Experiment and Non-experiment

Chong-ho Yu, Ph.Ds.

Experimental research and non-experimental research

"Experiment" is a widely misused term. When some people talk about their "experiment," indeed their study is non-experimental in nature. The following are the characteristics of experimental and non-experimental research designs.


  • Random sampling: a sampling method in which each member of a set has independent chances to be selected (the notion of "equal chances" is a theoretical ideal mentioned by many textbooks, but there are always some hidden bias or disposition in the real world).
  • Randomization: randomly assign subjects into the control group and the treatment group.
  • Experimenter manipulation: directly manipulate variables to test cause-and-effect relationships e.g. alter the amount of drug given to the patients. The researcher manuipulates the factor that she cares about.
  • Experimenter control: involves control of all other extraneous variables or conditions that might have an impact on the dependent variables. The researcher removes the effect that she doesn't care.

It is very common for even experienced researchers to be confused by random sampling and randomization. For example, Morse (2007) wrote,
What is wrong with randomization? Processes of saturation are essential in qualitative inquiry: saturation ensures replication and validation of data; and it ensures that our data are valid and reliable. If we select a sample randomly, the factors that we are interested in for our study would be normally distributed in our data, and be represented by some sort of a curve, normal or skewed. Regardless of the type of curve, we would have lots of data about common events, and inadequate data about less common events. Given that a qualitative data set requires a more rectangular distribution to achieve saturation, with randomization we would have too much data around the mean (and be swamped with the excess), and not enough data to saturate on categories in the tails of the distribution (p.234) (Emphasis added by the author).
Random sampling and randomization

As of August 7, 2017, the website of the Department of Statistics explained the role of sampling in statistical inference as follows:
The use of randomization in sampling allows for the analysis of results using the methods of statistical inference. Statistical inference is based on the laws of probability, and allows analysts to infer conclusions about a given population based on results observed through random sampling (para. 1) (Emphasis added by the author).
Again, randomization is concerned with assignment of group membership after the sample is drawn, whereas random sampling is a subject selection process.

Control and manipulation are very crucial to experimentation. Without them, the conclusion drawn from an observed phenomenon could be completely wrong even if it makes sense. Let's look at an everyday example: One of my friends has two TV sets. One of them is Japanese-made while the other is European-made. She insisted that the Japanese TV has a better quality than the European one because the former presents a sharper picture. Being skeptical to her claim, I conducted a small experiment: I simply swapped the locations of the two TV sets. As a result, the European TV set showed a clearer picture than the Japanese one. As you see, the factor here is the signal rather than the electronics. In an experiment, if I put all TVs under study in the same location, then location as a source of "noise" is under my control. If I alternating the location for each TV, then location becomes a variable under my manipulation.

Let's use herbs as another example: A Chinese friend maintained that some Chinese herbs could heal certain diseases. She even conducted an experiment to prove it. When her husband suffered a long-term illness, he took Chinese herbs for one week and his health condition improved substantively. The next week he stopped taking Chinese herbs and the condition reversed. I asked her how many types of Chinese herbs her husband took, she answered, "Ten." If I feed a patient with 10 vitamins, I am sure he will get better, too! Because of the lack of manipulation/partition of the chemical components of the herb, this "experiment" did not tell us which Chinese herb is helpful to which body function.

However, it is important to note that "control" is not the core essence of experimentation. The difference between controlled experiment and randomized experiment will be discussed in a later section.


A quasi-experiment is a research design that does not meet all the requirements necessary for controlling the influence of extraneous variables. Usually what is missing is random assignment.

For example, when a researcher studies gender difference in computer use, obviously he cannot randomly assign gender (I am happy as a man. I don't want to be re-assigned).

It is generally agreed that the primary demarcation criterion of experiments and quasi-experiments is random assignment of group membership. Nonetheless, some authors consider random selection as a criterion, too. For example, according to Plichta and Garzon (2009), "quasi-experimental designs may lack random selection, random assignments, or both" (p.13). In a similar vein, Moule and Hek (2012) suggested that convenience sampling is "a part of survey or quasi-experiment designs" (p.95).

Survey research

This type of research is very common in political sciences and communications, in which many variables are not controllable. For example, if you intend to study how wars affect people's perception to the quality of policy making, you cannot create a war or manipulate other world affairs, unless you are the villain in the movie "Tomorrow never dies." Because of this limitation, researchers send surveys to participants who are exposed to the real conditions.

Secondary analysis: Archival research

Archival research is a subset of secondary data analysis, but the two terms are not synonymous. Meta-analysis, in which results of prior research are synthesized, is also a form of secondary analysis. As the name implies, archival research utilizes existing raw data archived in databases, but meta-analysis extracts statistical results from previous studies. If you don't like the tedious IRB process, go for secondary data analysis. 

