Microsoft Word - Osterhaus et al_publication.docx
!
!
!
!
!
Frontline)Learning)Research)Vol.3)No.)4)(2015))56)=)94)
ISSN)2295=3159))
Children’s understanding of experimental contrast and
experimental control: an inventory for primary school
Christopher Osterhausa, Susanne Koerbera, Beate Sodianb
aFreiburg University of Education, Department of Psychology, Germany
bLudwig-Maximilians-University Munich, Department of Psychology, Germany
Article received 18 October / revised 29 November / accepted 4 December / available online 19 January
Abstract
Experimentation skills are a central component of scientific thinking, and many studies
have investigated whether and when primary-school children develop adequate
experimentation strategies. However, the answers to these questions vary substantially
depending on the type of task that is used: while discovery tasks, which require children
to engage in unguided experimentation, typically do not reveal systematic skills in
primary school, choice tasks suggest an early use of adequate experimentation strategies.
To acquire a more accurate description of primary-school experimentation, this article
proposes a novel multiple-select paper-and-pencil inventory that measures children’s
understanding of experimental design. The two reported studies investigated the
psychometric properties of this instrument and addressed the development of primary-
school experimentation. Study 1 assessed the validity of the item format by comparing
2 items and an interview measure in a sample of 71 third- and fourth-graders (9- and
10-year-olds), while Study 2 investigated the reliability and the convergent validity of the
inventory by administering it to 411 second-, third- and fourth-graders (8-, 9- and
10-year-olds) and by comparing children’s performance in the 11-item scale to
2 conventional experimentation tasks. The obtained results demonstrate the reliability
and validity of the inventory and suggest that a solid understanding of experimental
design first emerges at the end of primary school.
Keywords: experimentation skills; control-of-variable strategy (CVS); scientific thinking;
primary (elementary) school; paper-and-pencil assessment
Corresponding author: Christopher Osterhaus, Department of Psychology, Freiburg University of Education,
Kunzenweg 21, 79117 Freiburg, Germany. Phone: +49(0)761 / 682-164, Fax: +49(0)761 / 682 - 98164,
Email: osterhaus@ph-freiburg.de. DOI: http://dx.doi.org/10.14786/flr.v3i4.220
Osterhaus)et )al )
) ) )
)
)
57!
!
1. Introduction
Experimentation skills constitute a fundamental component of scientific thinking, and many
developmental research studies have investigated children’s acquisition of adequate experimentation
strategies (e.g., Case, 1974; Inhelder & Piaget, 1958; Kuhn & Phelps, 1982; Siegler & Liebert, 1975;
for a review see Zimmerman, 2007). Two central research questions have been whether and when
children begin to master the so-called control-of-variables strategy (CVS). This strategy, which is also
referred to as the “vary-one-thing-at-a-time” strategy (Tschirgi, 1980), requires informative
experiments to contrast a single (focal) variable while keeping all other (non-focal) variables constant.
Most studies of experimentation skills probe children’s use of CVS in two different task settings:
(1) discovery tasks with unrestricted variable configurations (production of CVS) and (2) choice tasks
with restricted response options (choice of CVS).
Discovery tasks typically require children to explore the causal relations between different
candidate causes and an outcome over a set of experimentation trials (e.g., Kuhn et al., 1995;
Schauble, 1990). In each of these trials, mature reasoners form hypotheses about the system of
independent variables (i.e., which variables are causal and which are non-causal), and based on these,
they use CVS to isolate a single variable, which they contrast in an experiment that controls all
non-focal variables (controlled-contrastive experiment). By updating their hypotheses and repeating
this procedure over multiple trials, mature reasoners arrive at a final theory of the causal standing of
each variable. Discovery tasks hence involve reasoners in multiple phases of constructing scientific
knowledge (i.e., reiterative formation of hypotheses, experimentation and data interpretation), and
therefore they offer a high ecological validity and are particularly well suited for microgenetic studies
that investigate strategy change in detail by repeatedly using the same task with a high density of
observations in the short period of time when change is assumed to occur.
