyze your data. We have not covered statistics yet, though we will provide a brief introduction to basic statistics in this textbook. For now, remember that some statistical tests have a minimum number of people that must be present in the sample in order to conduct the analysis. You will complete a data analysis plan before you begin your project and start sampling, so you can always increase the number of participants you plan to recruit based on what you learn in the next few chapters.

 

Exercises

• How many people can you feasibly sample in the time you have to complete your project?

 

 

Bias

One of the interesting things about surveying professionals is that sometimes, they email you about what they perceive to be a problem with your study. I got an email from a well-meaning participant in my LCSW study saying that my results were going to be biased! She pointed out that respondents who had been in practice a long time, before clinical supervision was required, would not have paid anything for supervision. This would lead me to draw conclusions that supervision was cheap, when in fact, it was expensive. My email back to her explained that she hit on one of my hypotheses, that social workers in practice for a longer period of time faced fewer costs to becoming licensed. Her email reinforced that I needed to account for the impact of length of practice on the costs of licensure I found across the sample. She was right to be on the lookout for bias in the sample.

One of the key questions you can ask is if there is something about your process that makes it more likely you will select a certain type of person for your sample, making it less representative of the overall population. In my project, it’s worth thinking more about who is more likely to respond to an email advertisement for a research study. I know that my work email and personal email filter out advertisements, so it’s unlikely I would even see the recruitment for my own study (probably something I should have thought about before using grant funds to sample the NASW email list). Perhaps an older demographic that does not screen advertisements as closely, or those whose NASW account was linked to a personal email with fewer junk filters would be more likely to respond. To the extent I made conclusions about clinical social workers of all ages based on a sample that was biased towards older social workers, my results would be biased. This is called selection bias, or the degree to which people in my sample differ from the overall population.

Another potential source of bias here is nonresponse bias. Because people do not often respond to email advertisements (no matter how well-written they are), my sample is likely to be representative of people with characteristics that make them more likely to respond. They may have more time on their hands to take surveys and respond to their junk mail. To the extent that the sample is comprised of social workers with a lot of time on their hands (who are those people?) my sample will be biased and not representative of the overall population.

It’s important to note that both bias and error describe how samples differ from the overall population. Error describes random variations between samples, due to chance. Using a random process to recruit participants into a sample means you will have random variation between the sample and the population. Bias creates variance between the sample and population in a specific direction, such as towards those who have time to check their junk mail. Bias may be introduced by the sampling method used or due to conscious or unconscious bias introduced by the researcher (Rubin & Babbie, 2017).^[6] A researcher might select people who “look like good research participants,” in the process transferring their unconscious biases to their sample. They might exclude people from the sampling from who “would not do well with the intervention.” Careful researchers can avoid these, but unconscious and structural biases can be challenging to root out.

 

Exercises

• Identify potential sources of bias in your sample and brainstorm ways you can minimize them, if possible.

 

Critical considerations

Think back to you undergraduate degree. Did you ever participate in a research project as part of an introductory psychology or sociology course? Social science researchers on college campuses have a luxury that researchers elsewhere may not share—they have access to a whole bunch of (presumably) willing and able human guinea pigs. But that luxury comes at a cost—sample representativeness. One study of top academic journals in psychology found that over two-thirds (68%) of participants in studies published by those journals were based on samples drawn in the United States (Arnett, 2008).^[7] Further, the study found that two-thirds of the work that derived from US samples published in the Journal of Personality and Social Psychology was based on samples made up entirely of American undergraduate students taking psychology courses.

These findings certainly raise the question: What do we actually learn from social science studies and about whom do we learn it? That is exactly the concern raised by Joseph Henrich and colleagues (Henrich, Heine, & Norenzayan, 2010),\(^8\) authors of the article “The Weirdest People in the World?” In their piece, Henrich and colleagues point out that behavioral scientists very commonly make sweeping claims about human nature based on samples drawn only from WEIRD (Western, Educated, Industrialized, Rich, and Democratic) societies, and often based on even narrower samples, as is the case with many studies relying on samples drawn from college classrooms. As it turns out, robust findings about the nature of human behavior when it comes to fairness, cooperation, visual perception, trust, and other behaviors are based on studies that excluded participants from outside the United States and sometimes excluded anyone outside the college classroom (Begley, 2010).\(^9\) This certainly raises questions about what we really know about human behavior as opposed to US resident or US undergraduate behavior. Of course, not all research findings are based on samples of WEIRD folks like college students. But even then, it would behoove us to pay attention to the population on which studies are based and the claims being made about those to whom the studies apply.