Archival research is popular in economics and educational research, especially when the research project involves trends or longitudinal data. For example, if the researcher wants to find out the correlation between productivity and school performance, he can contact the General Accounting Office and the Department of Education for obtaining the related data in the last twenty years. The following are some examples of archival data that are openly accessible:
Obviously, there are advantages of archival data analysis:
  • It saves time, efforts, and money, because the data are online available (Most online databases are free, but CCMH requires a full data access fee).
  • It provides a basis for comparing the results of secondary data analysis and your primary data analysis (e.g. national sample vs. local sample).
  • The sample size is much bigger than what you can collect by yourself. A small-sample study lacks statistical power and the result might not be stable across different settings. On the contrary, big data can reveal stable patterns.
  • Many social science studies are conducted with samples that are disproportionally drawn from Western, educated, industrialized, rich, and democratic populations (WEIRD; Henrich, Heine, & Norenzayan, 2010). Nationwide and international data sets alleviate the problem of WEIRD.
On the the hand, there are shortcomings and limitations. For example, you might be interested in analyzing disposable income, but the variable is gross income. In other words, your research question is confined by what you have at hand (Management Study Guide, 2016).

Additionally, it is important to point out that very often there are discrepancies between different sources of archival data, and thus researcher should exercise caution in drawing firm conclusions derived from a single data source. For example, GDP per capita is commonly used in many archival research studies. Nonetheless, there exist vast differences between the two different sources indicating GDP per capita of each country, namely, World Development Indicators (WDI) and Penn World Table 7.1 (PWT) (Ram & Ural, 2014). In addition, based on the 2005 UN Human Development statistics, Harris (n.d.) pointed out that the most atheistic societies, including many secular European nations, are the healthiest. However, in Happy Planet Index none of those secular European countries is ranked among the top 20. The table below shows the recent figures of UNHD and HPI side by side.


UN Happiness
(Human Development)

Happy Planet Index









Costa Rica




















El Salvador

































Costa Rica





New Zealand





United Arab Emirates















United States




















Both natural settings and laboratory-controlled experiments have pros and cons. On some occasions, things happen in the real life challenge artificial experiments. For example, in some lab-controlled benchmark tests, Windows outperforms Mac OS, Linux, and even UNIX! But computer users tell different stories in real settings.

It is common that experimentation is equated with scientific methodology, and thus is highly regarded. Actually, certain science subjects do not heavily reply on experimentation, such as Astronomy (Big bang, Quantum tunneling) and physics (e.g. M-theory). In classical astronomy the major source of knowledge is from observation rather than experimentation (Deese, 1972). For example, you cannot blow up Mars and see how the absence of Mars affects the gravitational force of the Solar system (With modern rocket and nuclear technologies, humans may be able to do so, but we shouldn't)! And the study of the origin of the universe could not count on even observation. Mathematics is another example. Although today with the aid of high-power computer, several mathematicians are able to conduct "mathematical experiments" by simulation (Chaitin, 1998), basically the origin of mathematical theorems are from logical deduction. Lack of experimentation can also be found in certain areas of biology such as evolution. Barkow (1989) pointed out that an evolutionary scenario is speculative in which the usual requirements for empirical verifiability are relaxed in favor of an emphasis on logic and plausibility.

Randomization and Simpson's Paradox

Randomization is the major difference between experiment and quasi-experiment. It is important to point out some common misconceptions regarding randomization.

Random sampling and randomization

As mentioned before, many people confuse random sampling and randomization. The former is a sampling process while the latter is concerned with assignment of group membership. Further, The purpose of random sampling is to enhance the generalizability of the results while the purpose of randomization is to establish the cause-effect interpretations of the results. In other words, random sampling counteracts the threat to external validity whereas randomization addresses the threat of internal validity. However, the above concepts are easily confused (May & Hunter, 1988). The topic of internal validity and external validity will be discussed in another write-up entitled Threats to validity of Research Design.

In practice, randomization plays a more important role than random sampling in research. Let's face it. How often can a researcher draw a random sample? If the target population consists of all university students, are you able to draw samples from campuses in states other than your own? As a matter of fact, most research studies recruit convenience subjects that are instantly available (Frick, 1998). If the requirement of random sampling is strictly followed, experiments are hardly implemented. In fact, Reichardt and Gollob (1999) found that in a randomized experiment, the use of a t test with a convenience sample can be justified without reference to a hypothetical infinite population, in which random samples are drawn.

To rectify the situation of non-random sampling, randomization is used to spread errors randomly among treatment groups (Fisher, 1971). Pitman (1937a, 1937b, 1938) went so far as to assert that random sampling is unnecessary for a valid test of the difference between treatments in a randomized experiment. Using an example of 40 convenience subjects, Babbie (1992) conceptualized randomization as treating convenience samples as probability samples: "It is as though 40 subjects in this instance are a population from which we select two probability samples-each consisting the characteristics of the total population, so the two samples will mirror each other." (p.243)

However, like random sampling, randomization also encounters difficulties in implementation. Berk (2005) used the following example to illustrate one of the problems: Even if the experimenter randomly assigns prisoners into different treatment programs, the inmate may fail to show up. This can turn the randomized experiment into an "intent-to-treat" experiment.

