09-ResearchDesign-Limitations.Rmd

# Study design limitations {#Interpretation}



<!-- Introductions; easier to separate by format -->

```{r, child = if (knitr::is_html_output()) {'./introductions/09-ResearchDesign-Limitations-HTML.Rmd'} else {'./introductions/09-ResearchDesign-Limitations-LaTeX.Rmd'}}
```



## Introduction {#Chap9-Intro}

The type of study and the study design determine how the results of the study should be interpreted.
Ideally, a study would be perfectly externally and internally valid; in practice this is very difficult to achieve.
Practically *every* study has limitations.
The results of a study should be interpreted in light of these limitations.

Limitations generally can be discussed through three components:

* *Internal validity* (Sect. \@ref(IntroInternalValidity)):
  Discuss any limitations to internal validity due to the study design (such as identifying possible confounding variables).
  This is related to the *effectiveness* of the study with the sample (Sect. \@ref(InterpretStudyDesign)).
* *External validity* (Sect. \@ref(IntroExternalValidity)):
  Discuss how well the sample represents the intended population.
  This is related to the *generalisability* of the study to the intended population (Sect. \@ref(InterpretGeneralisability)).
* *Ecological validity*:
  Discuss how well the study methods, materials and context approximate the real situation being studied.
  This is related to the *practicality* of the results to real life (Sect. \@ref(InterpretApplicability)).

All these issues should be addressed when considering the study limitations.


::: {.importantBox .important data-latex="{iconmonstr-warning-8-240.png}"}
Almost every study has limitations.
*Identifying* potential limitations, and *discussing the likely impact* they have on the interpretation of the study results, is important and ethical.
:::



::: {.example #ValidityBeer}
@delarue2019taking discuss studies where subjects rate the taste of new food products.
They note that taste-testing studies should (p. 78):

> ... allow generalizing the conclusions obtained with a consumer sample in one particular study to the general targeted population [i.e., external validity]... 
> In the same time, tests should be reliable in terms of accuracy and replicability [i.e., internal validity]. 

However, even when studies have good internal and external validity, these studies often result in a 'high rate of failures of new launched products'.
That is, the studies do not replicate the real world, and so lack *ecological validity*.
:::


## Limitations: internal validity {#InterpretStudyDesign}

Internal validity refers to how well the study design isolates the relationship of interest and eliminates all other possible explanations (Sect. \@ref(IntroInternalValidity)).
A discussion of the limitations of internal validity should cover, as appropriate: possible confounding variables; the impact of the Hawthorne, observer, placebo and carry-over effects; the impact of any other design decisions.

If any of these issues are likely to compromise internal validity, the implications on the interpretation of the results should be discussed.
For example, if the study design implies that the Hawthorne effect may be an issue (since the participants were not blinded), this should be clearly stated, and the conclusion should indicate that the individuals in the study may have behaved differently than usual because (for example) they knew the were in a study. 



<!-- ```{r echo=FALSE, StudyProsCons} -->
<!-- StudyPC <- array( dim = c(3, 6) ) -->
<!-- colnames(StudyPC) <- c("Study type", -->
<!--                        "C groups may be initially different", -->
<!--                        "Individuals may decide C", -->
<!--                        "Blinding researchers", -->
<!--                        "Blinding individuals", -->
<!--                        "Ecologically valid") -->

<!-- StudyPC[1, ] <-  c("Observational", -->
<!--                    "Yes", -->
<!--                    "Yes", -->
<!--                    "Difficulty", -->
<!--                    "Probably", -->
<!--                    "Likely") -->
<!-- StudyPC[2, ] <- c("Quasi-experimental", -->
<!--                   "Yes", -->
<!--                   "No", -->
<!--                   "Possible", -->
<!--                   "Possible", -->
<!--                   "Possible") -->
<!-- StudyPC[3, ] <- c("True experiment", -->
<!--                   "No", -->
<!--                   "No", -->
<!--                   "Possible", -->
<!--                   "Possible", -->
<!--                   "Possible") -->