Choice tasks, in contrast, typically present children with a specific and directed hypothesis
regarding one of the candidate causes. In addition, they include restricted answer options that represent
distinct variable configurations from which children are allowed to choose (multiple choice [MC];
e.g., Croker & Buchanan, 2011; Koerber, Mayer, Osterhaus, Schwippert, & Sodian, 2015; Mayer,
Sodian, Koerber, & Schwippert, 2014; Tschirgi, 1980). Mature reasoners understand that they need to
compare conditions, and their command of CVS is demonstrated by selecting the answer option in
which only the focal variable is varied. This requirement of a single choice makes choice tasks easier
to administer and hence well suited for large-scale, paper-and-pencil-based assessments, which are
important research tools for investigating the large interindividual differences that already exist in
primary-school experimentation (Bullock, Sodian, & Koerber, 2009).
Despite their common measurement focus on CVS and their same basic task requirement (i.e.,
understanding that conditions need to be compared and an appropriate comparison needs to be made),
discovery and choice tasks reveal a substantially different picture of primary-school children’s
experimentation skills. While discovery tasks tend to reveal that only a small number of students
produces experiments that are designed in accordance with CVS (e.g., 17% initially used CVS in a
sample of fifth- and sixth-graders; Schauble, 1990), choice tasks reveal that most primary-school
children prefer controlled-contrastive experiments over confounded experiments when they are
allowed to choose between different experimental designs (e.g., 54% correct CVS choices by fourth-
graders in Koerber, Mayer, et al., 2015; around 60% in Bullock & Ziegler, 1999). Although
differences between production and choice tasks (i.e., between open-answer and closed-response
items) are a common empirical finding for knowledge scales (e.g., Nehm & Schonfeld, 2008), the
large discrepancies between the two types of CVS tasks suggest that differences in performance are
not solely attributable to the increased probability of correct guessing in choice tasks.
Performance differences between discovery and choice tasks might be attributable to their
specific and discrepant task demands and solution strategies (see Table 1). While discovery tasks
Osterhaus)et )al )
) ) )
)
)
58!
!
require the repeated generation of hypotheses, which is an ability that may develop later than do
experimentation skills (Piekny & Maehler, 2013) and which requires extensive memory and
processing capacities, CVS choice tasks might not assess children’s understanding of experimentation
because they can be solved by lower-level heuristics, such as varying a single variable while keeping
the others constant, which does not need to be motivated by children’s full understanding of
experimentation. Other evidence, such as children’s justifications, is therefore required before
concluding whether children fully understand the rationale for CVS.
Table 1
Student performance in and characteristics of discovery tasks, choice tasks and understanding of
experimental design (UNEX).
Discovery tasks Choice tasks UNEX
Description iterative experimenta-
tion and use of CVS to
uncover causal effects
single choice of a (controlled-
contrastive) experimental
comparison
identification of design
errors in (controlled-)
contrastive experiments
Task format open;
often computerized
MC;
sometimes other formats that
do not favour guessing (e.g.,
Bullock & Ziegler, 1999)
MS
Use of CVS 17% initial use in at
ages 11 and 12
(Schauble, 1990)
8–68%
(Croker & Buchanan, 2011),
54% (Koerber, Mayer, et al.,
2015), and 60% (Bullock &
Ziegler, 1999) all at age 10
29% at age 10
(cf. Study 2)
Validity moderate
guessing is not a
problem
low
guessing is a problem
(especially for MC items)
high
guessing is not a
problem*
task requires non-
essential skills: tracking
of variables (memory
and processing skills),
hypothesis generation
may be solved by lower-level
heuristics
understanding features
of experimental contrast
and experimental
control required
Reliability high
performance estimates
stable across problems
low
performance estimates differ
across tasks and contexts
high
performance estimates
stable across problems*
Assessment time-consuming rapid; allows for large-scale
assessment
rapid; allows for large-
scale assessment
Note. *Characteristics investigated in the two studies presented in this article; CVS = control-of-
variables strategy; MC = multiple choice; MS = multiple select.
A further criticism of both discovery and choice tasks is that studies of scientific thinking—
with their narrow focus on CVS—have not explored the broader context of children’s understanding
Osterhaus)et )al )
) ) )
)
)
59!
!
of the experimental method. In addition to CVS, a basic feature of experimentation is that planned
comparisons are necessary to test for differences between conditions, as are randomization and an
understanding of local control, which is necessary to reduce variation due to extraneous factors and
which goes beyond the control of non-focal variables.