Simpson's Paradox

It is important to repeatedly emphasize that Randomization is not the silver bullet. In addition to the attrition issue mentioned above, randomization is subject to the threat of Simpson's Paradox, which was discovered by Dr. E. H. Simpson (1951), not O. J. Simpson or Bart Simpson. Simpson's Paradox is a phenomenon that the conclusion drawn from the aggregate data is opposite to the conclusion drawn from the contingency table based upon the same data.

If it is too abstract to you, let's look at an example: In England once a 20-year follow-up study was conducted to examine the survival rate and death rate of smokers and non-smokers. The result implied a significant positive effect of smoking because only 24% of smokers died compared to 31% of non-smokers. Phillip and Morris should celebrate, right? Not yet. When the data were broken down by age group in a contingency table, it was found that there were more older people in the non-smoker group (Appleton & French, 1996).

Another example of Simpson's Paradox can be found in a study regarding student retention conducted at Arizona State University. Although the initial analysis based on all data (Yu, DiGangi, Jannasch-Pennell, & Kaprolet, 2010) shows that among the students who stay at the university, the probability of being a resident (p=.67) is higher than that of non-residents (p=.33), a seemingly opposite conclusion emerges when observations are grouped by state in a GIS analysis, as shown in the Figure 1:

Figure 1. Retention rate mapped to student home states

How is Simpson's Paradox related to randomization? Obviously, the above study used non-experimental data. You cannot ask people to become smokers or non-smokers. Neither can age be assigned (I wish it can be. If so, I will request to be assigned to the young age group). As a result, two groups which were non-equivalent in age led to Simpson's Paradox. Although randomization is said to prevent this from happening, randomization is not 100% fool-proof. By simulation, Hsu (1989) found that when the sample size is small, randomization tends to make groups become non-equivalent and increase the possibility of Simpson's Paradox. Thus, after randomization with a small sample size, researchers should check the group characteristics on different dimensions (e.g. race, sex, age, academic year, ...etc.) rather than blindly trusting randomization.

Randomized and controlled experiments

Another area of confusion can be commonly found in the difference between randomized and controlled experiments. Today "randomized experiment" and "controlled experiment" are often used synonymously. One of the reasons is that usually an experiment consist of a controlled group and treatment group, and  group membership is randomly assigned into one of the groups. Since "control" and "randomization" are both perceived as characteristics of an experiment, it is not surprising that in many texts randomized experiment and controlled experiment are either used in an interchangeable fashion or the two terms are combined as one term such as  "randomized controlled experiment." The latter usage is legitimate as long as both control and randomization are implemented in the experiment. However, treating a randomized experiment as "a controlled experiment" and vice versa is misleading (e.g. "In controlled experiments, this is accomplished in part through the random assignment of participants to treatment and control groups" (Schneider et al., 2008)). Indeed, there is a subtle difference between the two.

Randomized experiment

R. A. Fisher is the pioneer of randomized experiment. In Fisher's view, even if there is a significant difference between the control and the treatment group, we may not be able to attribute the difference to the treatment when there exists many uncontrollable variables and sampling fluctuations. The objective of randomization is to differentiate between associations due to causal effects of the treatment and associations due to some variable that is a common cause to both the treatment and response variables. If there are influences resulted from uncontrolled variables, by randomization the influences would be randomly distributed across the control and treatment groups even though no control of those variables are made.

Controlled experiment

On the other hand, the logic of experimentation up to Fisher's time was that of controlled experiment. In a control experiment, many variables are experimentally fixed to a constant value. However, Fisher explicitly stated that it is an inferior method, because it is impossible to know what variables should be taken into account. For example, a careful researcher may assign equal numbers of males and females into each group, but she/he may omit the age and educational level of the subjects. In Fisher's view, instead of attempting to put everything under control, the researcher should let randomization take care of the uncontrollable factors. It is not to suggest that Fisher did not advocate controlling for other causes in addition to randomization. Rather he explicitly recommended that the researcher should do as much as control as he can, but he advised that randomization must be employed as "the second line of defense" (Shipley, 2000).


Following the same line of reasoning, the Canadian Task Force for Preventive Health Care (2003) prefers randomized experiments to controlled trials without randomization as clinical evidence, as shown in the following table.

Rating Research design


Evidence from randomized controlled trial(s)


Evidence from controlled trial(s) without randomization


Evidence from cohort or case-control analytic studies, preferably from more than one centre or research group


Evidence from comparisons between times or places with or without the intervention; dramatic results in uncontrolled experiments could be included here


Opinions of respected authorities, based on clinical experience; descriptive studies or reports of expert committees

Nonetheless, a randomized experiment is not necessarily superior to a controlled experiment. As mentioned before, when the sample size is small, randomization tends to make groups become non-isomorphic and thus may lead to a Simpson's Paradox (Hsu, 1989). Not surprisingly, when the sample size is small, a controlled experiment is more advisable.

Smoking does not cause lung cancer, really?