<!-- if( knitr::is_latex_output() ) { -->
<!--    kable(StudyPC, -->
<!--         format = "latex", -->
<!--         longtable = FALSE, -->
<!--         booktabs = TRUE, -->
<!--         linesep = c( "\\addlinespace"), # Add a bit of space between all rows.  -->
<!--         caption = "Study types and research questions", -->
<!--         align = c("r", "c", "c", "c", "c", "c"))   %>% -->
<!--    kable_styling(full_width = FALSE, font_size = 10) %>% -->
<!--    column_spec(column = 1, width = "3cm") %>% -->
<!--    column_spec(column = 2, width = "2cm") %>% -->
<!--    column_spec(column = 3, width = "2cm") %>% -->
<!--    column_spec(column = 4, width = "2cm") %>% -->
<!--    column_spec(column = 5, width = "2cm") %>% -->
<!--    column_spec(column = 6, width = "2cm") %>% -->
<!--    row_spec(0, bold = TRUE) # Columns headings in bold -->
<!-- } -->

<!-- if( knitr::is_html_output() ) { -->

<!--   out <- kable(StudyPC, -->
<!--                format = "html", -->
<!--                align = c("r", "c", "c", "c", "c", "l", "c", "c"), -->
<!--                longtable = FALSE, -->
<!--                caption = "Study types and research questions", -->
<!--                booktabs = TRUE)  -->

<!--   if ( knitr::is_html_output(excludes = "epub")) { -->
<!--     kable_styling(out,  -->
<!--                   full_width = FALSE) %>% -->
<!--     row_spec(row = 0,  -->
<!--              bold = TRUE)   -->
<!--   } else { -->
<!--     out   -->
<!--   } -->
<!-- } -->

<!-- ``` -->


<div style="float:right; width: 222x; border: 1px; padding:10px">
<img src="Pics/iconmonstr-door-7-240.png" width="50px"/>
</div>


::: {.example #LimitationsSeeds name="Study limitations"}
A study [@axmann2020access] randomly allocated Ugandan farmers to receive, or not receive, hybrid maize seeds to improve internal validity.
However, one potential threat to internal validity was that farmers receiving the hybrid seeds could share their seeds with their neighbours.

Hence, the researchers contacted the 75 farmers allocated to receive the hybrid seeds; none of the contacted farmers reported selling or giving seeds to other farmers. 
This extra step increased the internal validity of the study.
:::


The internal validity of *observational studies* is often compromised because confounding can be less effectively managed than for experimental studies (e.g., random allocation is not possible).
The internal validity of *experimental studies* involving people is often compromised because people must be informed that they are participating in a study.


<div style="float:right; width: 222x; border: 1px; padding:10px">
<img src="Illustrations/pexels-ketut-subiyanto-4308195.jpg" width="200px"/>
</div>


::: {.example #InternalHandHygiene name="Internal validity"}
In a study of the hand-hygiene practices of paramedics [@barr2017self], *self*-reported hand-hygiene practices were very different than what was reported by *peers*.

When participants knew they were being studied, their responses made their own behaviours appear better than their colleagues.
This is a study limitation that was necessary to discuss.
:::


::: {.exampleExtra  data-latex=""}
A study evaluated using a new therapy on elderly men, and listed some limitations of their study:

> ... the researcher was not blinded and had prior knowledge of the research aims, disease status, and intervention. 
> As such, these could all have influenced data recording [...]
> The potential of reporting bias and observer bias could be reduced by implementing blinding in future studies.
>
> --- @kabata2021effect, p. 10
:::


:::{.exampleExtra data-latex=""}
A study [@botelho2019effect] examined the food choices made when subjects were asked to shop for ingredients to make a last-minute meal.
Half were told to prepare a 'healthy meal', and the other half told just to prepare a 'meal'.
Part of the Discussion stated:

> Another limitation is that results report findings from a simulated purchase.
> As participants did not have to pay for their selection, actual choices could be different.
> Participants may also have not behaved in their usual manner since they were taking part in a research study, a situation known as the Hawthorne effect.
>
> --- @botelho2019effect, p. 436
:::



## Limitations: external validity {#InterpretGeneralisability}


<div style="float:right; width: 222x; border: 1px; padding:10px">
<img src="Pics/iconmonstr-share-11-240.png" width="50px"/>
</div>


External validity refers to the ability to *generalise* the results to the entire intended population from the sample (Sect.\ \@ref(IntroExternalValidity)).
For a study to be externally valid, it must first be internally valid: if the study of not effective in the sample studied (i.e., internally valid), the results may not apply to the intended population either.