Bullock et al. (2009) developed an interview that investigates children’s metaconceptual
understanding of experimental design (UNEX) in participants aged 12–22 years. This instrument asks
children to review a set of fictitious experiments that contain diverse design errors that violate the
principles of local control or experimental contrast. Children need to recognize that ill-designed
controlled-contrastive experiments do not provide an adequate test of hypothesis because variation due
to extraneous factors is not reduced or non-focal variables are not controlled for (violation of CVS);
they also must understand that hypothesis testing is impossible in ill-designed contrastive experiments
because a single observation is made (either due to the missing variation of the focal variable, or the
lack of an initial measurement in a pre-post design) and the principle of experimental contrast
therefore is ignored. This latter experiment type provides no information about the control of non-
focal variables, and the ‘violated’ design feature is simply the principle of experimental contrast,
which should be easier to understand than ‘experimental control’.
Primary-school children’s UNEX has not been investigated previously. However, Bullock and
her colleagues found that sixth-graders (the youngest age group interviewed with such an instrument)
solved around 50% of tasks correctly, by identifying the experimental design error and justifying their
opinion about whether the experiment was appropriately designed. It therefore appears that even very
young children can show UNEX if they are questioned in an age-appropriate format that offers
contextual support, such as providing a graphic representation of fictitious experiments and offering
answer alternatives in a closed-response format.
The present studies investigated whether a paper-and-pencil version of UNEX can provide
valid and reliable measurements in primary-school children, in order to determine whether early
abilities can be identified in this age group. Our inventory of primary-school UNEX uses 11 closed-
response, multiple-select (MS) items (examples are provided in Figure 1, and Appendix 1 provides the
full item set). For each item the children have to answer whether or not they consider a fictitious
experiment to be well designed, and whether or not they agree with each of three separate
justifications concerning the good or bad quality of the experiment (which includes one that identifies
the design error in question).
More conventional MC items provide a single choice, whereas the three separate decisions
required by our MS procedure have two important advantages: (1) the probability of correct guessing
is substantially reduced (i.e., 12.5% instead of 33% in the case of three statements), and (2) the MS
format investigates potential inconsistencies in children’s understanding of experimentation. MC items
only make it possible to conclude that children consider their chosen answer to be superior to the non-
selected alternatives, whereas MS items give additional information about their view on all response
options (i.e., although children may recognize the design error, they may still hold naïve beliefs
regarding the production of effects).
The correct and incorrect answer options used in the present studies are based on children’s
answers to open items in prestudies and they draw on a conceptual-development model of scientific
thinking (Koerber, Mayer, et al., 2015). Specifically, each item includes naïve, intermediate and
advanced-level answers. Naïve-level answers reveal no understanding of hypothesis testing, instead
referring to the production of an effect (cf. answer option 3 in Figure 1), intermediate-level answers
demonstrate a first understanding of the necessity of hypothesis testing and the existence of relevant
design features (e.g., testing and sample size are important; cf. answer option 2), and advanced-level
answers identify the specific design error and recognize that it restricts the information that can be
drawn from the experiment (cf. answer option 1).
Osterhaus)et )al )
) ) )
)
)
60!
!
Figure 1. Sample item.
While research has shown that children’s (dis-)agreement with conceptually different levels on
MS items matches the levels found in an interview conducted after exposure to the MS answer options
(MS item then open interview [I-after]; cf. Koerber, Osterhaus, & Sodian, 2015), little is known about
the relation between children’s MS choices and the beliefs they hold before being presented with the
MS item (open interview then MS item [I-before]). An instrument’s reliability and validity depends on
a close relation between initial beliefs and the levels identified in the MS item.
The present work therefore first investigated (in Study 1) whether the MS item format is valid,
resulting in ascriptions of levels that are significantly related with those levels found in I-before, and
then (in Study 2) addressed whether the inventory results in a reliable scale with satisfactory content
and convergent validities. Because primary-school children’s UNEX has not been investigated
previously, Study 2 also addressed the abilities and development of primary-school children.