Fisher smokingIt is important to point out that any dogmatic thinking is counter-productive to science, which is supposed to be an open system. R, A, Fisher, the inventor of randomized experiment, was dead wrong about the relationship between smoking and lung cancer. Between 1922 and 1947 the prevalent rate of deaths attributed to lung cancer surged 15 times across England and Wales. In 1947 Austin Bradford and Richard Doll were hired by the British Medical Research Council to investigate the possible cause of this pandemic. Obviously, it is unethical to conduct a randomized experiment, such as randomly assigning 3,000 healthy people to the smoking group and 3,000 to the control group. Alternatively, Hill and Doll conducted surveys in the hospitals of London. Doll was stunned by the fact that people who smoked tended to die of lung cancer, and in response he gave up smoking two-thirds of the way through the study. In 1950 Hill and Doll published their report in the British Medical Journal, suggesting that there was a causal link between smoking and lung cancer. In 1957 Fisher, who was a smoker, sent a letter to the journal to repudiate their conclusion. His reasoning is simple: without running a randomized experiment we cannot assert a cause and effect relationship between tobacco and lung cancer. Fisher insisted upon his position and kept counter-arguing his opponents until he died in 1962 (Christopher, 2016). Nonethless, at least Fisher was consistent by doing what he said. He kept smoking until his death!

Australian approach cannot work in America, really?

The previous example shows that the dogmas of randomized experimentation could hinder researchers from drawing a sound causal conclusion and delaying countermeasures against threats (e.g. the environmental hazard of DDT pointed out by Silent Spring and climate change suggested by IPCC). In addition, Berwick (2008) challenged the view that randomized experiments can be applied to all situations. Many years ago Rapid Response Team (RRT), an innovative preventative health care approach introduced by Australian doctors, in which a team of physicians and nurses monitor vital signals of patients and take proactive actions, was implemented in the United States. But, randomized experiments conducted by American researchers showed that there were no significant differences between RTT and non-RTT approaches in terms of reducing the number of unexpected deaths. Berwick questioned the validity of the conclusion, for it ignored the cultural context and the specific delivery mechanisms.

Similarly, Rawlins disputed the experimental "gold standard" in medical research by listing the limitations of randomized and controlled experiments. First, like social scientists, sometime medical researchers face a "mission impossible" scenario when the disease under investigation is extremely rare and thus the number of patients is very small. Second, on some occasions experimentation is unnecessary, especially when a treatment produces a "dramatic" benefit, such as Imatinib (Glivec) for chronic myeloid leukemia. In health science research there is a stopping rule. When the treatment shows healing effects, the trial should be stopped early so that the control group can switch to the more effective treatment. There is no consensus among statisticians as to how best to handle this situation, but treating this type of incomplete experiment as invalid would throw out valuable information (cited in Medical News Today, 2008).

Essock et al. (2003) also observed the discrepancy between the "real world" and the lab settings. Many drug treatment studies last about four to eight weeks only. Short-term drug tests may cost less to implement, but usually these studies do not yield the statistical significance that is found in long-term experiments. On the other hand, long-term drug trials have problems in retaining participants long enough to yield unbiased outcomes. In other words, the so-called causal conclusions produced in experiments may not reflect what would happen in the real world.

The dictator game in the real world

The dictator game, which is used very often for studying morality and cooperative behaviors, is another good example. In a typical experiment utilizing the dictator game, the participant is told to decide how much of a $10 pie he would like to give to an anonymous person who also signs up for the same experimental session. The game is so named because the decision made by the giver is final. Most experimental results are encouraging: Many participants were willing to share the wealth. However, the result is completely different when the dictator game is conducted in a naturalistic setting. In a study carried out by Winking and Nizer (2013) at a bus stop in Las Vegas, the researcher told some strangers that he was in a hurry to the airport and therefore he wanted to give away his $20 in casino chips. The researcher explicitly suggested to the receivers to share a portion of the money to another stranger at the bus stop, who was actually a member of the research team. In contrast to the experimental result, no one in the naturalistic study gave any portion of the endowment to the stranger. Thus, Winking and Nizer suspected that in the past the setting of the experimental context induced participants to choose prosocial options.

The Pepsi challenge

CokeThe preceding examples may be too remote to you. Let's look at some products that we consume everyday: Coke and Pepsi. In experimental settings, most participants prefer Pepsi to Coke. However, Gladwell (2007) disputed the result by presenting evidence that this so-called "Pepsi Challenge" is based on the unrealistic "sip test" method. Most tasters would favor the sweeter of two beverages when they make a single sip only, but the result is reversed when the entire can or bottle is consumed (I am skeptical of this type of taste tests, including wine tests, coffee tests, water tests...etc. Our limited sensation may not be able to distinguish one from another while the difference is very subtle. In an experiment the researcher tinted the white wine and asked the wine experts to rate the "red wine." Surprisingly, the experts did not recognize that it is not a glass of red wine! The following movies are some examples:

(Similar results are found in coffee tests and water tests).

