Conclusion-1:

The minor hypothesis-1 "IRT-LR method is more sensitive than the MH method",
tested in 4 analysis groups: (4400.40), (4400.32), (3000,40), and (3000,32).
In the groups: (4400.18) and (3000,18), the IRT-LT method did not prove to be
more sensitive than the MH method

This testing, more as an influence factor number of test participants, compared to

the number of items (length of test).
When the number of participants is constant and the number of item drops from 40
to 32, the IRT-LR remains more sensitive than MH. That means, the IRT-LR
method is more sensitive than the MH in the condition of the large number of
participants (also the number of large items).
This result is in accordance with the findings of Kubra Atalay Kabasakal, et.al.
where the IRT-LR method is more sensitive than the Mantel-Haenszel method
(Kabasakal et al., 2014).

In the groups: (4400.18) and (3000,18), the IRT-LT method did not prove to be
more sensitive than the MH method. Despite the large number of participants, the
sensitivity of the IRT-LR method is more affected and decreased compared to MH.
This can be caused by a factor of 18 items. Where, it has been below the minimum
recommended by Hulin, Lissak and Drasgow (1982), namely 1000 test participants
for 60 items, or 2000 test participants for 30 items (Wiberg, 2011) for 3 parameter
logistics model (3PL);

second, the nature of item (18) is also affected, as a result of IRT selection. On this
item the IRT requirements have been applied, though not strictly. This is not
beneficial for IRT-LR in estimating item parameters and participants. As a result,
the sensitivity of the IRT-LR method is more affected or decreased. If only the data
fit model, it is very possible that the IRT-LR is significantly more sensitive than
MH. In this position, the MH method is a nonparametric method, more suitable for
the number of small items (below 20) (Wiberg, 2011) and its sensitivity begins to
increase. These results are close to the results of Bartosz Kondratek's study,
Magdalena Grudniewska (Kondratek, 2014), where the MH method has a higher
power than IRT-LR when there is uniform DIF.
Conclusion-2:
The minor hypothesis-2 " IRT-LR method is more sensitive than the RL method" gives a lot of
results.
(a). tested in the group (4400.40), the IRT-LR method was significantly more sensitive than the
RL method. However in groups (3000,40), with the same test length (40), the sensitivity of the
IRT-LR method was not significantly different.

The larger number of test participants supports the IRT-LR method, which matches large data
(in Wiberg, 2007) (Camili, 2006). [96] Compared to the RL method which is more suitable for
small data.

(b). not tested in the group (4400.32) and (3000.32), IRT-LR method was not significant more
the sensitivive than the RL method.

In this case, the weakness of IRT-LR sensitivity to RL is not because it affects the number of
participants. Rather it is related to wheat received 32, which is the result of TTK findings, which
has not been saved by TRB. More supports the nonparametric RL method than IRT-LR.
Therefore, in both of these groups, it was plausible that IRT-LR sensitivity was not significantly
more sensitive than the RL method.

(c). Interesting things occur in groups: (4400,18) and (3000,18). The IRT-LR method again
proved significantly more sensitive than the RL method.

The causal factors can be drawn, more than the nature of the results of examination 18, namely
the results of the selection items TRB. This item has implemented the TRB requirements (though
not completely tight), so it supports the IRT-LR method, compared to the RL method. Also using
the 3PL logistic model, making it more in line with the IRT-LR method, compared to the RL
method that is more suited to the 1PL logistical model or assumes a linear relationship (Wiberg,
2007) [97]. Related to the plausibility of this second group analysis, the IRT-LR method is more
sensitive than the RL method. Regarding the number of items (18), the RL (nonparametric)
method is more suitable for long tests (more than 20 items) (Lopez, 2012) [98]. For shorter tests,
the performance of nonparametric methods is debatable.

By linking the MH and RL methods together to the IRT-LR method in the analysis group
(4400.18) and (3000,18), a new result was obtained. In this group, the MH method is relatively
stronger to compensate for the sensitivity of IRT-LR, compared to the RL method for IRT-LR.
In this group, the average number of grains detected by DIF by the IRT-LR method is still higher
than the MH method, but the difference is not significant. Meanwhile, the IRT-LR method is
significantly more sensitive than the RL method. This is due to MH method factors which are not
related to the model, but the RL method requires a linear model and contains elements of the
logistical model. The data uses the 3PL model, which does not support the RL method, but does
not reduce the sensitivity of the MH method, and supports the IRT-LR method. The IRT-LR and
RL methods are both related to the model. The RL method is more appropriate to use a linear
model or 1PL model, compared to 3PL. While the 3PL model is more suitable for the IRT-LR
method. So that it is acceptable, that in these two groups (TRB selection grain group), the IRT-
LR method was not significantly more sensitive than the MH method, but was significantly more
sensitive than the RL method. The opposite interesting thing happened in the analysis group
(4400.32) and (3000.32), where the RL method more strongly offset the sensitivity of the IRT-
LR method, compared to the MH method.

Conclusion-3:
The minor hypothesis-3 " RL method is more sensitive than the MH method", is consistently
significant in all analysis groups.

The results are consistent with the researchers' initial expectations that the Logistic Regression
(RL) method is more sensitive than the Mantel-Haenszel (MH) method of detecting differential
item functioning (DIF). Rogers and Swaminthan concluded that the RL method was as powerful
as MH in detecting uniform DIF, and more powerful than MH procedures in detecting
nonuniform DIF (Rezaee, Shabani, 2010).

[99] In Duncan's study (Rezaee, Shabani, 2010) [100] it was stated that RL is able to detect
nonuniform DIF better than the DIF method, and equally powerful detects uniform DIF with
MH, then inclusive of size influences will use RL as an option as DIF detection method. Jodoin
and Gierl produce that RL has power that can be compared with MH in detecting uniform DIF,
but is superior in detecting nonuniform DIF. In addition, Gierl, Rogers and Klinger found that
large influences (for MH, SIBTEST, and RL) were high correlations throughout the DIF
procedure except for nonuniform DIF sizes, which were only accessible by RL. Taken together,
all findings can provide confirmation of the superiority of RL over MH. Another point that can
be added is the results of Huang's research (Rezaee, Shabani, 2010) [101]. in testing the
reliability of different DIF methods, the comparison between standardization methods, MH and
RL, it was found that more grains were identified DIF by RL than MH and standardization.
Where items marked DIF by the MH method can be changed either as a uniform DIF or not
uniformly by the RL method. In other words, all items marked DIF by the MH method can also
be marked DIF by the RL method. More importantly in the work of Abbas Ali Rezaee and
Enayatollah Shabani, in the discussion section, recorded in verified DIF, the RL method is more
sensitive than the MH method. He took the MH method, the RL method was stated as the most
identified DIF. According to Gierl's findings, Khliq et al., Who found that in meetings with MH
that were referenced as conservative procedures, RL could accommodate many items that
showed DIF. Also Gierl, Jodoin, et al. Have found that RL has a good type I error rate (non-DIF
items are classified as DIF items). Overall, RL is easier than most with moderate DIF or large
DIF from the MH method. The results of this study are also in accordance with the results of a
study found by Chi-Yu Huang (Chi-Yu, 1998) [102], on the Iowa Basic Skills Test (ITBS) data,
where the RL method looks for more DIF grains through MH , each 40% and 11%