2. Study 1
Study 1 investigated the validity of the MS item format by comparing children’s performance
in two interview measures, which were conducted before (I-before) and after (I-after) the presentation
of the closed-response answer options. Study 1 also investigated the relation between the novel MS
and conventional MC formats, as well as the relation between MC and the interview, both before and
after the presentation of the MS task (i.e., I-before and I-after), to obtain a better understanding of the
relative performances of diverse testing formats.
2.1 Methods
2.1.1 Participants
The 71 included primary-school children comprised 56 third-graders (mean age 8 years,
8 months, SD=5 months; 31 girls) and 15 fourth-graders (mean age 10 years, 1 month; SD=4 months;
17
The hospital
On planet Mola, many Molans are sick.
In the hospital on planet “Mola”, a scientist
wants to find out whether the Molans recover
faster if they are allowed to receive visitors.
He performs an experiment:
20 Molans who have a heat disease are
allowed to receive visitors for 2 hours per day
during 2 weeks.
2 visiting hours
20 Molans who have a broken antenna are not
allowed to receive visitors during these 2
weeks. no visiting hours
He compares both groups and finds out:
After 2 weeks, all Molans with heat disease
have recovered.
Molans with a broken antenna, who have not
received visitors, are still sick after 2 weeks.
The scientist is convinced:
“Whether or not the Molans recover fast depends on the visiting hours.”
Was this a good experiment?
� Yes � No
18
on planet “Mola”
Susan, Lisa, and Vera think about whether this was a good experiment.
Who is right and who is not?
is right is not
right
1. Susi says: “It was not a good experiment
because the scientist should have compared
Molans with the same disease.”
� �
2. Lisa says: “It was a good experiment
because he investigated many Molans.” � �
3. Vera says: “It was a good experiment
because 20 Molans have recovered from
their disease.” � �
Which of the three girls has the best answer? No.______
Osterhaus)et )al )
) ) )
)
)
61!
!
7 girls). Children were recruited from three predominantly middle-class schools in Germany. Parental
informed consent was obtained for all children.
2.1.2. Materials
The children were presented with and interviewed about (see Procedure) two items from our
inventory: one contrastive and one controlled-contrastive experiment. Both experiments were set up in
an artificial context in order to reduce interferences of children’s content knowledge on design
evaluations.
Item 1 (contrastive experiment) presented the children with a story about a scientist who wants
to test the hypothesis that yellow fertilizer increases plant growth significantly more than blue
fertilizer. The scientist administers yellow fertilizer to 100 plants. He observes that all plants got
bigger blossoms and therefore concludes that the yellow fertilizer works better than the blue one.
Children were asked whether or not this was a good experiment, and evaluated three explanations for
their opinion (MS) on three hypothesized levels: (1) an explanation based on the production of effects
(naïve level; “It was a good experiment because all plants that received the yellow fertilizer became
bigger”), (2) an explanation that contained a design feature that was not the crucial to the experiment’s
validity (intermediate level; “It was a good experiment because he tested the yellow fertilizer on many
plants”) and (c) the correct explanation that identified the design error in question (advanced level; “It
was not a good experiment because he only tested the yellow fertilizer”). After evaluating each of the
three levels (MS), the children also indicated which of the three explanations they considered the best
(MC).
Item 2 (controlled-contrastive experiment) presented the children with a story about two
grandmothers who use lake or river water to water their plants. While those plants watered with lake
water grow well, plants that are watered with river water are withering. Children had to evaluate
whether the experimental data was sufficient to support the hypothesis that lake water causes plants to
grow well, or whether more information would be needed (i.e., control of non-focal variables).
Analogously to Item 1, the three explanations reflected the distinct levels; however, while the naïve
level as was the case for Item 1 referred to the production of effects and the advanced level identified
the design error in question, the intermediate level for this item included a reference to a potential
causal mechanism rather than to a non-crucial design feature (i.e., “Lake water contains more minerals
than lake water. Therefore, lake water is better for plants”).
2.1.3 Procedure
Both items were read out loud to the children in a one-on-one interview. Children marked their
answers in their own booklets. Interviews were conducted at two points during the presentation of the
item: (1) after the children’s initial design evaluation and before presenting the MS answer options (I-
before; “Why was the experiment good/bad?”), and (2) after presenting the answer options and the
children choosing the best one (I-after; “Why did you consider this answer to be the best one?”). Both
instances included follow-up questions such as “Why did this [the reason children named] make the
experiment a good/bad one?” or “Would you have done anything differently?”