Other elements and sample size

In educational research, What Works Clearinghouse (WCC) still adopts the conventional ranking of study type. Slavin is critical of this criterion by pointing out that in small, brief, and artificial studies random assignment does not necessarily guarantee validity; over-emphasizing randomized studies without taking sample size and other design elements into account might introduce bias that "can lead to illogical conclusions" (p.11).

Casual Inferences

Ruling out rival interpretations in quasi-experiment and observational studies

Some statisticians assert that one can never draw causal inferences without experimental manipulation (e.g. SAS Institute, 1999). Some researchers argued that causal inferences are weakened in quasi-experiments (e.g. Keppel & Zedeck, 1989). However, Christensen (1988) held a more liberal position:
Many causal inferences are made without using the experimental framework; they are made by rendering other rival interpretations implausible. If a friend of yours unknowingly stepped in front of an oncoming car and was pronounced dead after being hit by the car, you would probably attribute her death to the moving vehicle. Your friend might have died as a result of numerous other causes (a heart attack, for example), but such alternative explanations are not accepted because they are not plausible. In like manner, the causal interpretations arrived at from quasi-experimentation analysis are those that are consistent with the data in situations where rival interpretations have been shown to be implausible. (p.306)
I would go further than Christensen to assert that even some observational studies could yield valid causal conclusions. While the example of an car accident in Christensen's argument is hypothetical, we can see a similar example in the real life. Some researchers assert that we could still attribute causal factors to effects with observational data if virtually identical units in two different outcomes are observed. To attribute causal factors to accidents, in Georgia 300 accidents were compared to 300 non-accidents involving the same car, driver, weather condition, and lighting. The non-accidents occurred one mile back on the same road, a location passed by the driver minutes earlier en route to the crash site. Researchers found a substantial excess of roads that curved more than six degrees with downhill gradients. In another example, to answer the question of whether helmets reduce the risk of death in motorcycle crashes, virtually identical units were compared: Cases in which two people rode the same motorcycle, a driver and a passenger, one helmeted and the other was not. Researchers concluded a 40% reduction of risk resulted from wearing a helmet (Rosenbaum, 2005).

Kim Jong Un’s New Reign on the Cover of Southern Weekly Magazine in the PRC, April 23, 2012A similar scenario could be found in political and economics studies. During the Cold War era, the whole world was divided into three camps, namely, the Communist world led by the Soviet Union and the People's Republic of China, the Capitalist conglomerate led by the United States, and the non-aligned countries. Some countries were partitioned into two political entities due to unresolved ideological differences embraced by different local parties. Obvious examples include North Korea and South Korea, Mainland China and the Republic of China (Taiwan), East Germany and West Germany, as well as North Vietnam and South Vietnam. This division is not a result of randomization, of course. Nevertheless, the observational data about the two camps could still inform us about certain causes and effects. Many years ago philosopher Margaret Walker (person communication) argued that there is no causal relationship between Communist ideology and the horrible consequences in the Communist countries. I held a different view. As mentioned before, we could still attribute causal factors to effects with observational data when virtually identical conditions associated with the outcomes are observed. In terms of cultural heritage, language, and racial attributes, the two countries in each pair on the preceding list share a high degree of resemblances. The major difference is found in the political and economic system only. Owing to self-isolation and the containment policy performed by the West, the Communist blocs could "experiment" with central planning, class struggle, and so on without much outside influence. Needless to say, after half a century people were disenchanted by the broken economy and the lack of human rights in those Communist countries (Courtois et al., 1999). It would be difficult to deny a causal relationship between Communism and those undesirable consequences (Yu, 2009).

Another good example of natural experiment is racial diversity before and after Proposition 209. In 1996, the State of California passed Proposition 209, which prohibited public institutions from using race-based admission policies. After Proposition 209 there was a 50-percent reduction in black freshman enrollment and a 25-percent drop for Hispanics. Nonetheless, although the black and Hispanic enrollment was reduced at the most prestigious University of California campuses (-42% at UC Berkley; -37% at UCLA), other less competitive UC campuses increased their black and Hispanic enrollment (+22% at UC Irvine, +18% at UC Santa Cruz; +65% at UC Riverside) (Sander & Taylor, 2012).

Lurking variables, proxy measure, and theoretical casual variables in correlational studies

Archival research is also called correlational research because cause-and-effect inferences cannot be directly made. For example, even though the last twenty-year data shows a positive correlation between productivity and school performance, it would be a leap of faith to conclude that school performance gain is the cause of productivity gain or vice versa. Usually another variable, which may be the true cause, is "lurking" behind background. This variable is called lurking variable, and is easily undetected by a correlational study.