Another thing to keep in mind is that just because a sample may be representative in all respects that a researcher thinks are relevant, there may be relevant aspects that didn’t occur to the researcher when she was drawing her sample. You might not think that a person’s phone would have much to do with their voting preferences, for example. But had pollsters making predictions about the results of the 2008 presidential election not been careful to include both cell phone-only and landline households in their surveys, it is possible that their predictions would have underestimated Barack Obama’s lead over John McCain because Obama was much more popular among cell phone-only users than McCain (Keeter, Dimock, & Christian, 2008).\(^{10}\) This is another example of bias.

 

 

Putting it all together

So how do we know how good our sample is or how good the samples gathered by other researchers are? While there might not be any magic or always-true rules we can apply, there are a couple of things we can keep in mind as we read the claims researchers make about their findings.

First, remember that sample quality is determined only by the sample actually obtained, not by the sampling method itself. A researcher may set out to administer a survey to a representative sample by correctly employing a random sampling approach with impeccable recruitment materials. But, if only a handful of the people sampled actually respond to the survey, the researcher should not make claims like their sample went according to plan.

Another thing to keep in mind, as demonstrated by the preceding discussion, is that researchers may be drawn to talking about implications of their findings as though they apply to some group other than the population actually sampled. Whether the sampling frame does not match the population or the sample and population differ on important criteria, the resulting sampling error can lead to bad science.

We’ve talked previously about the perils of generalizing social science findings from graduate students in the United States and other Western countries to all cultures in the world, imposing a Western view as the right and correct view of the social world. As consumers of theory and research, it is our responsibility to be attentive to this sort of (likely unintentional) bait and switch. And as researchers, it is our responsibility to make sure that we only make conclusions from samples that are representative. A larger sample size and probability sampling can improve the representativeness and generalizability of the study’s findings to larger populations, though neither are guarantees.

Finally, keep in mind that a sample allowing for comparisons of theoretically important concepts or variables is certainly better than one that does not allow for such comparisons. In a study based on a nonrepresentative sample, for example, we can learn about the strength of our social theories by comparing relevant aspects of social processes. We talked about this as theory-testing in Chapter 8.

At their core, questions about sample quality should address who has been sampled, how they were sampled, and for what purpose they were sampled. Being able to answer those questions will help you better understand, and more responsibly interpret, research results. For your study, keep the following questions in mind

• Are your sample size and your sampling approach appropriate for your research question?
• How much do you know about your sampling frame ahead of time? How will that impact the feasibility of different sampling approaches?
• What gatekeepers and stakeholders are necessary to engage in order to access your sampling frame?
• Are there any ethical issues that may make it difficult to sample those who have first-hand knowledge about your topic?
• Does your sampling frame look like your population along important characteristics? Once you get your data, ask the same question of the sample you successfully recruit.
• What about your population might make it more difficult or easier to sample?
• How many people can you feasibly sample in the time you have to complete your project?
• Are there steps in your sampling procedure that may bias your sample to render it not representative of the population?
• If you want to skip sampling altogether, are there sources of secondary data you can use? Or might you be able to answer you questions by sampling documents or media, rather than people?

 

Key Takeaways

• The sampling plan you implement should have a reasonable likelihood of producing a representative sample. Student projects are given more leeway with nonrepresentative samples, and this limitation should be discussed in the student’s research report.
• Researchers should conduct a power analysis to determine sample size, though quantitative student projects should endeavor to recruit as many participants as possible. Sample size impacts representativeness of the sample, its power, and which statistical tests can be conducted.
• The sample you collect is one of an infinite number of potential samples that could have been drawn. To the extent the data in your sample varies from the data in the entire population, it includes some error or bias. Error is the result of random variations. Bias is systematic error that pushes the data in a given direction.
• Even if you do everything right, there is no guarantee that you will draw a good sample. Flawed samples are okay to use as examples in the classroom, but the results of your research would have limited generalizability beyond your specific participants.
• Historically, samples were drawn from dominant groups and generalized to all people. This shortcoming is a limitation of some social science literature and should be considered a colonialist scientific practice.