Warning Signs: Bad Research Ahead

Posted July 9, 2009

In an effort to keep this blog alive during this busy time, I am digging through my library of writings for relevant material. This summary was written as a reference for new scientists and science teachers and may be also be found on the CTEG website (Critical Thinking Education Group).

Some indications the research may not be of the highest quality

Science is never certain
Science accepts a few first principles on which all arguments rest such as the principle that observed effects have natural causes. If any one of these first principles is shown to be false, then everything we know is open to question. Therefore, science recognizes that 100% certainty is not possible. Science is open-minded.
That said, some respected scientists have used this word in the popular press. I suspect that they mean strong evidence, however, I feel this practice misleads the public and fuels common misunderstandings of science.

“The effects of…”
“The influence of…”
“The role of…”        …followed by ANYTHING that cannot be randomly assigned to subjects, such as:

  • Gender
  • Birth Order
  • Race/Ethnicity/Religious Affiliation
  • Intelligence
  • Height
  • Favorite Color… you get the idea

All of these statements imply causal relationships and nothing that cannot be randomly assigned can be said to cause anything (in a single experiment or study). The acceptance of causal relationships among variables that cannot be studied experimentally (meaning a “true” experiment) requires extraordinary converging evidence.

Like the titles that imply cause, the language used in conclusions can be misleading. This is particularly problematic in secondary sources such as the popular media. “A is related to B” does not mean that A causes B. For example, the years of education is related to number of teeth, but education does not cause tooth loss.

“It depends on what your definition of ‘is’ is…”
Studies that claim relationships among variables, but define those variables in vague or unique terms should be scrutinized. You may claim that your diet results in weight loss by adding up all of the pounds people in sample have lost in a given time period, but your definition is a deliberate misrepresentation of weight loss if you do not include the weight people gain.

“Eating Brand X reduces cholesterol…”
Compared to what? EVERYTHING in the world is relative. Effects are differences in measures among levels of a variable. A variable, by definition, is something that varies. A study may find that eating Brand X may result in lower cholesterol than eating Brand Y (given otherwise identical conditions), but no single value of a variable can do anything or affect anything.
When comparisons are missing in the communication of findings, the claim cannot be evaluated. For example, it is entirely possible that eating Brand Y raises cholesterol.

One-group, pretest-post test comparisons tell us absolutely nothing. For example, findings such as:

  • People with headaches are given aspirin and experience pain relief within 1 hour.
  • People enrolled in my weight-loss clinic lost an average of 10 pounds in 4 weeks.
  • Walls are cleaner after being washed with “SuperClean”.

May easily be explained a number of ways, including:

  • Headaches usually last about 45 minutes.
  • People who are publicly weighed tend to lose weight to avoid embarrassment.
  • Walls washed with anything are cleaner than they were before being washed.


This term has lost its meaning through repeated misuse.

  • We do not find significance or seek it.
  • Effects are not significant or non-significant. By definition, if there is an effect, a statistically-significant result was found. If no significant result was found, there is no effect.
  • Statistical significance does not mean significant findings. When a researcher says “Group A had significantly higher scores than Group B,” they mean that the difference in scores was probably due to the manipulated variable, not chance. We cannot conclude that the treatment results in significant differences in the dependent measure.


A report that starts with, “the purpose of this study is to find support for the hypothesis that…”
Scientists do not seek support for what they believe. Scientists seek the truth.

“The difference was not significant, however, the mean was slightly higher…”
In hypothesis testing, there are no “trends toward significance”. Lines are drawn at acceptable levels of error and moving those lines to suit one’s wishes defies the purpose of the scientific method – to remove human biases.

THE BOTTOM LINE: If the differences are not statistically significant, then we must conclude that there are no differences in the population, or we must discuss the methodological flaws or omissions that resulted in the null findings. We can NEVER reject the null hypothesis based on what we did NOT find.

Hypotheses without a theoretical foundation
Sometimes hypotheses seem to spring from nowhere, either somewhat unrelated to the statements made in the introduction, or even contradicting them. At times, the logic in the introduction is a bit twisted, or speculative, or requires a number of assumptions to go from the known literature to the hypothesis to be tested. This is a sign that the introduction was actually written after the study was completed.