Even if the researcher is aware of the existence of the lurking variables, he or she has no control of what data were collected. Rather, the researcher must go by the existing variables available in the data bank. Another limitation that hinders the researcher from drawing a valid causal inference from archival data is the problem of indirect measurement. On some occasions the variable chosen by the researcher is a proxy measure of what the researcher intends to study. For example, the researcher may be interested in studying the causal relationship between Christian spirituality and productivity. If the instrument is designed by the researcher, he or she might insert questions like "how often do you pray," "how often do you attend church activities" or other questions specific to Christian spirituality into the survey. However, when archival data are downloaded from the Internet, the researcher might use general demographics (e.g. religion affiliation) to indicate Christian spirituality. In other words, the researcher will make inferences based on inferences (proxy measure). Although the problems of lurking variables and proxy measure could also happen in other types of research methods, they are especially severe when the researcher is unable to customize the instrument.

There are many jokes about careless use of correlational studies. For example, once a study indicated that consumption of alcohol improves academic performance (the explanation may be something else: when the overall economy improves, both alcohol consumption and academic performance go up). A study in Taiwan during the 70s indicates that the more woks a household owned, the fewer children the family had. Thus, the government gave woks to households in an attempt to lower national birth rate. The moral of these stories: researchers should select theoretical casual variables even though the study is correlational.

Nevertheless, Luker, Luker, Jr., Cobb, and Brown (1998) defended the use of causal inference in correlation/regression frameworks:

In the social and behavioral sciences, experimental randomization and control are usually not possible. This has led to an awkward condition in which our work does not permit useful policy recommendations. The well-intentioned assertion that relationships do not mean causation, while useful in contesting gross simple-mindedness, is paralyzing and misleading in the social sciences. Or, as Dewey puts it, the critical characteristic of all scientific operations is revealing relationships. Relationships are a necessary condition of causation. We know that X cannot be a cause of Y unless X and Y are related. The causal analysis of nonexperimental data, therefore, can only go on through the analysis of relationships. Causal inference from non-experimental data, then, requires the testing of theoretical causal variables in a variety of quasi-experimental or multiple regression frameworks...Statistical failures of models suggest that we are not on the right track. Confirmation of the models suggests the possibility of ameliorative solutions.

It is noteworthy that the problems faced by experimentation can also be found in quasi-experiments. The main point is that the "real world" is more complicated than an experimental setting, in which the treatment and the outcome, or the cause and the effect has a one-on-one mapping. Murnane and Willet (2011) wrote, "Randomized experiments and quasi-experiments typically provide estimates of the total effect of a policy intervention on one or more outcomes, not the effects of the intervention holding constant the levels of other inputs" (p.31). This issue, which is concerned with internal validity and external validity, will be discussed in the write-up entitled Threats to validity of Research Design.

Explicit questions and selection bias in survey research

Whether causal inferences can be drawn from survey research is debatable. It is true that survey research does not implement any variable manipulation. However, when a questionnaire includes explicit questions concerning rationale and motivation, such as "Why do you choose Web-based instruction over conventional instruction?" it is difficult to explain that the answers provided by respondents do not indicate any cause and effect.

Generalizability always comes hand in hand with causal inferences. Survey research is not weaker than experiment in this regard. In many situations, survey research tends to obtain a more random sample than experimental research does. Usually subjects are required to be physically present in experiment studies, and thus only convenience samples are recruited from the local campus or the local town. On the other hand, survey research can break through this limitation by sending questionnaires to prospective subjects across the country. In the age of Internet, the researcher can even set up an online form to reach potential respondents all over the world.

Truman presidential election However, someone may argue that a "cyber-sample" is a self-selected sample rather than a random sample. In this case a systematic bias may affect who responds to the questionnaire and who doesn't. The prediction of "Dewey defeats Truman" by Chicago Daily Tribune in 1948 presidential election is a classic example of selection bias. The interviewees were polled by phone and thus the sample was confined to households who own a telephone. By the same token, when the survey is posted on the Web, it is likely that respondents are computer literate and have access to computer equipment. Indeed, the same problem can be found in experimental research. Subjects could refuse to participate in the experiment or withdraw from the study even though they start the process. In both survey research and experimental research, the question is not whether there are missing data. Rather, the question should be: "Are data completely missing at random?"

Nonetheless, if the subject matter to be studied is Web-based instruction, this should not be considered a selection bias. In an online survey concerning Web-based instruction, the researcher should expect that all respondents possess basic computer operation skills and have access to the Internet (Once I assisted a researcher to post an online survey on my database server. But several respondents, who used 2400 baud modems, complained that it took five to ten minutes to load a page).

Research design and statistical analysis

Traditionally, analysis of variance (ANOVA) is said to be appropriate for data collected in an experiment whereas regression analysis is considered a proper method for data collected in non-experimental designs. Keppel and Zedeck (1989) argued that both ANOVA and regression are suitable to experimental designs while only regression is fitful to most non-experimental designs. In other words, regression is applicable to both experimental and non-experimental deigns when the independent variables are continuous and/or categorical. For this reason, Pedhazur and Schmelkin (1991) asserted that regression is superior to ANOVA. However, Pedhazur and Schmelkin criticized that in non-experimental designs some researchers convert continuous variables into categorical variables in order to fit the data into an ANOVA framework as if it were experimental. This conversion not only leads to loss of information, but also changes the nature of the variables and the design.