::: {.importantBox .important data-latex="{iconmonstr-warning-8-240.png}"}
External validity refers to how well the sample is likely to represent the *target population* in the RQ.
If the population is Alaskans, then the study is externally valid if the sample is representative of Alaskans.
The results *do not* have to apply to people in the rest of the United States (though this can be commented on, too).
The intended population is *Alaskans*.
:::


External validity depends on *how* the sample was obtained.
Results from random samples are likely to generalise to the population and be externally valid.
(The analyses in this book assume all samples are *simple random samples*.)
Furthermore, results from approximately representative samples may generalise to the population and be externally valid if those *in* the study are not obviously different than those *not in* the study.


::: {.example #ExternalNZ name="External validity"}
A New Zealand study [@data:Gammon2012:B12] identified (for well-documented reasons) a *population* of interest: 'women of South Asian origin living in New Zealand' (p. 21).
The women in the *sample* were 'women of South Asian origin living in New Zealand [...] recruited using a convenience sample method throughout Auckland' (p. 21).

The results may not generalise to the *intended* population because all the women in the sample came from only one city (Auckland), and the sample was not a *random* sample (so the study may not be externally valid).

The results will not generalise to *all* New Zealand women, but this is *not* a limitation since this was not the target population.
The researchers did not intend the results to apply to *all* New Zealand women.
:::


::: {.example #ExternalBiochar name="Using biochar"}
A study of growing ginger using biochar [@farrar2018short] used one farm at Mt Mellum, Australia.
The results may only generalise to growing ginger at Mt Mellum,  but since ginger is usually grown in similar types of climates and soils, so the results *may* apply to other ginger farms also.
:::


## Limitations: ecological validity {#InterpretApplicability}

The likely *practicality* of the study results in the real world should also be discussed.
This is called *ecological validity*.


<div style="float:right; width: 222x; border: 1px; padding:10px">
<img src="Pics/iconmonstr-recycling-9-240.png" width="50px"/>
</div>


::: {.definition #EcologicalValidity name="Ecological validity"}
A study is *ecologically valid* if the study methods, materials and context closely approximate the real situation of interest.
:::


Studies don't *need* to be ecologically valid to be useful; much can be learnt under special conditions, as long as the potential limitations are understood when applying the results to the real world.
The ecological validity of experimental studies may be compromised because the experimental conditions are sometimes artificially controlled (for good reason).


<div style="float:right; width: 222x; border: 1px; padding:10px">
<img src="Illustrations/pexels-nappy-936019.jpg" width="200px"/>
</div>


::: {.example #EcologicalCups name="Ecological validity"}
Consider a study to determine the proportion of people that buy a coffee in a reuseable cup.
People could be *asked* about their *behaviour*.
This may not be ecologically valid, as how people *act* may not align with what they *say*.

An alternative study could *watch* people buy coffees at various coffee shops, and record what people *do* in practice.
This second study is more likely to be *ecologically valid*, as we are watching real-world behaviour.
:::


:::{.exampleExtra data-latex=""}
A study observed the effect of using high-mounted rear brake lights [@data:Kahgane1998:RearBrakeLights], which are now commonplace.
The American study showed that such lights reduced rear-end collisions by about 50%.
However, after making these lights mandatory, rear-end collisions reduced by only 5%. 
Why?
:::


## Study types and limitations {#StudyDesignsLimitations}

Experimental studies, in general, have higher *internal* validity than observational studies, since more of the study design in under the control of the researchers; for example, random allocation of treatments is possible to minimise confounding.


:::{.importantBox .important data-latex="{iconmonstr-warning-8-240.png}"}
Only well-conducted experimental studies can show cause-and-effect relationships.
:::


However, experimental studies may suffer from poor *ecological* validity; for instance, laboratory experiments are often conducted conducted under controlled temperature and humidity.
Many experiments also require that people be told about being in a study (due to ethical requirements), and so internal validity may be comprised due to the [Hawthorne effect](#HawthorneEffectExperimental).


:::{.example name="Retrofitting"}
In a study of retro-fitting house with energy-saving devices, @giandomenico2022systematic found large discrepancies in savings for observational studies (12.2%) and experimental studies (6.2%).
The authors say that 'this finding reinforces the importance of using study designs with high internal validity to evaluate program savings' (p. 692).
:::


## Summary {#Chap9-Summary}

The limitations in a study need to be identified, and may be related to:

* *internal validity* (effectiveness): how well the study is conducted with the sample, isolating the relationship of interest.
* *external validity* (generalisability): how well the sample results are likely to apply to the intended population.
* *ecological validity* (practicality): how well the results may apply to the real-world situation.