Osterhaus)et )al )
) ) )
)
)
62!
!
Table 2
Coding of MS and interviews
Levels children agreed to
Naïve Intermediate Advanced
Coding of MS (final level)
Naïve x --- ---
Naïve x x ---
Naïve x x x
Naïve x --- x
Naïve --- --- ---
Intermediate --- x ---
Intermediate --- x x
Advanced --- --- x
Coding of interviews Sample answers
Naïve “because the fertilizer made the plants look more beautiful” (production of
effect)
Intermediate “because he [the scientist] tried it on so many plants”
(reference to non-relevant design feature)
“because lake water is often dirty, whereas not much rubbish is thrown
into river water” (reference to mechanism)
Advanced “because he [the scientist] did not try both fertilizers”
(correct identification of design error)
2.1.4 Transcription and coding of children’s answers
Interviews were audiotaped, transcribed verbatim and coded by two independent raters (see
Table 2 for coding examples and the MS coding). Using a strict criterion, the lowest level which
children agreed to was taken as the final level in the coding of the MS item (e.g., if a child accepted
the naïve and intermediate levels simultaneously, the MS item was coded as naïve). The interrater
kappa reliability values for I-before and I-after were .94 and .88, respectively, for Item 1, and .73 and
.90 for Item 2. Since some children gave invalid answers (especially on I-before; e.g., they refused to
answer or only gave answers that were irrelevant to the question), some of the subsequent analyses
involved a smaller sample.
2.2. Results and discussion
2.2.1 Core performance
Core performance data for Items 1 and 2 (Table 3) and a Wilcoxon signed-rank test of I-before
revealed that, as expected, more children recognized the design error in the contrastive (Mdn=1) than
the controlled-contrastive (Mdn=0) experiment, Z=2.92, p<.01, r=.30.
Osterhaus)et )al )
) ) )
)
)
63!
!
2.2.2 Comparison of the different formats
To investigate whether meaningful relations between the different formats in the two items
existed (i.e., whether there was a high agreement in the levels assigned), we assessed Spearman
correlations (rho) and, in addition, computed Wilcoxon signed-ranks statistics (Z) when non-
significant correlations suggested low convergence, which may have been a result of over- or
underestimation in one of the formats.
For Item 1 (Table 4), correlations between formats were significant for all comparisons,
except for I-before and MC, where the correlation did not reach significance. However, as a Wilcoxon
signed-ranks test revealed, there was no significant difference in difficulty of MC or I-before, and
none of the formats led to a systematic over- or underestimation. Rather, 24% of all children
performed better on MC than on I-before while 27% were assigned a higher level in I-before than in
MC (cf. Table 4).
For Item 2 (Table 5), only the correlation between I-before and the MS task was significant. A
Wilcoxon signed-ranks test revealed that the MC question (Mdn=1) significantly overestimated
performance with respect to I-before. Similarly, MS was significantly more difficult than MC, just as
I-before was more difficult than I-after, suggesting that presenting answer options significantly
decreased the difficulty of identifying the design error in this second item with a more complex
experimental design.
Table 3
Core performance data for Items 1 and 2 (contrastive and controlled-contrastive experiment)
Item / format Naïve level Intermediate level Advanced level Total
Item 1
I-before 15 (33) 11 (24) 19 (42) 45 (100)
MS 62 (87) 1 (1) 8 (11) 71 (100)
MC 23 (32) 14 (20) 34 (48) 71 (100)
I-after 18 (26) 29 (41) 23 (33) 70 (100)
Item 2
I-before 25 (51) 16 (33) 8 (16) 49 (100)
MS 53 (75) 13 (18) 5 (7) 71 (100)
MC 10 (14) 26 (37) 35 (49) 71 (100)
I-after 21 (30) 26 (37) 23 (33) 70 (100)
Notes. Data are n (%) values. I-before = open interview then MS item; I-after = MS item then open
interview.