Further Reading

For beginners

Kerlinger (1986), Shadish, Cook and Campbell (2002) are two good books to get started with experimental design for neither book requires a strong mathematical or statistical background. Their books concentrate on the design aspect rather than the analysis aspect.

Montgomery (2012) is a very updated and comprehensive book though it is written for engineering majors. Readers should be able to follow the content after taking one or two introductory statistics courses. You may skip the chapter on response surface because it may not be applicable to educational and psychological research. Dr. Montgomery is a professor of Industrial Engineering at Arizona State University.

For intermediate users

Kennedy and Bush (1985)'s book was written for graduate students in education and psychology who have a modest background in both mathematics and statistics and who are interested in a subject-matter field rather than statistical methodology. One nice thing about the book is that it explains the mathematical notation symbols, which are confusing to many readers.

For beginner and intermediate users

Levine & Parkinson (1994) is a book for both beginners and research professionals. The first half of the book covers experimental methods for psychologists in general whereas the second half covers very detailed examples of experimental methods in cognitive psychology, social psychology, and clinical psychology. Levine and Parkinson are professors of psychology at Arizona State University.

For advanced users

Maxwell & Delaney (1990) and Winer, Brown, and Michels (1991) are considered classics in the field of experimental design. Their books cover both the design and the analysis aspects. However, their books require a very strong statistical background.

Last revised: 2017


  • Appleton, D. R. & French, J. M. (1996). Ignoring a covariate: An example of Simpson's paradox. American Statistician, 50, 340-341.

  • Babbie, E. (1992). The practice of social research (6th ed.). Belmont, CA: Wadsworth.

  • Barkow, J. H. (1989). Darwin, sex, and status: Biological approaches to mind and culture. Toronto: University of Toronto Press.

  • Berwick, D. (2008, August). Inference and improvement in health care. Paper presented at the 2008 Joint Statistical Meeting, Denver, CO.

  • Canadian Task Force on Preventive Health Care. (2003). Canadian Task Force on Preventive Health Care levels of evidence used to rate research design and quality of individual studies. Retrieved August 13, 2008, from http://www.ctfphc.org/

  • Christopher, B. (2016, September). Why the father of modern statistics didn’t believe smoking caused cancer. Priceonomics. Retrieved from https://priceonomics.com/why-the-father-of-modern-statistics-didnt-believe/

  • Berk, R. (2005). Randomized experiments as the bronze standard. UC Los Angeles: Department of Statistics, UCLA. Retrieved from http://escholarship.org/uc/item/2d57d1kq.

  • Chaitin, G. J. (1998). The limits of mathematics: A course on information theory and the limits of formal reasoning. Singapore: Springer-Verlag.

  • Christensen, L. B. (1988). Experimental methodology. Boston, MA : Allyn and Bacon.

  • Cohen, J. (1962). The statistical power of abnormal-social psychological research: A review. Journal of Abnormal and Social Psychology, 65, 145-153.

  • Courtois, S., Kramer, M., Werth, N., Panne, J. L. , Paczkowski, A., Bartosek, K., Margolin, J. L. (1999). The black book of Communism: Crimes, terror, repression. Boston, MA: Harvard University Press.

  • Deese, J. (1972). Psychology as science and art. New York, NY: Harcourt Brace Jovanovich, Inc.

  • Essock, S., Drake, R., Frank, R., & McGuire. T. (2003). Randomized controlled trials in evidence-based mental health care: Getting the right answer to the right question. Schizophrenia Bulletin, 29(1), 115-123.

  • Fisher, R. A. (1971). The design of experiments (9th ed.). New York, Hafner Publishing Company.

  • Frick, R. W. (1998). Interpreting statistical testing: Process and propensity, not population and random sampling. Behavior Research Methods, Instruments, & Computers, 30, 527-535.

  • Gladwell, M. (2007). Blink: The power of thinking without thinking. New York, NY: Black Bay Books.

  • Harris, S. (n.d.). The myth of secular moral chaos. Retrieved from http://www.samharris.org/site/full_text/the-myth-of-secular-moral-chaos

  • Henrich, J., Heine, S. J., & Norenzayan, A. (2010). The weirdest people in the world? Behavioral and Brain Sciences, 33, 61-135. http://dx.doi.org/ 10.1017/S0140525X0999152X

  • Hsu, L. M. (1989). Random sampling, randomization, and equivalence of contrasted groups in psychotherapy outcome research. Journal of Consulting and Clinical Psychology, 57, 131-137.

  • Keppel, G., & Zedeck, S. (1989). Data analysis for research designs: Analysis of variance and multiple research/correlation approaches New York: W. H. Freeman.

  • Kennedy, J. J. & Bush, A. J. (1984). An introduction to the design and analysis of experiments. Lanham, MD: University Press of America, Inc.

  • Keppel, G. & Zedeck, S. (1989). Data analysis for research design: Analysis of variance and multiple regression/correlation approaches. New York: W. H. Freeman.