Although it is a common practice to write after completion of the study, changing one’s hypothesis to match findings is not. In fact, it is a serious violation of ethical principles.

Researchers using proper scientific method have a priori hypotheses that are based on what is known (findings of other studies, mostly). Whether this information is written before or after the study is conducted is not the issue.
Hypotheses which do not match the method
Sometimes the method of a study does not seem to directly test the hypotheses. This is another indication that the hypotheses themselves have been changed to suit the findings, making it appear as though the conclusions are more meaningful than they actually are. Studies often produce interesting findings that are not directly related to the original hypotheses. These should be discussed, but should never be over-analyzed using the same data, nor should they be allowed to replace the original hypothesis.

We may have good explanations for what we have found, but that does not mean our explanations are correct (or even close). Instead of reconstructing a study to fit the findings, new hypotheses should be tested in future studies.

Too much analysis, too many factors, mediation & moderation
The canon of parsimony dictates that explanations that require the fewest assumptions are the most likely to be correct. The more assumptions that are required, the greater the probability that an error has been made. The more statistical tests that are conducted, the greater the probability of error. Studies can be overcomplicated in their procedure, analysis, or even hypotheses. Hypotheses regarding mediating or moderating variables must be tested with care.


  • Citing statements from the introductions of a research report usually involves hearsay (”Joe said that Bart said…”). Only the findings of studies can be evaluated as evidence.
  • Citations of review articles, textbooks, or popular press are always filtered information, either in the form of opinion or, again, hearsay.

Drawing conclusions across tests
A common mistake is drawing conclusions without direct comparisons. For example, a group of children were given a pretest on verbal ability, then received either “phonics” or “traditional” teaching in language arts. They were tested again after 4 weeks of classes. Separate tests showed that the scores of students in the “phonics” condition improved, but the scores of those in the “traditional” condition did not. The researchers concluded that “phonics” was more effective than “traditional” teaching. This conclusion does not logically follow, since the scores of those in “phonics” were never compared to those in “traditional” teaching methods. A proper analysis in this example would be to compared the pretest-post test differences of one condition to those the other condition.

Causal modeling does not permit causal conclusions
This kind of analysis has become a popular means of analyzing data to show complicated relationships among variables. It is often mistaken, however, for a means to overcome a lack of random assignment. Causal models are no different from other models. Finding a model which fits data does not mean a causal relationship exists. Correlation, which is required for cause, is sufficient for a good fit. Only experimental methods can result in valid causal conclusions.


Tom on July 11th, 2009 at 04:33:
Nice post, thanks for it. I had one comment. You wrote “THE BOTTOM LINE: If the differences are not statistically significant, then we must conclude that there are no differences in the population, or we must discuss the methodological flaws or omissions that resulted in the null findings. We can NEVER reject the null hypothesis based on what we did NOT find.”
I don’t concur that a lack of statistical significance means we have to conclude there is no difference. Differences that are small (though perhaps big enough to be important) might go undetected due to issues of statistical power. So, whether we find statistical significance or not, I believe our best course is to estimate the difference of interest and consider a confidence interval around that difference.

Administrator on July 11th, 2009 at 08:11:
Tom — obviously this is not a simple topic and I don’t want to try to go into it in detail; readers can look up the concepts of power and error. However, I would like to acknowledge your comment with a note that there is an ongoing debate among statisticians over the importance of hypothesis testing and the value of effect sizes. Confidence intervals are a part of that debate.
This issue is complicated, but my opinion is that we must have a decision criteria. The practical significance of a finding is extremely important, but the temptation to generalize an effect size to a population is great and it would be a huge mistake. Experimental research, for example, is controlled and effect sizes are maximized in order to find out if a treatment affects the dependent variable at all.
Given null findings, we can conclude that there are no differences in the population because we have assumed this all along. We start with the assumption that the null hypothesis is true and we retain that as truth until given enough evidence to decide otherwise. A confidence interval is not evidence of a difference, but rather an estimate of the range of values (including zero) within the which the difference is likely to fall.
Again, this is obviously a complicated issue that I should not try to discuss in a comment, but I wanted to acknowledge that your point is valid and discussion is certainly warranted.

Comments for this page are closed.

Print Friendly, PDF & Email