## Quick review questions {#Chap9-QuickReview}

::: {.webex-check .webex-box}
Are the following statements true or false?

1. When interpreting the results of studies, the steps taken to maximize internal validity should be considered. `r if( knitr::is_html_output() ) {torf( TRUE)}`
2. If studies are not externally valid, then they are not very useful. `r if( knitr::is_html_output() ) {torf( FALSE)}`
3. When interpreting the results of studies, the steps taken to maximize external validity do not need to be considered. `r if( knitr::is_html_output() ) {torf( FALSE)}`
4. When interpreting the results of studies, ecological validity is about the impact of the study on the environment.
`r if( knitr::is_html_output() ) {torf( FALSE)}`
:::




## Exercises {#InterpretationExercises}

Selected answers are available in Sect. \@ref(InterpretationAnswer).


::: {.exercise #ValidityLighting}
A research study examined how people can save energy through lighting choices [@gentile2022improving].
The study states (p. 9) that the results 'are limited to the specific study and cannot be easily projected to other similar settings'.

What type of validity is being discussed here?
:::


::: {.exercise #InterpretationExerciseValidities}
When interpreting the results of studies, we consider the practicality (`r if( knitr::is_html_output() ) {mcq( c(
                     "internal",
                     "external",
                     answer = "ecological") )} else {"________________"}` validity),  the generalizability 
(`r if( knitr::is_html_output() ) {mcq( c(
                     "internal",
                     answer = "external",
                     "ecological") )} else {"________________"}` validity)
and the effectiveness (`r if( knitr::is_html_output() ) {mcq( c(
                     answer = "internal",
                     "external",
                     "ecological") )} else {"________________"}` validity).

Internal validity refers to issues such as `r if( knitr::is_html_output() ) {mcq( c(
                     answer = "confounding",
                     "sampling") )} else {"________________"}`and the Hawthorne effect.

External validity refers to `r if( knitr::is_html_output() ) {mcq( c(
                     "confounding",
                     answer = "sampling") )} else {"________________"}` methods.
:::


::: {.exercise #InterpretationExerciseExternalValidity}
A student project at the university where I work asked the RQ:
   
> Among university students on-campus, is the percentage of word retention higher in male students than female students?

When discussing *external validity*, the students stated:

> We cannot say whether or not that the general public have better or worse word retention compared to the students that we will be studying. 

Why is the statement not relevant?
:::
<!-- 2019 SEM2: Kai M.; Will C.; Riley A. -->


::: {.exercise #InterpretationExerciseParachutes}
Despite their common use, no experimental scientific evidence has shown that parachutes are effective [@data:Smith2003:parachutes].
To obtain evidence, researchers conducted an experimental study [@data:Yeh2018:Parachutes] to 'determine if using a parachute prevents death or major traumatic injury when jumping from an aircraft'.

The researchers randomised $23$ volunteers into one of two groups: wearing a parachute, or wearing an empty backpack.
The response variable was a measurement of death or major traumatic injury upon landing.
From the study, death or major injury was the same in both groups (0% for each group).
However, the study used [@data:Yeh2018:Parachutes]:

> ... small stationary aircraft on the ground, suggesting cautious extrapolation to high altitude jumps [...]


1. Comment on the internal validity.
1. Comment on the external validity.
1. Comment on the ecological validity.
:::


::: {.exercise #InterpretationSleep}
A study examined how well hospital patients sleep at night set [@delaney2018they].
The researchers state (p. 2):

> Non-probability convenience sampling was used to recruit patients

Later, the researchers state (p. 7):

> ... the results may not be generalisable to all hospitals or all ward areas [...] 
> while most healthy individuals sleep primarily or exclusively at night, it is important to consider that patients requiring hospitalization will likely require some daytime nap periods. 
> This study looks at sleep only in the night-time period 22:00--07:00h, without the context of daytime sleep considered.

Discuss these issues using the *language* introduced in this chapter.
:::




<!-- QUICK REVIEW ANSWERS -->
`r if (knitr::is_html_output()) '<!--'`
::: {.EOCanswerBox .EOCanswer data-latex="{iconmonstr-check-mark-14-240.png}"}
**Answers to in-chapter questions:**

- \textbf{\textit{Quick Revision} questions:}
**1.** Only the first statement is true.
:::
`r if (knitr::is_html_output()) '-->'`