  • Kerlinger, F. N. (1986). Foundations of behavioral research. New York, NY: Holt, Rinehart and Winston.

  • Levine, G., & Parkinson, S. (1994). Experimental methods in psychology. Hillsdale, N.J.: L. Erlbaum.

  • Luker, B., Luker, B. Jr., Cobb, S. L., & Brown, R. (1998). Postmodernism, institutionalism, and statistics: Considerations for an institutionalist statistical method. Journal of Economic Issues, 32, 449-457.

  • Management Study Guide. (2016). Secondary data. Retrieved from http://managementstudyguide.com/secondary_data.htm

  • Maxwell, S. E., & Delaney, H. D. (1990). Design experiments and analyzing data: A model comparison perspective. Belmont, CA: Wadsworth Publishing company.

  • May R. B., & Hunter, M. A. (1988). Interpreting students' interpretations of research, Teaching of Psychology, 15, 156-158.

  • Medical News Today. (2008). Attack traditional ways of assessing the evidence of therapeutic interventions. Retrieved from http://www.medicalnewstoday.com/articles/126043.php

  • Montgomery, D. C. (2012). Design and analysis of experiments (8th ed.). New York, NY: Wiley.

  • Morse, J. (2007). Sampling in grounded theory. In A. Bryant, & K. Charmaz (Ed.), Sage handbook of grounded theory (pp. 229-244). Los Angeles, CA: Sage.

  • Moule, P., & Hek, G. (2012). Making sense of research: An introduction for health and social care practitioners. Thousand Oak, CA: Sage.

  • Murnane, R. J. & Willet, J. B. (2010) Methods Matter: Improving causal inference in educational and social science research. NY: Oxford University Press.

  • Pedhazur, E. J. & Schmelkin, L. P. (1991). Measurement, design, and analysis : An integrated approach. Hillsdale, N.J. : Lawrence Erlbaum Associates.

  • Pitman,E. J. G. (1937a). Significance tests which may be applied to samples from any populations. Journal of Royal Statistical Society B, 4, 119-130.

  • Pitman,E. J. G. (1937b). Significance tests which may be applied to samples from any populations II: The correlation coefficient. Journal of Royal Statistical Society B, 4, 225-232.

  • Pitman,E. J. G. (1938). Significance tests which may be applied to samples from any populations III: The analysis of variance test Journal of Royal Statistical Society B, 29, 322-335.

  • Plichta, S., & Garzon, L. (2009). Statistics for nursing and allied health. New York, NY: Lippincott Williams & Wilkins.

  • Ram, R., & Ural, S. (2014). Comparison of GDP per capita data in Penn World Table and World Development Indicators. Social Indicators Research, 116, 639-646.

  • Reichardt, C. S., & Gollob, H. F. (1999). Justifying the use and increasing the power of a t test for a randomized experiment with a convenience sample. Psychological Methods, 4, 117-128.

  • Rosenbaum, P. (2005). Heterogeneity and causality: Unit Heterogeneity and design sensitivity in observational studies. American Statistician, 59, 147-152.

  • Sander, R., & Taylor, S. (2012). Mismatch: How affirmative action hurts students it's intended to help, and why universities won't admit it. New York, NY: Basic Books.

  • SAS Institute. (1999). Comments on interpreting regression statistics. Retrieved from http://www.sas.com

  • Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and Quasi-experimental designs for genealized causal inference. New York, NY: Wadsworth Publisher.

  • Shipley, B. (2000). Cause and correlation in biology: A user's guide to path analysis, structural equations and causal inferences. Cambridge: Cambridge University Press.

  • Schneider, B., Carnoy, M., Kilpatrick, J., Schmidt, W. H., & R. J. Shavelson. (2008). Estimating causal effects using experimental and observational designs: A think tank white paper. Washington D. C.: American Educational Research Association.

  • Simpson, E. H. (1951). The interpretation of interaction in contingency tables. Journal of the Royal Statistical Society, Ser. B., 13, 238-241.

  • Slavin, R. (2008). Perspectives on evidence-based research in education. Educational Researcher, 37(1), 5-14.

  • Winer, B. J., Brown, D. R., & Michels, K. M. (1991). Statistical principles in experimental design. New York: McGraw-Hill, Inc.

  • Winking, J., & Mizer, N. (2013). Natural-field dictator game shows no altruistic giving. Evolution and Human Behavior, 34, 288-293. http://dx.doi.org/10.1016/j.evolhumbehav.2013.04.002

  • Yu, C. H. (2009). Causal inferences and abductive reasoning: Between automated data mining and latent constructs. Saarbrücken, Germany: VDM-Verlag.

  • Yu, C. H., DiGangi, S., Jannasch-Pennell, A., & Kaprolet, C. (2010). A data mining approach for identifying predictors of student retention from sophomore to junior year. Journal of Data Science, 8, 307-325. Retrieved from http://www.jds-online.com/
Last updated: 2018, January.

Go up to the main menu



Other courses

Search Engine

Contact me