Osterhaus)et )al )
) ) )
)
)
64#
#
Table 4
Convergence between different item formats for Item 1 (contrastive experiment)
Notes. Data are n (%) values. Int. = intermediate; adv. = advanced; Z = Wilcoxon signed-rank test. Over- and underestimation of the second comparison (b) relative to
the first (a). Rho and Z are computed for children with complete observations on both measures for each test pair (i.e., children who did not give a valid answer on a measure that
was part of the respective test pair were excluded from the respective analysis).
* p<.05. ** p<.01.
Table 5
Convergence between different item formats for Item 2 (controlled-contrastive experiment)
Notes. Data are n (%) values.
** p<.01. *** p<.001 . 1 r=.29. 2 r=.39. 3 r=.53
Same level (a=b) Overestimation (b>a) Underestimation (ba) Underestimation (b60%), and lower frequencies of intermediate and advanced conceptions (all <20% and <25%,
respectively).
3.2.2 Scale analysis
The reliability of the scale was determined by fitting a partial credit model (Masters, 1982) to the
children’s responses. This model assumes that developmental progression is unidirectional and that
categories (i.e., naïve, intermediate and advanced) are hierarchical, reflecting the assumption of our
conceptual-development model. All but one item (Item U5, a contrastive experiment) had a good fit to the
model (i.e., 0.85> infit mean-square statistic [MNSQ] <1.15; see Table 6). Removing this poorly fitting item
yielded a scale with an expected a posteriori estimate based on a plausible values (EAP/PV) reliability of .82
(weighted-likelihood estimator person separation reliability=.55, Cronbach’s α=.85; scale mean=5.35,
SD=5.85, minimum=0, maximum=22). All items satisfied the partial credit model’s requirements of
increasing point-biserial correlations and ability estimates per category. However, none of our items fulfilled
a third criterion (ordered delta parameters). It is debated in the literature as to whether this is a mandatory
requirement for model fit to hold (Adams, Wu, & Wilson, 2012), but the unordered delta parameters are
consistent with children’s low frequency of choosing the intermediate level. We therefore dichotomized the
data by collapsing the naïve and intermediate levels. A Rasch model fitted to the resulting binary data
revealed a good fit for all items (0.85< infit MNSQ <1.15) and an EAP/PV reliability of .71 (weighted-
likelihood estimator person separation reliability=.62, Cronbach’s α=.72). All subsequent analyses are based
on these binary data.
!
! ! !
!
!
!
67!
!
Table 6
Percentage of answers on the naïve, intermediate and advanced level for the 11 UNEX items of Study 2
overall, and the item difficulty, discrimination and item fit.
Notes. Difficulty, discrimination and infit mean-square statistic (MNSQ) are based on an analysis of the partial
credit model; large negative item difficulties indicate easy items, while large positive values indicate difficult ones; infit
MNSQ should be between .85 and 1.15 for an item to show a good fit to the partial credit model; # indicates an item
with a poor fit according to infit MNSQ. Item numbering applies to Study 2 and the full item set given in the Appendix.
3.2.3 Developmental patterns and interindividual differences
A univariate analysis of variance revealed a significant main effect in UNEX for grade,
F(2,402)=21.75, p<.001, partial η2=.10. Interestingly, while fourth-graders recognized significantly more
design errors than did third-graders, F(1,402)=24.14, p<.001, partial η2=.06 (see Figure 2), there was no
difference between grades 2 and 3, F(1,402)=1.89, p>.05. There were also no differences between boys and
girls, F(1,402)=3.18, p>.05.
3.2.4 Content validity
An explanatory item response model (De Boeck & Wilson, 2004) with a fixed person effect (age)
revealed that controlled-contrastive experiments were, as hypothesized, more difficult than contrastive
experiments (see Table 7 for model fit, Table 8 for parameter estimates).
Figure 2. Percentage of correct answers for contrastive and controlled-contrastive experimental designs per
grade. Error bars indicate 95% confidence intervals.
Level
Item Experiment type Naïve
(0)
Inter.
(1)
Adv.
(2)
Difficulty Discri-
mination
Infit
MNSQ
U1 Contrastive 62.4 19.9 17.7 –.14 .57 1.10
U2 Contrastive 67.8 9.3 22.9 –.32 .66 1.03
U3 Contrastive 74.2 7.6 18.2 .20 .69 0.92
U4 Contrastive 73.2 4.4 22.4 –.14 .66 1.03
U5 Contrastive 65.2 10.4 24.3 –.48 .80 0.74#
U6 Contrastive 73.2 5.6 21.2 .01 .45 0.96
U7 Controlled-contrastive 73.7 15.1 11.2 .56 .43 1.04
U8 Controlled-contrastive 72.7 4.6 22.7 –.15 .63 1.05
U9 Controlled-contrastive 72.3 11.5 16.2 .69 .49 0.90
U10 Controlled-contrastive 68.1 12.9 19.9 –.04 .69 0.91
U11 Controlled-contrastive 74.0 5.6 20.44 –.06 .72 0.89
!
! ! !
!
!
!
68!
!
Table 7
Model comparisons for the explanatory item response model
Notes. M0 = reference model (1-parameter logistic model); AIC = Akaike information criterion; BIC =
Bayesian information criterion; -2LL = deviance; df = degrees of freedom; LR test = Likelihood ratio test.
*** p<.001.
Table 8
Coefficients for M2
3.2.5 Convergent validity
One-third (130) of the children chose CVS for both the cars and slopes tasks; however, only 55
children (14%) chose CVS consistently across the 2 items. The cars task was performed correctly by 20%,
33% and 44% of second-, third- and fourth-graders, respectively; the corresponding rates for the slopes task
were 18%, 35% and 46%. The overall performance was thus slightly worse than that reported by Bullock
and Ziegler (1999), where approx. 40% and 60% of third- and fourth-graders, respectively, correctly solved a
CVS choice task. These differences are probably due to differences in sociodemographic characteristics
between the previous urban sample and our more rural sample (cf. Koerber, Mayer, et al., 2015).
Interestingly, whereas the change-all strategy (vary all—including non-focal—variables) was the
third most frequently chosen strategy for the cars task (20%), it was the least popular strategy for the slopes
task (4%). In contrast to the cars task, where low performance resulted mostly from children disregarding the
necessity of experimental control, errors in the slopes task were primarily due to children choosing an
incorrect focal variable (i.e., size or position of the marble instead of slope steepness). While both these
errors reflect unsuccessful coordination of hypothesis and evidence (cf. Kuhn, 2011), they suggest that the
content domain of the specific task and children’s knowledge thereof influence the expression of this
interference between children’s hypotheses and their construction of evidence.
A binomial regression revealed that performance in the cars task was significantly predicted by
controlled-contrastive UNEX, χ2(1)=4.67, p=.03. Specifically, children with a more profound controlled-
contrastive UNEX were more likely to use CVS than any other strategy, β=.13, t(1)=4.62, p=.03, odds
ratio=1.13. However, contrastive-UNEX experiments did not predict strategy choice in the cars task, β=.02,
t(1)=.11, p=.74. In contrast, the use of CVS in the slopes task was predicted by contrastive UNEX,
χ2(1)=8.16, p=.004. Specifically, children with a high contrastive UNEX showed an increased use of CVS,
β=.13, t(1)=8.02, p=.005, odds ratio=1.14, while there was no positive effect of controlled-contrastive
Model Effects
(fixed)
Effects
(random)
AIC BIC -2LL df LR test
M0 (1PL) --- Intercept 3493.3 3506.1 3489.3
M1 Experiment
type
Intercept 3483.3 3502.6 3477.4 2 11.87***
M2 Experiment
type + age
Intercept 3465.4 3491.0 3457.4 1 20.01***
Model B SE p
Intercept –7.20 1.14 < .001
Age .54 .12 < .001
Controlled-contrastive (reference: contrastive) –.33 .01 < .001
!
! ! !
!
!
!
69!
!
UNEX, β=–.08, t(1)=2.03, p=.15. This finding is consistent with the descriptive data for the slopes task,
which suggest that few children choose an experiment in which all non-focal variables are varied. These
results might be due to children only contrasting non-focal variables when their individual influences are
unknown to them and they want to find out about them in a single test (cf. Schauble, 1990).
While 68% and 56% of the children with mastery in the controlled-contrastive UNEX (≥four items
correctly solved) solved the cars and slopes tasks correctly, 70% and 68% of the incompetent children (